Texture Classification Via Convex-and-Concave Decoded Pattern

Abstract— This paper proposes an effective LBP-like method, SIFT descriptor. Guo et al. [28] developed a completed
referred to as convex-and-concave decoded pattern (CCDP) for modeling of LBP to produce CLBP (Completed Local Binary
texture modeling and classification. The essence of CCDP is Pattern) which combines the magnitudes of local differences
to divide local features into two distinct parts, i.e., convex
decoded pattern (CvDP) and concave decoded pattern (CxDP) as well as their signs. The authors in [14] proposed a family
by encoding relationships between the pixels in triplets around of descriptors referred to as binary gradient contours (BGC1,
the central pixel of a 3×3 square neighborhood. To make BGC2 and BGC3) based on pairwise comparison of adjacent
CvDP and CxDP more invariant to monotonic gray scale pixels belonging to one or more closed route traced along
transformation, two supplementary triplets are added in the the periphery of 3×3 grayscale image patch. Recently, Kaya
modeling. The two triplets are formed using the central pixel
and four virtual pixels set to the average local and global et al. [10] proposed local binary patterns by neighborhoods
gray levels and the median of the grayscale values of the 3×3 (nLBPd ) where the comparison between the peripheral pixels
square neighbourhood and the whole image, respectively. The is done with sequential neighbors and/or inside neighbors
histograms from single scale analysis realized with CvDP and defined by a distance parameter d. More recently, Issam et
CxDP are combined into hybrid distributions to form the final al. [21] proposed local directional ternary pattern (LDTP) for
CCDP feature descriptor. Extensive experiments performed on
eight challenging texture datasets show that the CCDP operator texture classification. The LDTP operator consists in encod-
can achieve classification performance which is better than 29 ing both contrast information and directional pattern features
recent state-of-the-art descriptors. in a compact way based on local derivative variations. LDTP
Index Terms— LBP, CCDP, Feature extraction, Texture clas- conveys valuable information about the nature of textures by
sification. capturing local structures using both LTP’s [19] and LDP’s
[11] concepts simultaneously.
I. I NTRODUCTION Although significant progress has been made and most
Texture modeling and classification plays a substantial role extensions of LBP have improved the performance of LBP,
in image processing, computer vision. In the past decades, there still remain some potential flaws of LBP. To address
a variety of texture feature extraction approaches have been the weakness of several LBP-like descriptors and keep the
proposed in the literature. One can cite human perception- simplicity and effectiveness of the basic LBP, we designed
based features [7], random features [5], filter-based methods conceptually simple and robust texture representation model,
such as Gaussian Markov random fields [3], wavelet [4] and referred to as convex-and-concave decoded pattern (CCDP)
Gabor [1], methods based on the ranklet transform [6], use for texture modeling and classification. The proposed de-
of fractal analysis [8], etc. scriptor, which has low computational complexity, offers bet-
Recently, LBP-based approaches have emerged as one ter texture description while allowing considerably enhancing
of the most effective texture operators, because of their both the discriminative power of LBPs and their robustness
interesting performance. As a simple and efficient statistical to small variations.
descriptor, the local binary pattern (LBP) proposed in [2], The remaining sections are structured as follows: Section
has been successfully applied in a wide range of applications II briefly review some existing texture descriptors. Section
including dynamic texture recognition, face description and III presents the proposed CCDP descriptor. Comprehensive
recognition, motion detection, medical image analysis image experimental results and comparative evaluation are given
retrieval, biomedical image analysis, object detection, finger- in Section IV. Section V concludes the paper and proposes
print matching and background subtraction. The efficiency some future research directions.
and usability of the LBP method and its success in various
computer vision applications has inspired the development II. R EVISITING THE E XISTING M ETHODS
of much new effective LBP variants. Indeed, after Ojala’s In this section we review LBP and its variant nLBPd from
work, several renowned extensions and modifications of LBP which we are inspired to develop our own LBP-like method.
