theses copyright undertaking · 2020. 6. 29. · applications. however, most of the existing...
TRANSCRIPT
i
Abstract
Lines provide important information in images and line detection is crucial in many
applications. However, most of the existing algorithms focus only on the extraction of line
positions, ignoring line thickness. In this thesis, we aim to address this issue. We describe a
general method for wide line detection and apply this method to solve two practical line
detection problems.
In the first section, we propose a novel wide line detector to extract the line entirely. In
contrast with the traditional directional-derivative-based edge and line extraction method,
our wide line detector is based on isotropic nonlinear filtering without any derivatives and
consequently is more robust to noise. We develop an approach for dynamic selection of
parameters of the proposed wide line detector. We also introduce a general scheme for
analyses of this wide line detector further. A sequence of tests is conducted on a variety of
image samples. The experimental results demonstrate that the proposed wide line detector
works very well for a range of images containing lines of different widths, especially for
those where the width of lines varies greatly and where the lines run close together or cross
each other.
We then address the first application of our wide line detector: palm-line based palmprint
recognition. We present a novel palm-line feature extraction method for personal
identification. As compared to previous work on palm line extraction, the proposed wide line
detector-based method extracts not only structure features but also strength features of palm
lines. We introduce a translation-invariant scheme for palm-line feature matching. We also
develop an experimental scheme to find out the optimal combination of parameters for the
proposed palm-line feature extraction method. An extensive test is conducted on a public
palmprint database. Experimental results show that the performance of the proposed
palm-line feature extraction method is comparable with the state-of-the-art algorithms of
palmprint identification and thereby palm-line features can be used to recognize palmprints.
Finally, we for the first time attempt to extract tongue cracks, one of pathological features
in tongue diagnosis. We propose a framework for automatic tongue crack extraction. We
derive a tongue crack detection scheme based on the wide line detector which extracts the
whole of the line by employing an isotropic nonlinear filter. The wide line detector describes
the relationship between the size of the isotropic filter, i.e. the scale of this detector, and the
width of detected lines. Due to the large range of widths of tongue cracks, the maximum
widths of cracks vary greatly with different tongue images and consequently the sizes of the
ii
isotropic filters should be very different. To implement the proposed scheme totally
automatically, we design an adaptive algorithm of line width estimation. The proposed
scheme has been tested on a set of typical cracked tongue samples and our experimental
results show the promising performance of our automatic tongue crack extraction scheme.
iii
List of Publications
The following technical papers have been published or are currently under review based on
the result generated from this work.
1. Laura Liu, David Zhang, and Jane You, “Detecting Wide Lines Using Isotropic
Nonlinear Filter,” IEEE Trans. Imag. Processing, vol. 16, no. 6, pp. 1584-1595,
2007;
2. Laura Liu and David Zhang, “Adaptive Detection of Cracks in Tongue Images,”
submitted to IEEE Trans. Medical Imaging,;
3. Laura Liu and David Zhang, “Palm-Line based Personal Recognition,” submitted to
Pattern Recognition;
4. Laura Liu and David Zhang, “Tongue crack extraction,” accepted by 1st Int’l Conf.
Medical Biometrics, Hong Kong, Jan. 4-5, 2008;
5. Laura Liu and David Zhang, “Palm-Line Detection,” ICIP 2005, pp. 269-272,
Genova, September 11-14, 2005;
6. Li Liu and David Zhang, “A Novel Palm-Line Detector,” the 5th International
Conference on Audio- and Video-based Biometric Person Authentication, pp.
563-571, New York, July 20-22, 2005;
7. Li Liu and David Zhang, “A Novel Palm-Line Detector,” the 6th ACM Postgraduate
Research Day, pp. 24-29, March 12, 2005.
i
Acknowledgements
I would like to take this opportunity to express my sincere gratitude to everyone who
generously gave me their support.
On the first place, I would like to express my gratitude and appreciation to my Supervisor,
Prof. David Zhang, for his endlessly patient guidance throughout my research studies. My
Ph.D. studies would have never been completed without his encouragement, help, and good
advice. His immense enthusiasm for research is extremely motivating. The discussions with
him have not only contributed significantly to this work, but also have made lasting
influences in me.
I was fortunate to share my office with Guang Ming Lu, Adams Kong and Michael Wong
during my initial days, from whom I have learned quite a lot. I also admire the selfless
attitudes of them. They were always ready to advice and help.
I really appreciate the open and friendly environment at Biometrics Lab. Specifically, I
would like to thank Dr. Jane You, Dr. Xiang Qian Wu and Dr. Wang Meng Zuo for the
support and constructive comments they gave me. It was wonderful to be with encircled by
so many nice and smart people like Dr. Lei Zhang, Dr. Da Cheng Tao, Dr. Jian Yang, Dr.
Ajay Kumar, Denis Guo, Jerry Zhao, Kenneth Li and Vivek.
My special thanks goes to all my friends in Hong Kong, specifically Annie, Qinqin, Dayu,
Helen and Byron. They made my time here very lively. I will definitely miss the discussions
and the fun.
I would like to dedicate this work to my parents. Their love, constant support and
encouragement made me what I am.
ii
Table of Contents
Abstract......................................................................................................................... i
List of Publications..................................................................................................... iii
Acknowledgements....................................................................................................... i
1 Introduction ............................................................................................................1 1.1 Motivation...................................................................................................1 1.2 Statements of Originality ............................................................................2 1.3 Organization of this thesis...........................................................................3
2 Literature Review...................................................................................................6 2.1 Overview of line detection..........................................................................6
2.1.1 Hough transform ..............................................................................6 2.1.2 Parallel edge extraction....................................................................8 2.1.3 Hessian-based ridge detection..........................................................9
2.2 Review of wide line detection.....................................................................9 2.2.1 Edge-based line finder ...................................................................10 2.2.2 Ridge-based line detector...............................................................11 2.2.3 Line operator ..................................................................................12
2.3 Review of palmprint recognition ..............................................................13 2.3.1 Palmprint features ..........................................................................14 2.3.2 Palmprint recognition and problem statement ...............................15 2.3.3 Representative methods of palmprint recognition .........................19
2.4 Review of tongue crack detection .............................................................23 2.4.1 Review of computerized tongue diagnosis ....................................23 2.4.2 Tongue crack ..................................................................................25
2.5 Conclusions...............................................................................................26
3 Wide Line Detection.............................................................................................28 3.1 Introduction...............................................................................................28 3.2 Model design.............................................................................................29
3.2.1 1D line profile model .....................................................................29 3.2.2 2D line detection ............................................................................30
3.3 Line detection method...............................................................................33 3.4 Parameter selection ...................................................................................36
3.4.1 Radius of circular mask, r ..............................................................37 3.4.2 Brightness contrast threshold, t ......................................................43
3.5 Experimental results..................................................................................48 3.5.1 Synthesized image..........................................................................48 3.5.2 Real image......................................................................................48
3.6 Conclusion and discussions ......................................................................51
4 Analyses of the Wide Line Detector....................................................................56
iii
4.1 Introduction...............................................................................................56 4.2 Background ...............................................................................................58
4.2.1 Signal-to-noise ratio .......................................................................58 4.2.2 ROC curve......................................................................................58
4.3 Core function.............................................................................................60 4.3.1 SECH v.s. EXP...............................................................................60 4.3.2 The power of the hyperbolic secant function.................................64
4.4 Radius of circular mask.............................................................................66 4.4.1 Operating radius v.s. real radius.....................................................66 4.4.2 Relationship of mask radius and detected line width.....................68
4.5 Line orientation .........................................................................................72 4.6 Time complexity .......................................................................................74 4.7 Robustness to impulse noise .....................................................................74 4.8 Summary ...................................................................................................79
5 Palmprint Recognition by Using Line Features ................................................80 5.1 Introduction...............................................................................................80
5.1.1 Palm lines .......................................................................................82 5.1.2 Review ...........................................................................................83
5.2 Line-like feature extraction.......................................................................84 5.3 Palm-line feature matching .......................................................................90 5.4 Experimental results..................................................................................91
5.4.1 Palmprint database .........................................................................91 5.4.2 Parameters for the palm-line feature based palmprint recognition 91 5.4.3 Palmprint matching and parameter selection .................................92 5.4.4 Verification.....................................................................................98 5.4.5 Identification ................................................................................100
5.5 Conclusion and discussions ....................................................................101
6 Automatic Extraction of Cracks on Tongue Images .......................................102 6.1 Introduction.............................................................................................102 6.2 Methodology of crack extraction ............................................................105
6.2.1 Preprocessing ...............................................................................107 6.2.2 Extracting tongue cracks ..............................................................107 6.2.3 Post-processing ............................................................................112
6.3 Adaptive line width estimation ...............................................................114 6.4 Experimental results and analysis ...........................................................118
6.4.1 Performance evaluation................................................................119 6.4.2 Experimental results.....................................................................120
6.5 Conclusion and discussion ......................................................................129
7 Conclusion...........................................................................................................130 7.1 Main contributions ..................................................................................130 7.2 Future Work.............................................................................................132
Bibliography .............................................................................................................134
iv
Lists of Figures
Fig. 1.1 The framework of this thesis. ................................................................................. 5
Fig. 2.1 Operation of the line operator. (left) Average intensity, L (indicated by the various
shades of gray), is determined for a number of orientations. (right) Line strength,
S, is obtained as the difference between the maximum value of L (dark-gray
shaded points) and the average intensity, N, of a square region aligned with the
direction of maximum L (light-gray shaded points). [4] ...................................... 13
Fig. 2.2 Palmprint features. (a) Principal lines and datum points. (b) Wrinkles, ridges and
delta points............................................................................................................ 14
Fig. 2.3 Block diagram of palmprint recognition architecture........................................... 18
Fig. 2.4 The main steps of preprocessing. (a) Original image, (b) binary image, (c)
boundary tracking, (d) building a coordinate system, (e) extracting the central part
as a subimage, and (f) preprocessed result. [50]................................................... 20
Fig. 2.5 Flowchart of the automatic tongue diagnosis system in TCM. [8,83].................. 24
Fig. 3.1 Four circular masks at different positions on a line image based on the 1D ideal
line model IL as in (3-1). ...................................................................................... 31
Fig. 3.2 Five circular masks at different positions of a line image based on the 1D
common line profileGL as in (3-2). ..................................................................... 31
Fig. 3.3 Illustration of gray-level relations between three regions. IGr , IIrG and IIIGr
are gray levels of the line part, the edge region and the background, respectively.
The difference of the gray levels between the line ( IGr ) and the background
( IIIGr ) is larger than the brightness contrast threshold t (a) with or (b) without
larger than t×2 . (i), (ii), (iii) and (iv) show the corresponding relations in (3-4),
respectively. .......................................................................................................... 32
Fig. 3.4 Brightness difference versus the similarity functions defined in (3-3) and (3-11),
respectively. Here the brightness contrast threshold t is set to 10. ....................... 35
Fig. 3.5 Illustration of Proposition 1 and its proof............................................................. 39
Fig. 3.6 Illustration of the inequality relation about the ratio of the width of the line
detected to the radius of a circle mask with normalized constant weighting........ 41
Fig. 3.7 (a) A segmented palmprint image. Palm-line response images obtained using
v
brightness contrast thresholds t of (b) 6, (c) 7, (d) 8, (e) 9, (f) 10, (g) 11, (h) 15, (i)
20, and (j) 25......................................................................................................... 46
Fig. 3.8 (a) A test image including straight lines and curves of different widths. The line
detection results obtained using (b) the edge-based line finder, (c) the ridge-based
line detector, and (d) the proposed wide line detector. ......................................... 47
Fig. 3.9 (a), (b) Aerial images and (c) an X-ray image taken from [3]. (d)-(f) The
extraction results of line positions and line widths reported in [3] where line
positions are displayed in white with the corresponding edges displayed in black.
(g)-(i) The corresponding line detection results obtained using our wide line
detector are drawn in white. ................................................................................. 53
Fig. 3.10 The segmented palmprint images (first row) and the palm-line detection results
obtainedusing the edge-based line finder (second row), the ridge-based line
detector (third row), and our wide line detector (last row). Detected lines are
displayed in black. ............................................................................................... 54
Fig. 3.11 Segmented tongue images (first row) and crackle detection results obtained using
the edge-based line finder (second row), the ridge-based line detector (third row)
and the proposed wide line detector (last row). Detected crackles are displayed in
red. ....................................................................................................................... 55
Fig. 4.1 The ROC space and plots of the four prediction examples. [97].......................... 59
Fig. 4.2 Brightness difference versus the similarity formulae defined in (4-2), (4-3), and
(4-4), respectively. Here the brightness contrast threshold t is set to 10, 5=JS for
(4-3), and 6=JE for (4-4).................................................................................. 61
Fig. 4.3 (a) A test image including straight lines and curvilinear structures of different
widths. (b) The ROC curves of (a) for the 0-1 function, the exp to 6th power
formula, and the sech to 5th power formula. ......................................................... 63
Fig. 4.4 The ROC curves of Fig. 4.3a for different functions and different weighting. .... 65
Fig. 4.5 (a) A synthesized image with a bright line of 5 pixels wide. The pixel values are
marked. (b) and (c) The line response images by using the 0-1 function and the
SECH to 5th power function, respectively. The line responses on the center pixel
and at the most left pixel are shown. .................................................................... 65
Fig. 4.6 An example of circular mask approximated by a square kernel of 1515× .......... 67
Fig. 4.7 The plots of WMSB mass of lines with width of (a) 5 pixels, (b) 15 pixels, and (c)
25 pixels, respectively........................................................................................... 71
vi
Fig. 4.8 The line with orientation of (a) 15 degree, (b) 30 degree, (c) 60 degree, and (d) 75
degree, respectively. ............................................................................................. 71
Fig. 4.9 The test image (Fig. 4.3a) is added impulse noise with the noise density of (a)
0.01 and (b) 0.1, respectively. Line strength images by using our wide line
detector to (a) and (b) are shown in (c) and (d), respectively. (e) and (f) are final
results of line detection by performing morphological operations to (c) and (d),
respectively. .......................................................................................................... 75
Fig. 4.10 Line detection results of the noisy image Figure 4.9a using (a) Koller’s
edge-based line finder [26] and (b) Steger’s ridge-based line detector [3]........... 76
Fig. 4.11 Filter the noisy image Figure 4.9(a) using median filters with three different
neighbors: (a) a 33× square, (b) a 31× horizontal line, and (c) a 51× horizontal
line. ....................................................................................................................... 78
Fig. 5.1 The main patterns in a palmprint. [13] ................................................................. 82
Fig. 5.2 Brightness difference versus the similarity defined in the core functions (5-1),
(5-4), and (5-5), respectively. Here the brightness contrast threshold t is set to 10.
.............................................................................................................................. 85
Fig. 5.3 A 3D plot of the WMSB mass given a test image with a 3 pixels wide line. ....... 86
Fig. 5.4 The palmprint sub-images (the first row) with the size of 128128× , the
corresponding line-strength images (the second row) using the wide line detector,
the palm-line feature images (the third row) via thresholding the line-strength
images, and the post-processed palm-line feature images (the last row) with the
size of 6464× .................................................................................................... 87
Fig. 5.5 Palm-line feature extraction results using the WLD-based palm-line feature
extraction method with different brightness contrast thresholds t and different
radii of the circular mask r. For 0=t , the core function defined in (5-5) is
employed. ............................................................................................................. 89
Fig. 5.6 EER (%) values for different combinations of brightness contrast threshold and
radius of circular mask. Here the standard deviation of Gaussian smoothing filter
is 75.0=σ and the translation range is 2=s . ................................................... 93
Fig. 5.7 EER (%) values for different parameter combinations of circular mask radius and
translation range given the standard deviation of Gaussian smoothing filter (a)
0.1=σ and (b) 25.1=σ , respectively. Here the brightness contrast threshold
is 0=t . ................................................................................................................. 96
vii
Fig. 5.8 Genuine and impostor distributions for the optimal parameter combination. ...... 98
Fig. 5.9 Verification performance comparison on PolyU Palmprint Database. ................. 99
Fig. 6.1 The diagram of tongue crack extraction procedure using the WLD................... 106
Fig. 6.2 Preprocessing of tongue crack extraction. (a) Captured tongue image, (b) contour
tongue image with the centroid and corresponding contour, (c) tongue body mask,
and (d) extracted tongue body image. ................................................................ 108
Fig. 6.3 Tongue crack extraction. (a) The ground truth for Fig. 6.2d, (b) The line-strength
image using the wide line detector (WLD), (c) tongue crack image obtained by
postprocessing (b), and tongue crack extraction results using (d) the Line Operator
(LO) and (e) Steger’s method, respectively. ....................................................... 109
Fig. 6.4 Illustration of the line width estimate. The first row is the three dark line models.
The second row shows the corresponding Gaussian smooth profiles. The last row
is the first-order derivatives of Gaussian (DoG) profiles shown in the second row.
The line width can be obtained by calculating the distance between the pair of
minimum and maximum of DoG filtered results along the line profile. ............ 115
Fig. 6.5 Line width estimation algorithm......................................................................... 117
Fig. 6.6 (a) The vertical cracks with (b) the ground truth and extraction results by using (c)
WLD, (d) LO, and (e) Steger’s method, respectively. ........................................ 121
Fig. 6.7 (a) The horizontal cracks with (b) the ground truth and extraction results by using
(c) WLD, (d) LO, and (e) Steger’s method, respectively.................................... 123
Fig. 6.8 (a-b) Two irregular cracks with (c-d) the corresponding ground truth and
extraction results by using (e-f) WLD, (g-h) LO, and (i-j) Steger’s method,
respectively. ........................................................................................................ 126
Fig. 6.9 Box-and-whisker plots of the PM values of three methods for all test cracked
tongue images. .................................................................................................... 128
viii
Lists of Tables
Table 2.1 Edge and Line detection methods. ..................................................................... 7
Table 2.2 Comparison of biometric techniques [46,56]. A low ranking indicates poor
performance in the evaluation criterion whereas a high ranking indicates a very
good performance............................................................................................. 16
Table 2.3 Comparison of technology risk rating [57]. ..................................................... 17
Table 2.4 Seven types of tongue cracks. [75]................................................................... 27
Table 3.1 The deviation of function PtyxIyxIh )/),(),((sec 00− from (3-3). Here c is
the subtraction of (3-11) from (3-3), and c is the absolute value of c.......... 35
Table 3.2 Relationship between radii of Gaussian masks and approximately critical
widths of lines (ACLW) detected. .................................................................... 42
Table 4.1 Look up table of core functions 1s and 2s with different powers and the
deviation of the two core functions from 0s ..................................................... 62
Table 4.2 Square kernel and the corresponding operating radius and real radius. Here the
operating radius is half the number of pixels along x or y-direction excluding
the center pixel. ................................................................................................ 67
Table 4.3 The values of x given different widths and different radii. Here the operating
radius is half the number of pixels along x or y direction excluding the center
pixel. 0L is the line pixel where WMSB mass of line pixels reaches the
maximum, while 0B is the background pixel where the WMSB mass of
background pixels reaches the minimum. ........................................................ 69
Table 4.4 The calculated φ . Here opR is the operating radius which is half the number of
pixels along x or y-direction excluding the center pixel................................... 73
Table 4.5 Comparisons of calculations per pixel for different line detection methods. The
size of the convolving kernel is NN × . .......................................................... 74
Table 5.1 Equal error rates (%) for different brightness contrast thresholds (t) and
different radii of circular masks (r). Here the standard deviation of Gaussian
smoothing filter is 0.75 and the translation range is 2. For t larger than 0, the
ix
core function defined in (5-1) is applied. For 0=t , the core function defined
in (5-5) is applied. ............................................................................................ 93
Table 5.2 Equal error rates (%) for different translation ranges (s) and different radii of
circular masks (r). Here 0=t and the standard deviation of Gaussian
smoothing filter takes two values, 1 and 1.25. ................................................. 95
Table 5.3 Equal error rates (%) for different radii of circular masks (r) and different
scales of Gaussian smoothing filter with 0=t and 5=s . The core function
defined in (5-5) is used..................................................................................... 97
Table 5.4 Comparison of our method with four representative palmprint verification
methods on PolyU database. ............................................................................ 99
Table 5.5 Comparison of identification accuracy on two registration databases. .......... 101
Table 6.1 Comparison of the relationship between radii of different profiles and
approximately critical widths of detected lines (ACWL). ............................. 113
Table 6.2 The estimation results of maximum widths of cracks by applying the algorithm
described in Figure 6.5. The true widths of cracks are also listed for reference.
........................................................................................................................ 117
Table 6.3 Comparisons of performance evaluation between different methods used in Fig.
6.3.................................................................................................................... 120
Table 6.4 Comparisons of performance evaluation between different methods used in Fig.
6.6.................................................................................................................... 122
Table 6.5 Comparisons of performance evaluation between different methods used in Fig.
6.7.................................................................................................................... 124
Table 6.6 Comparisons of performance evaluation between different methods used in Fig.
6.8a. ................................................................................................................. 127
Table 6.7 Comparisons of performance evaluation between different methods used in Fig.
6.8b.................................................................................................................. 127
Table 6.8 Comparisons of average performance evaluation between different methods.....
........................................................................................................................ 128
Table 6.9 Paired t-test results of different methods. The p values are calculated using the
paired t-test of the hypothesis that two matched samples in the two vectors
come from distributions with equal means. p<0.05 was considered statistically
significant. ................................................................................................................ 128
1
Chapter 1
Introduction
The objective of this chapter is to introduce the general concepts of line detection, including
the characteristics of lines in digital images and applications of line detection. We also
address the originality and the organization of this thesis.
1.1 Motivation
Lines, a generic term without making distinction between straight lines and curvilinear
structures, usually convey much relevant information of an image like edges [1,2]. Hence it
is important to detect lines reliably and effectively.
Line detection in digital images is an important low-level operation in image analysis and
computer vision. In photogrammetric and remote sensing tasks, it can be used to extract
linear features, such as roads, railroads, and rivers, from satellite or low-resolution aerial
imagery, which can be used for the capture or update of data for geographic information
systems [3]. It is also extensively used in medical imaging for extracting anatomical features
such as linear structures in mammograms [4], blood vessels from an X-ray angiogram [5] or
a retina image [6], the bones in the skull from a CT or MR image [7], and cracks on a tongue
surface in TCM [8]. In addition, line detection is useful in, for example, extracting strokes in
character recognition [9,10], in estimating motion and structure from multiple images [11,12],
and in detecting biometric traits for personal authentication [13].
A large number of line detection algorithms have been proposed, most of which focused
on locating line positions without considering line widths. However, a digital image line
generally appears as a straight line or a curvilinear structure with one or more pixels wide, i.e.
a thin/narrow or thick/wide line. Since width is one of the predominant and most informative
features that characterize a line, it is important not only to extract line locations but also to
2
detect the complete line. In this thesis, we address this issue by proposing a general method,
the wide line detector (WLD), which can extract the full cross-sectional extent of lines at
each location along them. We also study two applications of the proposed method: palm-line
based palmprint recognition and adaptive tongue crack detection.
1.2 Statements of Originality
Figure 1.1 illustrates the framework of this thesis as well as the corresponding contributions
which are claimed to be original as follows:
1. A novel wide line detector using an isotropic nonlinear filter is presented. This detector
is based on the isotropic responses via circular masks. Unlike most existing edge and
line detectors which use directional derivatives, our proposed wide line detector applies
a nonlinear filter to extract a line completely without any derivative.
2. A model for the common line profile, named the edge-based bar-shaped line, is defined
to design the wide line detector.
3. The dynamic selection of parameters is developed to automatically determine the two
parameters – the brightness contrast threshold and the radius of circular mask – of the
proposed wide line detector for an input image.
4. A general scheme is introduced for analyses of the proposed wide line detector. The
optimal core function is determined for this detector by evaluating the performance of
various formulae.
5. A novel palm-line feature extraction method based on our wide line detector is proposed
for palmprint recognition. This palm-line feature extraction method can detect line-like
features which convey both structure features and strength features of palm lines.
6. A translation-invariant similarity measure is introduced for palm-line feature matching
which can be implemented reliably and effectively.
7. An experimental scheme is designed to determine the optimal combination of
3
parameters for the proposed palm-line feature based palmprint recognition methodology,
which are the brightness contrast threshold and the radius of circular mask for the wide
line detector, the Gaussian smoothing filter scale for post-processing, and the translation
range for palm-line feature matching.
8. A framework is proposed for adaptive tongue crack detection.
9. An adaptive tongue crack detection scheme is derived based on the proposed wide line
detector. This scheme allows totally automatic extraction of tongue cracks
10. An algorithm of maximum line width estimation is designed to determine the maximum
width of tongue cracks in a tongue image, which is related to the radius of the circular
mask, one of parameters for the proposed wide line detector.
1.3 Organization of this thesis
Following the global introduction, the thesis proceeds in Chapter 2 with an overview of line
detection techniques, palmprint recognition, and tongue cracks. Special attention is given to
the classic wide line detection approaches and palmprint recognition algorithms which are
compared to our proposed methods in the thesis.
The subsequent parts (Chapters 3-6) present the scientific contributions of the research.
All the chapters correspond to work that has been published or is currently under review in
peer-reviewed journals.
In Chapter 3, we concentrate on the detection of the complete line. We present a wide line
detector to extract the full cross-sectional extent of lines at each location along them. In
contrast with the traditional directional-derivative-based edge and line detector, the proposed
wide line detector applies an isotropic nonlinear filter to extract a line completely without
any derivative. A dynamic selection of parameters is developed by the analysis of robustness
of the proposed wide line detector. In addition, this chapter investigates the relationship
between the size of circular masks and the width of detected lines.
4
We then analyze the proposed wide line detector in Chapter 4. The optimal core function
for the proposed wide line detector is determined by comparing the performance of various
formulae and the results show the power of the hyperbolic secant function which is
employed as the core function in Chapter 3. The relationship between the real radius and
operating radius is described and the restriction on the real radius presented in Chapter 3 is
illustrated. The calculation of line orientation defined in Chapter 3 is also illustrated to show
its reasonableness. In addition, a comparative study is conducted to demonstrate the
robustness of our method to impulse noise.
