Texture classification via attractive-and-repulsive

decoded gradient contours

1st I. El khadiri 2nd M. Kas
Laboratoire LASTID, Département de Physique Laboratoire LASTID, Département de Physique
Faculté des Sciences, Université Ibn Tofail, BP 133, 14000 Faculté des Sciences, Université Ibn Tofail, BP 133, 14000
Kénitra, Maroc Kénitra, Maroc
[email protected]

3rd Y. El merabet 4th Y. Ruichek

Laboratoire LASTID, Département de Physique Le2i FRE2005, CNRS, Arts et Métiers
Faculté des Sciences, Université Ibn Tofail, BP 133, 14000 Univ. Bourgogne Franche-Comté, UTBM, F-90010
Kénitra, Maroc Belfort, France
[email protected]

5th R. Touahni
Laboratoire LASTID, Département de Physique
Faculté des Sciences, Université Ibn Tofail, BP 133, 14000
Kénitra, Maroc

Abstract—This paper presents a novel LBP-like descriptor Recently, active researches in computer vision and pattern
for texture representation. The proposed method, referred to recognition applications focus on pattern based features, due
as attractive-and-repulsive decoded gradient contours (ARDGC), to their effectiveness and the ease of extracting them from
consists in dividing local features into two distinct categories,
attractive and repulsive decoded gradient contours (ADGC and an image. LBP based operators, due to their interesting per-
RDGC) thanks to the flexibility of attractive-repulsive charac- formance, have emerged as one of the most effective texture
teristics of pixels in a 3×3 grayscale image patch. Unlike some descriptors. The LBP descriptor developed by Ojala et al. [1],
existing methods which are based on pairwise comparison of although originally designed for texture analysis, has been
adjacent pixels, the essence of the proposed ARDGC model is successfully applied in many research areas. Indeed, its low
to encode the differences between local intensity values within
triplets of pixels, along a route traced along the periphery of the computational cost and its invariance to contrast changes,
3×3 square neighborhood. In order to increase the robustness made it attractive not only to texture recognition, but also to
of the proposed operator, a new triplet, formed in addition to many other areas of computer vision including outdoor scene
the central pixel, by the average local and average global gray analysis, medical image analysis, biomedical image analysis,
levels, is incorporated in the modeling of ADGC and RDGC . dynamic texture recognition, face description and recognition,
The final multi-scale attractive-and-repulsive decoded gradient
contours ARDGC descriptor is obtained by linear concatenation image retrieval, motion detection, object detection, fingerprint
of ADGC and RDGC. Extensive experimental results on nine matching and background subtraction, remote sensing, etc. The
representative texture databases show that the proposed ARDGC success of the LBP in various applications gave birth of a lot
descriptor demonstrates superior performance to 30 recent state- of LBP variants, which have been developed and continue
of-the-art LBP variants. to be proposed. Indeed, since Ojala’s work [1] and due to
Index Terms—LBP, Attractive-repulsive characteristics,
ARDGC, Feature extraction, Texture classification its efficiency and flexibility, the overall LBP-like philosophy
has proven very popular, and a great variety of LBP variants
have been proposed in the literature to improve discriminative
I. I NTRODUCTION
power, robustness, and applicability of LBP. The authors in [2]
Texture, which is ubiquitous in natural images, is an impor- proposed a family of operators termed binary gradient contours
tant feature of the appearance of object surfaces and is widely (BGC1, BGC2 and BGC3) which are based on pairwise
used in object surface description and classification. There are comparison of adjacent pixels belonging to one or more closed
two critical issues in texture analysis: feature extraction and route traced along the periphery of 3×3 square neighborhood.
classifier designation. Designing an effective texture feature Abusham et al. [3], inspired by the dominating set and the
has been the subject of a wide range of applications including, graph theory, proposed a new extension of LBP known as
