2020 3rd IEEE International Conference on Information Communication and Signal Processing

Geometry-based Completed Local Binary Pattern for Texture Image Classification

Wei Wang1,2, Qiqi Kou3*, Shu Zhou1,2, Ke Luo1,2, Lifeng Zhang1,2

1st
Tiandi (Changzhou) Automation Co., Ltd., Changzhou 213015, China
2st
CCTEG Changzhou Research Institute, Changzhou 213015, China
3st
School of Computer Science and Technology, China University of Mining and Technology, Xuzhou 221116, China
e-mail: [email protected], [email protected], [email protected], [email protected], [email protected]

Abstract—In view of the fact that the accuracy of texture image texture feature extraction methods have been proposed one
classification is easily affected by changes in illumination and after another, generally divided into four types: structural
rotation, based on the analysis of geometric curvatures method, mathematical transformation method, model method
information of the image microscopic geometric surface and and statistical method. The structural method is for images
the completed local binary pattern (CLBP), this paper with more regular texture structure, by extracting and
proposed a new descriptor, named as Geometry-based defining more regular texture textons in the image block,
Completed Local Binary Pattern (GCLBP). Inspired by the defining their spatial arrangement rules and calculating the
continuous rotation invariance and illumination robustness of distribution parameters of the texture primitives, the texture
the geometric curvature information, principal curvatures
features of the image can be efficiently described and
(PCs) of all pixels are first calculated and then used to
represent the gradient magnitude information of the image,
characterized. Mathematical transformation method is
which are further exploited to replace the original gradient usually performed by designing different mapping methods,
magnitude information in CLBP. To further improve the such as multi-scale mapping, spatial domain frequency
accuracy of texture classification, a cross-scale joint coding domain mapping, etc. Then, the image is mapped from the
strategy is exploited to form the final GCLBP. The spatial domain to the required transformation domain.
experimental results on two standard texture databases Methods based on mathematical transformation mainly
demonstrate that the GCLBP algorithm proposed in this paper include Fourier transform, Gabor transform and wavelet
is not only far superior to the original CLBP in terms of decomposition, etc. In the model method, it can be
classification recognition accuracy and dimensionality of implemented mainly in three steps: establishing a statistical
feature vector, but also better than most existing advanced distribution model on the texture image, adjusting the
texture classification methods. appropriate model parameters according to the data
distribution of the image, using the statistical characteristic
Keywords-completed local binary pattern (CLBP); principal parameters of the model to represent the texture features.
curvatures (PCs); texture classification; geometric curvatures Model-based texture feature representation methods mainly
include markov random field, hidden markov model and
I. INTRODUCTION
simultaneous auto-regressive, etc. Based on the statistical
As an important visual cue for humans to perceive the distribution characteristics of the texture image, statistical
information of surrounding objects, texture mainly describes method uses gray statistical characteristic parameters to
the surface properties of the corresponding object, and represent the spatial distribution of pixels. Classical
reflects the gray-scale distribution of the image in the spatial statistical methods mainly include gray level co-occurrence
domain [1]. With the rapid development and progress of matrix, Semi-variogram, Tamura and local binary pattern
computer vision technology, the extraction and analysis of (LBP) [2], etc. Due to the advantages of simple calculation,
image texture features are becoming more and more fast speed and high effect, LBP has been widely used in
important, and have been widely used in the recognition of content-based image retrieval and other pattern recognition
faces and irises, cross-view re-identification and tracking of applications, such as dynamic texture recognition, medical
moving targets, stitching and reconstruction of 3D scenes, image analysis, pedestrian detection, palmprint recognition,
etc. target tracking, etc.
At present, many scientific research institutes and leading Since Ojala proposed the famous LBP algorithm in 2002,
enterprises in the field of artificial intelligence have many improved algorithms have been proposed on the basis
established corresponding image processing and pattern of LBP. For example, the original LBP is sensitive to noise.
recognition laboratories, and conducted in-depth research on In order to improve the robustness to noise, Tan proposed to
the extraction and description of texture feature information. divide the threshold of the central pixel and adjacent pixels
Through the continuous research and innovation of scientific into local ternary patterns (LTP) [3]. To improve the
research workers in recent years, a large number of image classification and recognition ability to gray scale and