In Chapter 5, we address the first application of the wide line detector, a novel palm-line
feature extraction method for palmprint recognition. This wide line detector based method
can extract both structure features and strength features of palm lines. An isotropic nonlinear
filter is employed to calculate the line strength of palmprint. Palm-line feature images are
obtained via binarilization of line strength images. A Gaussian smoothing filter serves as
post-processing for palm-line feature image denoise. A translation-invariant similarity
measure is introduced for palm-line feature matching. An experimental scheme is also
designed to determine the optimal combination of parameters for the proposed method.
In Chapter 6, we, for the first time, attempt to detect tongue cracks, one of pathological
features in tongue diagnosis. A framework is described for adaptive tongue crack detection
and consequently an adaptive tongue crack detection scheme is derived based on the
proposed wide line detector. This scheme can extract tongue cracks totally automatically. A
novel algorithm of maximum line width estimation is also designed to automatically
calculate the maximum width of cracks for an arbitrary input tongue image; the maximum
width of cracks determines the radius of circular mask, one of the parameters required for the
proposed wide line detector.
Finally, we give the conclusions of our work in Chapter 7, where some suggestions for
further development are also provided.
5
Fig. 1.1 The framework of this thesis.
Performance Evaluation
4. A general scheme is introduced for analyses of the proposed wide line detector.
Wide Line Detector
1. A novel wide line detector using an isotropic nonlinear filter is presented;
2. A model for the common line profile is defined;
3. The dynamic selection of parameters is developed.
Adaptive Tongue Crack Detection
9. An adaptive tongue crack detection scheme is derived;
10. An algorithm of maxi- mum line width estima- tion is designed.
Palm-Line based Palmprint Recognition
5. A palm-line feature extraction method is proposed based on WLD for palmprint recognition;
6. A translation-invariant similarity measure is introduced;
7. An experimental scheme is designed to determine the optimal parameter combination.
Our Proposed Method Applications Analyses
8. A framework for the adaptive tongue crack detection is proposed;
6
Chapter 2
Literature Review
In this chapter, we introduce the general concepts and methods of line detection. We briefly
review some well-known line detection techniques which consider line width. We also
review the methods for palmprint recognition which will be compared to our proposed
method in Chapter 5. We finally introduce the concepts of tongue cracks.
2.1 Overview of line detection
Line detection is considered one of the most fundamental tasks in image analysis and
computer vision. Consequently, much research has been done in this area and a considerable
body of literature has been accumulated. Table 2.1 illustrates the typical techniques used for
line detection. Surveys of edge and line detection may be found in [133, 135].
Most work on line detection can be classified into three primary categories: Hough transform
(HT), parallel edge extraction, and Hessian-based ridge detection. In this section, we will
introduce and discuss the three techniques for line detection.
2.1.1 Hough transform
The Hough transform [14-15, 92-93], which is a widely used line detection method [136], is
the classical technique for finding analytically defined curvilinear structures (e.g., straight
lines, circles, ellipses etc.). It transforms a binary image into Hough parameter counting
space. Any straight line in the image is represented by a single point in the parameter space
and any part of this straight line is transformed into the same point. The main idea of the
Hough transform is to determine all the possible line pixels in the image, to transform all
lines that can go through these pixels into corresponding points in the parameter space, and
to detect the points in the parameter space that frequently resulted from the Hough transform
7
Table 2.1 Edge and Line detection methods.
Category Methods Description
Hough Transform Hough Transform Voting in the dual plane.
Roberts operator ⎥⎦
⎤⎢⎣
⎡−
=10
011h , ⎥
⎦
⎤⎢⎣
⎡−
=0110
2h
Prewitt operator⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−=
111000111
1h , ⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
=101101101
1h
Sobel operator ⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−=
121000121
1h , ⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
=101202101
1h
1st derivative
Canny edge detection
Optimal for step edges by the first derivative of Gaussian smoothing filter with standard deviationσ .
Laplacian of Gaussian (LoG)
A second derivation of a smoothed 2D function, non-directional (isotropic)
Mexican hat The inverted LoG operator
Difference of Gaussians (DoG)
The difference of two Gaussian averaging masks with substantially different σ .
2nd derivative
Ridge detection Local maximum or minimum in the direction of the main principal curvature, the interior of elongated
objects in the image domain.
Thresholding Adaptive thresholding
Segmentation using variable thresholds, the threshold value varies over the image as a function of local image
characteristics.
8
of lines in the image [16]. Although the original Hough transform [14] was designed to
detect straight lines and curves, this original method can be used to detect any analytically
defined shape. The Hough transform is particularly useful when the curves one is looking for
are sparsely digitized, have “holes” and/or the images are noisy.
If the analytic expression of the desired curves is known, the Hough transform works very
well. Unfortunately, this is not often true in practice. In this case, a generalized Hough
transform [17-19] can offer the solution. This method constructs a parametric curve
description based on sample situations detected in the learning stage and can be used to
detect arbitrary shapes. However, when a complex curvilinear structure needs to be detected,
the Hough transform becomes quickly so high dimensional that it is not feasible. Moreover,
although the generalized Hough transform can be used to detect arbitrary curvilinear
structures in theory, it requires the complete specification of the exact shape of the target
object to achieve precise segmentation [16], which is often difficult and not available for
complex curvilinear structures in practice.
2.1.2 Parallel edge extraction
The second technique is based on edge extraction. It treats lines as objects with parallel
edges [20-26,94,134] and accordingly adopts edge detection for preprocessing. It is common
to first find all edges in the image and then analyze the edge image to find the line by
detecting parallel opposing edges. The edge detection algorithm used in the early stage in
most cases [22,26] is the first derivative of the Gaussian, a well known edge detector [27].
Actually, this technique converts the problem of line detection to edge detection and thus
strongly depends on reliable edge detection. Due to the first derivative of the Gaussian used
for edge extraction, this line detection technique is sensitive to image noise. For the parallel
edge extraction based line detection technique, the detection of the line is now a two step
process: edge detection and line finding. If an error is made during the edge detection stage,
it cannot be corrected in the line finding stage.
9
2.1.3 Hessian-based ridge detection
For the final technique, lines are regarded as ridges and valleys in the image [28]. A ridge
point is differential-geometrically defined as location for which the gray value assumes a
local maximum or minimum in the direction of the main principal curvature which
corresponds to the maximum eigenvalue of the Hessian [29,30]. Hence, this technique
extracts ridges by computing the eigen-decomposition of the Hessian at each image pixel and
find line points by selecting pixels that have a high second directional derivative
perpendicular to the line direction [30-35]. Ridge strength can be measured by the LA norm−γ
index [30], which is designed to minimize the response to blobs.
The advantage of the Hessian-based ridge detection technique is that lines can be detected
with a high accuracy [29]. However, this technique usually leads to multiple responses to a
single line and returns wrong line locations if the line has different lateral contrast [3].
Although this technique can be iterated through scale-space to obtain a coarse estimate of the
line width [30], the line position at the optimal scale for width estimation will in general be
severely biased. In addition, this approach is sensitive to noise due to the use of second order
of derivatives.
2.2 Review of wide line detection
The importance of detecting wide lines, i.e. the full cross-sectional extent of lines, is often
underestimated in line detection, but it may be as significant as the location of line positions.
In the past decade, there has been increasing interest in the wide line detection [3,4,26,36-39]
and the techniques developed in the literature can be divided into the three categories
introduced in Section 2.1.
For the Hough transform, since the analytic expression of desired lines is required, the
wide/thick line detection based on HT only focuses on straight lines/bands [37,38]. Such
approaches are expected to be slow due to a large number of votes and require a large
10
amount of memory. Also, it does not seem to be able to handle straight lines of non-uniform
brightness and width as well as curvilinear structures.
One possible powerful approach for wide line detection is to use edge detection followed
by a method for detecting parallel opposing edges [25,26]. The other is to apply ridge
extraction at a proper scale to detect line positions as well as line width [3]. However,
dedicated line detectors are generally more successful [4,36].
The adaptive thresholding technique, which is widely used for document image analysis
[128-131], can also be used to detect edges and segment wide lines [132]. However, it
appears that none could extract the wide line well with a set of operating parameters when
the background is too complex, i.e., the illumination changes dramatically and there are too
much structured noise.
In this section, we introduce three specific wide line detection approaches for our analysis,
[3], [4], and [26], which have been applied in a number of papers [39-44] and will be
compared to our proposed methods in the following Chapters.
2.2.1 Edge-based line finder
In their work, Koller et al. [26] aimed to propose an algorithm which is independent of scale
and does not involve the specification of width parameters to detect curvilinear structures of
different widths and orientations. This algorithm first calculates the line direction by
analyzing the directional derivatives at the appropriate scales and then employs an edge
detector, the first derivative of the Gaussian, perpendicularly to the line direction to detect
the parallel opposing edges of the line. To overcome the multiple line response and the
sensitivity to edges, the response of the two specially tuned edge detection filters is
combined in a nonlinear way. This edge-based line finder algorithm can be implemented by
five steps:
Step 1: Create the scale-space image ss GLL ⊗= , where L is the input image and sG is the
Gaussian filter with its variance set equal to the scale parameter s;
11
Step 2: Calculate the gradient sL∇ ;
For all points x :
Step 3: Determine the direction )sin,(cos αα=od , where
)(2)2(tan 1yyxxxy LLL −=− α ;
Step 4: Obtain the edge responses ddsxLR sl ⋅+∇= )( and ddsxLR sr ⋅+−∇= )( ;
Step 5: Calculate the filter response ))(,)(min(),( rrllrl RRsignRRsignRRF ⋅⋅= , where
1)( =xsign for 0>x .
The advantage of this algorithm is that, since derivatives of Gaussian kernels are used, it
can detect lines of arbitrary widths by iterating in scale space and selecting as the line width
the scale that yields the maximum of a scale-normalized response. However, since special
filters need to be constructed for several or even all possible line directions, the edge-based
line finder algorithms are computationally very expensive. Furthermore, it obtains a coarse
estimate of the line width by iteration through the scale space, which makes the algorithm
even more expensive.
2.2.2 Ridge-based line detector
A well-founded algorithm for the line detection is given by Steger [3]. He proposes an
explicit model for lines and their surroundings and carries out a scale-space analysis for this
line model. This analysis leads to an algorithm which can extract lines and line widths with
high precision. In this algorithm, an input image is first convolved with a Gaussian kernel,
and then each pixel is assumed to be at the center position of a curvilinear structure. To test
whether the assumption is correct for the pixel under consideration, the first and second
directional derivatives of the Gaussian smoothed image are computed along the local line
direction which is the main principal curvature corresponding to the maximum eigenvalue of
the Hessian. A line point is detected at the zero-crossing of the first derivative of the image
taken in the direction of the eigenvector of the Hessian matrix, the Hessian being given by
12
the local second order derivatives:
⎥⎦
⎤⎢⎣
⎡=
⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢
⎣
⎡
∂∂
∂∂∂
∂∂∂
∂∂
=yyyx
xyxx
LLLL
yL
xyL
yxL
xL
yxH
2
22
2
2
2
),( . (2-1)
The width (w) of the curvilinear structure to be found depends on the scale (σ ) of the
Gaussian kernel which should satisfy
3w
≥σ . (2-2)
This ridge-based line detector overcomes the problem that the ridge-based line detection
approach will return inaccurate line locations when the contrast on one side of the line is
different from the contrast on the other side of the line. Since Gaussian kernels are used to
measure the derivatives of the image, the line detector can scale to lines of arbitrary widths.
However, since this approach extracts lines as the maxima of the magnitude of the second
directional derivatives, it can detect only salient lines. Moreover, this line detector will fail
when the curvilinear structures cross each other because the direction of a line is estimated
by an eigenvector corresponding to the maximal absolute value of the Hessian matrix
eigenvalue. In addition, once the selected σ becomes so large that neighboring lines start
to influence each other, the line model will not hold and the results will deteriorate.
2.2.3 Line operator
Taylor et al describe a line operator [4,45], the principle of which is illustrated in Fig. 2.1.
The Line Operator improves the line signal to background noise ratio by taking the average
gray level, L, of the pixels lying on an orientated local line passing through the target pixel
and subtracting the average intensity, N, of all the pixels in the locally orientated
neighborhood. The values of pixels falling on the border of the line or neighborhood are
weighted according to the area of the relevant pixel falling within the line or neighborhood.
The line-strength S is calculated by S = (L-N) and compared for n orientations. Line
direction is obtained from the orientation producing the maximum line strength. Line scale
13
can be obtained by applying this algorithm at multiple scales either by subsampling the
image or changing the local area size. For each pixel, the scale producing the maximum line
strength is taken as the detected line scale.
This algorithm is simple and effective for line detection. Its sensitivity makes the Line
Operator suitable to detect low-contrast lines and narrow lines. However, the detection
results may be corrupted by image noise or changes of contrast.
2.3 Review of palmprint recognition
In this section, we first introduce the palmprint features and then provide a short review on
palmprint recognition followed by the description of problem statement. We also briefly
review on some well-known palmprint recognition methods which will be compared to our
methods in Chapter 5.
Fig. 2.1 Operation of the line operator. (left) Average intensity, L (indicated by the various
shades of gray), is determined for a number of orientations. (right) Line strength, S, is
obtained as the difference between the maximum value of L (dark-gray shaded points) and
the average intensity, N, of a square region aligned with the direction of maximum L
(light-gray shaded points). [4]
14
(a) (b)
Fig. 2.2 Palmprint features. (a) Principal lines and datum points. (b) Wrinkles, ridges and
delta points.
2.3.1 Palmprint features
Biometrics, which deals with the problem of identifying a person based on human physical
and behavioral characteristics, is a better alternative to conventional authentication
approaches due to the inherency, universality, and uniqueness of biometric features [46].
Palmprpint, one of biometric features, is relatively stable and unique physical characteristic
[47]. As shown in Fig. 2.2, in a palmprint, there exists rich and useful information which can
be categorized as follows:
Principal Lines: Principal lines consist of hear line, head line and life line. They are the
most visible palm lines in a palmprint. Both the location and form of principal lines are
very important physiological characteristics for identifying individuals because they
vary little over time [48].
Wrinkles: Wrinkles are generally thinner and more irregular than principal lines.
15
They are classified as strong wrinkles and weak wrinkles so that more detail features
can be acquired.
Datum points: Datum points are two endpoints of head line and heart line which
intersect the both sides of palm. These provide a stable way to register palmprints.
Delta points: Delta point is defined as the center of a delta-like region in the palmprint
and provides stable and unique measurement for palmprint recognition [48].
Ridges: The palmprint is basically composed of ridges which are also a significant
measurement [48].
Compared with fingerprint, the most widely used biometric feature, palmprint contains not
only fingerprint-like features such as delta points and ridges but also additional distinctive
features such as principal lines and wrinkles. Therefore, it is possible to build a highly
accurate biometric system by palmprint.
2.3.2 Palmprint recognition and problem statement
Palmprint recognition, as a relatively new biometric technique, has drawn more and more
attention in the past few years. The comparison of various biometric techniques summarized
in Table 2.2 indicates that palmprint, on average, is more suitable than other biometrics for
personal recognition. Furthermore, the hand-scan technology is low risk rating in general, as
shown in Table 2.3. Moreover, compared with other hand-based biometrics such as
fingerprint and hand geometry, palmprint is more user-friendly, more cost effective, and
requires fewer data signatures but has a higher accuracy [13].
Palmprint recognition systems can be classified into two categories: offline and online
[50]. An offline system usually processes previously captured palmprint images, which are
often obtained by inked palms before digitization. An online system captures palmprint
images using a palmprint capture sensor that is directly connected to a computer for
real-time processing. Much research has done on offline palmprint images and some useful
results have been obtained [48,51-55]. Recently, several researchers have started working on
online palmprint recognition.
16
Table 2.2 Comparison of biometric techniques [46,56]. A low ranking indicates poor
performance in the evaluation criterion whereas a high ranking indicates a very good
performance.
17
Table 2.3 Comparison of technology risk rating [57].
Biometrics Verification
/Identification
Overt
/Covert
Behavioral
/Physiological
Give
/Grab Risk
Rating
Facial-scan high high medium high high
Finger-scan high medium high medium high
Hand-scan low low medium low low
Iris-scan high low high medium medium
Keystroke-scan low medium low low low
Retina-scan high low high low medium
Signature-scan low low low low low
Voice-scan low high low medium medium
A biometric system may operate either in verification mode (one-to-one matching) or
identification mode (one-to-many matching) [56]. In this thesis, we use the generic term
recognition without making a distinction between verification and identification. In a
palmprint recognition system, there are two stages: enrolment and recognition. Both stages
comprise three subsystems: palmprint image acquisition, palmprint image preprocessing and
feature extraction. Recognition stage consists of an additional subsystem - palmprint
matching. Figure 2.3 illustrates the block diagram of palmprint recognition architecture.
Problem statement
The key issue of palmprint recognition is how to characterize a palmprint stably and
effectively. Various palmprint representations have been proposed for recognition, such as
point features [51,58], line features [48,59-62], energy features [55,63], Fourier-domain
representation [64,65], algebraic representation [66-68], texture features [50,69-71], wavelet
signatures [13], etc. Since Zhang et al [50] extracted the texture features to describe a
palmprint and established an efficient and effective on-line palmprint recognition system,
texture feature has become the most popular representation used for palmprint recognition
and have achieved good performance in terms of speed and accuracy.
18
Fig. 2.3 Block diagram of palmprint recognition architecture.
Actually, it is the line feature that plays the most important role in palmprint recognition.
In [60,61], Wu et al for the first time made use of palm line structural features for
recognition. Observing palmprint, we can find that principal lines are very strong and look
wider than wrinkles. That is, many palm lines are not one pixel wide lines and have different
strength or width. The strength feature, or width information, of palm lines is very important
to characterize a palmprint clearly especially when different palmprints have similar line
structures. Hence, not only the structure feature but also the strength feature of palm lines is
necessary and important for palmprint recognition. In Chapter 5, we will describe our
proposed method for palmprint recognition by extracting both structural features and
strength features of palm lines.
19
2.3.3 Representative methods of palmprint recognition
In this subsection, we will briefly introduce the representative palmprint recognition
approaches which will be compared to our proposed method in Chapter 5.
PalmCode
PalmCode [50] is a 2D Gabor phase coding scheme for palmprint recognition. In the
preprocessing, a finger-gap-based algorithm is proposed to segment a normalized subimage
for reliable feature measurements. The five major steps (see Fig. 2.4) of the preprocessing
are [50]:
Step1: Apply a lowpass filter to the original image and convert the convolved image to a
binary image, as shown in Fig. 2.4b.
Step 2: Extract the boundaries of the holes, )2,1(),,( =iyFxF jiji , between fingers, as
shown in Fig. 2.4c.
Step 3: Compute the tangent of the two gaps. If there is a line ( cmxy += ) passing through
a pair of points on ),( 11 jj yFxF and ),( 22 jj yFxF satisfies the inequality,
cxmFyF jiji +≤ , for all i and j (see Fig. 2.4d), then the line is considered to be the tangent
of the two gaps.
Step 4: Line up the pair of points determining the tangent of the two gaps to get the y-axis of
the palmprint coordinate system. The origin of the coordinate system is the midpoint of the
two points (see Fig. 2.4d).
Step 5: Extract a subimage of a fixed size (normally 128128× ) based on the coordinate
system (see Fig. 2.4e). The subimage should be located at a certain area of the palmprint
image for feature extraction. The preprocessed palmprint image (the normalized subimage) is
shown in Fig. 2.4f).
20
(a) (b)
(c) (d)
(e) (f)
Fig. 2.4 The main steps of preprocessing. (a) Original image, (b) binary image, (c)
boundary tracking, (d) building a coordinate system, (e) extracting the central part as a
subimage, and (f) preprocessed result. [50]
The preprocessing aligns different palmprint images for matching and accordingly
required by any palmprint recognition method. In Chapter 5, we apply this preprocessing
algorithm to obtain the normalized subimage for our proposed method and the palmprint
recognition approaches employed to compare with our method.
In the stage of feature extraction, the normalized subimage is convolved with an adjusted
circular Gabor filter [50]
21
{ })sincos(2exp2
exp2
1),,,,( 2
22
2 θθπσπσ
σθ uyuxiyxuyxG +⎭⎬⎫
⎩⎨⎧ +−= , (2-3)
where 1−=i , 0916.0=u is the frequency of the sinusoidal wave, 4πθ = controls
the orientation of the function, 6179.5=σ is the standard deviation of the Gaussian
envelope. The signs of the filtered images are coded as a feature vector, which has been used
for iris recognition [72].
In palmprint matching, two PalmCodes are measured using the normalized hamming
distance. Due to imperfect preprocessing, translation is needed to get the final matching
score.
Competitive code (CompCode)
Competitive coding scheme [70] attempts to utilize the orientation information of palm lines
for palmprint recognition. Six orientations are specified for the competition according to the
neurophysiological findings [73]. The normalized subimage is convolved with the real parts
of the neurophysiology-based Gabor functions [73] along the six orientations, respectively,
to obtain the corresponding filter response of the palm line which is modeled in an
upside-down Gaussian shape. The orientation information of the palm line is determined by
applying a winner-take-all rule on the filtered results.
In the stage of palmprint matching, an angular distance is designed for comparing two
Competitive Codes. To implement real-time palmprint recognition, three bits are used to
represent an orientation. Based on this bit representation, a more effective and efficient
angular distance is defined for matching the Competitive Code.
Ordinal code
Sun et al [71] propose the orthogonal line ordinal features based on ordinal measures. This
method qualitatively compares two elongated, line-like image regions which are orthogonal
in orientation. The weighted average intensity of a line-like region is obtained using 2D
22
Gaussian filter [71]
⎥⎥
⎦
⎤
⎢⎢
⎣
⎡
⎟⎟⎠
⎞⎜⎜⎝
⎛ +−−⎟⎟
⎠
⎞⎜⎜⎝
⎛ +−=
22sincossincosexp),,(
yx
yxyxyxfδ
θθδ
θθθ , (2-4)
where θ denotes the orientation of 2D Gaussian filter, xδ and yδ denote the filter’s
horizontal scale and vertical scale, respectively. The scale ratio yx δδ should be higher
than 3 to make its shape like a line.
The orthogonal line ordinal filter, comparing two orthogonal line-like palmprint image
regions, is specially designed as follows [71]:
)2
,,(),,()( πθθθ +−= yxfyxfOF . (2-5)
For each local region in normalized subimage, three ordinal filters, )0(OF , )6(πOF ,
and )3(πOF , are performed on it to obtain three bit ordinal codes based on the sign of
filtering results. Finally, three ordinal templates named as ordinal code are obtained as the
feature of the input palmprint image. The matching metric is also based on Hamming
distance.
Palm-line structural features
In [61], Wu et al extract palm line structural features by using a set of directional line
detectors. They consider palm line as a kind of roof edge of which the position is determined
by the zero-crossing of their first-order derivative and the strength can be reflected by the
magnitude of their second derivative. Consequently, theθ -directional line detectors consist
of two filters which are convolution results of a 1D Gaussian smoothing filter in θ
direction with the first-order and the second-order derivatives of a 1-D Gaussian function in
2/πθ + direction, respectively. Four sets of directional line detectors in °0 -, °45 -, °90 -,
and °135 -directions are employed to extract the corresponding directional line images.
Finally, the palm line image is obtained by combining the four directional line images.
23
For palm-line matching, the palm line image is represented by the chain code [74]. The
chain code of a palm line is obtained by line tracing. The similarity of two palm line images
is measured by the ratio of the number of matched line points to the total number of line
points in two palm line images.
This palm-line extraction method achieves a good performance on a large-scale palmprint
database and considers only the structure features of palm lines. As discussed above, the
strength features, or width information, of palm lines is also very important to characterize a
palmprint clearly. Therefore, the performance can be improved further when considering the
strength features of palm lines, which will be addressed in Chapter 5.
2.4 Review of tongue crack detection
In this section, we first review the computerized tongue diagnosis and then introduce some
concepts of tongue cracks.
2.4.1 Review of computerized tongue diagnosis
Traditional Chinese Medicine (TCM) spans thousands of years and its practitioners have
accumulated very rich practical experiences in diagnostic methods. Tongue diagnosis [75,76]
ranks as one of the most important and widely used diagnostic methods in traditional
Chinese medicine. This method has proved, in practice, to be very valuable in clinical
applications and self-diagnosis. Furthermore, tongue diagnosis is a non-invasive technique
that is in accord with the most promising direction in the 21st century: no pain and no injury.
Traditional tongue diagnosis, however, has a very limited application in clinical medicine
due to its qualitative, subjective, and experience-based nature. Recently, researchers have
developed various methods and systems [8,77-86] to circumvent these problems. An
automatic tongue diagnosis system has been presented in [8,83] based on modern techniques
of image processing and pattern recognition, as illustrated in Fig. 2.5. A set of quantitative
and objective features and measurements have been developed according to the theories of
traditional tongue diagnosis. As many descriptive features in tongue diagnosis indicate some
24
implicit connections between color-related and texture-related features, various chromatic
features and textural features have been investigated for computerized tongue diagnosis.
In the aspect of extraction of chromatic features, Chiu [78] choose the HSL color model to
represent the color of tongue using saturation and luminance of the pixels inside the tongue
image. Yang [82] converts the tongue image from the color space of RGB to L*a*b so that
the color feature can be represented with only two parameters, *a and *b. Li and Yuen [80]
segment the tongue image into square blocks of 3636× , each represented by the mean of
its color pixels as three attributes using color values only in RGB color space and in CIE
LUV color space. Further, they investigate the color matching of tongue images by defining
a total of 10 different metrics in four color spaces (RGB, HSV, CIELUV, and CIELAB) [81].
Pang et al. [8,83-84] extract quantitative color features from four color spaces (RGB,
CIEYxy, CIELUV, and CIELAB) and use the mean and standard deviation of the colors of
pixels within the whole tongue region in all the four color spaces as the metrics for color.