texture classification, scene understanding, material classifica- local graph structure (LGS) for face recognition. Mohd et al.
tion, face detection and recognition, background subtraction, [4] proposed the concept of symmetrical local graph structure
pedestrian detection, etc. algorithm (SLGS), which took equalized advantage of the
pixels in the left side and right side of target pixel in LGS. 
Recently, the authors in [5] developed local binary patterns 1, if α1 >= α2
=(α1 , α2 ) = (2)
by neighborhoods (nLBPd ) where the comparison between 0, otherwise
the peripheral pixels is done with sequential neighbors and/or where Ip (p ∈ {0, 1, ..., P-1}) denotes the gray levels of the
inside neighbors defined by a distance parameter d. More peripheral pixels, P corresponds to the number of neighboring
recently, Chakraborty et al. proposed [6] local quadruple pixels (P=8) and =(·) is the Heaviside step function.
pattern (LQPAT) for facial image recognition and retrieval. The basic LBP method comes with limitations and disad-
LQPAT encodes relations amongst the neighbors in quadruple vantages: 1) even a small change in the input image would
space where two micro patterns are computed from the local cause a change in the LBP output; 2) LBP is being sensitive
relationships to form the LQPAT descriptor. The same authors to image rotation; 3) it loses local textural information due to
in [7] proposed center symmetric quadruple pattern (CSQP) its quantization procedure; 4) it is being highly sensitive to
for facial image recognition and retrieval. CSQP encodes noise due to its thresholding scheme.
larger neighborhood with optimal number of binary bits. Issam
et al. [8] proposed local directional ternary pattern (LDTP) for III. P ROPOSED METHOD
texture classification. The LDTP operator consists in encoding
The construction concepts of LBP, the family of binary
both contrast information and directional pattern features in
gradient contours (BGC1, BGC2 and BGC3) [2] and many
a compact way based on local derivative variations. LDTP
other LBP-like methods motivate us to propose attractive-and-
conveys valuable information about the nature of textures
repulsive decoded gradient contours for texture classification.
by capturing local structures using both LTP’s and LDP’s
In order to give a compact mathematical formulation of the
concepts simultaneously.
proposed model, it is convenient to define, beforehand, the
In this paper, keeping the simplicity and effectiveness of the ¯ and
attractive and repulsive binary thresholding indicators (ξ(·)
basic LBP, we designed conceptually simple and robust texture
ξ(·)) which constitute the backbone of the proposed method.
representation model referred to as attractive-and-repulsive ¯ and ξ(·) indicators is to partition all of
The essence of ξ(·)
decoded gradient contours (ARDGC) for texture modeling and
the triplets of pixels belonging to a closed path traced along
classification. The main advantages of the proposed model are:
the periphery of the 3×3 grayscale image patch (hence the
1) implementation simplicity; 2) low computational complex-
name contours) around the central pixel Ic into two categories,
ity; 3) free of tuning parameters setup; 4) high performance
attractive and repulsive patterns and this independently of
compared to old and recent state-of-the-art descriptors; 5)
the distribution of pixels around Ic . These patterns can be
considerable enhancement of both discriminative power of
viewed as a gray-scale local micro-patterns which describe
LPB-like descriptors and their robustness to small variations.
local undulation characteristics of the image. The attractive-
The remainder of this paper is organized as follows. Section ¯ and ξ(·)) are
and-repulsive binary thresholding indicators (ξ(·)
II briefly presents the basic LBP operator. Section III presents defined as below (cf. Eq. 3 and 4).
the proposed ARDGC descriptor. Comprehensive experimental
results and comparative evaluation are given in Section IV.
Section V concludes the paper and proposes some future ξ(α1 , α2 , α3 ) ==(α1 , α2 ) ⊗ =(α3 , α2 )
 
research directions. 1, if α1 >= α2 1, if α3 >= α2
= ⊗
0, otherwise 0, otherwise (3)