978-1-7281-8823-2/20/$31.00 ©2020 IEEE 274

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:36:01 UTC from IEEE Xplore. Restrictions apply.
rotation-variable images, Guo et al [4] proposed a completed A. Brief Review of CLBP
local binary pattern (CLBP), which uses the global threshold As mentioned above, the CLBP feature descriptor mainly
to convert the gray level of the center pixel into a binary includes three parts: sign information (CLBP_S), magnitude
code, namely CLBP-Center (CLBP_C). Then, the local information (CLBP_M) and center pixel information
difference information of the adjacent pixels are divided into (CLBP_C). By decomposing the local grayscale difference
two complementary components: sign and magnitude, and information into sign and magnitude information, it
defines them as two operators, CLBP-Sign (CLBP_S) and effectively solves the problem that different texture
CLBP-Magnitude (CLBP_M). By encoding these three information is divided into the same encoding mode.
binary strings and concatenating the corresponding three Furthermore, by combining the central pixel information
histograms, the final CLBP algorithm is proposed. Recently, with the sign and magnitude information, the discrimination
some other LBP variants have also been proposed, such as ability and robustness of the image texture feature descriptor
dominant rotated LBP (DRLBP)[5], sorted local gradient are improved. The original LBP code at a central point c and
pattern (SLGP) [6], extended mapping CLBP (EMCLBP) [7], p points spaced equidistantly around a circle of radius R is:
etc. Although the above methods have made certain progress, P −1
1, x  0
the classification accuracy of images with continuous CLBP _ S P , R =  s ( g p − g c )2 p , s ( x) =  (1)
rotation changes, uneven illumination and viewpoint changes p =0 0, x  0
still need to be further improved. g g
In this paper, to increase the robustness and efficiency of where c and p are the gray values of the central
texture classification with continuous rotation changes, sampling point and its p-th neighbor. Similar to the format of
uneven illumination and viewpoint changes, we propose the CLBP_S, the CLBP_M can be formulated as
geometry-based completed local binary pattern (GCLBP). In P −1
1, x  c
the proposed method, considering the continuous rotation CLBP _ M P ,R =  t (m p , c)2 p , t ( x, c) =  (2)
p =0 0, x  c
invariance and illumination robustness of the geometric
curvature information, principal curvatures (PCs)[8-10] of all where c is an adaptive threshold, mp is the magnitude
pixels are first calculated. Then, the calculation results are information at the image sampling point. In the original
further encoded and quantized, and finally a binary string CLBP algorithm, Guo sets the value of c to the average value
similar to CLBP-M is formed. By replacing the CLBP_M in of the mp of the whole image.
the traditional CLBP with the calculated binary string with For the central pixel information CLBP_C, to make it
PCs, the preliminary feature descriptor proposed in this easy to combine with the above-mentioned sign information
paper is obtained. To further improve the robustness and CLBP_S and magnitude information CLBP_M, its coding
classification accuracy of the feature descriptor, a cross-scale strategy is as follows
joint feature description strategy is exploited to form the CLBP _ CP,R = t ( gc , cI ) (3)
final GCLBP descriptor. Fig. 1 presents the flowchart of the
proposed GCLBP method. In the Equation 3, the definition of the function t is the
same as that in Equation 3, and the threshold value cI is the
GCLBP(P11, R1 ) average gray-scale value of the whole image. By
sign difference sign-based binary map concatenating the above three operators CLBP_S, CLBP_M
and CLBP_C, the CLBP is finally obtained.
Principal
principal Curvatures
curvatures PCs-based binary map
Although the CLBP algorithm has good feature
center difference center-based binary map representation capabilities, its feature dimension is up to
2200 dimensions, which will consume a lot of time and
GCLBP(P12, R1 )
storage space in the later image processing, such as feature
matching and recognition. So, its practical application range
Image I GCLBP Feature
is also limited.
GCLBP(P12, R2 )
B. Geometry-based Completed Local Binary Pattern
GCLBP(P13, R2 ) Calculation of Principal Curvatures:The PCs of a point
Figure 1. Framework of the proposed GCLBP method. on the geometric surface of an image can usually be obtained
by calculating the eigenvalues of the Hessian matrix, which is
The rest of this paper is organized as follows. Section 2 a square matrix composed of the second-order partial
presents the proposed GCLBP descriptor in detail. The derivatives of a multivariate function, and it describes the
experimental results are presented in Section 3. Conclusions local curvature of the image. For a gray image I (x, y ) , the
and future work are finally summarized in Section 4. Hessian matrix of one point for a scale factor σ is given by:
II. THE PROPOSED METHOD Dxx Dxy I  Gxx I  Gxy (4)
H  ( x, y ) = =
In this section, we first briefly introduce the basic D yx D yy I  Gxy I  G yy
concepts of the original CLBP, and then introduce the
calculating process of the GCLBP algorithm in detail.