have been proposed with the aim to improve its applica-
bility, robustness and discriminative power. To improve the A. Brief review of Traditional Local Binary Patterns
robustness of LBP to noise, Tan et al. [19] proposed local The spatial structure of a local image texture is charac-
ternary pattern (LTP) for face recognition. LTP consists in terized by thresholding a 3×3 grayscale image patch with
extending original LBP to 3-valued codes using a threshold the value of the central pixel. The obtained values are tuned
interval for center and neighboring pixels. Heikkilä et al. [16] into binary values (0 or 1) forming a code of 8 bits that
proposed center-symmetric LBP (CS-LBP) which combines is transformed into decimal number. Once the LBP label
the strengths of LBP texture operator and the well-known of each pixel is generated, a histogram is built to represent
the texture image. Formally, given the gray value Ic of the value φ and the average local gray level δ (cf. Eq. 5) of
central pixel in the image I, the kernel function fLBP (·) of the grey-scale values of the whole 3×3 neighbourhood and
LBP operator for a 3×3 square neighborhood [I7 , I6 ,...., I1 , the median value φ and the average global gray level δ (cf.
I0 ] is given as follows: Eq. 6) of the grey-scale values of the whole image in the
modeling of the proposed operator.
P−1 
X 1,
if a >= b P−1
fLBP (x) = =(Ip , Ic )×2p , =(a, b) = 1 X
0,
otherwise δ= ( Ip + Ic ) (5)
p=0 P + 1 p=0
(1)
where Ip is the gray value of neighboring pixels, P PM PN
m=1n=1 I(m, n)
corresponds to the number of neighboring pixels and =(·) δ= (6)
M×N
is the Heaviside step function.
where I is the gray level function of the image of size
B. Brief review of Local Binary Patterns by neighborhoods M×N, Ic represents the gray value of the central pixel
(nLBPd ) and Ip (p ∈ {0, 1, ..., P-1}) denotes the gray levels of the
Local patterns by neighborhoods (nLBPd ) developed by peripheral pixels.
Yılmaz et al. [10] encodes the relationship between the
peripheral pixels in doublets around the central pixel Ic . The Based on both convex-and-concave binary thresholding
comparison between the surronding pixels is done inside indicators introduced previously, two operators termed con-
neighbors and/or with sequential neighbors defined by a vex decoded pattern (CvDP) and concave decoded pattern
distance parameter d. For d=1, the nLBPd operator becomes (CxDP) are constructed which can be expressed in the
equivalent to the BGC1 operator [14]. The kernel function following manageable and formal way (cf. Eq. (7) and (8):
fnLBPd (·) of nLBPd operator for a 3×3 square neighborhood
P−1
is defined by Eq. 2. X
fCvDP (x) = (=(Ip , Ic ) ⊗ =(I(p+d) mod P , Ic )) × 2P−1−p +
p=0
P −1
X =(δ, Ic ) ⊗ =(φ, Ic ) × 2P +
fnLBPd (x) = =(Ip , I(p+d) mod P ) × 2p (2)
=(δ, Ic ) ⊗ =(φ, Ic ) × 2P+1 − 1
p=0
P−1
X
III. P ROPOSED METHOD = ξ(Ip , Ic , I(p+d) mod P ) × 2P−1−p +
p=0
The construction concepts of LBP, nLBPd and many
others motivate us to propose convex-and-concave decoded ξ(δ, Ic , φ) × 2P + ξ(δ, Ic , φ) × 2P+1 − 1
pattern descriptor for texture classification. In order to give a (7)
compact mathematical formulation of the proposed operator,
P−1
it is convenient to define, beforehand, the convex and concave fCxDP (x) =
X
(=(Ip , Ic ) ⊗ =(I(p+d) mod P , Ic )) × 2P−1−p +
¯ and ξ(·)). The essence
binary thresholding indicators (ξ(·) p=0
¯
of ξ(·) and ξ(·) indicators is to partition all of the triplets =(δ, Ic ) ⊗ =(φ, Ic ) × 2P +
around the central pixel Ic into two categories, convex and
=(δ, Ic ) ⊗ =(φ, Ic ) × 2P+1 − 1
concave patterns and this independently of the distribution
P−1
of pixels around Ic . These patterns can be viewed as a gray- X
¯ p , Ic , I(p+d) mod P ) × 2P−1−p +
= ξ(I
scale local micro-patterns which describe local undulation p=0
characteristics of the image. The convex-and-concave binary ¯ Ic , φ) × 2P + ξ(δ,
¯ Ic , φ) × 2P+1 − 1
¯ and ξ(·)) are defined as below ξ(δ,
thresholding indicators (ξ(·) (8)
(cf. Eq. 4 and 3).
