A Completed Modeling of Local Binary Pattern Operator for Texture Classification

Zhenhua Guo,

Lei Zhang,

David Zhang

2012 IEEE Transactions on Image Processing

NTNU-CSIE Chen-Lin Yu , Total page:23

2

Outline

Introduction Brief Review of LBP Completed LBP (CLBP) Experiments Conclusion

3

INTRODUCTION

4

Texture Classification

Texture classification is an active research topic in computer vision and pattern recognition.

In texture classification, the goal is to assign an unknown sample image to one of a set of known texture classes.

Texture classification process involves two phases:

(1) learning phase and (2) the recognition phase

Divided texture analysis methods into four categories:

Statistical

Geometrical

Signal processing

Model-based

http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/OJALA1/texclas.htm

6

Goal:

Proposing a new local feature extractor to generalize and complete LBP(CLBP).

Some information is missed in LBP code. We attempts to address that how to effectively represent the missing information in the LBP so that better texture classification.

In CLBP, a local region is represented by its center pixel and a local difference sign-magnitude transform (LDSMT).

7

BRIEF REVIEW OF LBP

8

LBP

gc is the gray value of the central pixel

gp is the value of its neighbors

P : is the total number of involved neighbors

R : is the radius of the neighborhood

gp: (Rcos(2p / P), Rsin(2p / P))

Example:

9

Rotation variance

We define formula U which value of an LBP pattern is defined as the number of spatial transitions (bitwise 0/1 changes) in that pattern.

0 1 1

1 0

0 0 0

1 1 1

1 0

1 0 1

01110000U(LBPP,R) = 2

11110101U(LBPP,R) = 4

10

The uniform LBP patterns refer to the patterns which have limited transition or discontinuities (U<=2) in the circular binary presentation [13].

(superscript “riu2” means rotation invariant “uniform” patterns with U<=2)

[13] Multiresolution gray-scale and rotation invariant texture classification with Local Binary Pattern

11

COMPLETED LBP (CLBP)

12

CLBP Framework

OriginalImage

LocalDifference

LDSMT

Center GrayLevel

S

M

CLBP_S

CLBP_M

CLBP_C

CLBP MapCLBP

HistogramClassifier

(nearest neighborhood)

13

I. Local Difference Sign-Magnitude Transform (LDSMT)

a 3*3 sample block local differences

sign components magnitude components

14

Fig(b) local differences vector

Fig(c) sign vector Fig(d) magnitude vector

Ex: different vector is [3,9,-13,-16,-15,74,39,31]

after LDMST sign vector is [1,1,-1,-1,-1,1,1,1]

and magnitude vector is [3,9,13,16,15,74,39,31]

15

II. CLBP Map

By the LDSMT ,three operators, namely CLBP_C, CLBP_S and CLBP_M, are proposed to code the C, S and M features, respectively

Combine CLBP_S and CLBP_M• 1. Concatenation (CLBP_S_M)• 2. Joint (CLBP_S/M)

Combine three operators• 1. Joint (CLBP_S/M/C)• 2. Hybrid (CLBP_M_S/C or CLBP_S_M/C)

16

Dissimilarity Metric and Multi-scale CLBP

There are various metrics to evaluate between two histograms, such as histogram intersection, log-likelihood ratio, and chi-square statistic [13].

The nearest neighborhood classifier with the chi-square distance is used to measure the dissimilarity between two histograms

17

EXPERIMENTS

18

Outex Database

Includes 24 classes of textures , each texture available at the site is captured using three different simulated illuminants pro-vided in the light source:

– H : 2300K ( 陽光 左下 )

– Inca : 2856K ( 日光燈 右下 )

– TL84 : 4000K( 螢光燈 右上 )

and nine rotation angles

(0o,5º,10º,15º,30º,45º,60º,75º and 90º)

19

① CLBP_S better than CLBP_M② CLBP_M/C better than CLBP_M

CLBP_S/M/C better than CLBP_S/M③ LTP better than CLBP_S(more robust to noise)④ CLBP_M than VARP,R

⑤ Finally, CLBP_S/M/C achieves better and more robust results than the state-of-the-art methods LBP VARP,R and VZ_MR8

20

CONCLUSION

21

We analyzed LBP from a viewpoint of local difference sign-magnitude transform (LDSMT), and consequently a new scheme, namely completed LBP (CLBP)

By fusing CLBP_C 、 CLBP_S 、 CLBP_M codes, it will much better texture classification accuracy than the state-of-the-arts LBP.

22

ENDThank you.