275
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:36:01 UTC from IEEE Xplore. Restrictions apply.
where G xx , G yy , and Gxy denote the Gaussian second-order generate the GCLBPP, R, . what’s more, GCLBPP, R, is generated
1 2 2 3
1 1 1 2

derivative filters of the image along direction of x, y, and xy, in a similar way to GCLBPP, R, . Finally, GCLBPP, R, and 1 2 1 2

respectively. The Gaussian function with standard deviation σ

1 1 1 1

 2 , 3
GCLBPP1 , R2 are further concatenated to generate the final
is formulated as
GCLBP （ 1 , 2）+（ 2 , 3）
（P1 , R1）+（P1 , R2）
descriptor, as illustrated in Fig. 1. As far as
1  x2 + y2 
G ( x, y ,  ) = exp −  (5) we know, the feature dimension of the local binary pattern is
2 2
 2 2  mainly determined by the number of sampling points in the
neighborhood. Therefore, to reduce the feature dimension of
The eigenvalues of the Hessian matrix can be achieved by
the proposed GCLBP descriptor, P1 herein is set to 8. For the
solving the characteristic equation H − I = 0 for  [11]. The
other parameter variables, we empirically set R1 = 1 , R2 = 5 ,
eigenvalues can be formulated as follows:
1 = 1 ,  2 = 2 , and  3 = 4 . so, the dimensionality of the final
max =  Dxx + Dyy + (Dxx − Dyy )2 + 4(Dxy )2 
1 (6) （ , ）+（ , ） descriptor is 4 * (P+2) * (P+2) *2= 800,
2  GCLBP 1 2
（P , R ）+（P , R ）
1 1 1
2
2
3

