new approach to edge detection on different
TRANSCRIPT
Computing and Informatics, Vol. 38, 2019, 1067–1090, doi: 10.31577/cai 2019 5 1067
NEW APPROACH TO EDGE DETECTIONON DIFFERENT LEVEL OF WAVELETDECOMPOSITION
Vladimir Maksimovic, Branimir Jaksic, Mile PetrovicPetar Spalevic, Stefan Panic
Faculty of Technical ScienceUniversity of PristinaKneza Milosa 738220 Kosovska Mitrovica, Serbiae-mail: {vladimir.maksimovic, branimir.jaksic, mile.petrovic,
petar.spalevic}@pr.ac.rs, [email protected]
Abstract. This paper proposes a new approach to edge detection on the imagesover which the wavelet decomposition was done to the third level and consistingof different levels of detail (small, medium and high level of detail). Images fromthe BSD (Berkeley Segmentation Dataset) database with the corresponding groundtruth were used. Daubechies wavelet was used from second to tenth order. Gradientand Laplacian operators were used for edge detection. The proposed approach isapplied in systems where information is processed in real time, where fast imageprocessing is required and in systems where high compression ratio is used. Thatis, it can find practical application in many systems, especially in television systemswhere the level of details in the image changes. The new approach consists in thefact that when wavelet transform is applied, an edge detection is performed overthe level 1 image to create a filter. The filter will record only those pixels that canbe potential edges. The image is passed through a median filter that filters onlythe recorded pixels and 8 neighbors of pixel. After that, the edge detection withone of the operators is applied onto the filtered image. F measure, FoM (Figure ofMerit) and PR (Performance Ratio) were used as an objective measure. Based onthe obtained results, the application of the proposed approach achieves significantimprovements and these improvements are very good depending on the number ofdetails in the image and the compression ratio. These results and improvements canbe used to improve the quality of edge detection in many systems where compressedimages are processed, that is, where work with images with a high compression ratiois required.
1068 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. Panic
Keywords: Edge detection, wavelet decomposition, compression, F measure, figureof merit, performance ratio
1 INTRODUCTION
In recent years, multimedia systems have recorded tremendous growth and progress,which means that image resolutions got higher increasing the complexity of the sys-tem and the complexity of the image. It is necessary to achieve as much compressionas possible so these images can be streamed, stored and processed more easily. To-day’s trends require more technology to be involved, so in television systems wehave a virtual reality (VR), augmented reality (AR) and a combination of them.Since it is necessary to have lower bitrate and achieve a high compression ratio andto process an image in VR or AR environment, it is necessary to perform certainoperations over images such as segmentation, edge detection, etc. [1].
Edge detection is one of the fundamental processes in image processing. Thisalso means the segmentation of the image where the segmentation of the desiredobject is performed by detecting the edge. Edge detection is based on the fact thatthere are sudden changes in the gray intensity between the objects. Edge detectionsignificantly reduces the image analysis process by using less data and at the sametime storing all the necessary information. Many edge detection techniques havebeen tried, but gradient and Laplacian methods have proved to be the best [2, 3].Gradient methods for edge detection find the maximum and minimum gradient ofthe image intensity, all in the first derivative of the image. Laplacian methods arebased on finding zero crossing in the second derivative of the image [4]. The gradientof the image can be calculated as [4, 5]:
∇f (x, y) =∂f
∂xi+
∂f
∂yj (1)
where f (x, y) represents the image at the location (x, y) where x and y are thecoordinates of the row and column. The gradient ∇f (x, y) contains the informationabout gray change. The gradient of ∇f (x, y) can be calculated [4, 5]:
e (x, y) =√f 2x + f 2
y (2)
where e (x, y) can be used as an edge detector and can also be defined as the sumof the absolute values of the partial derivative fx and fy [4, 5]:
e (x, y) = |fx (x, y)|+ |fy (x, y)| . (3)
Based on these theoretical principles, the gradient and Laplacian methods areproposed. The gradient methods include Sobel, Prewitt, Robert, while the Laplacianinclude LoG (Laplacian of Gaussian). The classic use of the Canny operator is
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1069
a gradient method, but in some parts, it also contains elements of the Laplaceapproach [6, 7, 8].
The analysis of the frequency domain using Fourier transform is very useful forsignal analysis because the frequency domain is very important for the considerationof the nature of the signal and the influence of the noise over it. The disadvantageof this technique is the loss of time domain information. When viewing the Fouriertransform in a frequency domain, it is difficult to say when a certain event occurred,or when some frequencies occurred. To overcome this problem, only a small part(“window”) of the signal is analyzed at a time, and this window slides over the signalapplying the Fourier transform on it. In this way, part of the frequency informationis lost in order to get an information about the time when a certain frequency hasoccurred. Wavelet transformation is used to solve this problem. It allows the useof different window sizes for different frequencies. The basic difference between thewavelet transformation and the short-time Fourier transform is that the windowlength changes at the wavelet transform and in this way the frequency and timeresolution can be changed. The basic idea of each wavelet transform is to present anarbitrary function x(t) as a superposition of a wavelet set or basic functions. Thebasic functions are derived from a prototype called mother wavelet, by scaling ortranslating this function. The wavelet transformation of the x(t) signal is given inEquation (4) [9, 10]:
ψ (τ, s) =1
s
∫ ∞−∞
x (t) ∗ ψ∗(t− τs
)dt (4)
where τ is a translation, and s is a scaling, while ψ is a “mother” wavelet, and ψ∗ isconjugally complex of ψ. When processing images in a spatial domain, the operationis discretized. Discrete wavelet transformation (DWT) is defined as [9, 10]:
DWTψx (τ, s) =1√|s|
∫ ∞−∞
x (t) ∗ ψ∗(t− τs
)dt. (5)
The discrete wavelet transformation (DWT) can be presented as a matrix ψ (c)and DWT coefficients can be obtained by taking an internal product between thesignal and the wavelet matrix [9, 10]:
DWTψx [n, s] =1√|s|
∑n
x [n]ψ∗n,s [n] . (6)
Special family of wavelet functions has been developed for DWT. They aregenerally divided into orthogonal or biorthogonal and are characterized by a high-pass and low-pass filters [11]. Daubechies wavelet belongs to a family of orthogonalfunctions and the main feature is the possibility of the maximum number of vanishingmoments for a predefined supported length. Types of Daubechies (db) wavelets thatare most commonly used in practical applications are dbN, where N represents theorder as well as the number of vanishing moments in the supported interval from 0
1070 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. Panic
to 20. By increasing the order of the Daubechies wavelet, better characteristics areobtained, but the complexity of the implementation, the price of the system anderrors in the calculations rise. In practical applications, orders 2 to 10 are mostcommonly used [10, 12, 13].