10
Note that, CvDP and CxDP operators produce 1023 (2 -
¯ b, c) ==(a, b) ⊗ =(c, b) 1) possible patterns in a 3×3 neighborhood. Once the CvDP
ξ(a,
and CxDP labels are constructed for every pixel of the
(3)

1, if a <= b and c <= b image, the histograms representing the texture are generated
=
0, otherwise as follows (cf. Eq. (9) and (10):
X
ξ(a, b, c) ==(a, b) ⊗ =(c, b) hCvDP (λ) = γ(fCvDP (x)), λ) (9)

1, if a >= b and c >= b (4) x
=
0, otherwise
X
hCxDP (λ) = γ(fCxDP (x)), λ) (10)
x
where =(a, b) = =(b, a) and ⊗ denotes multiplication
operator. where λ ∈ [0, Nbins ] is the bin number, Nbins =210 -1 is
Aiming at finding a code which is invariant to monotonic the number of bins and the gamma function γ(·) is defined
gray scale transformation, we adopted to include the median as below (cf. Eq. (11)):
 (dataset 3 in Table III) where several methods as well

1, if a = b; as the proposed one manage to differentiate all classes
γ(a, b) = (11) perfectly.
0, otherwise
• It is noteworthy that there is a significant performance
To achieve a reliable representation of texture which is drop for several tested methods on OuTeX TC 00001,
more robust, the coarse and fine information can be captured XU HR and CUReT, (datasets 5, 6 and 8). Descriptors
by multi-scale which can be made through a linear combi- like LNDP, ID-LBP, dLBPα and CS-LBP are some
nation of different features obtained by several operators. In examples that highlight this remark.
this study, the two descriptors CvDP and CxDP are combined • It is easily noticed that the proposed multi-scale CCDP,
into hybrid distributions to form the final CCDP model (cf. which exploits complementary information of fused
Eq. (12)). This hybrid texture description model is powerful operators (i.e., CvDP and CxDP), provides better clas-
because it permits to enhance the discriminative power of sification results on all the tested datasets than other
CvDP and CxDP descriptors through the complementary tested state-of-the-art descriptors and it is in the top 1
information of both convex and concave representations. descriptors, except for Jerry Wu dataset (dataset 2 in
CCDP operator produces 2046 (2(210 -1)) possible patterns Table III) for which it is ranked at the 7th position.
in a 3×3 neighborhood. Furthermore, for this dataset, as shown in Table III
(column 2), the CCDP method allowed to reach the
hCCDP = hCvDP ∪ hCxDP (12) classification rate of 98.18%, which is considered as
IV. E XPERIMENTS AND DISCUSSIONS satisfactory classification result as it is very close to the
rate of the descriptors ranked before it (98.91% for the
In this section, we investigate the performance of the
top 1).
proposed CCDP method against the 29 recent state-of-the-
art texture descriptors, summarized in Table I. For that, The striking performance of CCDP for the classification
eight publicly available datasets, shown in Table II, are of texture of the eight tested datasets clearly demonstrates
used to highlight the improvement that our method provide. that the concatenated marginal joint distributions of CvDP
As regarding the evaluation protocol, half of samples of and CxDP turns out to be a very powerful representation of
each dataset were randomly selected for training and the image texture and robust to gray scale variations.