II. B RIEF REVIEW OF T RADITIONAL L OCAL B INARY 1, if α1 >= α2 and α3 >= α2
=
PATTERNS (LBP) 0, otherwise

The basic LBP descriptor, proposed by Ojala et al. [1] in

ξ(α1 , α2 , α3 ) ==(α1 , α2 ) ⊗ =(α3 , α2 )
1996, encodes the pixel-wise information in texture images.  
It consists in characterizing the spatial structure of a local 1, if α1 <= α2 1, if α3 <= α2
= ⊗
image patch by thresholding a 3×3 grayscale image patch 0, otherwise 0, otherwise (4)

with the value of the central pixel. The resulting values are 1, if α1 <= α2 and α3 <= α2
=
tuned into binary values (0 or 1) forming a code of 8 bits that 0, otherwise
is transformed into decimal number. Once the LBP code is
generated for each pixel in the image, a histogram is built to where =(α1 , α2 ) = =(α2 , α1 ) and ⊗ denotes multiplication
represent the texture image. Formally, given the gray value Ic operator.
of the central pixel in the image I, the kernel function fLBP (·)
of the LBP descriptor for a 3×3 square neighborhood is given Aiming at finding a code which is insensitive to noise and
as follows: invariant to monotonic gray scale transformation, and thus
improve robustness and discriminative power of the traditional
P−1 LBP descriptor and its variants, we add, in the modeling, the
comparison of the central pixel Ic to the average local gray
X
fLBP (x) = =(Ip , Ic ) × 2p (1)
p=0
level ℵ (cf. Eq. (5)) of the whole 3×3 neighborhood and the
average global gray level ℵ (cf. Eq. (6)) of the whole image of ADGC and RDGC descriptors through the complementary
I. information of both attractive and repulsive representations.
ARDGC operator produces 1022 (210 -2)) possible patterns in
P−1
1 X a 3×3 neighborhood.
ℵ= ( Ip + Ic ) (5)
P + 1 p=0
PM PN hARDGC = hADGC ∪ hRDGC (12)
α1 =1 α2 =1 I(α1 , α2 ) IV. E XPERIMENTS AND DISCUSSIONS
ℵ= (6)
M×N
In this section, we investigate the performance of the pro-
By using ξ(·) (cf. Eq. (3)) and ξ(·) (cf. Eq. (4)) indicators, posed ARDGC descriptor against the 30 recent state-of-the-art
two operators termed attractive and repulsive decoded gradient texture methods summarized in Table I. For that, nine publicly
contours (ADGC and RDGC) are built and expressed in the available datasets, shown in Table II, are used to highlight
following formal way: the improvement that our model provide. As regarding the
evaluation protocol, half of samples of each dataset were
P−1
X randomly selected for training and the remaining half for
fADGC (x) = (=(Ip , I(p+d) mod P ) ⊗ =(I(p+d) mod P+1 , ... testing. The images of the testing subset are then classified
p=0 through the nearest-neighbor rule (1-NN) with L1 distance
I(p+d) mod P )) × 2p + =(ℵ, Ic ) ⊗ =(ℵ, Ic ) × 2P -1 (cf. Eq. (13)). The procedure is repeated 100 times, each time
P−1
X with a new subdivision being conducted of the training and
= ξ(Ip , I(p+d) mod P , I(p+d) mod P+1 ) × 2p testing sets, and the accuracy is then averaged over all these
p=0 tests.
+ ξ(ℵ, Ic , ℵ) × 2P -1 X
(7) dL1 (hi , hj ) = |hi (k) − hj (k))| (13)
k
P−1
X where hi ={h1i ,h2i ,...,hK i } is the query feature vector and
fRDGC (x) = (=(Ip , I(p+d) mod P ) ⊗ =(I(p+d) mod P+1 , ... hj ={h1j ,h2j ,...,hK
j } the target feature vector.
p=0