A very important advantage of the wavelet transform is that it uses a mul-tiresolution analysis that allows analysis of different signal frequencies at differentfrequency resolutions. For high-frequency parts, shorter windows are used, whichensures good time domain resolution, while longer parts of the lower frequencies havelonger windows, thus giving good information about the frequencies. If the signalis passed through a set of two filters, low-pass and high-pass, its frequency contentwill be split into two equal-width ranges. The output from these filters contains halfthe frequency of the original signal and the same number of samples as the originalsignal. By decimating, or by passing the input signal through the low-pass filter, thenumber of samples is halved so that the time resolution is also halved while the fre-quency resolution increases. The high-pass filter transmits high-frequency content,i.e. signal details. The low-pass signal transmits low-frequency content, or signal ap-proximation. This approximation signal can be further fed through two filters andthe process can be repeated until the desired decomposition level is reached. Thecomplete information on the original signal is contained in the last approximationsignal and all the detail signals [14, 15].
Higher compression ratio disrupts the quality of the image, resulting in a largeloss of information, and hence makes it difficult to process them, for example, todo an edge detection or facial recognition [16, 17]. The decomposition significantlydegrades the image quality, but besides this degradation there are various types ofnoise that further impair the image quality. To solve this problem or to control it,the various types of filters have been developed [18, 19].
In this paper we used the characteristics of wavelet transform in order to imagecompression, or as a method on which some algorithms for image compression isbased, such as JPEG2000 and SPIHT. Images from the BSD database are com-pressed to the third decomposition level which resulted in a high degree of compres-sion.
2 SYSTEM MODEL AND PROPOSED APPROACH
In this paper, the BSD (Berkeley Segmentation Dataset) image database was usedfor analysis, with its corresponding ground truth images [20]. Table 1 lists theselected images from the BSD database meeting the complexity criteria [16], thatis, each image consists of a different level of detail: small, medium, and high levelof detail. The number of details was calculated by making DCT and DWT on thehigh-frequency components (details), which are divided into four quadrants, alongboth directions (x and y). After that, the mean absolute value of the amplitude ofthe components belonging to the quadrants is calculated [16]: DCT in quadrant 1(dctd); DCT in quadrants 2 and 3 (dctm); DWT in quadrant 1 (dwtd); DWT in
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1071
quadrants 2 and 3 (dwtm).
Images dctd dctm dwtd dwtm
Criterion L #135 069 1.544 2.517 0.181 0.354
Criterion M #35 010 3.838 6.197 1.199 2.048
Criterion H #8 143 7.868 15.241 3.181 6.336
Table 1. Criteria
Figure 1 gives a flow diagram of the proposed approach. In the proposed ap-proach, firstly, the BSD and the ground truth images are read on which the DWTis applied to the third decomposition level using the Daubechies wavelet (optionallyfrom 2 to 20). Since grayscale images are needed, a conversion is made, and variablesare created for filters and new images. In order to assign input values to a filter, it isnecessary to perform edge detection on level 1 images. Values stored in the filter areinformation about location of pixels pointing at the potential edge. The compressedimage on which the detection is performed is a binary image and passing throughthe loop, each pixel is examined. If it is equal to the 1, that is, there is an edge,the median filter is used on the pixel and 8 neighbors of pixel. In other words,only those pixels that are relevant for the edge detection are passed through thefilter. After that, the detection is applied and the results are compared, as shownin Figure 1. The algorithm is applied to all edge detection operators (Canny, LoG,Sobel, Prewitt, Robert).
The algorithm consists of the following steps:
Step 1 (Read image): Read the original image. If it is a color image, converts itto a grayscale image.
Step 2 (DWT): DWT is applied to the third decomposition level onto a readimage, whereby as a result, three images are obtained: image from level 1,image from level 2 and image from level 3. The next steps in the algorithm areapplied separately for an image from each level.
Step 3 (Edge detection): On level 1 image, an edge detection is used by selectingone of detectors (Canny, LoG, Sobel, Prewitt, Roberts). The image with de-tected edges serves to create a filter that will contain only those pixel coordinateswhere the edges are located.
Step 4 (Filtering image): Every pixel in the image is analyzed and if it is equalto 1, the filter records its coordinates.
Step 5 (Filtering 8-neighbors of pixel): If the condition in step 4 is fulfilled,filtering is done by filtering 8 neighbors of pixel relative to the current pixelusing the median filter (Figure 2). So, only those image pixels that can be edgesare filtered. The neighbors of pixel are extracted using the following formula:
1072 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. Panic
If current pixel P (i, j) is edge then use following neighbor pixels:
f (Pi,j) = f (Pi−1,j−1) , f (Pi−1,j) , f (Pi−1,j+1) , f (Pi,j−1) , f (Pi,j) , f (Pi,j+1) ,
f (Pi+1,j−1) , f (Pi+1,j) , f (Pi+1,j+1) (7)
and use median filter only on those pixels.
Step 6 (Edge detection on filtered image): Edge detection is applied on a fil-tered image.
For filtering level 2 and level 3 images, coordinates from level 1 image areused.
The proposed algorithm has a quadratic complexity and can be represented as:
O (row , col) = 8 (row − 2) (col − 2) + 1. (8)
In Figure 3, the complexity of the algorithm is given, depending on the numberof pixels in the row (row) and the number of pixel in columns (cols).