remaining half for testing. The images of the testing subset V. C ONCLUSION
are then classified through the nearest-neighbour rule (1-NN)
with L1 distance (cf. Eq. (13)). The procedure is repeated In this paper, we have developed, based on convex-and-
100 times, each time with a new subdivision being conducted concave binary thresholding indicators, convex-and-concave
of the training and testing sets, and the accuracy is then decoded patterns (CCDP), proved to be a conceptually and
averaged over all these tests. computationally simple yet efficient framework of LBP for
image texture modeling. The proposed CCDP descriptor is
defined through central pixel based comparison of pixels
X
dL1 (hi , hj ) = |hi (k) − hj (k))| (13)
k belonging to closed path traced along the periphery of
3x3 neighborhood. To make the descriptor more robust to
where hi ={h1i ,h2i ,...,hK i } is the query feature vector and gray scale variation, several virtual pixels are added in the
hj ={h1j ,h2j ,...,hK
j } the target feature vector. modeling. The proposed CCDP model was evaluated com-
The experimental results are summarized in Tables III and prehensively on eight challenging texture datasets against a
IV. Table III illustrates the average accuracy (i.e., over the total of 29 state-of-the-art descriptors. Experimental results
100 subdivisions) of each descriptor and for each texture highlighted that the proposed model has high discriminative
dataset. Table IV reports the ranking results for each dataset power to classify texture images with better classification
based on the average accuracy reachted by the evaluated accuracy than the evaluated 29 state-of-the-art methods. For
descriptors. In light of these results, the following findings future work, we plan to extend the evaluation of the proposed
can be drawn: descriptor to other challenging computer vision applications,
• It can be observed that the CS-LMP method produces such as emotional state classification, face recognition, writer
the worst overall classification accuracy on almost all identification, etc.
the tested datasets.
• It is apparent from Table III that, except CS-LMP which R EFERENCES
does not produce good result, all the other evaluated de- [1] B.S. Manjunath and W. Y. Ma. Texture features for browsing and
scriptors get very promising results on Brodatz dataset retrieval of image data. IEEE Trans. Pattern Anal. Mach. Intell., vol.
(dataset 1 in Table III) where their average accuracy 18, no. 8, pp. 837–842, Aug. 1996.
[2] T. Ojala, M. Pietikainen, and T. Maenpaa. Multiresolution gray-scale
is above 96%. Some descriptors like RALBGC, LDTP, and rotation invariant texture classification with local binary patterns.
BGC1, LBP, LTP, as well as the proposed CCDP IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 971-987,
descriptor manage to differentiate all classes perfectly Jul. 2002.
[3] F.S. Cohen, Z.G. Fan, M.A. Patel. Classification of rotated and scaled
(average accuracy equal to 100%) on Brodatz dataset. textured images using Gaussian Markov random field models. IEEE
We can express the same remark for KTH-TIPS dataset Trans. Pattern Anal. Mach. Intell. 13(2), (1991), pp. 192–202.