I(p+d) mod P )) × 2p + =(ℵ, Ic ) ⊗ =(ℵ, Ic ) × 2P -1 Tables III and IV illustrate the obtained experimental re-
P−1
X sults where the first one reports the classification accuracies
= ξ(Ip , I(p+d) mod P , I(p+d) mod P+1 ) × 2p obtained by all tested methods on each texture dataset, while
p=0 the second one gathers the accuracy based ranking results on
+ ξ(ℵ, Ic , ℵ) × 2P -1 each dataset. The following findings can be drawn from the
(8) analysis of these two tables:
• It emerges from Table III that some methods like LDENP,
Given an N×M image I, ADGC and RDGC patterns can LNDP and many other descriptors produce the worst per-
be computed at each pixel through its 3×3 neighborhood. An formance on several tested datasets where their average
image can be characterized by the probability distribution of accuracy is always below 90%.
RDGC and ADGC patterns. Formally, the whole image I is • It is noteworthy that there is a significant performance
represented by ADGC and RDGC histogram vectors hADGC drop for all the tested descriptors on OuTeX TC 00013
and hRDGC (cf. Eq. (10) and (9): dataset (dataset 8 in Table III). Keep in mind that the
X classification is performed using 1-NN classifier, and
hADGC (λ) = ψ(fADGC (x)), λ) (9) more sophisticated classifiers like SVM should improve
x
X the performance.
hRDGC (λ) = ψ(fRDGC (x)), λ) (10) • It is also apparent from Tables III and IV that none of
x the tested state-of-the-art methods performs well over all
where λ ∈ [0, Nbins ], Nbins =210 -1 is the number of bins the tested datasets. Considering for example the LESTP
and the delta function ψ(·) is defined as below (cf. Eq. (11)): descriptor, it shows good performance on Brodatz (dataset
2), KTH-TIPS (dataset 3), OuTeX TC 00000 (dataset 6)
 and OuTeX TC 00001 (dataset 7) datasets, but performs
1, if α1 = α2 ; worse on the other tested datasets. The same remark
ψ(α1 , α2 ) = (11)
0, otherwise occurs for many other descriptors like SLGS, DRLBP,
To achieve a reliable representation of texture which is more LECTP.
robust, the two histogram vectors ADGC and RDGC are com- • Clearly, it can be observed that the proposed multi-scale
bined into hybrid distributions to form the final ARDGC model ARDGC, which exploits complementary information of
(cf. Eq. (12)). This hybrid texture description model is pow- fused descriptors (ADGC and RDGC) gets very promis-
erful because it permits to enhance the discriminative power ing classification results on almost all the tested datasets.
 No. Name Acronym Dim. Ref. Year
1 Local Binary Patterns LBP 28 [1] 1996
2 Binary Gradient Contours (1) BGC1 28 -1 [2] 2011
3 Binary Gradient Contours (2) BGC2 28 -1 [2] 2011
4 Binary Gradient Contours (3) BGC3 (24 -1)2 [2] 2011
5 Radial Difference LBP RDLBP 26 -5 [9] 2012
6 Local Maximum Edge Binary Patterns LMEBP 8.29 [10] 2012
7 Multi-Scale Densely Sampled Complete LBP DCLBP 2.(26 -5) [11] 2013
8 Local Graph Structure LGS 28 [3] 2013
9 Symmetric Local Graph Structure SLGS 28 [4] 2014
10 Adjacent Evaluation LBP AELBP 28 [12] 2015
11 Adjacent Evaluation LTP AELTP 28 [12] 2015
12 Sign Maximum Edge Position Octal Pattern SMEPOP 2*(28 -5)-1 [13] 2015
13 Magnitude Maximum Edge Position Octal Pattern MMEPOP 2*(28 -5)-1 [13] 2015
14 LBP by Neighborhoods nLBPd 28 [5] 2015
15 Directional LBP dLBPα 28 [5] 2015
16 eXtended Center-Symmetric LBP XCSLBP 24 [15] 2015
17 Local Extreme Complete Trio Pattern LECTP 3.28 [14] 2015
18 Local Extreme Sign Trio Patterns LESTP 29 [14] 2015
19 Median Robust Extended Local Binary Pattern MRELBP 200 [20] 2015
20 Difference Symmetric Local Graph Structure DSLGS 28 [24] 2015
21 Local Quantization Code Histogram LQCH 104 [21] 2016
22 Dominant Rotated Local Binary Patterns DRLBP 28 [22] 2016
23 Extended Local Graph Structure ELGS 29 [16] 2016
24 Quad Binary Pattern QBP 24 [17] 2016
25 Complete Eight Local Directional Patterns CELDP 28 [18] 2016
26 Local Neighborhood Difference Pattern LNDP 26 − 1 [26] 2017
27 Local Quadruple Pattern LQPAT 29 [6] 2017
28 Centre Symmetric Quadruple Pattern CSQP 28 [7] 2018
29 Local Diagonal Extrema Number Pattern LDENP 24 − 1 [23] 2018
30 Local Directional Ternary Pattern LDTP 2(29 − 1) [8] 2018