In order to accurately and precisely present how well the edge detection is done,it is necessary to calculate Precision, Specificity, Sensitivity and Accuracy or F mea-sure [21]. F measure (F1 score) is a harmonic mean of precision and recall and itcombines precision and recall according to the formula [21]:
F =2 ∗ Precision ∗ Recall
Precision + Recall× 100. (9)
F is within the limits of 0 ≤ F ≤ 1, ideally, F is equal to 1. In the results, F ismultiplied by 100 and represents a percentage value.
Figure of Merit (FoM) was proposed by Pratt [22] and is a measure for estimatingthe accuracy of detected edges. In other words, it represents the deviation of theactual (calculated) point of the edge from the ideal point of the edge, and is definedas:
FoM =1
max [Id, Ii]
Id∑k=1
1
1 + δe2 (k)(10)
where Id is the number of points on the detected edge, and Ii is the number ofpoints on the ideal edge, e (k) represents the distance between the detected edgeand the ideal edge, and δ is scaling constant and is usually 1/9. FoM is within0 ≤ FoM ≤ 1. In this case, FoM is multiplied by 100 and represents a percentagevalue. The higher the FoM is, the better the detected edge is [22].
As an objective measure of the edge detection credibility, Performance Ratio(PR) was used too. The PR is ideally equal to infinity. The PR is calculated as theratio of the true edges to false ones [23, 24]:
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1073
PR =True Edge(Edge pixels identified as Edges)
False Edges(Non Edge pixels identified as Edges)+(Edge pixels identified as Non Edge pixels)× 100.
(11)
3 RESULTS
3.1 F Measure
Tables 2, 3 and 4 show the F values before and after the application of the pro-posed approach on image with a small, medium and high number of details, re-spectively, over which db (from 2nd to 10th order) wavelet transform to the thirddecomposition level was applied. These tables show the values obtained for the fiveoperators.
From Table 2 it can be seen that based on the obtained results, in the firstdecomposition level, using the proposed approach, improvements were achieved withthe Canny operator in almost all cases, except for db4. A similar situation occurswhen a LoG operator is used, with the exception that improvements have not beenachieved in the case of db6. Using the proposed approach and image with lowdetails, significant improvements have been achieved at the second level, where alloperators record improvements in F values, except in the case of db8 with Prewittand Sobel operators. In the third level, only Canny and LoG record improvementsin F values using the proposed approach.
In the image with a medium number of details, very small improvements havebeen achieved in some operators at the first level, or the values are generally similar,without major deviations, as can be seen in Table 3. The situation is differentwith the second and third decomposition levels, where improvements are achievedby all operators, with the difference that improvements are higher in the thirdlevel.
In the case where the image consists of a high number of details, using theproposed approach, a similar or slightly better F value is obtained. At the secondlevel, improvements are generally achieved with gradient operators, while at thethird level, using the proposed approach, better values are obtained for all opera-tors.
1074 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. Panic
Read image
DWT
(up to Level 3)
Edge detection
i = 2:row-1
j = 2:col-1
Filtering only
8-neighbors of Pixel
with median filter
Edge detection on
filtered image
Filtering level 1 image
if edge (i,j) == 1
Figure 1. Proposed approach
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1075
Pi,j
Pi-1,j-1
Pi-1,j
Pi-1,j+1
Pi,j+1
Pi+1,j+1
Pi+1,j
Pi+1,j-1
Pi,j-1
Figure 2. Extraction neighbors of pixel
Figure 3. Complexity of proposed algorithm
1076 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. PanicW
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
0.36
50.378
0.37
90.
364
0.36
80.380
0.36
30.365
0.37
10.375
LoG
0.31
10.314
0.31
30.317
0.31
10.
299
0.30
00.303
0.31
10.315
Pre
witt
0.36
20.
342
0.35
50.
346
0.36
60.364
0.32
20.345
0.38
00.3
62
Sobel
0.36
50.
336
0.35
60.
349
0.37
10.
363
0.32
50.346
0.37
70.3
62
Roberts
0.53
60.563
0.50
30.579
0.51
90.576
0.55
10.566
0.51
40.582
Level2
Canny
0.27
80.305
0.24
60.269
0.20
30.225
0.18
90.194
0.19
40.211
LoG
0.25
60.274
0.23
50.256
0.19
10.211
0.20
50.207
0.18
60.202
Pre
witt
0.26
90.302
0.28
50.339
0.28
50.321
0.33
80.
301
0.27
90.351
Sobel
0.26
40.306
0.28
40.320
0.28
10.314
0.33
30.
303
0.27
80.333
Roberts
0.29
90.416
0.35
00.453
0.39
70.456
0.44
80.486
0.43
30.495
Level3
Canny
0.10
80.160
0.11
30.169
0.09
60.170
0.10
00.158
0.10
60.172
LoG
0.12
80.173
0.14
80.190
0.13
20.187
0.14
80.174
0.14
40.176
Pre
witt
0.14
70.
203
0.18
90.
241
0.18
40.
247
0.18
00.
251
0.19
40.2
56
Sobel
0.14
60.
201
0.18
70.
229
0.18
20.
235
0.17
90.
240
0.19
50.2
44
Roberts
0.18
30.
276
0.20
80.
255
0.21
80.
297
0.25
60.
335
0.26
80.3
48
Tab
le2.
Fva
lues
for
anim
age
wit
ha
low
num
ber
ofdet
ails
(LD
)
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1077W
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
0.31
80.
316
0.32
20.
318
0.32
10.
317
0.32
10.
319
0.32
10.3
20
LoG
0.23
40.
234
0.23
50.236
0.23
50.236
0.23
60.
236
0.23
70.2
36
Pre
witt
0.29
90.
222
0.21
90.220
0.22
50.
225
0.21
80.219
0.22
60.2
26
Sobel
0.22
90.
223
0.22
00.
219
0.22
60.
226
0.21
70.219
0.22
70.2
25
Roberts
0.25
00.
249
0.26
70.268
0.26
80.
258
0.27
40.