 No. Name Acronym Dim. Ref. Year
1 Local Binary Patterns LBP 28 [18] 1996
2 Center-Symmetric LBP CS-LBP 24 [16] 2006
3 Local Phase Quantization LPQ 28 [31] 2008
4 Improved Center-Symmetric LBP (D) D-LBP 24 [20] 2009
5 Improved Center-Symmetric LBP (ID) ID-LBP 24 [20] 2009
6 Local Quinary Patterns LQP 4·28 [17] 2010
7 LBP Variance Histogram LBPV 26 -5 [27] 2010
8 Binary Gradient Contours (1) BGC1 28 -1 [14] 2011
9 Binary Gradient Contours (2) BGC2 28 -1 [14] 2011
10 Binary Gradient Contours (3) BGC3 (24 -1)2 [14] 2011
11 Gradient-based LBP GLBP 28 [15] 2011
12 Local Maximum Edge Binary Patterns LMEBP 8.29 [26] 2012
13 Improved binary gradient contours (1) IBGC1 2·(28 -1) [9] 2012
14 Multi-Scale Densely Sampled Complete LBP DCLBP 2.(26 -5) [25] 2013
15 Local Graph Structure LGS 28 [30] 2013
16 Symmetric Local Graph Structure SLGS 28 [29] 2014
17 Sign Maximum Edge Position Octal Pattern SMEPOP 2*(28 -5)-1 [24] 2015
18 Magnitude Maximum Edge Position Octal Pattern MMEPOP 2*(28 -5)-1 [24] 2015
19 LBP by Neighborhoods nLBPd 28 [10] 2015
20 Directional LBP dLBPα 28 [10] 2015
21 Local Extreme Complete Trio Pattern LECTP 3.28 [33] 2015
22 Local Extreme Sign Trio Patterns LESTP 29 [33] 2015
23 Extended local graph structure ELGS 29 [32] 2016
24 Quad Binary Pattern QBP 24 [23] 2016
25 Symmetric Local Multilevel Pattern CS-LMP 24 [34] 2016
26 Local Neighborhood Difference Pattern LNDP 26 -1 [12] 2017
27 Local Quadruple Pattern LQPAT 29 [13] 2017
28 Local Directional Ternary Pattern LDTP 2(29 − 1) [21] 2018
29 Repulsive-and-Attractive Local Binary Gradient Contours RALBGC 2(29 − 1) [22] 2018

TABLE I
S UMMARY OF TEXTURE DESCRIPTORS TESTED AND COMPARED WITH OUR PROPOSED DESCRIPTORS .

No. Name Classes Samples Total samples Sample resolution Image format Predefined
per class (pixels) train/test sets?
1 Brodatz 13 16 208 256×256 Monochrome (TIFF) No
2 Jerry Wu 39 4 156 256×256 Colour (BMP) No
3 KTH-TIPS 10 4 40 100×100 Monochrome (PNG) No
4 OuTeX TC 00000 24 20 480 128×128 Monochrome (RAS) Yes
5 OuTeX TC 00001 24 88 2112 64×64 Monochrome (RAS) Yes
6 XU HR 25 40 1000 1280×960 Monochrome (PNG) No
7 Kylberg 28 160 4480 576×576 Monochrome (BMP) No
8 CUReT 61 92 5612 200×200 Monochrome (PNG) No

TABLE II
I MAGE DATASETS CONSIDERED IN THE EXPERIMENTS .

[4] R. Porter, N. Canagarajah. Robust rotation invariant texture classifica- [12] Verma, M., Raman, B. (2017). Local neighborhood difference pattern:
tion 1997. IEEE International Conference on Acoustics, Speech, and A new feature descriptor for natural and texture image retrieval.
Signal Processing, I-V, 1997, pp. 3157-3160. Multimedia Tools and Applications, 1-24.
[5] Liu, L. and Fieguth, P. (2012). Texture classification from random fea- [13] Chakraborty, S., Singh, S. K., Chakraborty, P. (2017). Local quadruple
tures. IEEE Transactions on Pattern Analysis and Machine Intelligence, pattern: A novel descriptor for facial image recognition and retrieval.
34(3), 574-586. Computers & Electrical Engineering.
[6] Bianconi, F., Fernández, A., González, E. and Armesto, J. (2009). [14] Fernández, A., Álvarez, M.X., Bianconi, F. Image classification with
Robust color texture features based on ranklets and discrete Fourier binary gradient contours. Opt. Lasers Eng. 49(9-10), 1177-1184 (2011).
transform. Journal of Electronic Imaging, 18(4), 043012-043012. doi:10.1016/[Link].2011.05.003.