min =  Dxx + Dyy − (Dxx − Dyy )2 + 4(Dxy )2 

1 (7) which is a relatively small value for image representation.
2 
where max and min are the minimum and maximum values III. EXPERIMENTAL RESULTS AND DISCUSSIONS
in the eigenvalues of the Hessian matrix, and they are also the A. Image Databases
two extreme components in the PCs of a point on the image To validate the superiority of the proposed GCLBP, two
surface. Normally, the two components of max and min are standard public texture databases, Outex and CUReT, are
rotation-invariant, and also contain the macro- and micro- used to experimental testing. In the Outex database,
structure information of the image. Outex_TC_00010 (TC10) and Outex_TC_00012 (TC12) are
As far as we know, The absolute value ratio between the designed for verifying the rotation and illuminant invariance
eigenvalues of the Hessian matrix can eliminate the influence of various descriptors. The two test suites include the same 24
of the nonlinear structure. Therefore, the scalar quantization classes of textures, which were collected under three different
method via ratio and arctangent function are formulated as illuminants (“horizon”, “inca”, and “t184”) and nine different
follows rotation angles (0°, 5°, 10°, 15°, 30°, 45°, 60°, 75°, and 90°).
max The CUReT consists of 61 texture classes collected under
= (8) different viewpoints and illumination angles, resulting in a
min
significant intra-class variation. More details about these two
 = arctan( ) (9) databases can be found in [12] and [13]. and we also follow
In order to facilitate the fusion with the sign-based binary the training and testing scheme in [4]. To ensure fair
string and center-based binary string in the original CLBP, comparison with other methods, all experiments on CUReT
we encode the variable  as follows database are randomly repeated one hundred times.
P −1
1, xk B. Dissimilarity Metric
GCLBP _ PCP, R =  t ( p , k )2 p , t ( x, k ) =  (10)
p =0 0, otherwise In this paper, to measure the similarity between training
histogram V and testing histogram U , the nearest-
where k represent the mean value of  p from the whole neighborhood classifier (NNC) with the chi-square distance
image. Similar to CLBPPriu, R2 , the rotation invariant version of are used as the dissimilarity metric. The chi-square distance
herein is defined as
GCLBP _ PCP, R is defined as GCLBP _ PCPriu, R2 _  .
N
D(U ,V ) =  (U i − Vi ) （U i +V i ）
2
GCLBP Generation: Note that there are three parameter (11)
i =1
variables, namely standard deviation  , number of
neighborhood sampling points P and sampling radius R, where N is the number of bins, U i and Vi are the values of U
which involved into sign-based binary string, center-based and V at the i th bin, respectively.
binary string and PCs-based binary string. Therefore, we
coordinately change the above three parameters to get a more C. Experimental Results
effective feature descriptor, as shown in figure 1. In the To verify the performance of the GCLBP descriptor
proposed GCLBP method, sign-based binary map, center- against rotation, illumination, and viewpoint changes,
based binary map and PCs-based binary map with  1 R1 and comparative evaluations with fourteen existing algorithms on
P1 are first obtained, and then combined jointly to build a 3-D Outex and CUReT are implemented. The overall best
joint histogram of them, denoted by GCLBPP, R . Similarly, 1 classification scores obtained by each algorithm for any
1 1
parameter setting are tabulated in Table 1. As can been seen,
GCLBPP1 ,2R1 are obtained by setting the standard deviation as the classification accuracy of the GCLBP algorithm proposed
 2 . Then, the GCLBPP1 ,1R1 and GCLBPP1 ,2R1 are concatenated to in this paper on TC10 test suite is higher than most of other
comparative algorithms, and only slightly lower than PCLBP