267
0.28
20.2
70
Level2
Canny
0.29
50.302
0.29
90.312
0.29
70.312
0.29
80.312
0.30
00.309
LoG
0.21
30.
211
0.20
70.213
0.21
10.213
0.20
90.214
0.20
70.212
Pre
witt
0.20
50.209
0.17
00.182
0.16
00.184
0.15
80.182
0.16
40.185
Sobel
0.20
50.209
0.16
90.180
0.16
10.182
0.15
80.178
0.16
30.183
Roberts
0.15
30.154
0.15
71.196
0.15
80.183
0.15
50.177
0.16
20.179
Level3
Canny
0.18
60.238
0.19
90.246
0.21
50.252
0.20
30.250
0.20
40.250
LoG
0.13
40.153
0.11
10.135
0.08
80.111
0.08
10.104
0.07
60.100
Pre
witt
0.11
80.138
0.09
20.118
0.07
80.115
0.07
20.110
0.07
50.110
Sobel
0.11
80.135
0.09
20.114
0.07
80.109
0.07
20.104
0.07
50.105
Roberts
0.07
80.081
0.06
50.088
0.04
20.054
0.03
40.046
0.03
50.047
Tab
le3.
Fva
lues
for
anim
age
wit
ha
med
ium
num
ber
ofdet
ails
(MD
)
1078 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. PanicW
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
0.22
20.
219
0.21
90.220
0.22
00.
219
0.21
90.220
0.22
00.221
LoG
0.19
20.
192
0.19
00.
189
0.19
20.
192
0.19
20.193
0.19
20.1
91
Pre
witt
0.17
10.
167
0.16
40.166
0.17
80.
175
0.16
50.166
0.17
20.1
69
Sobel
0.17
00.
165
0.16
40.165
0.17
70.
174
0.16
30.165
0.17
00.1
68
Roberts
0.16
10.
152
0.16
60.
163
0.16
30.
157
0.17
00.
162
0.16
40.1
57
Level2
Canny
0207
0.207
0.21
60.224
0.20
50.215
0.21
50.220
0.21
40.220
LoG
0.17
70.
176
0.18
20.
183
0.17
20.176
0.17
40.175
0.17
90.182
Pre
witt
0.15
90.166
0.13
90.145
0.12
70.143
0.13
50.141
0.13
30.150
Sobel
0.15
70.166
0.13
90.144
0.12
70.142
0.13
30.140
0.13
30.147
Roberts
0.12
30.
123
0.12
90.167
0.12
50.143
0.13
50.149
0.14
20.1
59
Level3
Canny
0.14
50.184
0.15
40.175
0.15
40.197
0.16
10.190
0.16
20.190
LoG
0.12
10.148
0.10
40.133
0.09
40.114
0.08
40.108
0.07
70.102
Pre
witt
0.11
00.136
0.07
70.102
0.07
80.114
0.07
30.103
0.07
70.111
Sobel
0.10
90.133
0.07
70.098
0.07
80.112
0.07
30.100
0.07
80.110
Roberts
0.08
70.089
0.06
40.088
0.06
70.084
0.06
40.079
0.07
00.087
Tab
le4.
Fva
lues
for
anim
age
wit
ha
hig
hnum
ber
ofdet
ails
(HD
)
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1079
3.2 Figure of Merit
Tables 5, 6 and 7 show the FoM values before and after the application of theproposed approach on image with a small, medium and high number of details,respectively, over which db (from 2nd to 10th order) wavelet transform to the thirddecomposition level was applied. These tables show the values obtained for the fiveoperators.
From Table 5 it can be seen that in the first decomposition level, the best valuesare obtained for the Canny operator, where in almost all cases it gives better values,except in the case of db4. By comparing Table 5 with Table 2, it can be seen thatusing FoM objective measurements, improvements have been made in the same orsimilar cases. By increasing the compression, that is, in Levels 2 and Levels 3, usingthe proposed approach, better FoM values were obtained, and greater improvementswere achieved, especially in the third level.
In the case of an image with a medium number of details, in the first level, Cannygave the best results, in other words, better values were achieved. In the second level,in the case of db2, improvements were achieved only with the Canny operator, whileother operators gave similar, but lower values. In other cases, the proposed algorithmrecords substantially better FoM values. For images with a medium number ofdetails, the best edge detection enhancements have been achieved in the third levelusing the proposed algorithm.
Since the compression ratio is higher in the image with a high number of details,based on this fact, it can be concluded that the detection will be much worse. Also,in the case of images with a high number of details, the best improvements areachieved with the Canny operator at all levels. Based on the results obtained, itcan be seen that the best improvements are achieved in the third level.
3.3 Performance Ratio
Tables 8, 9 and 10 show PR values before and after applying the proposed approachusing five edge detection operators and three wavelet decomposition levels. Table 8shows the values for an image with a low number of details. Based on the resultsobtained, it appears that improvements have been made with certain operators.However, based on the values obtained in Table 8, it can be seen that the oldPR values and new PR values obtained using the proposed approach have greatestdifference when using the Robert operator.
Table 9 contains the PR values obtained for an image with medium number ofdetails. From Table 9 can be seen that the new values obtained are best in thesecond and third decomposition levels. In other words, the best improvements areachieved in the second and third levels, depending on the operator used and theorder of the db wavelet.
Table 10 contains PR values obtained for an image with a high number of details.Based on these results, it can be seen that improvements have been achieved usingthe proposed approach, especially in the third decomposition level.