[7] Tamura, H., Mori, S. and Yamawaki, T. (1978). Textural features [15] He, Y., Sang, N. Robust illumination invariant texture classification
corresponding to visual perception. IEEE Transactions on Systems, using gradient local binary patterns. International Workshop on Multi-
Man, and Cybernetics, 8(6), 460-473. Platform/Multi-Sensor Remote Sensing and Mapping, China, pp. 1-6
[8] Y. Xu, H. Ji, C. Fermuller. A projective invariant for texture. Interna- (2011).
tional Conference on Computer Vision and Pattern Recognition, 2006, [16] Heikkilä, M., Pietikäinen, M., Schmid, C. (2006, December). Descrip-
pp.1932-1939. tion of interest regions with center-symmetric local binary patterns. In
[9] A. Fernández, M. X. Álvarez, F. Bianconi. Texture Description Through ICVGIP (Vol. 6, pp. 58-69).
Histograms of Equivalent Patterns. Journal of Mathematical Imaging [17] Nanni, L., Lumini, A., Brahnam, S. Local binary patterns variants as
and Vision. Vol. 45, pp 76-102. 2013. texture descriptors for medical image analysis. Artif. Intell. Med. 49(2),
[10] Yılmaz, K., Ertuğrulb, Ö. F., Tekinc, R. Two novel local binary pattern 117-125 (2010).
descriptors for texture analysis. Applied Soft Computing, Vol. 34, 2015, [18] Ojala, T., Pietikäinen, M., Harwood, D. A comparative study of texture
pp. 728-35. measures with classification based on feature distributions. Pattern
[11] Zhang, B., Gao, Y., Zhao, S., Liu, J. (2010). Local derivative pattern Recognit. 29(1), 51-59 (1996).
versus local binary pattern: face recognition with high-order local [19] X.Y. Tan, B. Triggs. Enhanced local texture feature sets for face
pattern descriptor. IEEE transactions on image processing, 19(2), 533- recognition under difficult lighting conditions, IEEE Trans. Image
544. Process. 19(6). (2010), pp. 1635-1650.
 Dataset
Texture descriptor 1 2 3 4 5 6 7 8
CCDP 100.00 98.18 100.00 99.97 99.68 96.63 99.69 94.45
RALBGC 100.00 97.92 100.00 99.83 99.22 95.14 99.22 93.82
LDTP 100.00 98.23 100.00 99.76 99.20 91.78 97.74 91.81
LESTP 99.98 98.63 100.00 97.69 94.41 93.11 99.45 89.64
LECTP 99.65 96.69 99.60 97.55 97.48 93.32 99.42 86.71
CS LMP 46.38 19.69 50.60 17.18 12.40 42.38 45.71 11.83
LGS 99.92 97.18 99.35 99.67 98.40 91.89 98.15 92.60
SLGS 99.71 98.50 97.75 99.83 98.31 92.38 98.54 93.76
LNDP 98.47 87.63 96.15 96.74 88.52 84.75 92.60 72.66
LQPAT 99.93 97.28 99.60 99.81 97.67 89.45 96.92 90.46
ELGS 99.97 97.73 100.00 99.68 98.73 92.14 98.48 92.91
LMEBP 99.97 98.91 100.00 97.58 93.60 94.31 99.49 89.33