Fig. 2.5 Flowchart of the automatic tongue diagnosis system in TCM. [8,83]
25
In the aspect of extraction of textural features, Chiu et al. [77] propose a structural texture
recognition algorithm which used certain features such as spatial gray-tone dependency
matrices (SGTDM) and Fourier power spectrum to verify or identify certain properties of
coating on the tongue. They further develop some SGTDM-based textural features such as
angular second moment, contrast, correlation, variance and entropy to determine the grimy
coating of a tongue [78]. Yuen et al. [79] apply the Gabor wavelet opponents color features
for tongue texture analysis, which contained both the multi-channel frequency properties and
color features. Yang [82] represents the texture features using Entropy and Energy functions
of a gray-level co-occurrence matrix. Pang et al. [8,83-84] employ two feature-based texture
descriptors, which are the second-order moment and the contrast measure of the gray-level
co-occurrence matrix, to extract different textural features from tongue images.
All of above methods have obtained encouraging results. However, none of them
considered physical and quantitative features in tongue diagnosis which have significant
pathological meanings. In Chapter 6, we will attempt to extract such feature and focus on
detecting tongue cracks, one of pathological features in tongue diagnosis.
2.4.2 Tongue crack
A cracked tongue [75] is frequently seen in clinical practice. The cracks on the tongue’s
surface resemble those that develop in soil after a prolonged period of drought. These cracks
can vary greatly in number and depth. They can be barely visible lines or extremely deep
fissures.
In general, there are no cracks in a normal tongue. The formation of cracks in the tongue
body may represent a lengthy pathological process. The most common cause of cracks is
dryness from exhaustion of the body fluids or yin. The clinical significance of cracks
depends on the tongue body color, the location of the cracks and their shape and depth. For
example, cracks in a pale tongue body can originate from a deficiency of blood. In the case
of deep cracks, besides deficiency of blood, there may also be injury to the fluids. A single,
small crack on the tongue body is less significant than cracks that are distributed over the
26
entire body of the tongue. In [75], tongue cracks are divided into seven types according to
the location and shape. Table 2.4 illustrates the seven types and their clinical significance in
tongue diagnosis. A common type of cracked tongue has a deep crack in the center reaching
to the tip, reflecting hyperactivity of Heart fire.
In spite of the clinical significance of the tongue crack, little work has been done on it.
Therefore, in Chapter 6, we will address the issue of tongue crack detection for tongue
diagnosis.
2.5 Conclusions
This chapter has introduced the primary techniques for line detection. We have reviewed
some well-know wide line detection approaches such as Koller’s algorithm [26], Steger’s
approach [3], and the line operator [4]. We have overviewed palmprint recognition and
reviewed the state-of-the-art palmprint recognition methods. We have also overviewed the
computerized tongue diagnosis and introduced the concepts of tongue cracks. In the
following chapters, we will present our proposed methods and compare them to those
existing approaches described in this chapter.
27
Table 2.4 Seven types of tongue cracks. [75]
Crack Type Clinical Significance Illustration
Horizontal cracks yin deficiency
Cracks like ice floes
yin deficiency from old age
Irregular cracks Stomach yin deficiency
Transverse cracks on sides
Spleen qi deficiency
Vertical cracks in center
Heart yin deficiency or blazing Heart fire
Transverse cracks behind tip
Lung yin deficiency
Very deep central crack with other small cracks
Kidney yin deficiency with heat
28
Chapter 3
Wide Line Detection
In this chapter, we present a novel wide line detector using an isotropic nonlinear filter.
Unlike most existing edge and line detectors which use directional derivatives, our proposed
wide line detector applies a nonlinear filter to extract a line completely without any
derivative directly used. The detector is based on the isotropic responses via circular masks.
A general scheme for the analysis of the robustness of the proposed wide line detector is
introduced and the dynamic selection of parameters is developed. In addition, this chapter
investigates the relationship between the size of circular masks and the width of detected
lines. A sequence of tests has been conducted on a variety of image samples and our
experimental results demonstrate the feasibility and effectiveness of the proposed method.
3.1 Introduction
The analysis of images in the fields of pattern recognition and computer vision generally
requires the detection of lines, also called curvilinear structures, from grayscale images. Line
detection plays an important role for the success of higher-level processing such as matching
and recognition [40,87-91]. Most of existing line detection methods focus on locating line
positions. Although the line position detection is important, it would be useful if all the line’s
pixels (i.e., the pixels that comprise the full cross-sectional width) were also extracted.
A line, mathematically, is a one-dimensional figure without thickness, but an image line
generally appears as a line of one or several pixels wide, i.e., as a thin/narrow or thick/wide
line, having linear or curvilinear structures. In image processing and pattern recognition
applications, line thickness or a line’s full cross-section is important in, for example,
segmenting multiple orientation lines [39], in recognizing roads, railroads, or rivers from
29
satellite or aerial imagery [20,34], in extracting anatomical features in medical imaging for
diagnoses [4,8,28], and in detecting biometric traits for personal authentication [13]. It is
thus important to detect not simply the line edges but rather the whole of the line.
In this chapter, we present a robust method to detect the whole of the line, which we call
the wide line detector. This wide line detector is implemented by employing a nonlinear
filter without the direct use of any derivative. Isotropic responses are obtained by using
circular masks which contain either a normalized constant weighting or a Gaussian profile.
We restrict the sizes of circular masks so as to ensure that lines of different widths can be
extracted in their entirety. We also design a line model which can analyze the robustness of
the proposed wide line detector and which allows the automatic selection of parameters.
The chapter is organized as follows. Section 3.2 introduces the line model and the relevant
design issues. Section 3.3 presents the isotropic nonlinear filter based line detection method.
In Section 3.4, we analyze the robustness of the proposed method and show how to
automatically select parameters. Section 3.5 describes the results of our experiments. Section
3.6 offers our conclusion.
3.2 Model design
In this section, we first introduce the models for line profiles in 1D and then describe the
design issues for the proposed line detection method with particular reference to 2D line
detection.
3.2.1 1D line profile model
Many line detection approaches model lines in 1D as bar-shaped [26], i.e., the ideal line of
width w×2 and height h is assumed to have a profile given by [3]
⎩⎨⎧
>≤
=wxwxh
xIL,0,
)( . (3-1)
However, the flatness of this profile is rarely found in real images. Generally speaking, in
30
terms of different gray levels, there are three regions in a line image: the line part having a
gray level IGr , the edge region having a gray level IIrG , and the background having a gray
level IIIGr . As an example, consider a bright line on a dark background, then we get
IIIIII rr GGGr >> . In this chapter, therefore, a model for the common line profile in 1D, the
edge-based bar-shaped line, is defined as
⎪⎩
⎪⎨
⎧
><<
≤×=
e
el
l
wx wxwb
wxhxGL
,0,,1
)( , (3-2)
where 0>> le ww and ]1,0[∈b and h is a scale or a vector, which represents height in the
1D case and intensity in the 2D case. For a bright line, also called a positive line, 0>h .
For a dark line, also called a negative line, 0<h .
3.2.2 2D line detection
We now address the design issues for the proposed line detection method. Fig. 3.1 shows a
dark line ( 0<h ) in 2D based on the 1D ideal line model IL as in (3-1). A circular mask is
shown at four image positions in Fig. 3.1. The detector groups pixels whose brightness is
similar to the brightness at the center of the mask into a weighted mask having similar
brightness (WMSB). This similarity can be measured by:
⎩⎨⎧
>≤
=tyx-IyxIiftyx-IyxIif
tyxyxs),(),( ,0),(),( ,1
),,,,(00
0000 , (3-3)
where ),( 00 yx is the coordinate of the center, ),( yx is the coordinate of any other pixel
within the mask, ),( yxI is the brightness of the pixel ),( yx , and t is the brightness contrast
threshold. The summation of the outputs s within the circular mask gives the mass of WMSB.
According to (3-3), when the center of the mask moves to a line on the image, the WMSB
reaches the global maximum as the mask lies in a flat region of the image (as the mask a
shown) and decreases when the center of the mask is very near a straight edge (as shown in
the mask b) and decreases even further when very near the straight edge while remaining
31
Fig. 3.1 Four circular masks at different positions on a line image based on the 1D ideal line
model IL as in (3-1).
Fig. 3.2 Five circular masks at different positions of a line image based on the 1D common
line profileGL as in (3-2).
nonetheless in the line region (as shown in the mask d). Therefore, the smaller the WMSB
mass, the larger the feature response. This is similar to the idea used for edge extraction and
corner detection in [49]. Hence, to detect a line completely, the WMSB mass of any pixel on
the line should be less than that of any background pixel.
Now let us consider a general situation: a line with an edge region in 2D based on the
common 1D line profileGL as in (3-2). Fig. 3.2 provides an illustration in which a dark line
a
b c
d e
Dark line
Gray edge
Bright background
Circle mask
Mask center a
bc d Dark line
Bright background
32
Fig. 3.3 Illustration of gray-level relations between three regions. IGr , IIrG and IIIGr
are gray levels of the line part, the edge region and the background, respectively. The
difference of the gray levels between the line ( IGr ) and the background ( IIIGr ) is larger than
the brightness contrast threshold t (a) with or (b) without larger than t×2 . (i), (ii), (iii) and
(iv) show the corresponding relations in (3-4), respectively.
is bounded by a gray band against a white background. In proximity to the lines are five
circular masks. According to (3-3), in order to completely detect the line it is necessary that
tGrGr >− IIII . It is obvious that no matter what value b and h take in (3-2), the three
different gray level regions (Fig. 3.3) have only four relations:
(i)⎪⎩
⎪⎨⎧
>−
≤−
tGrGr
tGrGr
IIIII
III (ii)⎪⎩
⎪⎨⎧
>−
>−
tGrGr
tGrGr
IIIII
III
(iii)⎪⎩
⎪⎨⎧
≤−
≤−
tGrGr
tGrGr
IIIII
III (iv)⎪⎩
⎪⎨⎧
≤−
>−
tGrGr
tGrGr
IIIII
III .
(3-4)
Now the question is how the edge region affects the WMSB mass of the line and the
background pixels and thereby influences line detection. These matters will be considered in
Section 3.4.2.
(ii) IIGr
(i) IIGr
IIIGr
IGr
t
t (iv) IIGr
(iii) IIGr
(i) IIGr
IIIGr
t
t (iv) IIGr
IGr
(a) (b)
33
3.3 Line detection method
The wide line detector is implemented to give isotropic line responses by applying a set of
rules in a circular mask. Using a square kernel, the circular mask is digitally approximated
by either with a constant weighting
⎩⎨⎧ ≤−+−
=otherwise
ryyxxifryxyx
0)()(1
),,,,(22
02
000ω , (3-5)
or with a Gaussian profile
2
20
20
2)()(
00 ),,,,( ryyxx
eryxyx−+−
−=ω , (3-6)
where r is the radius of the circular mask. The normalization of the circular mask is obtained
by the rule
∑+≤≤−+≤≤−
=
ryyryrxxrx
ryxyx
0000 ,
000 ),,,,(ω
ωω . (3-7)
As is usual when locally processing an image, the mask is centered on each pixel in the
image and the brightness of any other pixel within the mask is compared with that of the
center pixel. The comparison is determined by the rule defined in (3-3) along with a
weighting function 0ω :
),,,,(),,,,(),,,( 0000000 tyxyxsryxyxyxyxc ×= ω , (3-8)
where c is the output of the weighting comparison. This comparison is done for each pixel
within the mask. The WMSB mass of the center ),( 00 yx is given by
∑+≤≤−+≤≤−
=
ryyryrxxrx
yxyxcyxm
0000 ,
0000 ),,,(),( . (3-9)
The initial line response L is the inverse WMSB mass obtained by
34
⎩⎨⎧ <−
=otherwise
),(m if0
),(),( 0000
00
gyxyxmgyxL . (3-10)
Here g is the geometric threshold and 2/maxmg = , where maxm is the maximum value
which m can take (usually 2max rm π= ). The fixed threshold g for (3-10) is the theoretical
optimum which is shown in the later analyses.
According to (3-3), the comparison s varies dramatically when a slight change of the
brightness difference occurs very near the brightness contrast threshold t. In order to produce
a smooth profile near the brightness contrast threshold, a hyperbolic secant function was
used to give a much more stable and sensible version of (3-3) and is defined as:
500
00),(),(
sec),,,,( ⎟⎠⎞
⎜⎝⎛ −
=t
yxIyxIhtyxyxs , (3-11)
where xx eexh
−+=
2)(sec . The use of the 5th power introduces the minimum difference
from (3-3), as shown in Table 3.1. This equation is plotted along with (3-3) in Fig. 3.4 where
t is set to 10. It can be seen that, compared with (3-3), (3-11) produces a smoother profile
and does not have too large an effect on s as a pixel’s brightness changes slightly. This
equation makes a trade-off between stability about the threshold and the original requirement
of the function, which is to take pixels having intensities similar to that of the center together
into the mass of the circular mask.
Although the wide line detector is isotropic and can detect the whole of a line without the
need to find the direction of line pixels, line direction is still necessary either for
post-processing or for application requirements. From (3-10), the direction of a pixel with a
non-zero line response is determined by finding the longest axis of symmetry:
⎪⎪⎩
⎪⎪⎨
⎧
−
≥=
−
−
otherwisetan
0 ftan
1
1
m
m
mm
m
xy
dixy
φ , (3-12)
35
Fig. 3.4 Brightness difference versus the similarity functions defined in (3-3) and (3-11),
respectively. Here the brightness contrast threshold t is set to 10.
Table 3.1 The deviation of function PtyxIyxIh )/),(),((sec 00− from (3-3). Here c
is the subtraction of (3-11) from (3-3), and c is the absolute value of c.
P C c
2 2.6690 2.6690
3 1.0693 1.0693
4 0.4173 0.4173
5 0.0705 0.0705
6 -0.1427 0.1427
7 -0.2863 0.2863
8 -0.3893 0.3893
9 -0.4666 0.4666
10 -0.5264 0.5264
36
∑+≤≤−+≤≤−
−=
ryyryrxxrx
m yxyxcxxx
0000 ,
002
0 ),,,()( , (3-13)
∑+≤≤−+≤≤−
−=
ryyryrxxrx
m yxyxcyyy
0000 ,
002
0 ),,,()( , (3-14)
∑+≤≤−+≤≤−
−−=
ryyryrxxrx
m yxyxcyyxxd
0000 ,
0000 ),,,())(( , (3-15)
where ⎥⎦⎤
⎜⎝⎛−∈
2,
2ππφ . The value of φ is exactly correct only for lines parallel to one of
the coordinate axes or at an angle of 4π to a coordinate axis. In other cases, deviations of
the calculation of φ will occur. However, such deviation of φ from the line orientation is
small and can be ignored.
In some cases, only dark or bright lines are required, for example, with reference to blood
vessels in X-ray images (dark lines generally required) and aerial images (bright lines
generally required). Introducing a step function, )(∆Θ , we can implement the wide line
detector to extract bright or dark lines according to need:
5
00)(sec),,,,( ⎟⎠⎞
⎜⎝⎛ ∆∆Θ
=t
htyxyxs , (3-16)
where ⎩⎨⎧ >∆
=∆Θelse,0
0,1)( and
⎩⎨⎧
−−
=∆linebright if),(),(
linedark if),(),(
00
00
yxIyxIyxIyxI
.
3.4 Parameter selection
The proposed line detection method requires two parameters – the radius of the circular
mask, r, and the brightness contrast threshold, t. In this section, we provide analyses to show
how the two parameters affect the line detection result and then present approaches for
automatically selecting the two parameters so that the proposed method is robust.
37
3.4.1 Radius of circular mask, r
Obviously complete line detection requires that the circular mask should at least be bigger
than the line width. The radius of the circular mask, r, must therefore be restricted so as to
ensure that the whole of the line can be detected using the inverse WMSB mass. In this
subsection, we analyze the relationship between the radius of a circular mask, r, and the
width of the line detected, w×2 , first with regard to circular masks with a normalized
constant weighting and then those with a Gaussian profile.
Circular Mask with Normalized Constant Weighting
In this subsection, we employ the line mode IL to discuss the relationship between the size of
circular mask with constant weighting and the width of detected line.
Definition 1 Given an image region I and a line L of width w×2 , for any given
point Ip∈ , the distance of p to L, Lpd , is defined by the rule
{ }22 )()(),(,|),(minmLmLmmm ppppLmLL
Lp yyxxppdLpppdd −+−=∈∀= , (3-17)
where mL is the middle axis of L in the line direction.
Definition 2 Let C denote a circle region with radius r in an image. The mass of point 0p ,
the center of the circle region, is defined as
∫∫=C
pp dxdysM0
, (3-18)
⎪⎩
⎪⎨⎧
>
≤=
t-IIift-IIif
spp
ppp
0
0
,0 ,1
, (3-19)
where p is any other point in the circle region C and pI is the intensity of point p.
Proposition 1 Denote the background { }LyxIyxyxB ∉∈= ),(,),(|),( . For each
,Lpl ∈ ,Bpb ∈ if tIIbl pp >− , there exist
38
(1) lpM is monotonically decreasing relative to L
pld and }|max{ LpMM lpp lL
∈= for
each mL Lp ∈ .
(2) bpM is monotonically increasing relative to L
pbd and }|min{ BpMM bpp bB
∈= for
each { }wdpp LpbB b=∈ | .
Proof. Assume that the center of circular region C is the coordinate origin. Rotate the
coordinate axes about the origin so that the y-axis is parallel to mL , as shown in Fig. 3.5. We
define 11 pxw = and
22 pxw = , where { }LCpxx pp I∈= |max1
and
{ }LCpxx pp I∈= |min2
are the x-coordinates of the right- and left-most intersection
points of the circle region C and the line L, respectively. As there exists tIIbl pp >− , for
each BpLp bl ∈∈ , , according to (3-18) and (3-19), we can get ∫∫=LC
p dxdyMl
I
1 and
∫∫=BC
p dxdyMb
I
1 .
(1) For each Lpl ∈ , when the circle C is centered on lp (as shown in Fig. 3.5a), we
have
∫∫∫ ∫ −+−==−
−−
211
2
22
22 0
22
0
22 221wwx
x
xr
xrp dxxrdxxrdydxM p
pl
⎪⎪⎪⎪⎪
⎩
⎪⎪⎪⎪⎪
⎨
⎧
−≤+⎥⎥
⎦
⎤
⎢⎢
⎣
⎡⎟⎠⎞
⎜⎝⎛−+
−>
⎥⎥
⎦
⎤⎟⎠⎞
⎜⎝⎛ −
−−
+
⎢⎢
⎣
⎡ −+⎟
⎠⎞
⎜⎝⎛−+
=
rwwrr
wr
wr
wr
rww
rww
rww
rww
rw
rw
rwr
2211arcsin
2212
2arcsin1arcsin
12
21112
12
11
12
1112
π
(3-20)
where ww ≤≤ 10 . For convenience, denote
39
Fig. 3.5 Illustration of Proposition 1 and its proof.
21arcsin)( xxxxf −+= , (3-21)
where 10 ≤≤ x . Then the equation (3-20) can be rewritten as
⎪⎪⎩
⎪⎪⎨
⎧
−≤⎥⎦⎤
⎢⎣⎡ +
−>⎥⎦⎤
⎢⎣⎡ −+
=12
21)(
12)2()(
2
2
rwxxfr
rwxx
rwfxfr
Mlp
π,
where rwx 1= . It is obvious that the derivative 0≥′lpM . Therefore,
lpM is
monotonically decreasing relative to Lpl
d . Hence, when mL Lp ∈ , 0=LpL
d and LpM
takes the maximum value
⎟⎠⎞
⎜⎝⎛=−+=
rwfrwrw
rwrM
Lp2222 22arcsin2 . (3-22)
(2) For each Bpb ∈ , when the circle C is centered on bp (as shown in Fig. 3.5b), we
have bpb L
BCp SrdxdyM −π== ∫∫ 21
I
and
x
y
1w2w
1px2px
mL
L
C bpM
bpLS
y mL
x
1w
2w
1px2px
L
ClpM
(a) (b)
40
⎪⎪⎩
⎪⎪⎨
⎧
−<⎥⎦⎤
⎢⎣⎡ −+
−≥⎥⎦⎤
⎢⎣⎡ −
=
−+== ∫ ∫−
−−
rwxxfx
rwfr
rwxxfr
ww
xrxrxrdxdyS p
pbp
x
x
xr
xrL
21)()2(
21)(21
arcsin1
2
2
1
22221
2
22
22
π ,
where rwx 1= . It is obvious that 0<′bpLS . Therefore,
bpM is monotonically increasing
relative to Lpb
d . Hence, for each { }wdpp LpbB b=∈ | , L
pbd reaches the minimum value w,
and bpM takes the minimum value
⎪⎪⎩
⎪⎪⎨
⎧
>⎥⎦⎤
⎢⎣⎡ −
≤=−=
wrrwfr
wrrSrM
bpB Lp2)2(
221
2
2
2
π
ππ . (3-23)
□
The meaning of Proposition 1 is that, no matter what size of the circular mask, the smaller
the distance of the pixel to L, the larger the WMSB mass if the pixel is on the line, while the
smaller the WMSB mass if the pixel on the background. Consequently, pixels on the middle
axis of a line take the local maximum value of the WMSB mass, whereas pixels on the
background very near the edges of the line take the local minimum WMSB mass. As
mentioned in Section 3.2.2, in order to detect a line completely, the maximum of the WMSB
mass that the line pixel can take should be less than the minimum that the background pixel
can take. Therefore, according to Proposition 1 and its proof, from (3-22) and (3-23), we get
⎪⎩
⎪⎨
⎧
<≤<
<<<+
121
21)(2
210)2()(2
xxf
xxfxf
π
π, (3-24)
where rwx = .
41
Fig. 3.6 Illustration of the inequality relation about the ratio of the width of the line detected
to the radius of a circle mask with normalized constant weighting.
Figure 3.6 illustrates the inequality relations defined by (3-24). We can see that only
43.0≤x meets the inequality. Given the line width, the smaller the x value, the larger the
mask size r. There is a trade-off to be made here as a smaller ratio means a slower detection
yet the use of a larger ratio and therefore a smaller mask will undermine line detection.
Hence, we set 4.00 =x by experience. The relationship between the width of detected line,
w×2 , and the radius of a constant weighting circular mask, r, is
wr 5.2≥ . (3-25)
According to (3-22), the maximum of WMSB mass which line pixels can take is
)(2 02 xfrM
Lp = . Therefore, the geometric threshold g used in (3-10) is equal to
5.0)(2)(2 02
02 ≈π=π= xfrxfrg .
42
Table 3.2 Relationship between radii of Gaussian masks and approximately critical
widths of lines (ACLW) detected.
Circular Mask with Gaussian Profile
Now we discuss the relationship between the width of the line detected and the size of a
circular mask with a Gaussian profile. Although Proposition 1 is for a constant weight, it is
evident that we can get the same conclusion for a Gaussian weight. That is, the WMSB area
of a line reaches a local maximum when the line passes through the center of the circular
mask. Assume that a circle C with a radius r has a density )2/()( 222 ryxe +− and a line of
width w×2 traverses the center of the circle. Let CL denote the part of the line within the
circle. According to the definition of line response [see (3-10)], if a line of width w×2 is to
be completely detected by using a Gaussian mask with radius r, it requires
∫∫∫∫+
−+
−<
C
ryx
L
ryx
dxdyedxdyeC
2
22
2
22
22
21
, (3-26)
which is equivalent to,
Radius of Gaussian
mask ACLW Digital
ACLW
3 2.6 2
4 3.5 3
5 4.4 4
6 5.3 5
7 6.2 6
8 7.1 7
9 7.9 8
10 8.8 8
11 9.7 9
12 10.6 10
13 11.5 11
14 12.4 12
15 13.2 13
43
∫ ∫∫ ∫−
+−−
+−
<r xr
ryx
w xrr
yx
dxdyedxdye0 0
20 0
222
2
2222
2
22
21
, (3-27)
The right hand side can be simplified as
)1(2
21220 0
20 0
2 2
222
2
22
−−−
+−
−== ∫ ∫∫ ∫ erdeddxdyer
rr xr
ryx πρρθ
π ρ
, (3-28)
Therefore,
)1(4
212
0 02
222
22
−−+
−−<∫ ∫ erdxdye
w xrr
yx π. (3-29)
This equation determines the relationship between the width of the line detected and the
radius of a circular mask with a Gaussian profile. Given a mask with a radius r, the critical
width of the line detected is obtained when the left and right arguments of (3-29) are equal.
As the analytic form of the left function is not available, we provide only the approximate
critical width of the line detected as shown in Table 3.2, as well as the corresponding digital
approximations. Therefore, given a Gaussian profile mask of radius r, a line is definitely
detected if it is not wider than the digital approximation of the corresponding critical width.
3.4.2 Brightness contrast threshold, t
Ideally, a line is defined as the profile of two distinctive regions as described in Section 3.2.1
(see Fig. 3.1). However, in general, there are shaded areas between the two distinctive
regions, as shown in Fig. 3.2. Here, we call such area as the edge region. Therefore, we have
to take into account the influence of the edge region on the WMSB mass, which is related to
the brightness contrast threshold that defines the minimum contrast of the detected features.
In this subsection, we first analyze the relationship between the brightness contrast threshold
and the WMSB mass and then give the automatic selection of a proper t.
Proposition 2 Given a line L of width w×2 bounded by an edge region of width e denoted
by { }weewdwIyxyxE Lyx <<+≤≤∈= ,,),(|),( ),( and a circle C of radius r with the
constraint rw 4.0≤ . Define the background { }ELyxIyxyxB ∪∉∈= ),(,),(|),( . For each
44
Lpl ∈ , Epe ∈ , Bpb ∈ , if (i) tIIbl pp >− and (ii)
bl pp MM < , there exists
⎪⎩
⎪⎨⎧
>>−<
otherwiseMMtIIifMM
le
bele
pp
pppp . Here the sign “<< ” means much less than.
Proof. From Proposition 1, condition (ii) can be rewritten as BL pp MM < . As discussed
in Section 3.2.2, there are only four relations between three different gray-level regions:
Case 1: tIIbe pp >− and tII
el pp ≤− . Suppose ELL U=' , then 'L is a line with a
width of )(2 ew +× and mm LL =' . According to Definition 1, we always have '' L
pLp el
dd < .