1080 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. PanicW
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
88.9
5591.928
89.4
6189
.272
85.9
8291.960
86.0
2586.914
88.7
67
89.626
LoG
89.1
6389
.024
90.2
0990.514
86.9
1585
.582
83.6
0585.108
89.7
79
88.5
56Pre
witt
91.7
8491.937
92.3
4192
.327
92.3
5892.572
92.6
6492
.632
92.8
68
92.5
40Sobel
91.8
9191
.712
92.6
7892.695
92.4
2392.466
93.0
3192
.934
92.8
91
92.7
42Roberts
95.1
5095.387
89.4
1995.816
86.6
8095.561
93.8
3795
.613
83.2
59
95.653
Level2
Canny
61.7
2766.006
51.3
0852.452
40.2
0741.320
31.9
7434.924
36.0
49
37.174
LoG
75.0
0679.220
60.4
7266.126
47.1
8751.753
45.9
5550.186
44.3
79
49.053
Pre
witt
82.1
8276
.052
88.8
2492.094
88.6
5491.490
89.0
7491.115
88.6
76
91.615
Sobel
82.7
8577
.091
88.8
1891.599
88.5
8191.498
88.9
7990.786
88.7
16
91.575
Roberts
82.4
9485.980
89.8
3288
.389
91.0
1688
.300
91.2
8191.386
91.8
77
92.731
Level3
Canny
6.04
59.491
34.3
1638.640
27.0
3534.575
25.1
5833.672
26.9
26
33.476
LoG
42.3
8548.432
49.2
0153.907
47.3
1250.844
48.1
3251.458
48.6
46
52.246
Pre
witt
52.0
3655.246
78.6
7571.700
78.1
3685.231
76.1
9685.186
75.8
41
84.392
Sobel
52.1
1355.991
78.6
0673.264
78.0
8886.561
76.1
9084.761
75.8
37
84.193
Roberts
57.3
8474.222
78.8
2181.909
77.9
5181.587
78.8
3982.599
78.9
01
82.152
Tab
le5.
FoM
valu
esfo
ran
imag
ew
ith
alo
wnum
ber
ofdet
ails
(LD
)
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1081W
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
83.3
3482.917
83.5
9383.835
83.6
3383.926
83.5
9683.773
83.7
15
84.068
LoG
76.3
5876
.351
76.9
8476
.892
77.0
0376
.875
77.2
1676
.957
77.2
30
77.0
50Pre
witt
65.0
9664
.526
63.4
0363
.351
63.6
9463
.564
63.0
5862
.834
63.6
96
63.3
35Sobel
65.0
6164
.397
65.5
5363
.420
63.9
4863
.449
63.0
7662
.941
63.7
05
63.4
44Roberts
58.4
8458.800
60.4
7758
.903
56.1
6055
.423
58.0
4656
.431
57.5
63
56.5
11
Level2
Canny
80.2
7380.650
80.2
6080.889
80.9
6381.348
80.9
2181.384
81.2
00
81.295
LoG
73.4
3373
.272
73.3
0073.418
73.4
7473.511
73.2
6173
.002
73.5
19
73.5
04Pre
witt
60.8
3060
.702
55.9
9156.815
55.3
1756.128
54.3
4655.447
55.1
18
56.2
28Sobel
60.8
4360
.732
56.0
5756.975
55.3
5456.375
54.4
6355.598
55.1
77
56.389
Roberts
54.3
2750
.858
50.2
6551.582
46.0
3346.891
43.3
0043.996
43.9
72
44.406
Level3
Canny
70.6
2875.211
68.8
3573.034
68.8
4773.556
66.7
9471.193
66.9
68
71.698
LoG
61.5
0361.749
51.3
2353.911
44.3
1647.050
41.3
1043.968
39.2
84
41.870
Pre
witt
49.6
8050.908
40.5
0944.000
36.6
6740.527
34.8
4028
.934
35.0
81
38.936
Sobel
49.6
1851.109
40.5
4343.917
36.6
5440.182
34.9
1038
.611
35.0
98
38.686
Roberts
42.2
9838
.790
31.9
9034.236
19.1
4320.251
15.9
7916.983
16.2
16
17.198
Tab
le6.
FoM
valu
esfo
ran
imag
ew
ith
am
ediu
mnum
ber
ofdet
ails
(MD
)
1082 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. PanicW
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
70.1
1171.230
69.1
2771.555
67.7
8770.466
68.0
5769.417
68.4
20
70.459
LoG
80.6
1280
.029
80.7
1380
.131
81.0
1680
.593
80.8
9480
.399
81.0
20
80.5
43Pre
witt
64.3
3863
.895
58.0
6358.180
62.1
6561
.852
59.8
0459.969
60.5
13
60.4
54Sobel
64.2
5963
.817
58.4
3558.711
62.3
4062
.180
59.9
9260.167
60.9
24
60.6
59Roberts
52.6
9252
.351
56.1
2153
.588
49.4
4647
.979
53.2
0451
.154
49.3
77
47.7
62
Level2
Canny
79.2
4179.417
80.7
6680.868
79.8
1179.977
80.1
9580.628
79.9
64
81.130
LoG
74.0
3473
.853
74.9
8874
.765
74.4
9874
.087
73.8
2673
.502
75.3
16
74.9
86Pre
witt
57.1
6456
.225
51.8
6251
.670
50.1
9851.235
47.0
2047.824
50.5
19
51.654
Sobel
57.1
8656
.503
51.5
9451
.833
50.4
0951.258
47.0
4047.927
50.5
45
51.554
Roberts
48.5
1645
.265
45.5
1146.339
42.0
7541
.858
37.1
9037.934
39.9
90
40.475
Level3
Canny
69.1
4873.049
66.3
0270.065
67.7
0471.391
66.0
9669.790
66.0
69
70.312
LoG
61.9
5463.395
51.6
1954.322
45.8
4048.129
41.8
2444.563
39.6
93
41.655
Pre
witt
48.6
0450.094
36.4
8739.421
32.0
2835.209
29.1
7632.407
29.6
50
33.184
Sobel
48.6
7050.230
36.4
0739.104
32.0
8035.073
29.3
0932.294
29.6
56
32.994
Roberts
41.7
2437
.448
29.3
6831.066
21.0
3821.889
19.9
5120.950
20.8
29
21.832
Tab
le7.