276
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:36:01 UTC from IEEE Xplore. Restrictions apply.
[14], EMCLBP [7] and FbLBP [15], indicating that our are easily available for different training samples, are used as
GCLBP descriptor has outstanding rotation invariance. competitors. Seen from the figure 2, the classification
Moreover, the scores of the GCLBP under different accuracy of our GCLBP algorithm 46, 23, 12 and 6 training
illuminants (“horizon” and “t184”) on TC12 test suite samples are 97.56%, 94.90%, 90.41% and 81.93%,
exceeds the classification results of all the 14 algorithms, respectively. Therefore, we can conclude that our GCLBP
which also fully proves that GCLBP not only has rotation descriptor performs better than the other methods, even when
invariance but also has remarkable illumination invariance. the number of training samples is small.
As for the classification results on CUReT, the scores of
the fifteen algorithms are obtained with 46 training samples. IV. CONCLUSION
As shown in Table I, the classification result of the proposed Based on the intrinsic geometrical characteristic of PCs,
GCLBP on CUReT is slightly lower than the score achieved this paper has proposed a new feature descriptor GCLBP for
by FbLBP, but it is significantly higher than the other robust texture classification. By calculating, encoding and
algorithms, indicating that our GCLBP has excellent quantizing PCs, a binary string is obtained and used to replace
discriminating ability for texture images with large intra-class the original magnitude information. Then, by collaboratively
variation caused by viewpoint change. changing the sampling radius and Gaussian filter radius, the
corresponding histograms under different filtering results are
TABLE I. THE CLASSIFICATION RESULTS OF DIFFERENT METHODS ON
OUTEX AND CURET obtained. Finally, a cross-scale joint feature description
strategy is adopted to obtain the final feature descriptor
TC12 GCLBP. Experimental results on two standard public
Algorithm TC10 CUReT Published in Bins
“t” “h” databases demonstrate that the GCLBP algorithm proposed in
（1, 2）+（2 , 4）
GCLBP 99.48 98.26 98.87 97.56 This paper 800
this paper is significantly better than other existing algorithms
（8,1）+（8, 5）
in terms of robustness in illumination, rotation and viewpoint
PCLBP 99.61 97.52 96.16 95.34 Optik 2019 400
changes. In future work, We will explore optimal parameter
SLGP 97.79 84.17 83.82 94.83 SPL 2018 168 optimization strategies to improve the robustness of our
Yang [16] - 97.08 95.99 - Access 2018 - algorithm under other conditions.
EMCLBP 99.61 96.37 95.32 95.17 MTAP 2018 8712
ACKNOWLEDGMENT
FbLBP 99.87 97.38 96.60 97.86 ESWA 2017 432
DRLBP 99.19 97.15 95.37 90.77 PRL 2016 256 This work was supported in part by the Science and
Technology Innovation and Entrepreneurship Special Fund
NISP [17] 99.19 96.97 96.93 95.81 MTAP 2016 -
[18]
General Project of China Coal Technology and Engineering
LNP 94.97 80.93 83.77 91.25 TCSVT 2015 - Group under Grant No. 2018-TD-MS016 and in part by the
COV-LBPD[19] 98.78 95.72 97.62 - TIP 2014 289 Science and Technology Innovation Special Fund Key
LCP [20] 98.65 96.18 97.34 95.49 SPL 2014 180 Project of China Coal Technology and Engineering Group
PRICoLBP [21] 94.48 92.57 92.50 - TPAMI 2014 3540 under Grant No. 2018-TD-ZD005.
NRLBP [22] 93.44 86.13 87.38 83.08 TIP 2013 30
REFERENCES
CLBC 98.96 95.37 94.72 95.39 TIP 2012 1990
[1] Ribas. LC, Sa. JJDS, Scabini. LFS, Bruno. OM, “Fusion of complex
CLBP 99.14 95.18 95.55 97.39 TIP 2010 2200 networks and randomized neural networks for texture analysis,”
Pattern Recognit., vol. 103, UNSP: 107189, Jul. 2020.
[2] T. Ojala, M. Pietikäinen, and T. T. Mäenpää, “Multiresolution gray-
scale and rotation invariant texture classification with local binary
pattern,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no.7, pp.
971–987, July. 2002.
[3] X. Tan and B. Triggs, “Enhanced local texture feature sets for face
recognition under difficult lighting conditions,” IEEE Trans. Image
Process., vol. 19, no. 6, pp. 1635–1650, Jun. 2010.
[4] Z. Guo, L. Zhang, and D. Zhang, “A completed modeling of local
binary pattern operator for texture classification,” IEEE Trans. Image
Process., vol. 19, no. 6, pp. 1657–1663, Jun. 2010.
[5] R. Mehta, K. Egiazarian, “Dominant rotated local binary patterns
(DRLBP) for texture classification,” Pattern Recognit. Lett., nol. 71,
pp.16-22, 2016.
[6] T. Song and L. Xin, “Grayscale-inversion and rotation invariant
Figure 2. Average classification accuracy of six algorithms with various texture description using sorted local gradient pattern,” IEEE Signal
training samples on CUReT Process. Lett., vol. 25, no. 5, pp. 625–629, May 2018.
[7] MH. Shakoor and R. Boostani, “A novel advanced local binary
To further evaluate the classification effectiveness of our pattern for image-based coral reef classification,” Multimed. Tools
Appl., vol. 77, pp. 2561–2591, 2018.
GCLBP on CURET, the average classification accuracy with
[8] Kou Qiqi, Cheng Deqiang, Zhuang Hhuandong, Gao Rui, “Cross-
different training number N (N=6, 12, 23, 46) are presented in Complementary Local Binary Pattern for Robust Texture
Figure 2. Five representative algorithms, whose codes or data Classification,” IEEE Signal Processing Letters, 2019, 26(1): 129-133.