From Proposition 1, for each '' Lpl ∈ ,
'lpM is monotonically decreasing relative to
'
'
Lp
ld .
Therefore, for each Lpl ∈ , Epe ∈ , Bpb ∈ , we get ble ppp MMM << .
Case 2: tIIbe pp >− and tII
el pp >− . From Proposition 1, for each Lpl ∈ , when
{ }wdyxEp Lyxl ==→ −
),(|),( , Lpl
d approaches the maximum value w, and lpM reaches
the minimum value ( )rwfrME
p 22=−
. From we << , we always get
( ) ( )lEe ppp MMM minmax =<
−. Therefore, for each Lpl ∈ , Epe ∈ , Bpb ∈ , we have
ble ppp MMM << .
Case 3: tIIbe pp ≤− and tII
el pp ≤− . According to (3-18) and (3-19), we have
21 rdxdyMC
peπ== ∫∫ . Obviously, for each Lpl ∈ , Epe ∈ , Bpb ∈ , we always have
ebl ppp MMM ≤< .
Case 4: tIIbe pp ≤− and tII
el pp >− . Based on Proposition 1, when 0=Lpl
d , lpM
takes the maximum value ( )rwfrMLp
22= . Suppose EBB U=' , then for each '' Bpb ∈ ,
when { }ewdyxEp Lyxb +==→ +),(|),(' , L
pb
d'
reaches the minimum w, 'b
pM takes the
45
minimum value ( )[ ]rwfrME
p 22 −=+
π . Owing to rw 4.0≤ , from Fig. 3.6, we can get
+<
EL pp MM . Therefore, for each Lpl ∈ , Epe ∈ , Bpb ∈ , we have bel ppp MMM << .
Hence, from Case 1 and Case 2, if tIIbe pp >− , we have
le pp MM < ; from Case 3 and
Case 4, if tIIbe pp ≤− , we have
le pp MM > .
□
From Proposition 2 and its proof, we conclude that for a given image, if the contrast
between the edge region and the background is less than the threshold t, the edge region can
be regarded as one part of the line; otherwise, the edge region must be regarded as one part
of the line. That is, for a given image, a large brightness contrast threshold may result in a
‘narrow’ line being detected, while a small brightness contrast threshold must result in a
‘broad’ line being detected. Therefore, the brightness contrast threshold t qualitatively
determines the width of lines detected.
Fig. 3.7 gives an example of line response images with different t. Here, dark line pixels
are extracted directly from the input segmented palmprint image (see Fig. 3.7a) by using a
constant weighting mask with a radius of 8 pixels. It can be seen that as the brightness
contrast threshold increases, the number of false response pixels and true response pixels are
both decrease. It is thus necessary to select a proper t to guarantee the detection of the whole
line. By experience, we defined t as
))(( Istdroundt = , (3-30)
where std is the standard deviation function, round means the nearest integer, and I is the
input image. According to (3-30), the brightness contrast threshold t used in Fig. 3.7a is 9.
46
Fig. 3.7 (a) A segmented palmprint image. Palm-line response images obtained using
brightness contrast thresholds t of (b) 6, (c) 7, (d) 8, (e) 9, (f) 10, (g) 11, (h) 15, (i) 20, and (j)
25.
(a)
(b) (c) (d)
(e) (f) (g)
(h) (i) (j)
47
(a) (b)
(c) (d)
Fig. 3.8 (a) A test image including straight lines and curves of different widths. The line
detection results obtained using (b) the edge-based line finder, (c) the ridge-based line detector,
and (d) the proposed wide line detector.
48
3.5 Experimental results
Our line detection method is implemented for the synthesized image and real images in
several different applications to completely detect lines of different widths. For the purpose
of establishing the effectiveness and robustness of our line detection method, the output
images of our wide line detector are compared with those of line detection approaches based
on edge extraction [26] and ridge detection [3], which are designed to extract the line width
along with the line position. For speed, the sech to the 5th power formula in (3-11) is
implemented using a look up table.
3.5.1 Synthesized image
We synthesized an image including straight lines and curves (curvilinear structures)
with different widths and different intensities. Fig. 3.8a shows the synthesized image.
Applying the proposed wide line detector we obtain the result shown in Fig. 3.8d. Here
we employed the core function defined in (3-11). The brightness contrast threshold used
is 11 according to (3-30). The maximum width in Fig. 3.8a is 5 and thus the operating
radius of circular mask is 7 according to (3-25). Fig. 3.8b and Fig. 3.8c show the line
detection results by the edge-based line finder [26] and the ridge-based line detector [3],
respectively. Since our method detects lines based on extracting the whole line while
the other two approaches detect lines based on detecting the corresponding edges of
each line pixel, we show the line detection result in the corresponding way, as shown in
Fig. 3.8b-d. It can be seen that the line detection results by using our method is more
accurate than that obtained by using either the edge-based line finder or the ridge-based
line detector.
3.5.2 Real image
Figures 3.9a-c are taken from [3] including two aerial images and one X-ray image. Figures
3.9g-i show the line detection results obtained using our wide line detector. The brightness
contrast threshold for each input image is calculated by (3-30). A circular mask with a
normalized constant weighting is used for each input image to detect lines and the
49
corresponding radius of the circular mask is determined according to (3-25) based on the
width of the widest line expected to be detected. In post-processing, we discarded very short
linear structures (<10 pixels) and very “rounded” line structures (the eccentricity < 0.75).
Figures 3.9d-f display the corresponding line extraction results reported in [3].
Figure 3.9d shows the line detection result obtained by using the ridge-based line detector.
We can see that the unjustified edge points are reported in the junction area because the line
width here exceeds the range of widths that can be detected and, further, that the estimation
of the line width of the road object in the bottom of the image is too large due to the effect of
the nearby vegetation. Figure 3.9g shows the corresponding line detection result obtained by
using our wide line detector. It can be seen that the method is able to correctly detect the
wide lines (i.e., essentially all the pixels that comprise line cross-sectional extents), even at
the junction area in the middle of the image and the road close to vegetation in the bottom
part of the image (see red circles).
The aerial image at Fig. 3.9b is more of a challenge. It contains a large area where the
model of the line does not hold, but as can be seen in Fig. 3.9h, our wide line detector
nonetheless works well. Because bright lines are needed here, we employed the equation
(3-16) as the core function of the wide line detector to suppress the dark-line response.
Comparing the line detection results obtained by the ridge-based line detection approach (see
Fig. 3.9e) and our method, we can see that the narrow line in the left upper part, which has a
width close to two, is extracted correctly by using our method, while the corresponding line
detected in Fig. 3.9e is not fully extracted as its cross-sectional extent is too narrow at some
positions. Further, the wide line in the right upper part of the image is detected completely in
Fig. 3.9h, while it is missed in Fig. 3.9e.
Fig. 3.9c is a low contrast X-ray image. Figs. 3.9f and 3.9i show the line detection results
by using the ridge-based line detector and our method, respectively. It can be seen that our
method certainly performs as well as the ridge-based line detector in delineating the
vascularstenosis in the central part, and is also able to detect some very narrow and thin
50
arteries (see red circles) which do not appear in Fig. 3.9f.
The next example, Fig. 3.10, which shows four segmented 128128× palmprint images
(the first row), is from the domain of biometrics. There are a number of reasons for
employing palmprint images to test our line detection method: (1) palm lines, referring to
principal lines and wrinkles [13], are negative lines of varying widths; (2) there are many
details on palm lines such as corners, junctions and branches; (3) some palmprints contain
complex structures along with different line widths, as shown in the last image of the first
row. This example compares the line detection results by using our method (see the last row)
with those by the edge-based line finder (see the second row) and the ridge-based line
detector (see the third row), all of which are displayed in black. We can see that all three line
detectors can extract the principal lines well, however, our method better detects details such
as the ellipse on the principal line in the first column image (see the red circle), the
intersection of two thin lines in the right parts of images of the middle two columns (see the
red circles), and the branches on the principal lines in the last column image (see red
rectangles). Further, our line detection method also outperforms the two other approaches for
the last column image, which contains complex palm lines.
Figure 3.11 displays three segmented tongue images in the first row, which followed by
the output images of the edge-based line finder (the second row), the ridge-based line
detector (the third row) and our wide line detector (the last row), respectively. The line
detection results are all displayed in red. In the first column image, there is a crackle, which
refers to a dark line in a tongue image, very thick and broad. Our line detection method
extracts this crackle correctly (see yellow circles), whereas the other two approaches not
only fail to correctly extract this crackle but, because the width of the crackle varies greatly,
miss branches of the crackle. In the middle column, the most salient crackle in the middle of
the image has an irregular structure, especially in the upper part where the width of the
crackle changes dramatically and becomes discontinuous. Again, only our method correctly
extracts the wide line and the interruption to the line (see the yellow circle). The other two
51
approaches produce unjustified edge points, and a misleadingly continuous line. The last
tongue image is the most difficult because it is low contrast and contains many line segments
of different thicknesses and widths. Comparing the corresponding three output images, it can
be seen that our method outperforms the other two approaches both in the extraction of the
wide line and in the detection of the line structures (see the yellow circles and rectangle). In
addition, because our method uses no derivative and the implementation of the circular mask
decreases the influence of the directional noise, our wide line detector gives strong noise
rejection, that is, produces false crackles caused by reflecting points much less than the other
two approaches (see all blue rectangles).
3.6 Conclusion and discussions
We have presented a novel wide line detector for extracting a whole line by using an
isotropic nonlinear filter. Unlike existing approaches, our method employs a hyperbolic
secant formula based nonlinear filter to detect the whole of the line. Isotropic responses are
obtained by using a circular mask either with a normalized constant weighting or with a
Gaussian profile. The method is robust because of the automatic selection of two
parameters – the size of the circular mask and the brightness contrast threshold. The line
detection method works very well for a range of images containing lines of different widths,
especially for those where the width of lines varies greatly. This method also works well
when the lines run close together or cross each other due to no line direction required to
estimate. Because the wide line detector is not dependent on the Gaussian kernel to detect
lines, even narrow lines can be extracted well as long as the intensity difference between the
narrow lines and the background is larger than the brightness contrast threshold.
Although the proposed line detection method focuses on extracting a whole line, the line
position, however, can be easily obtained through a thinning process. Furthermore, the
localization of lines via the detection is independent of mask sizes. In addition, the wide line
52
detector is robust against noise because there is no derivative used directly. Hence, it is
effective for practical applications in which noise is inevitable.
Finally, it should be pointed out that the proposed method requires the maximum width of
lines be estimated before detection, which is not considered in this chapter. In our method, if
given the maximum width of interested lines, the parameter r is accordingly determined and
all lines not wider than this maximum width may be detected as long as they are strong
enough. In order to apply our method fully automatically, our future work is to automatically
estimate the maximum line width for each input image. In Chapter 6, we will attempt to
achieve this aim for tongue crack detection.
53
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Fig. 3.9 (a), (b) Aerial images and (c) an X-ray image taken from [3]. (d)-(f) The extraction
results of line positions and line widths reported in [3] where line positions are displayed in
white with the corresponding edges displayed in black. (g)-(i) The corresponding line
detection results obtained using our wide line detector are drawn in white.
54
(a) (b) (c) (d)
Fig. 3.10 The segmented palmprint images (first row) and the palm-line detection results
obtainedusing the edge-based line finder (second row), the ridge-based line detector (third row),
and our wide line detector (last row). Detected lines are displayed in black.
55
Fig. 3.11 Segmented tongue images (first row) and crackle detection results obtained using
the edge-based line finder (second row), the ridge-based line detector (third row) and the
proposed wide line detector (last row). Detected crackles are displayed in red.
56
Chapter 4
Analyses of the Wide Line Detector
In this chapter, we further analyze the wide line detector described in Chapter 3. We compare
the three choices of the core function by evaluating their performance on the synthetic image
and show the power of the hyperbolic secant function used as the core function in our wide
line detector. We describe the relationship between the real radius and operating radius and
illustrate the restriction on the real radius presented in Chapter 3. We also illustrate the
reasonableness of the calculation of line orientation given in Chapter 3. In addition, a
comparative study is conducted to demonstrate the robustness of the proposed wide line
detector to impulse noise.
4.1 Introduction
In Chapter 3, we have proposed a novel wide line detection method, wide line detector, and
proved its robustness and effectiveness by an investigation on a variety of image samples.
However, some issues should be addressed further.
First of all, the optimal core function for the wide line detector should be determined. The
core function, the measurement of brightness similarity, is the heart of the wide line detector.
In Chapter 3 a hyperbolic secant formula is used as the core function, while in our
preliminary work [95] an exponential formula is employed (3-11). Although both formulae
give a balance between the stability and sensibility of the measurement of similarity, it is
necessary to analyze the two types of core function, as well as the original brightness
similarity measurement (3-3), to determine which formula is optimal for the proposed wide
line detector. In signal detection theory, the signal-to-noise ratio is a good measure for
detection methods and ROC analysis provides tools to select possibly optimal models.
57
Therefore, we will employ the signal-to-noise ratio and the ROC curve to choose the optimal
core function for the proposed wide line detector.
Radius of circular mask is one of the two parameters of the proposed wide line detector
and plays a very important role in the detection of the complete line. In Chapter 3, we
investigate the relationship between the radius of circular mask and the width of detected
line. As this relationship is based on the real radius which is decided by the area of circular
mask and in general is not an integer, it is inconvenient in practice. Hence, we will study the
relations between the real radius and the size of square kernel which is employed to
approximate to the circular mask and thereby extend the relationship to the size of square
kernel.
Line orientation is not needed for our wide line detector. However, it is still necessary to
calculate line orientation either for post-processing or for application requirements. The
calculation of the orientation of line which is described in Chapter 3 is not exact but
approximate in most cases. We will illustrate the reasonableness of the approximation of the
calculation of line orientation.
Finally, the robustness of our wide line detector to noise should be considered further. Our
wide line detector is not based on derivatives of images and thus is robust against noise. We
will further demonstrate and analyze the robustness of our wide line detector to impulse
noise by comparing with two well-known line detection approaches [3,26].
The Chapter is accordingly organized as follows. Section 4.2 introduces the concepts of
performance evaluation which is used in the rest of this Chapter. Section 4.3 describes a
scheme to determine the optimal core function of the proposed wide line detector by
evaluating the performance of various formulae. Two types of circular mask radius and the
calculation of line orientation are discussed in Section 4.4 and Section 4.5, respectively.
Section 4.6 demonstrates the robustness of the proposed wide line detector to impulse noise.
Finally, we give conclusions in Section 4.7.
58
4.2 Background
In this section, we will introduce the two concepts: signal-to-noise ratio and ROC curve,
which will be used in the rest of this Chapter.
4.2.1 Signal-to-noise ratio
A good measure for a detection method is its output signal-to-noise ratio (S/N). In his
seminal work on computational edge detection, Canny [27] gave the definition of the output
signal-to-noise ratio of a detector for step edges as
∫
∫+
−
−=
W
W
W
dxxfn
dxxfA
)(
)(SNR
20
0
, (4-1)
where f is the response of the filter which is assumed to be zero outside of the region
],[ WW− , A is the amplitude of the step, and 20n is the mean-squared noise amplitude per
unit length.
The value of SNR is the output signal to noise ratio and should be as large as possible. The
signal-to-noise ratio is a measure of how the signal from the interested objects compares to
other background reflections (categorized as "noise"). Informally, signal-to-noise ratio refers
to the ratio of useful information to false or irrelevant data. In section 4.3, we will employ
the S/N measure to evaluate the performance of the proposed wide line detector based on
various core functions.
4.2.2 ROC curve
In signal detection theory, a receiver operating characteristic, or simply ROC curve [96], is a
graphical plot of the sensitivity vs. (1-specificity) for a binary classifier system as its
discrimination threshold is varied. The ROC curve can also be represented equivalently by
plotting the true positive rate (TPR) and against the false positive rate (FPR) for the different
possible cutpoints of a test. ROC analysis is related in a direct and natural way to cost/benefit
59
Fig. 4.1 The ROC space and plots of the four prediction examples. [97]
analysis of decision making. Widely used in medicine [98-100], radiology [101-102],
psychology [96] and other areas for many decades, it has been introduced relatively recently
in other areas like biometrics [50, 103], computer vision [104], and machine learning [105].
Figure 4.1 illustrate the ROC space which is defined by FPR and TPR as x- and y- axes,
respectively. In this space, an ROC curve demonstrates several things: (1) It depicts the
relative trade-offs between sensitivity (TPR) and specificity (1-FPR): any increase in
sensitivity will be accompanied by a decrease in specificity; (2) the closer the curve follows
the left-hand border and then the top border of the ROC space, the more accurate the test; (3)
the closer the curve comes to the 45-degree diagonal of the ROC space, the less accurate the
test. In section 4.3, we will employ the ROC curve to analyze the output signal-to-noise ratio
for the selection of optimal core function.
60
4.3 Core function
In the wide line detector, the form of the brightness similarity function is the heart of WMSB
determination, called core function. In this section, we compare the different formulae for the
core function by evaluating their performance to the synthetic image and show the power of
the hyperbolic secant function.
4.3.1 SECH v.s. EXP
The brightness similarity function should guarantee counting pixels that have similar
brightness to the center of the circular mask into the mass of WMSB and counting pixels
with dissimilar brightness out of the mass of WMSB:
⎩⎨⎧
>≤
=tyx-IyxIiftyx-IyxIif
tyxyxs),(),( ,0),(),( ,1
),,,,(00
00000 , (4-2)
where t is brightness contrast threshold. In the wide line detector presented in Chapter 3, a
hyperbolic secant function was used to measure the brightness similarity and is defined as
JS
tyxIyxI
htyxyxs ⎟⎠⎞
⎜⎝⎛ −
=),(),(
sec),,,,( 00001 , (4-3)
where 5=JS in WLD. An alternative to the hyperbolic secant function can be given by
⎥⎥⎦
⎤
⎢⎢⎣
⎡⎟⎠⎞
⎜⎝⎛ −
−=JE
tyxIyxItyxyxs ),(),(exp),,,,( 00
002 , (4-4)
where JE is an even number.
Two core functions defined by the formula 1s of the sech to 5th power and the formula
2s of the exp to 6th power are plotted in Fig. 4.2, as well as the 0-1 formula 0s . The
motivation for introducing a particular formula is to give the tradeoffs between the stability
and the sensibility of the core function of the wide line detector. That is, the formula should
give a smooth profile to avoid dramatic changes near the brightness contrast threshold.
61
Fig. 4.2 Brightness difference versus the similarity formulae defined in (4-2), (4-3), and
(4-4), respectively. Here the brightness contrast threshold t is set to 10, 5=JS for (4-3), and
6=JE for (4-4).
At the same time, the formula should guarantee the original requirement of the core function,
which is to take all pixels whose intensities are similar to that of the center.
Besides smoothness near the brightness contrast threshold, the formula used as the core
function of the WLD should introduce the minimum deviation from the 0-1 function (as
defined in (4-2)) rather than the minimum MSE (mean squared error) because the wide line
detector is based on the inverse of summation of the core functions. Table 4-1 lists the
similarities of the core functions 1s and 2s with different powers and the deviation of the two
core functions from 0-1 function 0s . The deviation reaches the minimum when the core
function is defined as in (4-3) with 5=JS . Therefore, the optimal core function for the
proposed wide line detector is the formula 1s which is employed for the proposed wide line
detection method described in Chapter 3.
62
Table 4.1 Look up table of core functions 1s and 2s with different powers and the
deviation of the two core functions from 0s .
EXP ( 2s ) SECH ( 1s ) JE/JS
0II − 4 6 8 4 5 6 0s
0 1 1 1 1 1 1 1
1 0.9999 1.0000 1 1 1 1 1
2 0.9984 0.9999 1.0000 1.0000 1.0000 1 1
3 0.9919 0.9993 0.9999 1.0000 1.0000 1.0000 1
4 0.9747 0.9959 0.9993 0.9997 0.9999 1.0000 1
5 0.9394 0.9845 0.9961 0.9981 0.9995 0.9999 1
6 0.8784 0.9544 0.9833 0.9917 0.9970 0.9989 1
7 0.7865 0.8890 0.9440 0.9719 0.9860 0.9931 1
8 0.6639 0.7694 0.8455 0.9216 0.9486 0.9666 1
9 0.5189 0.5878 0.6502 0.8176 0.8478 0.8737 1
10 0.3679 0.3679 0.3679 0.6481 0.6481 0.6481 1
11 0.2313 0.1701 0.1172 0.4391 0.3842 0.3306 0
12 0.1257 0.0505 0.0136 0.2476 0.1650 0.1007 0
13 0.0575 0.0080 0 0.1146 0.0488 0.0160 0
14 0.0215 0 0 0.0429 0.0092 0.0011 0
15 0.0063 0 0 0.0127 0.0010 0 0
16 0.0014 0 0 0.0028 0 0 0
17 0 0 0 0 0 0 0
18 0 0 0 0 0 0 0
19 0 0 0 0 0 0 0
20 0 0 0 0 0 0 0
02/1 ss − -2.8720 -2.4456 -2.1651 0.4173 0.0705 -0.1427
02/1 ss − 2.8720 2.4456 2.1651 0.4173 0.0705 0.1427
63
(a)
(b)
Fig. 4.3 (a) A test image including straight lines and curvilinear structures of different
widths. (b) The ROC curves of (a) for the 0-1 function, the exp to 6th power formula, and the
sech to 5th power formula.
64
4.3.2 The power of the hyperbolic secant function
Using the synthetic image in Chapter 3, we show the power of the hyperbolic secant function
by comparing S/N of the wide line detector using three formulae defined in (4-2), (4-3), and
(4-4), respectively. The 0-1 function 0s , as defined in (4-2), formulates the principle of our
method. The formula 1s of the sech to 5th power is the core function used in the proposed
wide line detector. The formula 2s of the exp to 6th power was used to detect edge and corner
in [49]. The signal-to-noise ratio is compared using the ROC curve [101]. The ROC curve
for the method was obtained thresholding the line-strength image at a series of values and
comparing the resulting binary images with the ground truth. Detected pixels were counted
as true positives if they coincided with the synthetic image, and false positives otherwise. We
used 100 threshold values to cover the true and false fraction ranges evenly. The resulting
ROC curves for the synthetic image (Fig. 4.3a) are shown in Fig. 4.3b. These results suggest
that the wide line detector using the sech of 5th power (SECH) produces the best
performance.
Fig. 4.4 is the ROC curves of the synthetic image (Fig. 4.3a) for the 0-1 function 0s and the
formula 1s of the sech to 5th power with two weightings defined in (3-5) and (3-6),
respectively. We can see the performance of the SECH based wide line detection method is
higher than that using the 0-1 function, no matter what kind of weighting method used.
Hence, the SECH formula is the optimal core function for the proposed wide line detector.
Fig. 4.5 gives an example to show the importance of introducing the SECH formula. Fig.
4.5a is a synthesized image which contains a bright line of 5 pixels wide. According to the
scheme of parameter selection described in Chapter 3, the brightness contrast threshold is 3
and the radius of circular mask is 7. Fig. 4.5b and Fig. 4.5c are line strength images by using
the 0-1 function and the SECH formula, respectively. We can see the 0-1 function fails to
detect the whole line whereas the SECH formula succeeds.
65
Fig. 4.4 The ROC curves of Fig. 4.3a for different functions and different weighting.
(a) (b) (c)
Fig. 4.5 (a) A synthesized image with a bright line of 5 pixels wide. The pixel values are
marked. (b) and (c) The line response images by using the 0-1 function and the SECH to 5th
power function, respectively. The line responses on the center pixel and at the most left pixel
are shown.
66
4.4 Radius of circular mask
The wide line detector is implemented using an isotropic nonlinear filter. In Chapter 3, we
analyzed the relationship between the radius of a circular mask, r, and the width of the line
detected, w×2 , and concluded wr 5.2≥ . In this section, we will firstly describe the
relationship between the real radius and operating radius and then illustrate the restriction on
the real radius, r. Finally, we will extend this restriction to the operating radius.
4.4.1 Operating radius v.s. real radius
The isotropic filter, also called circular mask, is approximated by square. We call the radius
decided by the area of circular mask (the number of mask pixels) real radius, i.e. the radius
of circular mask r as defined in (3-5). The radius of a square kernel is referred to operating
radius. Using operating radius, the circular mask can be redefined as
⎩⎨⎧ +≤−+−
=otherwise
dRyyxxifryxyx op
0)()(1),,,,(
2220
20
00ω , (4-5)
⎥⎦
⎥⎢⎣
⎢=
3opR
d , (4-6)
where d is a positive integer and opR is the operating radius. Figure 4.6 is an example of a
circular mask approximated by a square kernel of 1515× where the operating radius is 7
pixels and the real radius is 7.5 pixels. Table 4.2 gives operating radii of square kernels with
different sizes, as well as the corresponding real radii of circular masks.
67
Fig. 4.6 An example of circular mask approximated by a square kernel of 1515× .
Table 4.2 Square kernel and the corresponding operating radius and real radius. Here the
operating radius is half the number of pixels along x or y-direction excluding the center
pixel.
Mask size
Operating radius
Real radius
77× 3 3.43
99× 4 4.26
1111× 5 5.32
1313× 6 6.41
1515× 7 7.51
1717× 8 8.39
1919× 9 9.66
2121× 10 10.54
2323× 11 11.57
2525× 12 12.68
2727× 13 13.74
2929× 14 14.64
3131× 15 15.89
3333× 16 16.78
3535× 17 17.74
3737× 18 18.92
3939× 19 19.91
4141× 20 20.88
4343× 21 22.06
4545× 22 23.02
4747× 23 23.94
4949× 24 25.16
68
4.4.2 Relationship of mask radius and detected line width
In Section 3.4.1, analysis shows only 43.0≤= rwx , where r refers to the real radius, can
guarantee the complete detection of a line with w×2 pixels wide. Although any value of x
below 0.43 can satisfied the requirement of complete line detection, it is obvious there is a
balance between the time efficiency and detection ability of the wide line detector. Given the
line width, a smaller ratio means a slower the detection speed, while a larger ratio which
indicates a smaller mask will undermine line detection. Table 4.3 lists the values of x given
different widths and different radii. The complete detection of line L requires the subtraction
of the WMSB mass 0LW from
0BW gives a positive value, where 0LW is the maximum
WMSB mass of line pixels and0BW is the minimum WMSB mass of background pixels.