FoM
valu
esfo
ran
imag
ew
ith
ahig
hnum
ber
ofdet
ails
(HD
)
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1083W
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
28.6
9530.234
30.5
1928
.649
29.1
6430.704
28.5
2628.800
29.5
27
29.945
LoG
22.5
3222.923
22.7
9823.225
22.5
8921
.305
21.4
7321.741
22.5
73
23.038
Pre
witt
28.3
6726
.024
27.5
5326
.478
28.8
5628
.676
23.7
3926.379
30.7
06
28.4
26Sobel
28.7
9525
.315
27.6
6726
.757
29.5
3728
.484
24.1
1326.399
30.2
01
28.3
88Roberts
57.8
1364.519
50.5
4368.677
53.8
5767.983
61.2
4365.303
52.9
37
69.606
Level2
Canny
19.2
2221.983
16.3
2018.378
12.7
4214.538
11.6
6712.021
12.0
19
13.333
LoG
17.1
6518.843
15.3
4517.218
11.8
0313.387
12.8
5513.068
11.4
43
12.642
Pre
witt
18.3
8021.641
19.9
7825.675
19.8
9923.619
25.5
3521
.490
19.3
00
26.996
Sobel
17.9
1422.085
19.8
1123.516
19.5
8722.875
24.9
2221
.730
19.2
24
24.930
Roberts
21.2
2735.587
26.9
4841.347
32.9
2541.833
40.5
6047.205
38.1
56
48.936
Level3
Canny
6.04
59.491
6.35
610.141
5.28
610.217
5.54
99.375
5.9
09
10.356
LoG
7.35
810.470
8.71
711.731
7.60
011.502
8.70
210.541
48.6
46
52.246
Pre
witt
8.61
612.775
11.6
3015.895
11.2
5616.399
11.0
1016.723
12.0
19
17.248
Sobel
8.51
712.598
11.4
8514.844
11.1
3515.399
10.8
7115.795
12.0
77
16.171
Roberts
11.2
2119.060
13.1
0717.118
13.9
7821.112
17.2
1425.186
18.2
70
26.741
Tab
le8.
PR
valu
esfo
ran
imag
ew
ith
alo
wnum
ber
ofdet
ails
(LD
)
1084 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. PanicW
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
23.3
5523
.137
23.7
9123
.282
23.6
0923.231
23.6
0323
.428
23.6
12
23.4
80LoG
15.2
4615.296
15.3
9915.421
15.3
5015.459
15.4
2615
.413
15.5
38
15.4
31Pre
witt
14.8
1414
.285
14.0
2414.098
14.4
8514
.477
13.9
2014.028
14.6
34
14.5
95Sobel
14.8
7214
.320
14.1
0814
.003
14.5
5814.568
13.8
8814.006
14.6
51
14.5
36Roberts
16.6
9416
.557
18.2
4518.317
18.3
0617
.373
18.8
8718
.219
19.5
95
18.4
54
Level2
Canny
20.9
0621.648
21.3
7622.659
21.1
6622.678
21.2
4622.690
21.3
87
22.407
LoG
13.5
3913
.402
13.0
2013.546
13.3
6613.549
13.2
3413.610
13.0
59
13.478
Pre
witt
12.8
8313.215
10.2
1711.158
9.52
511.262
9.40
411.097
9.7
81
11.356
Sobel
12.8
6513.208
10.1
7811.005
9.56
811.110
9.39
010.842
9.7
30
11.182
Roberts
9.01
39.132
9.33
712.206
9.36
911.179
9.16
710.777
9.3
62
10.936
Level3
Canny
11.4
2815.604
68.8
3573.034
13.6
6016.821
12.7
4916.686
12.7
76
16.643
LoG
7.45
49.010
6.24
07.788
4.80
66.272
4.41
05.796
4.1
04
5.568
Pre
witt
6.70
58.030
5.08
56.670
4.24
56.477
3.86
76.152
4.0
40
6.210
Sobel
6.70
07.825
5.08
86.452
4.23
76.091
3.87
05.821
4.0
40
5.867
Roberts
4.25
04.427
3.47
24.805
2.18
52.855
1.76
52.435
1.8
16
2.471
Tab
le9.
PR
valu
esfo
ran
imag
ew
ith
am
ediu
mnum
ber
ofdet
ails
(MD
)
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1085W
avelet
db2
db4
db6
db8
db10
Opera
tor
Old
New
Old
New
Old
New
Old
New
Old
New
Level1
Canny
70.1
1171.230
69.1
2771.555
67.7
8770.466
68.0
5769.417
68.4
20
70.459
LoG
80.6
1280
.029
80.7
1380
.131
81.0
1680
.593
80.8
9480
.399
81.0
20
80.5
43Pre
witt
64.3
3863
.895
58.0
6358.180
62.1
6561
.852
59.8
0459.969
60.5
13
60.4
54Sobel
64.2
5963
.817
58.4
3558.711
62.3
4062
.180
59.9
9260.167
60.9
24
60.6
59Roberts
52.6
9252
.351
56.1
2153
.588
49.4
4647
.979
53.2
0451
.154
49.3
77
47.7
62
Level2
Canny
79.2
4179.417
80.7
6680.868
79.8
1179.977
80.1
9580.628
79.9
64
81.130
LoG
74.0
3473
.853
74.9
8874
.765
74.4
9874
.087
73.8
2673
.502
75.3
16
74.9
86Pre
witt
57.1
6456
.225
51.8
6251
.670
50.1
9851.235
47.0
2047.824
50.5
19
51.654
Sobel
57.1
8656
.503
51.9
5451
.833
50.4
0951.258
47.0
4047.927
50.5
45
51.554
Roberts
48.5
1645
.265
45.5
1146.339
42.0
7541
.858
37.1
9037.934
39.9
90
40.475
Level3
Canny
69.1
4873.049
66.3
0270.065
67.7
0471.391
66.0
9669.790
66.0
69
70.312
LoG
61.9
5463.395
51.6
1954.322
45.8
4048.129
41.8
2444.563
39.6
93
41.655
Pre
witt
48.6
0450.094
36.4
8739.421
32.0
2835.209
29.1
7632.407
29.6
50
33.184
Sobel
48.6
7050.230
36.4
0739.104
32.0
8035.073
29.3
0932.294
29.6
56
32.994
Roberts
41.7
2437.448
29.3
6831.066
21.0
3821.889
19.9
5120.950
20.8
29
21.832
Tab
le10
.P
Rva
lues
for
anim
age
wit
ha
hig
hnum
ber
ofdet
ails
(HD
)
1086 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. Panic
4 CONCLUSIONS
This paper proposed a new approach to edge detection in the images on which thewavelet decomposition was applied to the third level. As mother wavelet, Daubechieswas used from the second to the tenth order. The analyzed images are categorizedinto three complexity criteria, so they consist of a small, medium and high numberof details. F measure, FoM, and PR were used for an objective measure. Theproposed approach provides significant improvements in edge detection for almostall operators (Canny, LoG, Prewitt, Sobel).