TABLE I
S UMMARY OF STATE - OF - THE - ART TEXTURE DESCRIPTORS TESTED AND COMPARED TO OUR PROPOSED DESCRIPTORS .

No. Name Classes Samples Total samples Sample resolution Image format Predefined
per class (pixels) train/test sets?
1 CUReT 61 92 5612 200×200 Monochrome (PNG) No
2 Brodatz 13 16 208 256×256 Monochrome (TIFF) No
3 KTH-TIPS 10 4 40 100×100 Monochrome (PNG) No
4 KTH-TIPS2b 11 16 176 100×100 Colour (PNG) No
5 MondialMarmi 12 64 768 136×136 Colour (BMP) No
6 OuTeX TC 00000 24 20 480 128×128 Monochrome (RAS) Yes
7 OuTeX TC 00001 24 88 2112 64×64 Monochrome (RAS) Yes
8 OuTeX TC 00013 68 20 1360 128×128 Colour (BMP) No
9 USPTex 191 12 2292 128×128 Colour (PNG) No

TABLE II
I MAGE DATASETS CONSIDERED IN THE EXPERIMENTS .

• Considering the ranking between the tested descriptors that the hybrid distribution turns out to be a powerful repre-
within each dataset (cf. Table IV), the best method sentation of image texture and robust to gray scale variations.
in terms of accuracy is ARDGC which performs sig-
nificantly and consistently the best on eight datasets V. C ONCLUSION
(i.e., CUReT, Brodatz, KTH-TIPS, KTH-TIPS2b, Mon-
dialMarmi, OuTeX TC 00001, OuTeX TC 00013 and In this paper, based on attractive-and-repulsive binary
USPTex datasets). Keep in mind that for the other dataset thresholding indicators, two distinct operators are built by
(i.e., OuTeX TC 0000 dataset), even if ARDGC operator dividing local features into two distinct parts, i.e., attractive
is ranked at the 4th position, it allowed, as shown in Table and repulsive decoded gradient contours (ADGC and RDGC),
III (column 6), to achieve a score of 99.83%, which is according to the relationship defined between the pixels around
considered as satisfactory classification result as it is close the central pixel of 3×3 grayscale image patch. More pre-
to the rate of the descriptors ranked before it (99.90% for cisely, ADGC and RDGC, are based on comparison within
the top 1). triplets of adjacent pixels belonging to closed route traced
along the periphery of the 3×3 grayscale image patch. The
In light of these findings, one can state that the proposed discriminative power of ADGC and RDGC has been enriched
LDTP descriptor presents a significant performance stability by incorporating the triplet formed, in addition to the central
over the evaluated state-of-the-art methods on all the tested pixel, by the average local and global gray levels in the mod-
datasets. The performance of the ARDGC descriptor for tex- eling. The histograms from the single scale analysis realized