Table 4.3 indicates that this requirement will be satisfied when the value of x is equal to or
less than 0.4, no matter what the values of the width and radius. Hence, we set 4.00 =x and
thus get the relationship of the radius of circular mask and width of detected line, wr 5.2≥ .
However, this relationship is for real radius of a circular mask and not convenient in
practice where is usually expected to know the proper operating radius directly. According to
(4-5), (4-6), and (3-5), we have
22
2
3r
RR op
op =⎟⎟⎠
⎞⎜⎜⎝
⎛+ . (4-7)
Considering (3-25), we can get the relationship of the operating radius and the detected line
width
wRop 3.2≥ . (4-8)
Figure 4.7 shows the WMSB mass of wide lines with different widths. We call the
minimum interger satisfying equation (4-8) critical operating radius. In Fig. 4.7, the values
of critical operating radii are 6, 17, and 28 pixels for lines of 5, 15, and 25 pixels wide,
respectively.
69
Table 4.3 The values of x given different widths and different radii. Here the operating
radius is half the number of pixels along x or y direction excluding the center pixel. 0L is
the line pixel where WMSB mass of line pixels reaches the maximum, while 0B is the
background pixel where the WMSB mass of background pixels reaches the minimum.
Line width (pixel)
Operating Radius (pixel)
Real Radius (pixel)
0LW 0BW
00 LB WW − x
1 1.2616 0.60 0.80 0.20 0.3963 1
2 2.0342 0.38 0.77 0.38 0.2458
2 2.0342 0.62 0.69 0.07 0.4916
3 3.4318 0.38 0.68 0.30 0.2914 2
4 4.2595 0.32 0.72 0.40 0.2348
2 2.0342 0.85 0.69 -0.15 0.7374
3 3.4318 0.57 0.59 0.03 0.4371
4 4.2595 0.47 0.63 0.16 0.3522 3
5 5.3226 0.37 0.67 0.30 0.2818
3 3.4318 0.70 0.59 -0.11 0.5828
4 4.2595 0.60 0.58 -0.02 0.4695
5 5.3226 0.47 0.60 0.12 0.3758 4
6 6.4080 0.40 0.64 0.24 0.3121
4 4.2595 0.72 0.58 -0.14 0.5869
5 5.3226 0.57 0.56 -0.01 0.4697
6 6.4080 0.50 0.59 0.09 0.3901 5
7 7.5061 0.42 0.62 0.20 0.3331
5 5.3226 0.67 0.56 -0.11 0.5636
6 6.4080 0.59 0.55 -0.04 0.4682
7 7.5061 0.50 0.57 0.07 0.3997 6
8 8.3873 0.45 0.60 0.15 0.3577
6 6.4080 0.67 0.55 -0.12 0.5462
7 7.5061 0.57 0.54 -0.03 0.4663
8 8.3873 0.52 0.56 0.04 0.4173
9 9.6574 0.45 0.59 0.14 0.3624
7
10 10.5399 0.42 0.62 0.20 0.3321
70
(a)
(b)
71
(c)
Fig. 4.7 The plots of WMSB mass of lines with width of (a) 5 pixels, (b) 15 pixels, and (c)
25 pixels, respectively.
(a) (b) (c) (d)
Fig. 4.8 The line with orientation of (a) 15 degree, (b) 30 degree, (c) 60 degree, and (d) 75
degree, respectively.
72
4.5 Line orientation
As our wide line detector is based on an isotropic nonlinear filter, no line orientation is
required for line detection. However, the calculation of line orientation is necessary either for
post-processing or for application requirements. In Chapter 3, we give the way to calculate
the orientation of a line pixel, φ . According to the definition of (3-12), the calculation of φ
is correct only for lines either parallel to one of the coordinate axes or at an angle of 4/π
to a coordinate axis. In this section, we will illustrate although the error of the calculation of
φ will occur in other cases, such error is very small and thus can be ignored.
One pixel wide lines with different orientations are shown in Fig. 4.8. Figure 4.8a and
4.8b are more horizontal lines with orientation of 15 degree and 30 degree, respectively.
While Figure 4.8c and 4.8d are more vertical lines with orientation of 60 degree and 75
degree, respectively. Table 4.4, according to the definition (3-12), gives the calculated φ of
lines in Fig. 4.8 with different widths. Two masks are employed to calculate line orientations,
one is of 77× (the operating radius is 3) and the other is of 1313× (the operating radius
is 6). The brightness contrast threshold used here is 10. We can see that errors of φ due to
the definition of (3-12) are small. Moreover, the calculation of φ keeps the characteristic of
the line orientation. That is, the error of the calculation of φ from line orientation makes a
more horizontal line (as shown in Fig. 4.8a and Fig. 4.8b) more horizontal, while a more
vertical line (as shown in Fig. 4.8c and Fig. 4.8d) more vertical. Hence, the error of the
calculated φ can be ignored and the calculation of φ can satisfy the requirement of
post-processing and applications.
73
Table 4.4 The calculated φ . Here opR is the operating radius which is half the number of
pixels along x or y-direction excluding the center pixel.
Calculated Orientation (degree)Line Orientation
(degree)
Line Width (pixel)
3=opR 6=opR
Digital Orientation (degree)
1 8.1301 6.8924
2 14.9951 8.8999
3 --- 9.5644
4 --- 12.8451
5 --- 14.7746
15
6 --- ---
15.9453
1 23.1986 17.1027
2 28.9677 18.9528
3 --- 19.5607
4 --- 22.5304
5 --- 24.2539
30
6 --- ---
29.7449
1 66.8014 72.8973
2 61.0323 71.0472
3 --- 70.4393
4 --- 67.4696
5 --- 65.7461
60
6 --- ---
60.2551
1 81.8699 83.1076
2 75.0049 81.1001
3 --- 80.4356
4 --- 77.1549
5 --- 75.2254
75
6 --- ---
74.0546
74
Table 4.5 Comparisons of calculations per pixel for different line detection methods. The
size of the convolving kernel is NN × .
Line detection method Additions Multiplications
Koller’s approach 142 +N 82 +N
Steger’s approach* > 23N > 123 2 +N
Line Operator 142 −+ NN 2N
Wide Line Detector 32 2 −N 12 −N
* The calculations of convolving the image with the 2D Gaussian partial derivative kernels without considering the calculations introduced by computing eigenvector and eigenvalue of Hessian matrix.
4.6 Time complexity
In computational complexity theory, the time complexity of a problem is the number of steps
that it takes to solve an instance of the problem as a function of the size of the input. For a
given input image, the time complexity is related to the calculations per pixel, i.e., the
number of additions and multiplications for each pixel. Table 4.5 shows the time complexity
of the proposed wide line detector compared with there typical line detection methods,
Koller’s line finding approach, Steger’s line detection approach, and the Line Operator. In
terms of Big O notations, all the four line detection methods have the same polynomial time
complexity )( 2nO .
4.7 Robustness to impulse noise
Our proposed wide line detector applies an isotropic nonlinear filter to extract a line
completely without any derivative. Because the detection is not based on derivatives of
images, the wide line detector is robust against noise. In this section, a comparative study is
conducted to demonstrate the robustness of our method.
75
(a) (b)
(c) (d)
(e) (f)
Fig. 4.9 The test image (Fig. 4.3a) is added impulse noise with the noise density of (a) 0.01 and
(b) 0.1, respectively. Line strength images by using our wide line detector to (a) and (b) are
shown in (c) and (d), respectively. (e) and (f) are final results of line detection by performing
morphological operations to (c) and (d), respectively.
76
(a)
(b)
Fig. 4.10 Line detection results of the noisy image Figure 4.9a using (a) Koller’s edge-based
line finder [26] and (b) Steger’s ridge-based line detector [3].
The testing results are illustrated in Fig. 4.9. The test image is the synthetic image
employed in Chapter 3 (see Fig. 4.3a), which contains straight lines and curvilinear
structures of different widths with varying background. Fig. 4.9a and 4.9b are noisy images
by adding impulse noise to the test image with noise density of 0.01 and 0.1, respectively.
Applying our wide line detector to the two noisy images, we get line strength images Fig.
4.9c and 4.9d and then binarize them by proper thresholds. The threshold is the minimum
value which makes false negatives equal to zero. Here false negatives refer to true line pixels
not yet detected when thresholding line-strength image at a series of values. Fig. 4.9e and
4.9f demonstrate the final detection results after performing morphological operations (a
77
morphological opening operation followed by a removing spur operation) on binary images.
The final detection results show that line structures and widths of two noisy images are
preserved very well. Therefore, the robustness of our method against noise is not to remove
the impulse noise before detection but to prevent the impulse noise from deforming or
destroying the line structures and widths.
For comparison, Fig. 4.10 displays the line detection results of the noisy image Fig. 4.9a
using Koller’s edge-based line finder [26] and Steger’s ridge-based line detector [3]. Due to
derivatives employed to extract lines, both approaches are very sensitive to noise. In the line
detection results, line structures are destroyed where noise is present, as well as line widths.
Normally, before using a derivative-based detection method, a filter is employed to reject
noise and preserve image structures. However, it doesn’t always work well. The median
filter is regarded as an excellent rejector of impulse noise and is able to remove the noise and
replace the bad pixels with reasonable values [106]. Figure 4.11 displays the median filtering
results of the noised image Fig. 4.9b using three different neighborhoods: a 33× square,
a 31× horizontal line, and a 51× horizontal line. The one pixel wide line disappears in Fig.
4.11a using the median filter with a 33× square neighborhood and the filtering result image
is degraded. A median filter with a horizontal line neighborhood can keep the one pixel wide
line and yet much more noise, as shown in Fig. 4.11b and 4.11c. Also, directional distortion
occurs in the two filtering result images due to horizontal neighborhoods used. Therefore, no
matter what kind of neighborhoods used, the median filtering fails to offer the
derivative-based detection method a noise free and structure preserved image.
78
(a)
(b)
(c)
Fig. 4.11 Filter the noisy image Figure 4.9(a) using median filters with three different
neighbors: (a) a 33× square, (b) a 31× horizontal line, and (c) a 51× horizontal line.
79
4.8 Summary
In this Chapter, we have analyzed the proposed wide line detector further. The optimal core
function, the hyperbolic secant function of the 5th power, is determined by evaluating the
performance of various formulae. The relations between the real radius and the operating
radius are investigated to facilitate the parameter selection. The restriction on the real radius
is illustrated and is extended to the operating radius. The calculation of line orientation is
analyzed to show the reasonableness of the approximation results. The time complexity of
the proposed method is analyzed and shown that the wide line detector has the same time
complexity )( 2nO as the typical line detection methods. The robustness of our wide line
detector to impulse noise is also demonstrated by a comparative study.
80
Chapter 5
Palmprint Recognition by Using Line
Features
In this chapter, we address the first application of the wide line detector, a novel palm-line
feature extraction method for palmprint recognition, which is discussed in the introductory
chapter (cf. Fig. 1.1). This palm-line feature extraction method can detect line-like features
which convey both structure features and strength features of palm lines. An isotropic
nonlinear filter is employed to calculate the line strength of palmprint. Palm-line feature
images are obtained via binarilization of line strength images. A Gaussian smoothing filter
serves as post-processing for palm-line feature image denoise. A translation-invariant
similarity measure is introduced for palm-line feature matching. To determine the optimal
combination of parameters for the proposed palm-line feature extraction method, an
experimental scheme is designed. An extensive test is conducted on the PolyU Palmprint
Database [107]. Experimental results show that the performance of the proposed palm-line
feature extraction method is comparable with the state-of-the-art algorithm of palmprint
identification, which indicates that the palm-line feature is appropriate for palmprint
recognition.
5.1 Introduction
Biometrics, an important and powerful technique for personal recognition, is becoming more
and more popular in today’s information and networked society. The biometric
computing-based approach is concerned with identifying a person by his/her characteristics
[47,108]. These characteristics can be divided into two categories. One is physiological
characteristics including fingerprints [109-110], face [111-113], iris [72], hand geometry
81
[114], and palmprints [50]. The other is behavioral traits such as voice [115-116], signature
[117], and gait [118-119]. Based on the intrinsic features of a human being, biometric “keys”
or “passwords” are always available and reliable for automatically recognizing a person’s
identity.
Fingerprint identification [109-110] is the most well-known and widespread biometric
method. However, manual workers and elderly people may have great difficulty to provide
clear fingerprint for minutia extraction. Iris-based recognition has been successfully
developed [120], but it suffers from the inconvenience of iris picture capturing and even
breaks down due to unstable eyes. Thus, there is a demand for a new automatic personal
identification system.
Recently, there has been a great deal of attention devoted to the study of palmprint-based
personal identification [13,48,50-51,55,58-59]. Palmprint is one of the most reliable traits in
personal recognition due to its stability and uniqueness. Palmprint-based personal
verification is user-friendly and convenient. Moreover, palmprint-based identification can be
operated by a low cost and low resolution imaging device. In the palmprint identification
system developed by the Biometric Research Centre in the Hong Kong Polytechnic
University, a charge-coupled device (CCD) camera is used to capture low-resolution
(<100dpi) palmprint images. To limit the palm’s stretching, translation, and rotation, the
CCD-based palmprint capture device [50] is fitted with some pegs. These pegs separate the
fingers and thus form holes, one between the forefinger and the middle finger and the other
between the ring finger and the little finger. A typical palmprint image captured by this
system is shown in Fig. 5.1.
82
Fig. 5.1 The main patterns in a palmprint. [13]
5.1.1 Palm lines
The main patterns in a palmprint image, as illustrated in Fig. 5.1, can be generalized to
principal lines, wrinkles, and creases (also called ridges). Usually, there are three principal
lines in a palmprint: the heart line, the head line, and the life line. These lines vary little over
time, and their shapes and locations on the palm are the most important physiological
features for individual recognition. Most wrinkles are much thinner than the principal lines
and much more irregular. However, some wrinkles of some palms are as strong as the
principal lines and also stable and reliable for identification. Creases exist all over the palm
just like the ridges in a fingerprint and cannot be observed in low-resolution images.
Although some crease-based palmprint recognition methods [121] have been proposed, they
require rather high resolution imaging and consume a large amount of data. Generally
speaking, the principal lines and wrinkles, also called palm lines, are stable and reliable for
individual identification and can be exploited and derived from a low-resolution palmprint
image.
Palm lines have their own properties. Firstly, palm lines are negative lines (also called
dark lines) on which pixels’ intensities are lower than neighbours’. Secondly, palm lines,
Principal lines
Wrinkles
Creases
83
especially principal lines, consist of many short line segments and curves which form many
corners, junctions, branches, rings and chains. Finally, principal lines and some thick
wrinkles are very strong and look wider than other light palm lines. That is, palm lines have
different width which generally reflects strength information. The width feature of palm lines
is very important to describe a palmprint clearly especially when different palmprints have
similar line structures. Hence, not only structure features but also width information of palm
lines is necessary and important for palmprint recognition.
Although line features play an important role in palmprint identification, little work has
been done on line feature based palmprint recognition. In [50], Zhang et al pointed out that it
is difficult to obtain a high recognition rate using only principal lines because of their
similarity among different people. In this chapter, we will investigate on the ability of palm
lines to recognize palmprint images.
5.1.2 Review
Palm-line based palmprint identification schemes, such as interesting points, line segments
and line features, have been presented. In the offline palmprint verification, Zhang et al. [48]
approximated line features by using several straight-line segments. Duta et al. [51] extracted
a set of feature points along palm lines and the associated line orientation to represent line
features. You et al. [58] detected interesting points by applying the Plessey operator, a corner
detector, to achieve higher performance than edge points by eliminating the redundancy. Han
et al. [59] extracted the line-like features of a palmprint by using a morphological edge
detector.
In the past several years, more and more researchers have focused on the on-line palmprint
identification for the real-time application. You et al. [55] further extracted the “interest”
lines based on the interesting points by applying a fuzzy rule. Zhang and Zhang [13]
obtained a set of statistical features of palm lines to characterize a palmprint by employing
the overcomplete wavelet expansion. Wu et al. [60], for the first time, directly extracted palm
lines through morphological operations and trace each extracted line in a recursive process.
84
They further devised a set of directional line detectors to extract palm lines in different
directions and represented these irregular lines using chain codes [61]. However, none of
these offline and online palmprint recognition methods considers width features of palm
lines which indicate their strength characters.
In this Chapter, we focus on palmprint recognition by detecting wide lines which convey
palm-line structure and strength features. A palm-line feature extraction method based on the
wide line detector [62] is proposed to extract line-like features which contain width
information of palm lines. To give isotropic responses, circular masks of different radii are
digitally approximated via a constant weighting. By inverting the WMSB mass, the line
strength of palmprint is determined. Palm-line feature images are obtained via binarilization
of line strength images. A Gaussian smoothing filter serves as post-processing for palm-line
feature image denoise. For palm-line feature matching, a translation-invariant matching
method is introduced. An experimental scheme is developed to select the optimal
combination of parameters for the proposed palm-line feature extraction method.
The Chapter is organized as follows. Section 5.2 describes the proposed WLD-based
palm-line feature extraction method. The matching scheme for detected palm-line features is
provided in Section 5.3. Section 5.4 describes the palmprint database employed to test our
method and shows the experimental results of palm-line feature based palmprint verification
and identification. Finally, we make discussions and draw conclusions in Section 5.5.
5.2 Line-like feature extraction
In this section, we describe the palm-line feature extraction method based on the wide line
detector which is presented in Chapter 3. As palm lines are dark lines, according to the
equation (3-16), the core function of the wide line detector used to extract line-like features
is defined as
85
Fig. 5.2 Brightness difference versus the similarity defined in the core functions (5-1), (5-4),
and (5-5), respectively. Here the brightness contrast threshold t is set to 10.
⎪⎩
⎪⎨⎧
>⎟⎠
⎞⎜⎝
⎛ −×=
otherwise1
),(),( if),(),(
sec),,,,( 00
500
0001yxIyxI
tyxIyxI
htyxyxc ω , (5-1)
where t is the brightness contrast threshold and 0ω is the normalized circular mask which is
digitally approximated by a constant weighting
∑+≤≤−+≤≤−
=
ryyryrxxrx
ryxyx
0000 ,
000 ),,,,(ω
ωω , (5-2)
⎩⎨⎧ ≤−+−
=otherwise
ryyxxifryxyx
0)()(1
),,,,(22
02
000ω . (5-3)
where r is the radius of the circular mask. In [95], an alternative to the core function is
defined based on an exponential function
86
Fig. 5.3 A 3D plot of the WMSB mass given a test image with a 3 pixels wide line.
⎪⎩
⎪⎨
⎧>
⎥⎥⎦
⎤
⎢⎢⎣
⎡⎟⎠⎞
⎜⎝⎛ −
−×=
otherwise1
),(),( if),(),(
exp),,,,( 00
600
0002yxIyxI
tyxIyxI
tyxyxc ω (5-4)
Normally, the brightness contrast threshold t in (5-1) and (5-4) should be a positive integer.
However, when 0=t , both core functions will degrade to
⎩⎨⎧ >
×=otherwise1
),(),( if0),,,,( 00
0000
yxIyxItyxyxc ω . (5-5)
The core function 0c is plotted in Fig. 5.2, as well as 1c and 2c , where t is set to 10 for (5-1)
and (5-4).
The comparison measured by any core function is done for each pixel within a circular
mask. The sum of all comparisons in the circular mask is calculated as the WMSB mass of
the center (ref. Eq.3-9) and the palm-line strength can be obtained by inverting the WMSB
87
mass (ref. Eq.3-10). The principal of the WLD-based palm-line feature extraction method is
illustrated in Fig. 5.3, where a test image with a 3 pixels wide line has been processed to give
WMSB mass as output. All of WMSB mass can be divided into two groups: one is obtained
from pixels on the line and the other is from background. The complete extraction of the line
requires the local maximum of WMSB mass given by pixels on the line be less than the local
minimum from background. This can be realized by selecting proper parameters for the wide
line detector.
Fig. 5.4 The palmprint sub-images (the first row) with the size of 128128× , the
corresponding line-strength images (the second row) using the wide line detector, the
palm-line feature images (the third row) via thresholding the line-strength images, and the
post-processed palm-line feature images (the last row) with the size of 6464×
88
Due to the normalization of circular masks, the WMSB mass varies between 0 and 1 and
thus the palm-line strength between 0 and 0.5. The palm-line feature image can be obtained
by removing all pixels which line-strength is below a threshold. Some examples of palmprint
sub-images are shown in Fig. 5.4, as well as the corresponding line-strength images and
palm-line feature images. Here the brightness contrast threshold t and the circular mask
radius r are set to 8 and 12, respectively. The threshold to obtain the binary palm-line feature
image is set to 0.1. As shown in the third row of Fig. 5.4, not only the structure features but
the width characters of palm lines are derived from the palmprint sub-images by using the
proposed WLD-based palm-line feature extraction method. However, there exists some noise
in the detected palm-line feature images. For denoise, a Gaussian smoothing filter is
employed and then the smoothed palm-line feature image is resized to 6464× for matching,
as shown in the last row of Fig. 5.4, where the scale of the Gaussian smoothing filter is 1.
In the wide line detector, two parameters, the brightness contrast threshold t and the radius
of the circular mask r, are required for complete line extraction. Figure 5.5 shows the
influence of the two parameters on the palm-line feature extraction results. With the increase
of the brightness contrast threshold, detected palm lines become thinner and weak lines will
be missed. While with the increase of the mask size, detected palm lines become wider.
Notice that given the brightness contrast threshold, when the radius of the circular mask
changes from 12 to 20, the width of detected palm lines especially principal lines varies little.
The increase of the circular mask radius reduces the noise of the palm-line feature extraction
results and yet lowers the completeness of detected wrinkles. Therefore, the parameter
selection has a great influence on the palm-line feature extraction and consequently on the
palm-line feature based palmprint identification.
89
0=t 4=t 8=t 12=t
4=r
8=r
12=r
16=r
20=r
Fig. 5.5 Palm-line feature extraction results using the WLD-based palm-line feature extraction
method with different brightness contrast thresholds t and different radii of the circular mask r.
For 0=t , the core function defined in (5-5) is employed.
90
5.3 Palm-line feature matching
A matching algorithm is described in this section to measure the degree of similarity between
two palm-line feature images. Let P and Q denote two palm-line feature images, the
similarity between P and Q is defined as the proportion of matched bits to the total line bits
in P and Q, i.e.,
∑∑= =
∧+
=N
i
N
jQp
jiQjiPMM
M1 1
),(),(2 , (5-6)
where ∧ is the logical AND operation, NN × is the size of palm-line feature image, and
PM and QM are the number of line pixels in P and Q, respectively. Obviously, M is
between 0 and 1. For perfect matching, the matching score is 1. Because of imperfect
segmentation of ROI region of the palmprint image, there may still be little translation
between the palmprints captured from the same palm at different times. To overcome this
problem, we need to translate one of palm-line feature images vertically and horizontally and
then perform the matching again. The maximum of matching scores obtained at all translated
positions is taken as the final matching score.
To implement an effective matching, we encode each binary palm-line feature image to a
bitwise line vector. According to this bit representation of the palm-line feature image, a
more effective implementation of translation-invariant matching score can be defined as:
∑∑
∑ ∑
= =
−
−=
−
−=
≤≤+
++∩= N
i
N
jbb
sNN
si
sNN
sj bb
ssssjiQjiP
sjsiQjiPM
1 1
)1,min(
)11,1max(
)2,min(
)21,1max(
|2|,|1|max
),(),(
)2,1(),(2max , (5-7)
where bP ( bQ ) is the bitwise line vector of P(Q), ∩ is the AND operator detecting
agreement between any corresponding pair of bits, and s is the translation range in the
matching process. The bit representation of palm-line feature image is not only effective for
matching but also effective for storage. The size of a line vector is 512 bytes for a palm-line
feature image size of 6464× .
91
5.4 Experimental results
An extensive test has been conducted on a large-scale palmprint database. In this section, we
will first describe the large-scale palmprint database and parameters required in the proposed
line feature based palmprint recognition method and then will report and analyze the
experimental results.
5.4.1 Palmprint database
The proposed palm-line feature detector has been tested on the public palmprint database
[107] built by the Biometric Research Center at the Hong Kong Polytechnic University. This
database contains 7,752 grayscale palmprint images corresponding to 386 different palms.
Around twenty samples from each of these palms were collected in two sessions, where
around 10 samples were captured in the first session and the second session, respectively.
The average interval between the first and the second collection was two months. The size of
these images is 284384× . In our experiments, the region of interest (ROI) is the central
128128× part of the palmprint image which was cropped by using the preprocessing
technique described in Section 2.3.3. This central part represents the whole palmprint and is
used for palm-line feature extraction. In the stage of palm-line feature extraction, for (5-1)
and (5-4), a look up table is calculated for speed.
5.4.2 Parameters for the palm-line feature based palmprint
recognition
In the proposed WLD-based palm-line feature extraction method, there are three parameters,
two for the wide line detector, the brightness contrast threshold t and the radius of circular
mask r, and one for post-processing, the scale of Gaussian smoothing filterσ . In the
palm-line feature matching scheme, there is one parameter, the translation range s. Therefore,
there are four parameters for the palm-line feature based palmprint recognition. From Fig.
5.5, we can see that palm-line feature extraction results vary greatly with different parameter
92
combinations. Hence, our experiments aim at finding the optimal parameter combination for
the palm-line feature based palmprint recognition.
For line-like feature extraction, we used six brightness contrast threshold values,
{ }10,8,6,4,2,0∈t , and nine radii of circular masks, { }20,19,18,17,16,15,14,13,12∈r . In
the stage of post-processing, we employed six scales for Gaussian smoothing filter,
{ }00.2,75.1,50.1,25.1,00.1,75.0∈σ . For palm-line feature matching, we used six
translation range values, { }7,6,5,4,3,2∈s . This results in 1,944 parameter combinations for
palm-line feature based palmprint recognition. A natural way to determine the optimal
parameter combination is to find out the lowest equal error rates (EER) among all parameter
combinations. However, it is too time-consuming to obtain the EER of all parameter
combinations. Fortunately, the four parameters of our methods can be assumed independent.