MMEPOP 98.98 94.97 97.35 98.85 95.18 87.71 94.80 79.68
SMEPOP 99.99 96.78 99.45 99.61 97.48 90.89 97.49 89.28
LBPV 99.34 91.74 96.75 98.95 97.54 84.13 97.14 80.88
DCLBP 99.99 97.65 100.00 99.22 98.47 90.82 97.30 87.86
BGC1 100.00 97.59 100.00 99.90 98.53 91.53 97.91 92.67
BGC2 99.39 97.09 100.00 99.14 96.68 90.13 97.30 89.96
BGC3 99.82 98.21 99.10 98.87 96.31 89.30 96.69 86.34
D-LBP 99.93 92.94 97.30 97.00 90.81 84.86 94.80 79.34
GLBP 99.86 95.32 99.40 97.26 92.85 86.41 94.11 80.53
ID-LBP 97.18 88.51 99.65 95.90 88.49 79.64 89.44 75.43
LBP 100.00 97.26 100.00 99.68 98.44 91.45 97.07 90.71
LQP 99.97 97.69 99.90 99.13 98.50 92.14 99.33 87.04
dLBPα 96.13 97.49 100.00 95.19 88.28 91.13 98.43 89.81
nLBPd 100.00 97.59 100.00 99.90 98.53 91.53 97.91 92.67
QBP 97.15 90.85 100.00 95.30 91.33 82.59 92.95 66.50
CS-LBP 98.13 90.91 100.00 92.72 84.72 81.53 93.60 67.31
LTP 100.00 98.06 100.00 99.65 99.13 91.47 98.96 90.26
LPQ 99.80 98.76 98.10 99.75 99.20 91.64 96.77 88.45

TABLE III
AVERAGE CLASSIFICATION ACCURACY. T E PROPOSED METHOD IS HIGHLIGHTED IN SHADOW.

Datasets
1 2 3 4 5 6 7 8
CCDP LMEBP CCDP CCDP CCDP CCDP CCDP CCDP
RALBGC LPQ RALBGC BGC1 RALBGC RALBGC LMEBP RALBGC
LDTP LESTP LDTP nLBPd LDTP LMEBP LESTP SLGS
BGC1 SLGS LESTP RALBGC LPQ LECTP LECTP ELGS
LBP LDTP ELGS SLGS LTP LESTP LQP BGC1
nLBPd BGC3 LMEBP LQPAT ELGS SLGS RALBGC nLBPd
LTP CCDP DCLBP LDTP BGC1 ELGS LTP LGS
SMEPOP LTP BGC1 LPQ nLBPd LQP SLGS LDTP
DCLBP RALBGC BGC2 ELGS LQP LGS ELGS LBP
LESTP ELGS LBP LBP DCLBP LDTP dLBPα LQPAT
ELGS LQP dLBPα LGS LBP LPQ LGS LTP
LMEBP DCLBP nLBPd LTP LGS BGC1 BGC1 BGC2
LQP BGC1 QBP SMEPOP SLGS nLBPd nLBPd dLBPα
LQPAT nLBPd CS-LBP DCLBP LQPAT LTP LDTP LESTP
D-LBP dLBPα LTP BGC2 LBPV LBP SMEPOP LMEBP
LGS LQPAT LQP LQP LECTP dLBPα DCLBP SMEPOP
GLBP LBP ID-LBP LBPV SMEPOP SMEPOP BGC2 LPQ
BGC3 LGS LECTP BGC3 BGC2 DCLBP LBPV DCLBP
LPQ BGC2 LQPAT MMEPOP BGC3 BGC2 LBP LQP
SLGS SMEPOP SMEPOP LESTP MMEPOP LQPAT LQPAT LECTP
LECTP LECTP GLBP LMEBP LESTP BGC3 LPQ BGC3
BGC2 GLBP LGS LECTP LMEBP MMEPOP BGC3 LBPV
LBPV MMEPOP BGC3 GLBP GLBP GLBP MMEPOP GLBP
MMEPOP D-LBP LPQ D-LBP QBP D-LBP D-LBP MMEPOP
LNDP LBPV SLGS LNDP D-LBP LNDP GLBP D-LBP
CS-LBP CS-LBP MMEPOP ID-LBP LNDP LBPV CS-LBP ID-LBP
ID-LBP QBP D-LBP QBP ID-LBP QBP QBP LNDP
QBP ID-LBP LBPV dLBPα dLBPα CS-LBP LNDP CS-LBP
dLBPα LNDP LNDP CS-LBP CS-LBP ID-LBP ID-LBP QBP
CS LMP CS LMP CS LMP CS LMP CS LMP CS LMP CS LMP CS LMP

TABLE IV
R ANKING RESULTS ON EACH TESTED DATASET. T HE PROPOSED METHOD IS HIGHLIGHTED IN SHADOW.