Depending on the number of details in the image, the decomposition level aswell as db wavelet, the improvements are different. With a small number of details,the greatest improvements were achieved with the Canny operator. Other oper-ators also achieved improvement but depending on the db wavelet order. Basedon the obtained results, in the image with a small number of details, it can beseen that best improvements are achieved in the third level using Laplacian oper-ators. In the image with the medium number of details, in the first level similarresults are generally obtained, while in the second and the third levels improvementis made using the proposed approach. For images with a high number of details,better or similar values are obtained using the proposed approach. What can beconcluded is that in the second level, the best values are obtained by gradient op-erators.
Considering that the compression ratio is higher of the image with a highernumber of details, based on this fact, it can be concluded that the detection willalso be poorer, and consequently there will be a lower F values, FoM values andPR values. Since the proposed approach is intended for systems where compressionis used, i.e. compressed image processing, it can be concluded that increasing thedegree of compression also provides a better difference, or better value using theproposed approach.
Today’s systems require image quality to be as good as possible, with as muchcompression as possible in order to process these images in real time, such as edgedetection, segmentation, streaming, streaming in systems using augmented reality,etc. The proposed approach can find many practical applications in all systemswhere real-time information needs to be processed, especially in television systems,but it also provides a good basis for the direction of future research related tocompression and edge detection.
Acknowledgment
This work was done within the research project of the Ministry of Science andTechnological Development of Serbia TR35026 and III47016.
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1087
REFERENCES
[1] Chen, Y.—Wang, D.—Bi, G.: An Image Edge Recognition Approach Based onMulti-Operator Dynamic Weight Detection in Virtual Reality Scenario. Cluster Com-puting, Vol. 22, 2018, Suppl. 4, pp. 8069–8077, doi: 10.1007/s10586-017-1604-y.
[2] Srivastava, D.—Kohli, R.—Gupta, S.: Implementation and Statistical Compar-ison of Different Edge Detection Techniques. In: Bhatia, S., Mishra, K., Tiwari, S.,Singh, V. (Eds.): Advances in Computer and Computational Sciences. Springer, Sin-gapore, Advances in Intelligent Systems and Computing, Vol. 553, 2017, pp. 211–228,doi: 10.1007/978-981-10-3770-2 20.
[3] Maini, R.—Aggarwal, H.: Study and Comparison of Various Image Edge De-tection Techniques. International Journal of Image Processing, Vol. 3, 2009, No. 1,pp. 1–11.
[4] Chaganti, V. R.: Edge Detection of Noisy Images Using 2-D, Discrete WaveletTransform. Thesis, Florida State University Libraries, 2005.
[5] Dong, X.—Li, M.—Miao, J.—Wang, Z.: Edge Detection Operator for Under-water Target Image. 2018 IEEE 3rd International Conference on Image, Vision andComputing (ICIVC), 2018, pp. 91–95, doi: 10.1109/ICIVC.2018.8492749.
[6] Kuang, T.—Zhu, Q-X.—Sun, Y.: Edge Detection for Highly Distorted ImagesSuffering Gaussian Noise Based on Improve Canny Algorithm. Kybernetes, Vol. 40,2011, No. 5-6, pp. 883–893, doi: 10.1108/03684921111142430.
[7] Jun, L.: A Wavelet Approach to Edge Detection. Thesis, Sam Houston State Uni-versity, Huntsville, Texas, 2003.
[8] Muthukrishnan, R.—Radha, M.: Edge Detection Techniques for Image Seg-mentation. International Journal of Computer Science and Information Technology,Vol. 3, 2011, No. 6, pp. 259–267, doi: 10.5121/ijcsit.2011.3620.
[9] Yang, T.—Sun, G.—Duan, X.: A New Method of Wavelet Transform-Based EdgeDetection. 2011 IEEE 13th International Conference on Communication Technology,Jinan, 2011, pp. 789–792, doi: 10.1109/ICCT.2011.6157985.
[10] Yelampalli, P. K. R.—Nayak, J.—Gaidhane, V. H.: Daubechies Wavelet-Based Local Feature Descriptor for Multimodal Medical Image Registration. IETImage Processing, Vol. 12, 2018, No. 10, pp. 1692–1702, doi: 10.1049/iet-ipr.2017.1305.
[11] Wang, S.-H.—Zhang, Y.-D.—Dong, Z.—Phillips, P.: Wavelet Families andVariants. Chapter 6. In: Wang, S.-H. et al.: Pathological Brain Detection. Springer,Singapore, Brain Informatics and Health, 2018, pp. 85–104, doi: 10.1007/978-981-10-4026-9 6.
[12] Balachandran, A.—Ganesan, M.—Sumesh, E. P.: Daubechies Algorithm forHighly Accurate ECG Feature Extraction. 2014 International Conference on GreenComputing Communication and Electrical Engineering (ICGCCEE), Coimbatore,pp. 1–5, 2014, doi: 10.1109/ICGCCEE.2014.6922266.
[13] Rana, K.—Thakur, S.: Comparisons of Wavelets and Algorithms Based onWavelets and Comparing the Results with JPEG. 2017 International Conference on
1088 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. Panic
Energy, Communication, Data Analytics and Soft Computing (ICECDS), Chennai,2017, pp. 3871–3876, doi: 10.1109/ICECDS.2017.8390189.