Therefore, we can determine the optimal parameters one by one.
5.4.3 Palmprint matching and parameter selection
To determine the optimal parameter combination for the palm-line feature based palmprint
recognition, we test the performance of the proposed method for different parameter
selections on a training set. The training set contains 600 palmprint images from 100
different palms which randomly selected from the public palmprint database. In the training
set, each palmprint in the public database was matched with all of the others. A matching
was labeled correct if the matched palmprint was captured from the same palm, and incorrect
otherwise. A total of 179,700 )2/599600( × matchings have been performed on the
training set including 1,500 genuine matching attempts and 178,200 impostor matching
attempts. The performance of the proposed method for different parameter combinations is
measured by the equal error rates (EER) where the false match rate (FMR) is equal to the
false non match rate (FNMR). A lower EER means a higher performance.
93
Table 5.1 Equal error rates (%) for different brightness contrast thresholds (t) and
different radii of circular masks (r). Here the standard deviation of Gaussian smoothing filter
is 0.75 and the translation range is 2. For t larger than 0, the core function defined in (5-1) is
applied. For 0=t , the core function defined in (5-5) is applied.
r t 12 14 16 18 20
0 0.837 0.746 0.621 0.629 0.629
2 0.840 0.789 0.665 0.657 0.671
4 0.853 0.796 0.732 0.797 0.802
6 1.067 1.060 1.033 1.009 0.982
8 1.123 1.113 1.159 1.178 1.211
10 1.406 1.300 1.226 1.278 1.289
Fig. 5.6 EER (%) values for different combinations of brightness contrast threshold and
radius of circular mask. Here the standard deviation of Gaussian smoothing filter
is 75.0=σ and the translation range is 2=s .
94
Optimal brightness contrast threshold t
The PolyU’s palmprint database was built up under constant and uniform lighting conditions
[50]. We therefore firstly optimized the brightness contrast threshold t. Given the scale of
Gaussian smoothing filter and the translation range, we calculated the EER of different
combinations of brightness contrast thresholds and radii of circular masks. Table 5.1 shows
the results where the standard deviation of Gaussian smoothing filter is 0.75 and the
translation range is 2. All of brightness contrast thresholds given in Section 5.4.1 were used
and the radius of circular masks ranged from 12 to 20 with the step of 2. The core function
defined in (5-1) is employed for 0>t , while (5-5) is applied for 0=t .
Figure 5.6 plots the results grouped by the radius of circular mask. Obviously, for each
radius, EER takes the minimum value at 0=t . This means, with same post-processing and
matching, the brightness contrast threshold of zero always indicates the lowest EER, the best
performance, no matter what value of the radius of circular mask used. Therefore, the
optimal brightness contrast threshold is 0=t and consequently the core function defined in
(5-5) is optimal for the proposed palm-line feature extraction method.
Optimal translation range s
Translating the matched palm-line feature image is necessary for palm-line feature matching
due to the imperfect preprocessing. If the translation range is too small, the performance will
deteriorate. While if the translation range is too large, the matching of palm-line features will
become very time-consuming. So the translation range should guarantee the performance of
the palm-line feature based palmprint recognition and, at the same time, guarantee the time
effectiveness of palm-line feature matching.
95
Table 5.2 Equal error rates (%) for different translation ranges (s) and different radii of
circular masks (r). Here 0=t and the standard deviation of Gaussian smoothing filter takes
two values, 1 and 1.25.
14 16 18 r s 1=σ 25.1=σ 1=σ 25.1=σ 1=σ 25.1=σ
2 0.628 0.577 0.578 0.534 0.583 0.551
3 0.422 0.415 0.413 0.409 0.421 0.419
4 0.362 0.349 0.352 0.346 0.362 0.355
5 0.357 0.339 0.347 0.335 0.356 0.346
6 0.356 0.339 0.346 0.335 0.355 0.343
7 0.356 0.339 0.346 0.334 0.355 0.343
Given the optimal brightness contrast threshold 0=t , Table 5.2 shows the EER values for
different combinations of three other parameters: translation range s, circular mask radius r,
and Gaussian smoothing filter scaleσ . Here we used all translation ranges given in Section
5.4.2 and two scales of Gaussian smoothing filter, { }25.1,0.1∈σ . Three radii of circular
masks, { }18,16,14∈r , were employed because in Table 5.1 they gives the corresponding
minimum EER value (marked in bold) for each brightness contrast threshold.
The EER values in Table 5.2 are divided into two groups according to the two scales of
Gaussian smoothing filter and plotted in Figure 5.7(a) and 5.7(b), respectively. We can see
that, given the combination of r andσ , the EER values decrease with increasing the
translation range and reach the corresponding lowest value when { }7,6,5∈s . Considering
the balance between the verification performance and the time effectiveness, we select
5=s as the optimal translation range.
96
(a)
(b)
Fig. 5.7 EER (%) values for different parameter combinations of circular mask radius and
translation range given the standard deviation of Gaussian smoothing filter (a) 0.1=σ and
(b) 25.1=σ , respectively. Here the brightness contrast threshold is 0=t .
97
Table 5.3 Equal error rates (%) for different radii of circular masks (r) and different
scales of Gaussian smoothing filter with 0=t and 5=s . The core function defined in (5-5)
is used.
r σ 12 13 14 15 16
1.00 0.452 0.346 0.357 0.351 0.347 0.371
1.25 0.348 0.341 0.339 0.335 0.335 0.340
1.50 0.345 0.339 0.335 0.333 0.335 0.337
1.75 0.340 0.338 0.336 0.334 0.339 0.337
2.00 0.345 0.339 0.338 0.338 0.339 0.340
0.366 0.341 0.341 0.338 0.339
Optimal circular mask radius s and Gaussian smoothing filter scaleσ
Now that the optimal brightness contrast threshold and translation range have been
determined, we optimize the circular mask radius and Gaussian smoothing filter scale in this
subsection. The EER values of different radii of circular masks against different scales of
Gaussian smoothing filters are shown in Table 5.3 with the optimal combination of
brightness contrast threshold and translation range. Five radii values { }16,15,14,13,12∈r
and five scales { }0.2,75.1,5.1,25.1,0.1∈σ are employed in this table.
We calculated the average EER value for each σ which are listed in the last column and
also for each r which are listed in the last row. The average EER for Gaussian smoothing
filter scales reaches the minimum at 5.1=σ and for circular mask radii at 15=r . This
combination of the two parameters gives the lowest EER 0.333%, which is the best
verification performance we can get. Therefore, the optimal combination for parameters,
brightness contrast threshold t, circular mask radius r, Gaussian smoothing filter’s scaleσ ,
and translation range s, is )5,5.1,15,0(),,,( =srt σ .
98
Fig. 5.8 Genuine and impostor distributions for the optimal parameter combination.
5.4.4 Verification
Verification refers to one-to-one matching which seeks to answer the question “whether the
person is whom he or she claims to be” by examining his or her biometric traits. In palmprint
verification, a user indicates his or her identity and thus the input palmprint is matched only
against the corresponding stored template. To determine the verification accuracy of our
WLD-based palmprint recognition method, each palmprint was matched with all of the other
palmprints in the public database. All matching scores are thereby divided into two groups:
genuine matching scores and impostor matching scores. The probability distributions of
genuine and impostor for the optimal parameter combination )5,5.1,15,0(),,,( =srt σ , are
plotted in Fig. 5.8. To measure the separability of the distributions of genuine and impostor,
we introduced the decidability index 'd (d-prime) [120], which is defined as the separation
between the means of the two distributions, divided by the square-root of their average
variance. The decidability index of our WLD-based method with the optimal parameter
combination is listed in Table 5.4.
99
Table 5.4 Comparison of our method with four representative palmprint verification
methods on PolyU database.
Method 'd EER
PalmCode 4.973 0.647%
CompCode 4.839 0.409%
Ordinal code 5.418 0.400%
Wu’s method 6.335 0.413%
WLD 5.453 0.373%
Fig. 5.9 Verification performance comparison on PolyU Palmprint Database.
100
Table 5.4 compares the accuracy measure of our WLD-based method with three coding
approaches, PalmCode [50], Competitive Code [70] and Ordinal Code [71], and one
palm-line feature based approach, Wu’s method [61]. The verification performances of the
five methods are illustrated in Fig. 5.9 by the receiver operating characteristic (ROC) curves
which plots FNMR as a function of FMR [103]. We can see our WLD-based method gives
the lowest EER among the five methods and the largest decidability index except Wu’s
method. The experimental results indicate that width features of palm lines are useful and
helpful for palmprint recognition.
5.4.5 Identification
Identification, also called one-to-many (1-to-S) matching, involves answering the question
“who is this person?” In our experiments, two registration databases are setup with the same
scale 386=S , which is the total number of different palms in the public palmprint database.
The first registration database RegDB1 contains S templates and the second one RegDB2
contains S3 templates. To establish the identification accuracy of our method, all templates
in both registration databases are randomly selected from the palmprint images captured in
the first session and the identification database consists of all the palmprint images captured
in the second session. The two registration databases, RegDB1 and RegDB2, and the
identification database contain 386, 1158, and 3863 palmprint images, respectively. Each
sample in the identification database is compared with all templates in each registration
databases and the template giving the maximum matching score is labeled as the
identification result. Table 5.5 compares the recognition rates of our WLD-based method on
both registration databases with PalmCode, Competitive Code, Ordinal Code, and Wu’s
method. The number of mismatched samples of each method is also listed in Table 5.5. In
palmprint identification, the accuracy of our WLD-based method outperforms other four
methods on both registration databases.
101
Table 5.5 Comparison of identification accuracy on two registration databases.
RegDB1 RegDB2 Method
Recognition rate Incorrect # Recognition rate Incorrect #
PalmCode 97.98% 78 99.02% 38
CompCode 98.91% 42 99.74% 10
Ordinal code 98.99% 39 99.82% 7
Wu’s method 98.71% 50 99.43% 22
WLD 99.22% 30 99.84% 6
5.5 Conclusion and discussions
We have presented a WLD-based method to simultaneously extracting structure and width
features of palm lines for palmprint recognition. A palm-line feature extraction method based
on the wide line detector is proposed to extract line-like features which contain width
information of palm lines. In the post-processing stage, a Gaussian smooth filter is employed
for palm-line feature image denoise. An effective translation-invariant matching method is
described for palm-line feature matching. An experimental scheme is designed to determine
the optimal parameter combination, )5,5.1,15,0(),,,( =srt σ . An extensive test of the
proposed WLD-based method with the optimal parameter combination has been conducted
on PolyU Palmprint Database compared with four representative palmprint identification
methods. In palmprint verification, the proposed WLD-based method offers the lowest EER
0.373%. In palmprint identification, the proposed WLD-based method generates the least
number of incorrect matching (30 in 3863 on RegDB1 and 6 in 3863 on RegDB2) and
accordingly the highest recognition rate, 99.22% on RegDB1 and 99.84% on RegDB2,
respectively. The experimental results demonstrate that the performance of the proposed
WLD-based method is comparable with the state-of-the-art algorithms of palmprint
recognition and thereby palm-line features can be used to recognize palmprints.
102
Chapter 6
Automatic Extraction of Cracks on
Tongue Images
This chapter attempts to detect tongue cracks, one of pathological features in tongue
diagnosis. A framework for adaptive tongue crack detection is proposed. Based on the
presented wide line detector described in Chapter 3, an automatic tongue crack detection
scheme is derived to extract tongue cracks. The wide line detector extracts the whole of the
line by employing an isotropic nonlinear filter and describes the relationship between the
size of the isotropic filter, i.e. the scale of this detector, and the width of detected lines. Due
to the large range of widths of tongue cracks, the maximum widths of cracks vary greatly
with different tongue images and consequently the sizes of the isotropic filters are very
different. To implement the proposed scheme totally automatically, an adaptive algorithm of
line width estimation is designed. The proposed scheme has been tested on a set of typical
cracked tongue samples and our experimental results demonstrate its effectiveness.
6.1 Introduction
Tongue diagnosis [75,76] is one of the most important and valuable diagnostic methods in
traditional Chinese medicine (TCM) and has been widely used to clinical analyses and
applications for thousands of years. In TCM theory [122], the tongue is connected with a
number of viscera via the meridians and accordingly the essential of the viscera can ascend
to nourish the tongue and pathological processes are reflected on it. Whenever there is a
complex disorder full of contradictions, examination of the tongue can instantly clarify the
main pathological processes. Therefore, it is very valuable in both clinic applications and
self-diagnosis. Moreover, tongue diagnosis is a non-invasive technique that is in accord with
103
the most promising direction in the 21st century: no pain and no injury.
Some researchers recently have paid attention to computerized tongue diagnosis
[8,77-86,125] by developing a set of objective and quantitative features and measurements
based on the theories of tongue diagnosis. So far, most investigation has been focused on
extraction of chromatic features [8,78,80-82,84-86,125] and textural features
[8,77-79,82,84,86,125]. Actually, according to the theories of tongue diagnosis, some
features on tongue surfaces are pathological forms and thereby have clinical significance, for
example, cracks on the tongue surface.
Tongue crack is one kind of textural features and frequently seen in clinical practice.
Tongue cracks refer to the surface of the tongue covered with fissures or lines in deep or
shallow shape, which are induced by the fusion or separation of the ligular papillae [123].
Normally, the tongue’s surface should be smooth and soft and show no cracks. A cracked
tongue indicates exhaustion of blood: the fewer and more superficial the cracks, the miler the
disease; the deeper and more numerous the cracks, the more serious the disease [76]. The
clinical significance of cracks depends on the tongue body color, the location of the cracks
and their shape and depth.
In spite of the clinical significance of tongue cracks, there is by far no paper, to the best of
our knowledge, concerning extraction of cracks on the tongue surface. This is our motivation
to develop the research on tongue crack extraction. Now a question arises: what is tongue
crack? Although tongue cracks appear as curvilinear structures (also simply called lines) on
the tongue surface, not all lines on the tongue surface are cracks. In the theories of tongue
diagnosis, flat and short lines which are no deeper than 0.35mm and no longer than 0.55cm
on the tongue surface have little clinical significance [75,76]. However, in computerized
tongue diagnosis, since the captured tongue pictures are 2D images, the information of depth
has been lost. Fortunately, the depth of cracks is normally related to their widths. Generally
speaking, the deeper the cracks the wider they are, and vice verse. Therefore, based on the
investigation on a number of cracked tongue images and suggestions from expert, we give
104
the definition of tongue cracks in the sense of imaging: tongue crack is kind of lines on the
tongue surface not too short and not too narrow.
So far, many algorithms for line detection have been developed for different applications.
However, there exist some shortcomings make them unsuitable for tongue crack extraction.
Hough transform [15,92-93], which is a widely used line detection method, was initially
proposed to find analytically defined curvilinear structures (e.g., straight lines, circles,
ellipses etc.). Although the generalized Hough transform [17] can be used to detect arbitrary
curvilinear structures in theory, it requires the complete specification of the exact shape of
the curvilinear structure which is very difficult and even unfeasible for tongue cracks due to
their complex shapes. One powerful line detection approach is based on the derivative
operations [3,26,28,30-31,35,94]. Steger [3] proposed a ridge-based line detector using
two-dimensional Gaussian partial derivative kernels. This detector extracts the ridges as
points for which the intensities are maxima or minima in the main principal curvature
direction, i.e., the direction of the maximum eigenvalue of the Hessian matrix. Since
Gaussian kernels are used to estimate the derivatives of the image, the line detector can scale
to lines of arbitrary widths. However, this approach is sensitive to noise due to the use of
second order derivatives, which makes it unsuitable for tongue crack extraction because of
the rough surface of tongue and the reflection on the tongue surface. Another simple but
powerful line detection approach is the line operator [45] which is successfully used to detect
linear structures in mammographic images [4]. The line operator extracts linear features
based on the difference between the average gray level of the pixels lying on an orientated
local line passing through the target pixel and the average intensity of all the pixels in the
similarly orientated neighborhood. The line operator improves the line signal to background
noise ratio by averaging the grey levels. However, this approach tends to output results with
similar widths whatever the true width of the detected lines. This is a disadvantage to extract
tongue cracks which widths normally vary in a large range.
Recently, a wide line detector [62] is developed to extract a line completely. Unlike line
105
detectors which use directional derivatives, this method applies a nonlinear filter to extract
the whole of the line without any derivative, which makes this line detector insensitive to
noise. In the wide line detector, a definite relationship is given between the width of detected
lines and the size of an isotropic filter which is employed to obtain an isotropic response.
Due to no Gaussian kernel used to detect lines, even narrow lines can be extracted well as
long as the intensity contrast between the narrow lines and the background is large enough,
which indicates this wide line detector can keep the width of the detected line as what it is
and thereby can extract lines of different widths. Therefore, we employ the wide line detector
to extract tongue cracks.
Purpose of this chapter is to attempt to extract tongue cracks by applying and improving
the wide line detector based on tongue crack features. An automatic tongue crack detection
scheme has been developed by modifying the wide line detector according to the
characteristics of tongue cracks. The size of the isotropic filter which determines the widths
of detected lines is related to the maximum width of cracks expected to extract. The widths
of tongue cracks, however, change greatly and the maximum width of tongue cracks varies
with tongue images. To implement the proposed scheme automatically, we also design an
adaptive algorithm to estimate the maximum width of tongue cracks.
The chapter is organized as follows. Section 6.2 describes the automatic tongue crack
detection scheme. In Section 6.3, we present the adaptive line width estimation algorithm.
Section 6.4 describes the results of our experiments and gives the analyses. Section 6.5 offers
our conclusion.
6.2 Methodology of crack extraction
Fig. 6.1 provides an overview of the proposed framework which indicates the automatic
tongue crack detection scheme. Given a tongue image (see Fig. 6.3a), in the first step, we
extracts the tongue body and, at the same time, creates the tongue body mask. In the second
step, we first utilize the extracted gray-level tongue body image to calculate the two
106
parameters for the wide line detector (WLD): the brightness contrast threshold and the radius
of the circular mask which is obtained by estimating the maximum width of cracks in the
tongue body image. The WLD algorithm is then employed to give the crack response image.
Finally, the crack response image is post-processed to obtain the final results. The tongue
crack detection module represents a modified version of the wide line detector and the
adaptive estimation of the maximum width of cracks is our novel contribution. In this section,
we describe the methodology of tongue crack extraction except the adaptive algorithm of the
line width estimation which will be presented in Section 6.3.
Fig. 6.1 The diagram of tongue crack extraction procedure using the WLD.
Contour shrinking
Contour Tongue Image
Tongue Body Mask
Binarization
Initialization
Tongue body segmentation
Tongue
Crack
Extraction
(WLD)
Brightness contrast
thresholding
Radius of the circular
mask
Final result
Preprocessing Extracting Tongue Cracks i
Smoothing
Estimate maximum width of cracks
Post-processing
107
6.2.1 Preprocessing
The purpose of this subsection is to segment the region of interest, i.e. the tongue body, from
the captured tongue image which has a large background. We employ the segmentation
method described in [83] to obtain the contour of the tongue body and segment the captured
tongue image to the contour tongue image. The tongue contour is then shrunk so that the
influence of the uneven illumination caused by the deformation close to the boundary can be
eliminated. The shrunk distance should not be too large since some tongue cracks are very
near the boundary of the tongue body (see Fig. 6.2a). Therefore, each pixel of the tongue
contour is moved to the centroid of the tongue body with a certain distance (as shown in Fig.
6.2b) and the shrunk contour is calculated by
220)()( ncnc
nccs
yyxx
xxdxx
−+−
−−= , (6-1)
220)()( ncnc
nccs
yyxx
yydyy
−+−
−−= (6-2)
where ),( nn yx is the centroid of the tongue contour, ),( cc yx is the coordinate of the
tongue contour pixel, ),( ss yx is the coordinate of the shrunk contour pixel, and 0d is the
shrunk distance. The tongue body mask is created by binarizing the contour tongue image
according to the shrunk contour and the tongue body image is consequently obtained, as
shown in Fig. 6.2c and Fig. 6.2d, respectively.
6.2.2 Extracting tongue cracks
We modified the method proposed in [95] to detect the tongue crack. This method, also
called the wide line detector [62], extracts a line completely by using a nonlinear filter based
on the isotropic response via a Gaussian weighting mask. Here, we replaced the Gaussian
weighting mask G with an inverse-Gaussian profile
108
(a)
(b) (c)
(d)
Fig. 6.2 Preprocessing of tongue crack extraction. (a) Captured tongue image, (b) contour
tongue image with the centroid and corresponding contour, (c) tongue body mask, and (d)
extracted tongue body image.
109
(a)
(b) (c)
(d) (e)
Fig. 6.3 Tongue crack extraction. (a) The ground truth for Fig. 6.2d, (b) The line-strength
image using the wide line detector (WLD), (c) tongue crack image obtained by
postprocessing (b), and tongue crack extraction results using (d) the Line Operator (LO) and
(e) Steger’s method, respectively.
110
( ) ∑+≤≤−+≤≤−
−+−−
−+−−
−+
−=
ryyryrxxrx
ryyxx
ryyxx
er
eryxyxIG
0000
2
20
20
2
20
20
,
2)()(
2
2)()(
00
12
1),,,,( , (6-3)
where ),( 00 yx is the coordinate of the center, ),( yx is the coordinate of any other pixel
within the mask, and r is the radius of the circular mask. The equation (6-3) gives a
normalized version of the inverse-Gaussian profile.
The wide line detector first examines the intensity of the center of the mask and groups
pixels having similar intensities to the center into the weighted mask having similar
brightness (WMSB). Since tongue cracks are dark lines, this similarity in our method is
measured by:
⎪⎩
⎪⎨
⎧≥⎟
⎠
⎞⎜⎝
⎛ −•=
otherwise
),(),(if),(),(
sec),,,( 00
500
00IG
yxIyxIt
yxIyxIhIGyxyxwmsb , (6-4)
where xx eexh
−+=
2)(sec , ),( yxI is the brightness of the pixel ),( yx of the tongue
body image, t is the brightness contrast threshold and wmsb is the output of the weighting
comparison. This function gives a smooth profile and does not have too large an effect on
wmsb as a pixel’s brightness changes slightly, especially when it is near the threshold and
further, has proved to be optimal in Chapter 4. This function gives a trade-off between the
stability about the threshold and the requirement of the detector, which is to count pixels that
have similar intensity to the center into the mass of the circular mask.
This comparison is done for each pixel within the mask. The WMSB mass of the center
),( 00 yx is given by ∑ +≤≤−
+≤≤−=
ryyry
rxxrxyxyxcyxm 00
00),,,(),( 0000 .The line strength L is the
inverse WMSB mass obtained by using the following rule:
111
⎩⎨⎧ <−
=otherwise
gyxm ifyxmgyxL
),(0
),(),( 0000
00 . (6-5)
Here g is the geometric threshold and 2/maxmg = , where maxm is the maximum value
which m can take. As a normalized circular mask is used, maxm is not larger than but very
close to one and thereby the line strength ranges between zero and one. Fig. 6.3b shows the
line strength image of Fig. 6.2d.
The WLD requires two parameters – the brightness contrast threshold, t, and the radius of
the circular mask, r. In our method, the brightness contrast threshold is defined by
( ) 21
1
2),(),(
11
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛−
−= ∑
=
N
immii yxIyxI
Nt , (6-6)
where ∑=i
iimm yxIN
yxI ),(1),( and N is the number of non-zero pixels of the tongue
body image which is obtained through the preprocessing described in Section 6.2.1. The
brightness contrast threshold t is the standard deviation of the tongue body region.
In [62,95], the relationship between the width of lines detected and the size of Gaussian
weighting masks is given via analysis. In the same way, if a line of width w×2 is fully
detected by using an inverse-Gaussian weighting mask with radius r, it requires
⎟⎟⎠
⎞⎜⎜⎝
⎛−=−<−
−+
−+
−
∫∫∫∫ 211
211 2
1222 2
22
2
22
erdxdyedxdyeC
ryx
L
ryx
π , (6-7)
where C denotes the region of the circular mask and L the region of the line which passes
through the center of the circular mask. The relationship between the width of the detected
line and the radius of a circular mask with inverse-Gaussian profile is consequently
determined in terms of (6-7). Given the circular mask of radius r, the critical width of line
detected, cw×2 , is obtained when the left and right arguments of (6-7) are equal. As the
analytic form of the left function is not available, we only give the approximate results in
112
Table 6.1. We also show in Table 6.1 the relationship between the width of line detected and
the radii of circular masks with Gaussian profile ( )222 2/)(exp ryx +− and with constant
weighting, respectively. Comparing the results, we find out that the detection capacity of the
WLD using an inverse-Gaussian weighting mask is higher than that using the corresponding
circular mask with either constant weighting or Gaussian profiles. It means that, with the
same radius r, the circular mask of inverse-Gaussian profile can detect widest lines, while, in
order to detect lines of same width, the size of an inverse-Gaussian weighting mask needed
is smallest.
6.2.3 Post-processing
In the final stage, pixels of line strength below a threshold (=0.1) are removed firstly. Tongue
cracks are extracted by applying the technique of hysteresis thresholding to the line-strength
images, followed by some morphological operations. The hysteresis thresholding [126] uses
two thresholds: one is high and the other is low. Any pixel in the line-strength image that has
a value greater than the high threshold is presumed to be a line pixel. Then, any pixels that
are connected to any line pixel and that have a value greater than the low threshold are also
selected as line pixels. In the morphological operations, the crack will be discarded if (i) its
length, l, is very short (<20 pixels), (ii) its area, A, is too small (< 100 pixels), and (iii) its
eccentricity, e, is less than 0.9. The eccentricity is the ratio of the distance between the foci
of the ellipse and its major axis length. Here the major axis length of the ellipse is
approximated to l and the minor axis length of the ellipse is calculated by lAb = . So the
eccentricity e can be estimated by
4
21
lAe −= . (6-8)
Fig. 6.3c shows the tongue crack detection result of Fig. 6.2d
113
Table 6.1 Comparison of the relationship between radii of different profiles and
approximately critical widths of detected lines (ACWL).