[20] Xiaosheng, W., Junding, S. An effective texture spectrum descriptor. pearance description using dense sampling based local binary patterns.
In: Proceedings of the 5th International Conference on Information Asian Conference on Computer Vision. Springer Berlin Heidelberg,
Assurance and Security, (IAS 2009), Xi’an, China, vol. 2, pp. 361-364 2012. p. 375-388.
(2009). [26] Subrahmanyam M, Maheshwari R. P, Balasubramanian R. Local
[21] I. El khadiri, A. Chahi, Y. El merabet, Y. Ruichek and R. Touahni. Maximum Edge Binary Patterns: A New Descriptor for Image Retrieval
Local Directional Ternary Pattern: A New Texture Descriptor for and Object Tracking, Signal Processing, 2012, 92 (6), pp. 1467-1479.
Texture Classification. Computer Vision and Image Understanding, [27] Guo, Z., Zhang, L., Zhang, D. Rotation invariant texture classification
2018. [Link] using LBP variance (LBPV) with global matching. Pattern recognition,
[22] I. El khadiri, M. Kas, Y. El merabet, Y. Ruichek, R. Touahni. 2010, vol. 43, no 3, p. 706-719.
Repulsive-and-Attractive Local Binary Gradient Contours: New and [28] Guo, Z., Zhang, L., Zhang, D. A completed modeling of local binary
Efficient Feature Descriptors for Texture Classification. Information pattern operator for texture classification. IEEE Trans. Image Process.
Sciences, 2018. [Link] 19(6), 1657–1663 (2010).
[23] Zeng, H., Chen, J., Cui, X., Cai, C., Ma, K. K. Quad Binary Pattern [29] Abdullah, M. F. A., Sayeed, M. S., Muthu, K. S., Bashier, H. K.,
and Its Application in Mean-Shift Tracking. Neurocomputing, 2016. Azman, A., Ibrahim, S. Z. Face recognition with symmetric local graph
[24] Vipparthi, S. K., Murala, S., Nagar, S. K., Gonde, A. B. Local structure (slgs). Expert Syst Appl, 2014, vol. 41, no 14, p. 6131-6137.
Gabor maximum edge position octal patterns for image retrieval. [30] Sayeed, S., Yusof, I., Bashier, H. K., Hossen, J. Plant identification
Neurocomputing, 2015, vol. 167, p. 336-345. based on leaf shape and texture pattern using local graph structure
[25] Ylioinas, J., Hadid, A., Guo, Y., Pietikäinen, M. Efficient image ap- (LGS). Aust J Basic Appl Sci, 2013, vol. 7, no 11, p. 29-35.
[31] V. Ojansivu, J. Heikkila. Blur insensitive texture classification using
local phase quantization. Image and Signal Processing, Vol. 5099 of
Lecture Notes in Computer Science, 2008, pp. 236–243.
[32] Bashier, H. K., Hoe, L. S., Hui, L. T., Azli, M. F., Han, P. Y., Kwee,
W. K., Sayeed, M. S. Texture classification via extended local graph
structure. Optik-International Journal for Light and Electron Optics,
Vol.127, pp. 638-643, 2016.
[33] Vipparthi, S. K., Nagar, S. K. Local Extreme Complete Trio Pattern
for Multimedia Image Retrieval System. International Journal of Au-
tomation and Computing, 2015.
[34] Zeng, H., Wang, X., Gu, Y. Center Symmetric Local Multilevel Pattern
Based Descriptor and Its Application in Image Matching. International
Journal of Optics, 2016, vol. 2016.