[14] Zhao, D.—Sun, G.: A Novel Edge Detection Approach Using Multi-ResolutionImage. 2nd International Conference on Computer Engineering and Technology,Chengdu, 2010, pp. V2-692–V2-695, doi: 10.1109/ICCET.2010.5485706.
[15] Rao, G. S. B.—Prasad, P. M. K.—Kumar, M. N. V. S. S.: Investigation of Vari-ous Orthogonal Wavelets for Precise Analysis of X-Ray Images. Journal of Engineer-ing Research and Applications, Vol. 5, 2015, No. 2, Part 3, pp. 24–32.
[16] Ilic, S.—Petrovic, M.—Jaksic, B.—Spalevic, P.—Lazic, L.—Milose-vic, M.: Experimental Analysis of Picture Quality after Compression by DifferentMethods. Przeglad Elektrotechniczny, Vol. 89, 2013, No. 11, pp. 190–194.
[17] Delac, K.—Grgic, M.—Grgic, S.: Effects of JPEG and JPEG2000 Compressionon Face Recognition. In: Singh, S., Singh, M., Apte, C., Perner, P. (Eds.): PatternRecognition and Image Analysis (ICAPR 2005). Springer, Berlin, Heidelberg, Lec-ture Notes in Computer Science, Vol. 3687, 2005, pp. 136–145, doi:10.1007/11552499 16.
[18] Rani, K. S.—Rao, D. N.: A Comparative Study of Various Noise Removal Tech-niques Using Filters. Research and Reviews: Journal of Engineering and Technology,Vol. 7, 2018, No. 2, pp. 47–52.
[19] Wen, X.—Deng, Z.—Xue, H.: An Improved Median Filtering Image De-NoisingAlgorithm. Revista Tecnica de la Facultad de Ingenieria Universidad del Zulia,Vol. 39, 2016, No. 5, pp. 293–298, doi: 10.21311/001.39.5.38.
[20] The Berkeley Segmentation Dataset and Benchmark. Available at: https://www2.
eecs.berkeley.edu/Research/Projects/CS/vision/bsds/, accessed: 14.02.2019.
[21] Arbelaez, P.—Maire, M.—Fowlkes, C.—Malik, J.: Contour Detection andHierarchical Image Segmentation. IEEE Transactions on Pattern Analysis and Ma-chine Intelligence, Vol. 33, 2011, No. 5, pp. 898–916, doi: 10.1109/TPAMI.2010.161.
[22] Hagara, M.—Kubinec, P.: About Edge Detection in Digital Images. Radioengi-neering, Vol. 27, 2018, No. 4, pp. 919–929, doi: 10.13164/re.2018.0919.
[23] Khaire, P. A.—Thakur, N. V.: A Fuzzy Set Approach for Edge Detection. Inter-national Journal of Image Processing (IJIP), Vol. 6, 2012, No. 6, pp. 403–412.
[24] Jabbar, S. I.—Day, C. R.—Heinz, N.—Chadwick, E. K.: Using ConvolutionalNeural Network for Edge Detection in Musculoskeletal Ultrasound Images. Pro-ceeding of the 2016 International Joint Conference on Neural Networks (IJCNN),pp. 4619–4626, 2016, doi: 10.1109/IJCNN.2016.7727805.
New Approach to Edge Detection on Different Level of Wavelet Decomposition 1089
Vladimir Maksimovi�c received his B.Sc. and M.Sc. degrees inelectrical engineering from the Faculty of Technical Sciences inKosovska Mitrovica, University of Pristina, Serbia. He is Ph.D.candidate at the same faculty and his areas of research includeimage processing, telecommunications and multimedia systems.He has authored several scientific papers.
Branimir Jak�si�c received his B.Sc. and M.Sc. degrees in elec-trical engineering from the Faculty of Technical Sciences in Ko-sovska Mitrovica, University of Pristina, Serbia, and his Ph.D.degree in electrical engineering from the Faculty of ElectronicEngineering, University of Nis, Serbia in 2015. He is AssistantProfessor at the Faculty of Technical Sciences in Kosovska Mitro-vica. Areas of his research include telecommunications and mul-timedia systems. He has authored over 70 scientific papers onthe above subject.
Mile Petrovi�c is Full Professor at the Department of Elec-tronics and Telecommunications Engineering at the Faculty ofTechnical Sciences in Kosovska Mitrovica, Serbia. He has exten-sive experience in digital broadcasting systems and multimediaapplications, mobile technology and digital image processing. Heis the author of over 100 scientific peer-reviewed papers, initia-tor of a number of international Tempus projects in the countryand certified patents. He has participated in scientific researchprojects as well as in projects for the modernization and enhanc-ing quality of higher education. He is author of many textbooks
and scientific articles in the field of mobile radio and telecommunication networks.
Petar Spalevi�c received his B.Sc. degree from the Faculty ofElectronic Engineering, University of Pristina, in 1997 and hisM.Sc. and Ph.D. degrees from the Faculty of Electronic Engi-neering, University of Nis in 1999 and 2003, respectively. Heis Professor at the Faculty of Technical Sciences, Departmentof Telecommunications in Kosovska Mitrovica. His primary re-search interests are statistical communications theory, opticaland wireless communications, applied probability theory and op-timal receiver design.
1090 V. Maksimovic, B. Jaksic, M. Petrovic, P. Spalevic, S. Panic
Stefan Pani�c received his M.Sc. and Ph.D. degrees in electri-cal engineering from the Faculty of Electronic Engineering, Nis,Serbia, in 2007 and 2010, respectively. His research interestsin mobile and multi-channel communications include statisticalcharacterization and modelling of fading channels, performanceanalysis of diversity combining techniques, outage analysis ofmulti-user wireless systems subject to interference. Within dig-ital communication, his current research interests include theinformation theory, source and channel coding, and signal pro-cessing. He has published over 40 SCI indexed papers. Currently
he is Associated Professor at the Department of Informatics, Faculty of Natural Scienceand Mathematics, University of Pristina.