Critical Width of Line Detected ( cw×2 )
Constant weighting Gaussian weighting Inverse-Gaussian weighting
Radius of Gaussian mask (r)
ACWL Digital ACWL ACWL Digital ACWL ACWL Digital ACWL
3 2.4 2 2.6 2 4.0 4
4 3.2 3 3.5 3 5.3 5
5 4.0 4 4.4 4 6.6 6
6 4.8 5 5.3 5 7.9 8
7 5.6 5 6.2 6 9.2 9
8 6.4 6 7.1 7 10.6 10
9 7.2 7 7.9 8 11.9 12
10 8.0 8 8.8 9 13.2 13
11 8.8 9 9.7 10 14.5 14
12 9.6 9 10.6 10 15.8 16
13 10.4 10 11.5 11 17.2 17
14 11.2 11 12.4 12 18.5 18
15 12.0 12 13.2 13 19.8 20
16 12.8 13 14.1 14 21.1 21
17 13.6 13 14.9 15 22.4 22
18 14.4 14 15.8 16 23.7 24
19 15.2 15 16.7 17 25.0 25
20 16.0 16 17.6 17 26.4 26
21 16.8 17 18.5 18 27.7 28
22 17.6 17 19.4 19 29.1 29
23 18.4 18 20.3 20 30.4 30
24 19.2 19 21.2 21 31.7 32
25 20.0 20 22.1 22 33.0 33
26 20.8 21 22.9 23 34.3 34
114
6.3 Adaptive line width estimation
In line detection the scale of the filter is one very important factor to detect lines of arbitrary
widths. An optimal scale of the filter should be big enough but not too big to detect all lines
in an image. Since the widths of tongue cracks may vary largely with tongue body images,
the optimal scale of the filter will be quite different from image to image. Consequently we
should select the corresponding optimal scale of the filter for each tongue body image rather
than use only one filter-scale for all tongue body images. In terms of WLD, the scale of the
filter, i.e. the radius of the circular mask, is related to the maximum width of tongue cracks
expected to detect in a given tongue body image. Therefore, the estimation of the maximum
width of cracks in a tongue body image is essential and necessary for applying the wide line
detector to extract tongue cracks correctly and completely. Normally, the estimation of line
widths is done by human being. Adaptively estimating widths of curvilinear structures with
various widths and various orientations still remains a big challenge in image analysis and
pattern recognition due to complicated scenes. For the purpose of fully automatic extraction
of tongue cracks, in this section, we attempt to realize the adaptive estimation of the
maximum width of tongue cracks by utilizing their properties.
Tongue cracks can be considered as dark lines with a bar-shaped profile
⎪⎩
⎪⎨
⎧
>−<≤≤−
=wxbwxa
wxwx
0)(l , (6-9)
where w×2 is the line width, 0>a , and 0>b . Ideally, lines would have the same
contrast on both sides, i.e., ba = , as shown in the first row of Fig. 6.4a. However, this
assumption is not always true for tongue cracks due to the tongue coating and
illumination.The first rows of Fig. 6.4b and 6.4c show the asymmetrical bar-shaped lines
where ba < and ba > , respectively.
115
(a) (b) (c)
Fig. 6.4 Illustration of the line width estimate. The first row is the three dark line models.
The second row shows the corresponding Gaussian smooth profiles. The last row is the
first-order derivatives of Gaussian (DoG) profiles shown in the second row. The line width
can be obtained by calculating the distance between the pair of minimum and maximum of
DoG filtered results along the line profile.
If the bar-shaped profiles )(xl are convolved with a Gaussian kernel
( ) ( )σπσσ 2/2/exp 22xg −= , a smooth profile )(xs is obtained in each case, as shown in
the second row of Fig. 6.4. Thus, a useful criterion for dark lines can be employed to
estimate the line width. The left edge of the dark line, i.e., the decreasing edge, has the
minimum of )(xs′ , while the right edge of the dark line, i.e., the increasing edge, has the
maximum of )(xs′ , as shown in the last row of Fig. 6.4. Mathematically, )(xs′ can be
obtained by convolving the bar-shaped profiles )(xl with the first-order derivative of
Gaussian kernel (DoG) which is defined as:
Tongue crack models
Gaussian smooth profiles
DoG profiles
116
2
2
232
)( σσσπ
x
exxDoG−−
= , (6-10)
whereσ is the scale of Gaussian kernel. Therefore, the line width can be estimated by
calculating the distance between the pair of minimum and maximum of the DoG filtered
response.
A line width estimation algorithm is designed based on multi-scale analyses to adaptively
estimate the maximum width of cracks for each input tongue images. This estimation
algorithm is specified in Fig. 6.5. Since tongue cracks are normally horizontal and vertical,
we sample some rows and columns for analyses (step 01). In order to eliminate the influence
of reflecting points, one-dimensional median filter is first applied to each sampled vector
(step 02). For each filtered vector v, a multi-scale analysis is developed in the scale space
∑ to calculate the normalized DoG filtered response
σσ σ dvdv norm ×=− , (6-11)
σσ DoGvdv ∗= , (6-12)
and then find the maximal vertical distance vdσ among all pairs of local minimum and local
maximum of normdv −σ on each scale ∑∈σ (step 05-06). The candidates of the maximum
width of cracks consist of the maximum width of each vector vw , where the vertical distance
reaches the maximum in the scale space (step 07). The maximum width of cracks w is
determined by the maximum value of candidates which vertical distance is larger than a
threshold vVvd dratioT ∈∪×= max (step 08). Here we define ratio=0.8.
117
Fig. 6.5 Line width estimation algorithm.
Table 6.2 The estimation results of maximum widths of cracks by applying the algorithm
described in Figure 6.5. The true widths of cracks are also listed for reference.
Image # Estimated Max-Width (pixel)
True max-width (pixel)
Fig. 6.2d 12 12
Fig. 6.6a 14 14
Fig. 6.7a 13 13
Fig. 6.8a 17 16
Fig. 6.8b 21 18
Input: tongue body image I
setΣ of scales
Output: maximum width w of tongue cracks in I
01 Sample a set of vectors 0V in tongue body image I along the horizontal and vertical
direction;
02 Obtain filtered vectors V by apply the one-dimensional median filter to vectors 0V ;
03 For each Vv∈
04 For each Σ∈σ
05 Obtain the normalized filtering response normdv −σ of vector v for scaleσ ;
06 Calculate the max-vertical distance vdσ of normdv −σ for scaleσ ;
EndFor
07 Determine the maximum width vw of vector v, where the corresponding vertical
distance vv dd σσ ∑∈∪= max ;
EndFor
08 ( )vVvd
vv dratioTdVvww ∈∪×=>∈= max,|max .
118
The estimation of max-width of cracks in Fig. 6.2d is listed in Table 6.2. In WLD, the filter
scale is the radius of the circular mask, which is determined by the maximum width of cracks.
The relationship between the radius of circular mask and the max-width of crack is given in
Table 1 and thus, according to the estimated max-width of cracks, the size of circular mask
used for Fig. 6.2d is 4141× with the radius of 20 pixels wide. The extraction result obtained
by WLD using the corresponding circular mask is shown in Fig. 6.3c.
Actually, the max-width of cracks can determine the filter scale not only for the WLD but
also for other line detection methods as long as the relationship is definite between the filter
scale and the line width. In [3], Steger gave the restriction on the scale of Gaussian kernel to
ensure that the salient lines can be detected. In [4], the principle of the line operator (LO)
requires that the filter scale, i.e., the size of the square neighborhood, should be larger than
the maximum line width to guarantee the complete extraction. Here, we employ the two
popular wide line detection methods to extract cracks in Fig. 6.2d under the corresponding
scales determined by the estimated max-width, which are shown in Fig. 6.3d and Fig. 6.3e,
respectively. Compared with the ground truth as shown in Fig. 6.3a, the extraction results of
the wide cracks by using the three methods show that the scales for three methods
determined by the estimated max-width of cracks are proper for the tongue cracks and
thereby indicate that the proposed estimation algorithm can work well.
6.4 Experimental results and analysis
We have conducted a test of the described scheme on a set of typical tongue crack samples
purposely chosen by the experts for diagnosis. To establish the effectiveness of the wide line
detector for the tongue crack extraction, we employed two popular line detection methods to
extract tongue cracks by replacing the WLD in the stage of extracting tongue cracks and
compared the performance of the three methods. In this section, we will firstly describe the
measure of performance evaluation of tongue crack extraction methods and then present and
discuss the experiment results.
119
6.4.1 Performance evaluation
Performance evaluation of medical image segmentation is a challenging job due to the
complexity and difficulty of medical image segmentation. Normally, a line detector
performance is evaluated with specified ground truth [127]. Once the ground truth is given,
quantitative evaluation can be readily carried out by comparing detected features with the
ground truth features. For each tongue crack sample used in our experiment, the ground truth
was drawn by hand, which is a binary map where tongue cracks are marked in white.
Let GTC and TDC denote the set of tongue crack pixels of the ground truth and tongue
cracks detected by an approach, and GTB and TDB the set of background pixels,
respectively. The set of correctly detected tongue crack pixels is true positives
( GTTD CCTP I= ). The set of falsely detected tongue crack pixels is false positives
( GTTD BCFP I= ). The set of correctly marked non-crack pixels is true negatives
( GTTD BBTN I= ). The true positive rate (TPR) is established by dividing the number of
true positives by the total number of crack pixels in GT. The false positive rate (FPR) is
computed by dividing the number false positives by the total number of non-crack pixels in
GT. A measurement of accuracy (Ac) can be defined by the sum of true positives and true
negatives divided by the image size.
We also define a performance measure of a tongue crack extraction method as:
)()()(
)(FNnumFPnumTPnum
TPnumPM++
= , (6-13)
where num(X) denotes the number of the set X and FN is the false negatives which the set of
the ground truth missed by the detection results. The performance measure PM ranges from
0 to 1. For perfect tongue cracks detection, PM=1. For all other cases, the performance
measure is smaller than one. The definition of PM guarantees that the performance measure
is closer to zero with more tongue crack pixels missed and/or falsely detected by a line
120
Table 6.3 Comparisons of performance evaluation between different methods used in
Fig. 6.3.
Method TPR FPR Ac PM
Steger 0.7844 0.0109 0.9863 0.4419 LO 0.7738 0.0084 0.9886 0.4840
WLD 0.8181 0.0019 0.9956 0.7196
detector. The performance measure is used to evaluate the effectiveness of tongue crack
extraction methods.
Table 6.3 shows the comparisons of performance evaluation between different methods
used in Fig. 6.3. The proposed WLD-based scheme gives the highest value of true positive
rate and the lowest value of false positive rate, which resulted in achieving the highest
accuracy score and the highest performance measure among all extraction methods. In the
following experiments, we will employ the true positive rate, false positive rate, accuracy,
and performance measure to evaluate the performance of different methods.
6.4.2 Experimental results
In this section, we will show the effectiveness of the WLD-based scheme by comparing with
two popular line detection approaches, the line operator (LO) and Steger’s unbiased detector.
In WLD, the two parameters are the brightness contrast threshold which is calculated by (6)
and the radius of inverse-Gaussian weighting mask which is determined by the max-width of
tongue cracks maxw as shown in Table 6.1. The parameter for Steger’s approach isσ , the
standard deviation of a 2-D Gaussian partial derivative kernel, which is restricted by
3/w≥σ (w is half of the line width) and the parameter for LO is l , the length of line,
which should be at least larger than the line width expected to detect. In our experiments, we
set max3.0 w=σ and max2w=l . The max-width of tongue cracks for each sample is
automatically estimated by applying the algorithm presented in Section 6.3. Table 6.2 lists
the estimated max-widths of tongue crack samples used in our experiments, as well as the
corresponding true max-widths.
121
(a)
(b) (c)
(d) (e)
Fig. 6.6 (a) The vertical cracks with (b) the ground truth and extraction results by using (c)
WLD, (d) LO, and (e) Steger’s method, respectively.
122
Table 6.4 Comparisons of performance evaluation between different methods used in
Fig. 6.6.
Method TPR FPR Ac PM
Steger 0.7588 0.0134 0.9846 0.3014
LO 0.8025 0.0092 0.9892 0.3943
WLD 0.8563 0.0016 0.9972 0.7279
The first example is illustrated in Fig. 6.6a, which contains a vertical crack on the rough
tongue surface. The line width of this tongue crack sample is calculated by the algorithm
described in Fig. 6.5. The estimated max-width of this crack is 14 pixels and accordingly the
radius of inverse-Gaussian weighting mask used in WLD is 11 pixels wide, referring to
Table 6.1. The extraction result of this tongue crack sample using WLD is shown in Fig. 6.6c
and the ground truth for this tongue crack sample in Fig. 6.6b. The results obtained by LO
and Steger’s method are shown in Fig. 6.6d and Fig. 6.6e, respectively. Table 6.4 shows the
values of true positive and false positive rates reported by different methods. The wide line
detector generates the highest value of true positive rate and the lowest value of false
positive rate, which results in the highest accuracy and highest PM value among three
methods. Compared with the ground truth, the wide line detector output the best extraction
results without any piece of falsely detected cracks, while both LO and Steger’s method
produce spurious cracks where the shadow occurs. The spurious cracks lead to very high
false positive rates for both LO and Steger’s method, 0.92% and 1.34%, respectively. The
reason for the failure of the two line detectors is that both of them extract lines based on the
derivative of kernels in nature, which is sensitive not only for the linear structures but also
for edges. In WLD, no derivative operation is directly used and consequently produces much
more stable and sensible extraction results.
123
(a)
(b) (c)
(d) (e)
Fig. 6.7 (a) The horizontal cracks with (b) the ground truth and extraction results by using
(c) WLD, (d) LO, and (e) Steger’s method, respectively.
124
Table 6.5 Comparisons of performance evaluation between different methods used in
Fig. 6.7.
Method TPR FPR Ac PM Steger 0.8423 0.0285 0.9686 0.3787
LO 0.8435 0.0276 0.9695 0.3859 WLD 0.8734 0.0081 0.9892 0.6471
Figure 6.7 illustrates the strong noise rejection of the wide line detector as compared with
the LO and Steger’s method. The example shown in Fig. 6.7a is a typical sample of
horizontal tongue cracks, which ground truth is given by expert, as shown in Fig. 6.7b. The
estimated max-width of this tongue crack sample is of 13 pixels and accordingly the radius
of circular mask used for WLD is 10 pixels wide. Figure 6.7c-e display the extraction results
by applying the WLD, LO, and Steger’s method, respectively. The performance evaluation
results of the three methods are listed in Table 6.5. Notice that both LO and Steger’s method
produce false positive rates much higher than that of WLD. Compared with the ground truth,
most of the spurious crack pixels detected by the line operator and Steger’s method are found
as the branches of horizontal cracks which are induced by the reflection on the tongue
surface. In contrast, the WLD produces very few reflection-caused spurious crack pixels due
to the use of equation (6-4) which is defined specially for dark line detection. Therefore, the
WLD has strong noise rejection ability for tongue crack extraction.
The last clinical example is presented in Fig. 6.8 and consists of two samples (as shown in
Fig. 6.8a and Fig. 6.8b), both of which are typical irregular cracks which demonstrate more
clinical significance than vertical and horizontal cracks. The two samples contain many
pieces of cracks with different widths and directions and thereby are more of challenge. The
ground truth for both samples is labeled by expert and shown in Fig. 6.8c and Fig. 6.8d,
respectively. The estimated max-width values are 17 pixels and 21 pixels for Fig. 6.8a and
Fig. 6.8b and accordingly the corresponding radii of circular masks are 13 pixels wide and
16 pixels wide, respectively. The crack extraction results of two samples by using the three
125
(a) (b)
GT (c) (d)
WLD (e) (f)
LO (g) (h)
126
Steger (i) (j)
Fig. 6.8 (a-b) Two irregular cracks with (c-d) the corresponding ground truth and extraction
results by using (e-f) WLD, (g-h) LO, and (i-j) Steger’s method, respectively.
line detection methods are displayed in Fig. 6.8e-j and the corresponding performance
evaluation results are listed in Table 6.6 and Table 6.7, respectively. Again, our WLD
produces the best results for both samples. Compared these extraction results of tongue
cracks by the WLD, LO, Steger’s method, it can be observed that the detected cracks by
using WLD have different widths which are close to the true widths of cracks in both
samples, while the widths of detected cracks in each sample by either LO or Steger’s method
seem to be very similar, which results in the highest true positive rate of the wide line
detector among different methods. This illustrates that, given the scale for a method, our
WLD can remain the width of the detected line as it is, whereas the LO and Steger’s method
have the tendency to produce detected lines with similar widths. In addition, for the two
samples, both LO and Steger’s method respond many week linear structures which are not
regarded as cracks by expert and thereby disappear in the ground truth. The unexpected
response of the two line detection methods leads to the high values of false positive rates for
the two samples. The reason that our WLD has no output of weak lines is that this method
detects lines based on the comparison of brightness in a neighborhood (referring to the
equation (6-4)) and accordingly weak lines will give a weak response which tends to be
discarded when thresholding (referring to the equation (6-5)).
127
Table 6.6 Comparisons of performance evaluation between different methods used in
Fig. 6.8a.
Method TPR FPR Ac PM Steger 0.6712 0.0258 0.9553 0.4833
LO 0.6310 0.0292 0.9496 0.4383 WLD 0.7787 0.0085 0.9782 0.6902
Table 6.7 Comparisons of performance evaluation between different methods used in
Fig. 6.8b.
Method TPR FPR Ac PM
Steger 0.5089 0.0299 0.9405 0.3544 LO 0.6355 0.0445 0.9350 0.3853
WLD 0.7505 0.0207 0.9646 0.5761
Notice that the detected cracks by using our WLD (as shown in Fig. 6.8f) seems to be
wider than the labeled cracks in the corresponding ground truth (as shown in Fig. 6.8d)
which results in the highest false positive rate among all extraction results obtained using our
WLD. This is because the over estimated max-width (21 pixels vs. 18 pixels) of tongue
cracks for the sample shown in Fig. 6.8b and accordingly a circular mask about 30% larger
than required is used to detect tongue cracks. This illustrates that the use of the proper size of
a circular mask, i.e. the proper scale of a filter, is very important in our WLD and thus the
estimation algorithm of max-width of cracks, which determines the scale of a filter, plays a
significant role for tongue crack extraction.
Table 6.8 shows the average values and variance of performance evaluation results of
different methods for all cracked tongue samples used in this manuscript. Table 6.9 reports
the paired t-test results of the hypothesis that two matched samples come from distributions
with equal means. It can be seen that performance evaluation results of the wide line detector
are significantly different from the results of both the line operator and Steger’s method.
128
Table 6.8 Comparisons of average performance evaluation between different methods.
Method TPR FPR Ac PM
Steger 1297.00.7131± 0089.00.0217 ± 0195.00.9671± 0718.00.3919 ±
LO 0981.00.7372 ± 0152.00.0238 ± 0239.00.9664 ± 0431.00.4176 ±
WLD 0515.08154.0 ± 0077.00.0082 ± 0136.00.9850 ± 0623.00.6722 ±
Table 6.9 Paired t-test results of different methods. The p values are calculated using the
paired t-test of the hypothesis that two matched samples in the two vectors come from
distributions with equal means. p<0.05 was considered statistically significant.
Method TPR FPR Ac PM
WLD-Steger 0.0555 0.0040 0.0037 0.0020
WLD-LO 0.0261 0.0119 0.0185 0.0004
LO-Steger 0.4512 0.5710 0.7616 0.3190
0
10
20
30
40
50
60
70
80
90
Steger LO WLD
Per
form
ance
Mea
sure
(%
)
Q1
min
median
max
Q3
Fig. 6.9 Box-and-whisker plots of the PM values of three methods for all test cracked
tongue images.
129
We have also conducted a test of proposed scheme on a total of 286 cracked tongue
images. Figure 6.9 shows comparative statistical box-and-whisker plots of PM values for all
test images. The plots reveal a much better performance of the WLD algorithm. Hence, our
WLD are proper for tongue crack extraction.
6.5 Conclusion and discussion
We for the first time attempted to detect cracks of tongue images. An automatic tongue crack
detection scheme using the wide line detector is described in this chapter. An adaptive
algorithm of line width estimation is designed to implement the wide line detector totally
automatically. The error generated by the maximum line width estimation algorithm is small
enough. The tongue crack extraction method works well for all test tongue images. Because
no derivative is directly used in the wide line detector, the tongue crack extraction method
gives strong noise rejection. The detection results obtained using our scheme keep line width
information well due to no Gaussian smooth kernel applied. Compared to the classic line
detection methods, our WLD has been successful in improving the performance by
increasing PM 71.5% compared to Steger’s method and 61.0% to the line operator,
respectively. Therefore, the wide line detector is definitely proper for tongue crack
extraction.
130
Chapter 7
Conclusion
We have presented a general method for wide line detection. In contrast with the previous
work which detects the wide line as parallel opposing edges, the proposed wide line detector
extracts the line entirely. The line detection method works very well for a range of images
containing different widths of linear and curvilinear structures, especially for those where the
width of lines varies greatly. This method also works well when the lines run close together
or cross each other due to no line direction required to estimate.
We have applied this proposed general method to solve two practical line detection
problems: palm-line feature based palmprint recognition and adaptive tongue crack detection.
Both have obtained promising results. The main contributions of this thesis are listed below.
7.1 Main contributions
A novel wide line detector using an isotropic nonlinear filter: In contrast with the
traditional directional-derivative-based edge and line extraction method, our wide
line detector is based on isotropic nonlinear filtering without any derivatives and
consequently is more robust to noise. Since no Gaussian kernel is required for our
wide line detector, even narrow lines can be extracted well, which indicates the
multi-scale inherence of our wide line detector.
Dynamic selection of parameters for the wide line detector: We introduced a scheme
to investigate the relationship between the size of circular masks and the width of
detected lines and defined the calculation of brightness contrast threshold so that the
wide line detector can automatically determine the two parameters for the input
image.
131
Analysis of the proposed wide line detector: We presented a scheme for selecting the
optimal core function for the wide line detector by evaluating the performance of
various formulae. We clarified the restriction on the radius of circular masks and the
calculation of the line orientation. We also conducted a comparative study to
demonstrate that the wide line detector is more robust against noise than the
traditional line detection methods.
A novel palm-line feature extraction method for palmprint recognition: As compared
to previous work on palm line extraction [60,61], our wide line detector-based
method extracts not only structure features but also strength features of palm lines for
personal identification. We described a translation-invariant approach for palm-line
feature matching. We also developed an experimental scheme to find out the optimal
combination of parameters for the proposed palm-line feature extraction method. The
experimental results demonstrate that the performance of the proposed WLD-based
method is comparable with the state-of-the-art algorithms of palmprint recognition
and thereby palm-line features can be used to recognize palmprints.
An automatic tongue crack extraction scheme: We for the first time attempted to
extract cracks from tongue images. We proposed a framework for automatic tongue
crack extraction. We derived a tongue crack detection scheme based on our wide line
detector. We also designed an adaptive algorithm for line width estimation to
determine the maximum width of tongue cracks in an input tongue image
automatically; the adaptive selection of maximum width of tongue cracks makes the
tongue crack extraction totally automatically. Compared to the typical line detection
methods, the automatic tongue crack extraction scheme improves the performance by
increasing PM 71.5% compared to Steger’s method and 61.0% to the line operator,
respectively. The experimental results show that the wide line detector is definitely
proper for tongue crack extraction.
132
7.2 Future Work
We now discuss a few directions along which we plan to continue our work.
Line width extraction: The proposed wide line detector can be extended to estimate
line width. We plan to design an algorithm to extract the thickness at each point and
then to validate that this thickness is reasonable at each point.
3-D line detection: The principle introduced in Chapter 3 can be used to detect lines
in 3-D. We would like to design a 3-D line detection algorithm to extract wide lines
and estimate the width and volume of lines.
Proper size of palm-line feature image for palm-line feature based palmprint
recognition: At the moment, the size of the palm-line feature image for palm-line
feature matching is fixed to 64x64. In Coding methods, the feature images for
matching is 32x32. We plan to investigate which size of palm-line feature image is
proper for the WLD-based palm-line feature extraction method.
Palm-line model: Although palmprint recognition has achieved a high performance, it
is needed to build the model of palm lines because it can reveal the nature of
palmprints and accordingly improve the performance further. We would like to study
the characteristics of palm lines further and build a model for palm lines.
Tongue crack description: The destination of tongue crack detection is for diagnosis.
According to the knowledge in TCM, we plan to describe tongue cracks in length,
width, depth, location and orientation to facilitate the tongue diagnosis.
Heuristic algorithm for tongue crack detection: We plan to develop a heuristic
algorithm to improve the performance of tongue crack detection. Such heuristic
algorithm will be based on the wide line detector and employ prior knowledge of
tongue diagnosis.
133
Retina vessel extraction: We would like to apply the proposed wide line detector and
the line width extraction method for retinal vessel extraction, which requires the
precise estimation of vessel width at each vessel point.
134
Bibliography
[1] D. Burr and M.C. Morrone, “A nonlinear model of feature detection,” in
Nonlinear Vision, Determination of Receptive Field, Fuction and Networks,
Boca Raton, FL: CRC, 1992.
[2] F.A. Pellegrino, W. Vanzella, and V. Torre, “Edge detection revisited,” IEEE
Trans. SMC-B, vol. 34, no. 3, pp. 1500-1518, 2004.
[3] C. Steger, “An unbiased detector of curvilinear structures,” IEEE Trans. Pattern
Anal. Machine Intell., vol. 20, no. 2, pp. 113-125, 1998.
[4] R. Zwiggelaar, S.M. Astley, C.R.M. Boggis, and C.J. Taylor, “Linear structures
in mammographic images: detection and classification,” IEEE Trans. Medical
Imaging, vol. 23, no. 9, pp. 1077-1086, 2004.