277
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:36:01 UTC from IEEE Xplore. Restrictions apply.
[9] Hanbay. K, Alpaslan. N, “Continuous rotation invariant features for [16] P. Yang, F. Zhang and G. Yang, “Fusing DTCWT and LBP based
gradient-based texture classification,” Comput. Vis. Image features for rotation, illumination and scale invariant texture
Understand., vol. 132, pp. 87–101, Mar. 2015. classification,” IEEE Access, vol. 6, pp. 13336-13349, 2018.
[10] Kou Qiqi, Cheng Deqiang, Chen Liangliang, Zhao Kai, “A [17] N. Shrivastava and V. Tyagi, Noise-invariant structure pattern for
Multiresolution Gray-scale and Rotation Invariant Descriptor for image texture classification and retrieval, Multimedia Tools Appl.
Texture Classification,” IEEE Access, 2018, 6: 30691-30701. 75(18) (2016) 1–20.
[11] M. Do Carmo, “Differential Geometry of Curves and Surfaces,” [18] S. Wang, Q. Wu, J. Yang and J. Wang, “Local N-ary pattern and its
Computer Aided Engineering Design, vol. 2, no.4, 1976. extension for texture classification,” IEEE Trans. Circuits Syst. Video
[12] T. Ojala et al., “Outex—New framework for empirical evaluation of Technol., vol. 25, no. 9, pp. 1495–1506, 2015.
texture analysis algorithm,” in Proc. Int. Conf. Pattern Recog., 2002, [19] X. Hong, G. Zhao, M. Pietikainen, and X. Chen, “Combining LBP
pp. 701–706. difference and feature correlation for texture description,” IEEE Trans.
[13] K. J. Dana, B. van Ginneken, S.K. Nayar, J.J. Koenderink, Image Process., vol. 23, no. 6, pp. 2557–2568, Jun. 2014.
“Reflectance and texture of real-world surfaces,” ACM Trans. Graph., [20] T. Song and H. Li, “Noise-robust texture description using local
vol. 18, no. 1, pp. 1-34, 1999. contrast petterns via global measures,” IEEE Signal Process. Lett.,
[14] Qiqi Kou, Deqiang Cheng, Liangliang Chen, Yinping Zhuang. vol. 21, no. 1, pp. 93–97, Jan. 2014.
Principal Curvatures based Local Binary Pattern for Rotation [21] X. Qi et al, “Pairwise rotation invariant co-occurrence local binary
Invariant Texture Classification, OPTIK, 2019, (193): UNSP 162999. pattern,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 36, no. 11, pp.
[15] Z. Pan, Z. Li, H, Fan and X. WU, “Feature based local binary pattern 2199–2213, Nov. 2014.
for rotation invariant texture classification,” Expert Systems With [22] J. Ren, X. Jiang, and J. Yuan, “Noise-resistant local binary pattern
Applic., vol. 88, pp. 238–248, 2017. with an embedded error-correction mechanism,” IEEE Trans. Image
Process., vol. 22, no. 10, pp. 4049–4060, Oct. 2013.

278
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 28,2024 at 08:36:01 UTC from IEEE Xplore. Restrictions apply.