[5] C. Coppini, M. Demi, R. Poli, and G. Valli, “An artificial vision system for
X-ray images of human coronary trees,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 15, no. 2, pp. 156-162, 1993.
[6] A. Hoover, V. Kouznetsova, and M. Goldbaum, “Locating blood vessels in
retinal images by piecewise threshold probing of a matched filter response,”
IEEE Trans. Med. Imag., vol. 19, no. 3, pp. 203-210, 2000.
[7] J.B.A. Maintz, P.A. van den Elsen, and M.A. Viergever, “Evaluation of ridge
seeking operators for multimodality medical image matching,” IEEE Trans.
Pattern Anal. Machine Intell., vol. 18, no. 4, pp. 353-365, 1996.
[8] B. Pang and D. Zhang, “Computerized tongue diagnosis based on Bayesian
networks,” IEEE Trans. on Biomedical Engineering, vol. 51, no. 10, pp.
1803-1810, 2004.
[9] S. Mori and T. Sakakura, “Line filtering and its application to stroke
segmentation of handprinted Chinese characters,” Proc. Int’l Conf. Pattern
Recognition, pp. 366-369, 1984.
[10] L.H. Chen, J.Y. Wang, H.Y. Liao, and K.C. Fan, “A robust algorithm for
separation of Chinese characters from line drawings,” Image and Vision
135
Computing, vol. 14, no. 10, pp. 753-761, 1996.
[11] C.J. Taylor and D.J. Kriegman, “Structure and motion from line segments in
multiple images,” IEEE Trans. Pattern Anal. Machine Intell., vol. 17, no. 11, pp.
1021-1032, 1995.
[12] Z.Y. Zhang, “Estimating motion and structure from correspondences of line
segments between two perspective images,” IEEE Trans. Pattern Anal.
Machine Intell., vol. 17, no. 12, pp. 1129-1139, 1995.
[13] L. Zhang and D. Zhang, “Characterization of palmprints by wavelet signatures
via directional context modeling,” IEEE Trans. SMC-B, vol. 34, no. 3, pp.
1335-1347, June 2004.
[14] P. Hough, A method and means for recognizing complex patterns, U.S. Patent
3069654, 1962.
[15] J. Illingworth and J. Kittler, “A Survey of the Hough Transform,” Computer
Vision, Graphics, and Image Processing, vol. 44, pp. 87–116, 1988.
[16] M. Sonka, V. Hlavac, and R. Boyle, Image Processing, Analysis, and
Machine Vision, PWS Publishing, 1999.
[17] D.H. Ballard, “Generalizing the Hough transform to detect arbitrary shapes,”
Pattern Recognition, vol. 13, no. 2, pp. 111–122, 1981.
[18] L. S. Davis, “Hierarchical generalized Hough transforms and line segment
based generalized Hough transforms,” Pattern Recognition, vol. 15, no. 4, pp.
277-285, 1982
[19] J. Illingworth and J. Kittler, “The adaptive Hough transform,” IEEE Trans.
Pattern Anal. Machine Intell., vol. 9, no. 5, pp. 690-698, 1987.
[20] D. Geman and B. Jedynak, “An active testing model for tracking roads in
satellite images,” IEEE Trans. Pattern Anal. Machine Intell., vol. 18, no. 1, pp.
1-14, 1996.
[21] N. Merlet and J. Zerubia, “New prospects in line detection by dynamic
programming,” IEEE Trans. Pattern Anal. Machine Intell., vol. 18, no. 4., pp.
426-431, 1996.
136
[22] J.H. Jang and K.S. Hong, “Detection of curvilinear structures using the
Euclidean distance transform,” Proc. IAPR Workshop on Machine Vision
Applications, pp. 102-105, 1998.
[23] A.W.K. Loh, M.C. Robey, and G.A.W. West, “Analysis of the interaction
between edge and line finding techniques,” Pattern Recognition, vol. 34, pp.
1127-1146, 2001.
[24] D.S. Guru, B.H. Shekar, and P. Nagabhushan, “A simple and robust line
detection algorithm based on small eigenvalue analysis,” Pattern Recognition
Letters, vol. 25, pp. 1-13, 2004.
[25] W.P. Kegelmeyer, J. Pruneda, P. Bourland, A. Hills, M.Riggs, and M. Nipper,
“Computer-aided mammographic screening for speculated lesions,” Radiology,
vol. 191, pp. 331-337, 1994.
[26] T.M. Koller, G. Gerig, G. Szekely, and D. Dettwiler, “Multiscale detection of
curvilinear structures in 2-D and 3-D image data,” the 5th Int’l Conf. Computer
Vision, pp. 864-869, Boston, USA, 1995.
[27] J. Canny, “A computational approach to edge detection,” IEEE Trans. Pattern
Anal. Machine Intell., vol. 8, no. 6, pp. 679-698, 1986.
[28] D. Eberly, R. Gardner, B. Morse, S. Pizer, and D. Scharlah, “Ridges for image
analysis,” J. Math. Imaging and Vision, vol. 4, pp. 353-373, 1994.
[29] A. Busch, “A common framework for the extraction of lines and edges,” Int’l
Archives of Photogrammetry and Remote Sensing, vol. XXXI, part B3, pp.
88-91, 1996.
[30] T. Lindeberg, “Edge detection and ridge detection with automatic scale
selection,” Int’l J. Computer Vision, vol. 30, no. 2, pp. 117-154, 1998.
[31] R.M. Haralick, “Ridges and valleys in digital images,” Comput. Vision,
Graphics, Image Processing, vol. 22, pp. 28-38, 1983.
[32] J.B. Subirana-Vilanova and K.K. Sung, “Ridge-detection for the perceptual
organization without edges,” the 4th Int’l Conf. Computer Vision, pp. 57-64,
Berlin, Germany, 1993.
[33] L.A. Iverson and S.W. Zucker, “Logical/linear operators for image curves,”
137
IEEE. Trans. Pattern Analysis and Machine Intelligence, vol. 17, no. 10, pp.
982-996, 1995.
[34] A. Busch, “Revision of built-up areas in a GIS using satellite imagery and GIS
data,” Proceedings of IAPRS, vol. 32, part 4, pp. 91-98, 1998.
[35] M. Jacob and M. Unser, “Design of steerable filters for feature detection using
canny-like criteria,” IEEE Trans. Pattern Anal. Machine Intell., vol. 26, no. 8,
2004.
[36] J. Lowell, A. Hunter, D. Steel, A. Basu, R. Ryder, and R.L. Kennedy,
“Measurement of retinal vessel widths from fundus images based on 2-D
modeling,” IEEE Trans. Medical Imaging, vol. 23, no. 10, 2004.
[37] M.C.K. Yang, J.S. Lee, C.C. Lien, and C.L. Huang, “Hough transform
modified by line connectivity and line thickness,” IEEE Trans. Pattern Anal.
Machine Intell., vol. 19, no. 8, pp. 905-910, 1997.
[38] R.C. Lo and W.H. Tsai, “Gray-scale Hough transform for thick line detection in
gray-scale images,” Pattern Recognition, vol. 28, no. 5, pp. 647-661, 1995.
[39] J. Chen, Y. Sato, and S. Tamura, “Orientation space filtering for multiple
orientation line segmentation,” IEEE Trans. Pattern Anal. Machine Intell., vol.
22, no. 5, pp. 417-429, 2000.
[40] E.R. Davies, M. Bateman, D.R. Mason, J. Chambers and C. Ridgway, “Design
of efficient line segment detectors for cereal grain inspection,” Pattern
Recognition Letters, vol. 24, pp. 413-428, 2003.
[41] B.K. Jeon, J.H. Jang, and K.S. Hong, “Road detection in spaceborne SAR
images using a genetic algorithm,” IEEE Trans. Geosecience and Remote
Sensing, vol. 40, no. 1, pp. 22-29, 2002.
[42] A.M. Lopez, F. Lumbreras, J. Serrat, and J. Villanueva, “Evaluation of methods
for ridge and valley detection,” IEEE Trans. Pattern Anal. Machine Intell., vol.
21, no. 4, pp. 327-335, 1999.
[43] M. Sofka, and C.V. Stewart, “Retinal vessel centerline extraction using
multiscale matched filters, confidence and edge measures,” IEEE Trans.
Medical Imaging, vol. 25, no. 12, pp. 1531-1546, 2006.
138
[44] P.Y. Lau and S. Ozawa, “A simple method for detecting tumor in T2-weighted
MRI brain images: An image-based analysis,” IEICE Trans. Information and
Systems, vol. E89-D, no. 3, pp. 1270-1279, 2006.
[45] R.N. Dixon and C.J. Taylor, “Automated asbestos fiber counting,” Inst. Phys.
Conf. Ser., vol. 44, pp. 178-185, 1979.
[46] A. Jain, L. Hong, and S. Pankanti, “Biometric identification”, Communications
of the ACM, vol. 43, no. 2, pp. 90-98, 2000.
[47] D. Zhang, Automated Biometrics-Technologies and Systems. Kluwer Academic,
Dordrecht, 2000.
[48] D. Zhang, W. Shu, “Two novel characteristics in palmprint verification: datum
point invariance and line feature matching,” Pattern Recognition, 32, pp.
691-702, 1999.
[49] S.M. Smith and J.M. Brady, “SUSAN – A new approach to low level image
processing,” Int’l. J. Computer Vision, vol. 23, no. 1, pp. 45-78, 1997.
[50] D. Zhang, W.K. Kong, J. You, M. Wong, “Online Palmprint Identification,”
IEEE Trans. PAMI, vol. 25, pp. 1041-1050, 2003.
[51] N. Duta, A.K. Jain, K.V. Mardia, “Matching of palmprints,” Pattern
Recognition Letters, 23, pp. 477-485, 2002.
[52] W. Shu and D. Zhang, “Automated Personal Identification by Palmprint,”
Optical Engineering, vol. 37, no. 8, pp. 2659-2362, 1998.
[53] W. Shu, G. Rong, Z. Bian, and D. Zhang, “Automatic Palmprint Verification”,
International Journal of Image and Graphics, vol. 1, no. 1, pp. 135-152, 2001.
[54] P.S. Wu and M. Li, “Pyramid Edge Detection Based on Stack Filter,” Pattern
Recognition Letter, vol. 18, no. 4, pp. 239-248, 1997.
[55] J. You, W.K. Kong, D. Zhang, K.H. Cheung, “On hierarchical palmprint coding
with multiple features for personal identification in large databases,” IEEE
Trans. Circuits and Systems for Video Technology, vol. 14, pp. 234-243, 2004.
[56] A.K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric
recognition,” IEEE Trans. Circuits and Systems for Video Technology, vol. 14,
139
no. 1, pp. 4-20, 2004.
[57] S. Nanavati, M. Thieme, and R. Nanavati, Biometrics: Identity Verification in a
Networked World. John Wiley & Sons, New York. 2002.
[58] J. You, W. Li, D. Zhang, “Hierarchical palmprint identification via multiple
feature extraction,” Pattern Recognition, 35, pp. 847-859, 2002.
[59] C.C. Han, H.L. Cheng, K.C. Fan, C.L. Lin, “Personal authentication using
palmprint features,” Pattern Recognition, 36, pp. 371-381, 2003.
[60] X.Q. Wu, K.Q. Wang, D. Zhang, “A novel approach of palmline extraction,”
ICIG, pp. 230-233, 2004.
[61] X.Q. Wu, K.Q. Wang, D. Zhang, “Palm-line extraction and matching for
personal authentication,” IEEE Trans. SMC-A, vol. 36, pp. 978-987, 2006.
[62] L. Liu, D. Zhang, and J. You, “Detecting wide lines using isotropic nonlinear
filter,” IEEE Trans. Image Processing, vol. 16, no. 6, pp. 1584-1595, 2007.
[63] X.Q. Wu, K.Q. Wang, and D. Zhang, “Fuzzy directional element energy feature
(FDEEF) based palmprint identification,” Proc. of ICPR’02, pp. 95-98, 2002.
[64] W.X. Li, D. Zhang, and Z. Xu, “Palmprint identification by Fourier transform,”
Intel. J. of Pattern Recognition and Artificial Intelligence, vol. 16, no. 4, pp.
417-432, 2002.
[65] P. Hennings, M. Savvides, B.V.K. Vijayakumar, “Verification of biometric
palmprint patterns using optimal trade-off filter classifiers,” ICIAR, pp.
1081-1088, 2005.
[66] G.M. Lu, D. Zhang, and K.Q. Wang, “Palmprint recognition using eigenpalm
features,” Pattern Recognition Letters, vol. 24, no. 9-10, pp. 1463-1467, 2003.
[67] X.Q. Wu, K.Q. Wang, and D. Zhang, “Fisherpalms based palmprint
recognition,” Pattern Recognition Letters, vol. 24, no. 15, pp. 2829-2838, 2003.
[68] D.W. Hu, G.Y. Feng, and Z.T. Zhou, “Two-dimensional locality preserving
projections (2DLPP) with its application to palmprint recognition,” Pattern
Recognition, vol. 40, no. 1, pp. 339-342, 2007.
140
[69] W.K. Kong and D. Zhang, “Feature-level fusion for effective palmprint
authentication,” Proc. of the 1st ICBA, LNCS 3072, pp. 761-764, 2004.
[70] W.K. Kong and D. Zhang, “Competitive coding scheme for palmprint
verification,” Proc. of the 17th ICPR, vol. 1, pp. 520-523, 2004.
[71] Z.N. Sun, T.N. Tan, Y.H. Wang, and Stan Z. Li, “Ordinal palmprint
representation for personal identification,” Proc. of CVPR’05, pp. 279-284,
2005.
[72] J.G. Daugman, “High confidence visual recognition of persons by a test of
statistical independence,” IEEE Trans. Pattern Anal. Machine Intell., vol. 15,
no. 11, pp. 1148-1161, 1993.
[73] T.S. Lee, “Image representation using 2D Gabor wavelet,” IEEE Trans. Pattern
Anal. Machine Intell., vol. 18, no. 10, pp. 959-971, 1996.
[74] K.R. Castleman, Digital Image Processing, Englewood Cliffs, NJ: Prentice-Hall,
1996.
[75] G. Maciocia, Tongue Diagnosis in Chinese Medicine (Revised Edition). Seattle,
WA: Eastland, 1995.
[76] B. Kirschbaum, Atlas of Chinese Tongue Diagnosis. Seattle, WA: Eastland,
2000.
[77] C.C. Chiu, H.S. Lin and S.L. Lin, “A structural texture recognition approach for
medical diagnosis through tongue,” Biomedical Engineering, Application, Basis
and Communication, vol. 7, no. 2, pp. 143-148, 1995.
[78] C.C. Chiu, “A novel approach based on computerized image analysis for
traditional Chinese medical diagnosis of the tongue,” Computer Methods and
Programs in Biomedicine, vol. 61, no. 2, pp. 77-89, Feb. 2000.
[79] P C Yuen, Z Y Kuang, W Wu and Y T Wu, "Tongue Texture Analysis using
Opponent Colour Features for Tongue Diagnosis in Traditional Chinese
Medicine", Texture Analysis in Machine Vision, (Ed. M. Pietikäinen). Series in
Machine Perception and Artificial Intelligence - vol. 40, pp. 179-188, World
Scientific, November 2000.
[80] C.H. Li and P.C. Yuen, “Regularized color clustering in medical image
141
database,” IEEE Trans. Medical imaging, vol. 19, no. 11, pp. 1150-1155, 2000.
[81] C H Li and P C Yuen, "Tongue image matching using color content", Pattern
Recognition, vol. 35, No. 2, pp. 407-419, 2002.
[82] C.Yang, “A novel imaging system for tongue inspection,” 19th IEEE Proc.
Instrumentation and Measurement Technology Conference, vol. 1, pp. 159-163,
2002.
[83] B. Pang, D. Zhang, and K.Wang, “The bi-elliptical deformable contour and its
application to automated tongue segmentation in Chinese medicine,” IEEE
Trans. Medical Imaging, vol. 24, no. 8, pp. 946-956, Aug. 2005.
[84] B. Pang, D. Zhang and K. Wang, “Tongue image analysis for appendicitis
diagnosis,” Information Sciences, vol. 175, no. 3, pp. 160-176, Oct. 2005.
[85] H.Z. Zhang, K.Q. Wang, X.S. Jin, and D. Zhang, “SVR based color calibration
for tongue image,” Proc. Intl. Conf. Machine Learning and Cybernetics, pp.
5065-5070, Aug. 2005.
[86] H.Z. Zhang, K.Q. Wang, D. Zhang, B. Pang, and B. Huang, “Computer aided
tongue diagnosis system,” 27th Intl. Conf. IEEE-EMBS, pp. 6754-6757, Sept.
2005.
[87] D. Behar, J. Cheung, and L. Kurz, “Contrast techniques for line detection in a
correlated noise environment,” IEEE Trans. Image Processing, vol. 6, no. 5, pp.
625-641, 1997.
[88] G.J. Genello, J.F.Y. Cheung, S.H. Billis, and Y. Saito, “Graeco-latin squares
design for line detection in the presence of correlated noise,” IEEE Trans.
Image Processing, vol. 9, no. 4, pp. 609-622, 2000.
[89] G.M. Schuster and A.K. Katsaggelos, “Robust line detection using a weighted
MSE estimator,” Proc. Int’l Conf. Image Processing (ICIP’03), vol. 1, pp.
293-296, 2003.
[90] H. Li, W. Hsu, M.L. Lee and H. Wang, “A piecewise Gaussian model for
profiling and differentiating retinal vessels,” Proc. Int’l Conf. Image Processing
(ICIP’03), vol. 1, pp. 1069-1072, 2003.
[91] Z.Y. Liu, K.C. Chiu and L. Xu, “Strip line detection and thinning by
142
RPCL-based local PCA,” Pattern Recognition Letters, vol. 24, no. 14, pp.
2335-2344, 2003.
[92] R.O. Duda and P.E. Hart, “Use of the Hough Transformation to Detect Lines
and Curves in Pictures,” Comm. ACM, vol. 15, no. 1, pp. 11-15, 1972.
[93] V.F. Leavers, “Which Hough Transform?” CVGIP: Image Understanding, vol.
58, no. 2, pp. 250–264, 1993.
[94] J.H. Jang and K.S. Hong, “Linear band detection based on the Euclidean
distance transform and a new line segment extraction method,” Pattern
Recognition, vol. 34, pp. 1751-1764, 2001.
[95] L. Liu and D. Zhang, “A Novel Palm-Line Detector,” Audio- and Video-Based
Biometric Person Authentication: 5th International Conference, AVBPA 2005,
Hilton Rye Town, NY, USA, Lecture Notes in Computer Science, vol. 3546,
Springer, pp. 563-571, July 20-22, 2005.
[96] D.M. Green and J.M. Swets, Signal detection theory and psychophysics. New
York: John Wiley and Sons Inc., 1966.
[97] http://en.wikipedia.org/wiki/Image:ROC_space.png
[98] M.H. Zweig and G. Campbell, “Receiver-operating characteristic (ROC) plots:
a fundamental evaluation tool in clinical medicine,” Clinical chemistry, vol. 39,
no. 8, pp. 561-577, 1993.
[99] M.S. Pepe, The statistical evaluation of medical tests for classification and
prediction. New York: Oxford, 2003.
[100] X. He, C.E. Metz, J.M. Links, B.M. Tsui, and E.C. Frey, “Three-class ROC
analysis – A decision theoretic approach under the ideal observer framework,”
IEEE Trans. Medical Imaging , vol. 25, no. 5, pp. 571-581, 2006.
[101] J.A. Hanley and B.J. McNeil, “The meaning and use of the area under a
receiver operating characteristic (ROC) curve,” Radiology, vol. 143, pp. 29-36,
1982.
[102] N.A. Obuchowski, “Receiver operating characteristic curves and their use in
radiology,” Radiology, vol. 229, no.1, pp. 3-8, 2003.
143
[103] D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, and A. Jain, “FVC2000:
Fingerprint verification competition,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 24, no. 3, pp. 402-412, 2002.
[104] T. Kanungo, M.Y. Jaisimha, J. Palmer, and R.M. Haralick, “A methodology for
quantitative performance evaluation of detection algorithms,” IEEE Trans. Imag.
Processing, vol. 4, no. 12, pp. 1667-1674, 1995.
[105] K.A. Spackman, “Signal detection theory: Valuable tools for evaluating
inductive learning,” Proceedings of the Sixth International Workshop on
Machine Learning, San Mateo, CA: Morgan Kaufman, pp. 160–163, 1989.
[106] J.C. Russ, The Image Processing Handbook, 3rd ed., CRC Press LLC, 1999.
[107] PolyU Palmprint Database, http://www.comp.polyu.edu.hk/~biometrics/.
[108] A. Jain, R. Bolle, S. Pankanti (ed.), Biometrics: Personal Identification in
networked Society, Kluwer Academic, Dordrecht, 1999.
[109] A.K. Jain, L. Hong, and R. Bolle, “On-line fingerprint verification,” IEEE
Trans. Pattern Anal. Machine Intell., vol. 19, no. 4, pp. 302-314, 1997.
[110] A. Jain, S. Pankanti, and L. Hong, “A multichannel approach to fingerprint
classification,” IEEE Trans. Pattern Anal. Machine Intell., vol. 21, no. 4, pp.
348-359, 1999.
[111] C.J. Liu, “Gabor-based kernel PCA with fractional power polynomial models
for face recognition,” IEEE Trans. Pattern Anal. Machine Intell., vol. 26, no. 5,
pp. 572-581, 2004.
[112] X.G. Wang, X.O. Tang, “A unified framework for subspace face recognition,”
IEEE Trans. Pattern Anal. Machine Intell., vol. 26, no. 9, pp. 1222-1228, 2004.
[113] X.F. He, S.C. Yan, Y.X. Hu, P. Niyogi, and H.J. Zhang, “Face recognition using
Laplacianfaces,” IEEE Trans. Pattern Anal. Machine Intell., vol. 27, no. 3, pp.
328-340, 2005.
[114] R. Sanchez-Reillo, C. Sanchez-Avila, and A. Gonzalez-Marcos, “Biometric
identification through hand geometry measurements,” IEEE Trans. Pattern Anal.
Machine Intell., vol. 22, no. 10, pp. 1168-1171, 2000.
144
[115] J. Campbell, Jr., “Speaker recognition: A tutorial,” Proc. IEEE, vol. 85, no. 9,
pp. 1437-1462, 1997.
[116] K. Chen, “Towards better making a decision in speaker verification,” Pattern
Recognition, vol. 36, no. 2, pp. 329-346, 2003.
[117] L.L. Lee, T. Berger, and E. Aviczer, “Reliable online human signature
verification systems,” IEEE Trans. Pattern Anal. Machine Intell., vol. 18, no. 6,
pp. 643-647, 1996.
[118] L. Wang, T.N. Tan, H.Z. Ning, and W.M. Hu, “Silhouette analysis-based gait
recognition for human identification,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 25, no. 12, pp. 1505-1518, 2003.
[119] Z.Y. Liu, S. Sarkar, “Improved gait recognition by gait dynamics
normalization,” IEEE Trans. Pattern Anal. Machine Intell., vol. 28, no. 6, pp.
863-876, 2006.
[120] J.G. Daugman, “The importance of being random: statistical principles of iris
recognition,” Pattern Recognition, vol. 36, no. 2, pp. 279-291, 2003.
[121] J. Chen, C. Zhang, and G. Rong, “Palmprint recognition using crease,” in Proc.
Int. Conf. Image Processing, pp. 234-237, Oct. 2001.
[122] H. Yiu, Fundamental of Traditional Chinese Medicine, Foreign Language Press,
Beijing, 1992.
[123] Y. Xin, X.Z. Guo, L.S. Zhang, et al, “Tongue Diagnosis (Chinese-English),” the
Tian Jin Science and Technology Translation Publisher, June 2001.
[124] G. Li and Y. Cai, “Texture analysis for tongue analysis,” Technical Report
BV-2003-2, School of Computer Science, Carnegie Mellon University, May
2003.
[125] B. Pham and Y. Cai, “Visualization techniques for tongue analysis in traditional
Chinese medicine,” Proc. SPIE, vol. 5367, pp. 171-180, 2004;
[126] J.R. Parker, Algorithms for Image Processing and Computer Vision. John Wiley
& Sons, 1997.
[127] K. Bowyer, C. Kranenburg, and A. Dougherty, “Edge detector evaluation using
145
empirical ROC curves,” Comput. Vis. Image Understand., vol. 84, no. 1, pp.
77-103, 2001.
[128] O.D. Trier and A.K. Jain, “Goal-directed evaluation of binarization methods,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 17, no. 12, pp. 1191-1201, 1995.
[129] O.D. Trier and T. Taxt, “Evaluation of binarization methods for document
images,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 17, no. 3, pp. 312-315,
1995.
[130] M. Kamel and A. Zhao, “Extraction of binary character/graphics images from
grayscale document images,” CVGIP: Graphical Models Image Process. vol.
55, no. 3, pp. 203-217, 1993.
[131] Y.B. Yang, H. Yan, “An adaptive logical method for binarization of degraded
document images,” Pattern Recognition, vol. 33, pp. 787-807, 2000.
[132] R.R. Rakesh, P. Chaudhuri, C.A. Murthy, “Thresholding in edge detection: A
statistical approach,” IEEE Trans. Image Processing, vol. 13, no. 7, pp. 927-936,
2004.
[133] M. Basu, “Gaussian-based edge-detection methods – A survey,” IEEE Trans.
Systems, Man, and Cybernetics, Part C, vol. 32, no. 3, pp. 252-260, 2002.
[134] F. van der Heijden, “Edge and line feature extraction based on covariance
models,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 17, no. 1, pp. 16-33,
1995.
[135] D. Ziou, S. Tabbone, “Edge detection techniques – An overview,” Int’l J.
Pattern Recognition & Image Analysis, vol. 12, pp. 1-41, 1998.
[136] T. Asano, N. Katoh, “Number theory helps line detection in digital images,”
Proc. 4th Int’l Symposium on Algorithms and Computation, ISAAC’93, pp.
313-322, 1993.