ture classification on the nine tested datasets clearly highlights with RDGC and ADGC are concatenated together to build the
 Dataset
Texture descriptor 1 2 3 4 5 6 7 8 9
ARDGC 93.82 100.00 100.00 93.25 90.51 99.83 99.22 82.69 86.94
LDTP 91.81 100.00 100.00 90.47 86.00 99.76 99.20 80.32 83.00
LNDP 72.66 98.47 96.15 77.75 70.79 96.74 88.52 70.06 62.93
LQPAT 90.46 99.93 99.60 83.76 81.58 99.81 97.67 76.89 77.13
LDENP 64.82 95.40 88.80 65.80 57.87 95.12 81.66 63.61 44.99
CSQP 88.51 99.93 100.00 80.12 74.87 99.10 96.92 74.90 72.61
MRELBP 92.90 99.94 97.85 90.51 86.35 98.90 97.10 82.02 83.33
CELDP 70.71 98.37 100.00 85.73 83.28 97.71 95.68 77.42 74.43
DSLGS 93.76 99.71 97.75 85.17 81.20 99.83 98.31 81.06 79.98
LESTP 89.64 99.98 100.00 88.80 76.17 97.69 94.41 77.93 82.49
LECTP 86.71 99.65 99.60 88.59 80.76 97.55 97.48 76.26 81.09
LGS 92.60 99.92 99.35 84.19 81.97 99.67 98.40 80.10 81.13
SLGS 93.76 99.71 97.75 85.17 81.20 99.83 98.31 81.06 79.98
DRLBP 90.77 100.00 98.40 89.41 76.67 97.78 95.66 78.13 79.83
AELBP 88.72 99.30 99.05 79.64 76.49 99.57 96.06 77.46 73.24
ELGS 92.91 99.97 100.00 88.03 86.55 99.68 98.73 81.05 84.19
XCSLBP 77.42 97.93 95.00 85.41 73.02 95.10 93.16 52.28 61.67
LMEBP 89.33 99.97 100.00 87.81 76.78 97.58 93.60 76.00 82.17
MMEPOP 79.68 98.98 97.35 81.72 73.04 98.85 95.18 69.08 70.36
SMEPOP 89.28 99.99 99.45 86.08 77.10 99.61 97.48 76.92 78.53
RDLBP 89.31 99.98 99.70 86.87 82.30 99.40 97.98 77.33 79.32
DCLBP 87.86 99.99 100.00 89.00 89.35 99.22 98.47 77.91 81.26
BGC1 92.67 100.00 100.00 90.38 82.37 99.90 98.53 79.13 82.28
BGC2 89.96 99.39 100.00 82.84 75.36 99.14 96.68 76.80 78.31
BGC3 86.34 99.82 99.10 84.72 76.84 98.87 96.31 75.88 78.07
LBP 90.71 100.00 100.00 89.67 83.28 99.68 98.44 78.22 81.43
dLBPα 89.81 96.13 100.00 78.69 65.71 95.19 88.28 72.37 58.74
nLBPd 92.67 100.00 100.00 90.24 82.32 99.90 98.53 79.12 82.30
QBP 66.50 97.15 100.00 87.85 76.06 95.30 91.33 61.07 63.11
LQCH 86.61 98.44 100.00 90.51 82.83 96.54 94.69 76.34 79.28
AELTP 90.64 99.89 98.50 88.49 82.47 99.85 98.95 79.81 84.08
TABLE III
AVERAGE CLASSIFICATION ACCURACY. T HE PARAMETRIC METHODS AND THE PROPOSED METHODS ARE HIGHLIGHTED , RESPECTIVELY, IN LIGHT GRAY
AND SHADOW.

Datasets
1 2 3 4 5 6 7 8 9
ARDGC ARDGC ARDGC ARDGC ARDGC BGC1 ARDGC ARDGC ARDGC
DSLGS LDTP CSQP MRELBP DCLBP nLBPd LDTP MRELBP ELGS
SLGS DRLBP LDTP LQCH ELGS AELTP AELTP DSLGS AELTP
ELGS BGC1 CELDP LDTP MRELBP ARDGC ELGS SLGS MRELBP
MRELBP LBP LESTP BGC1 LDTP DSLGS BGC1 ELGS LDTP
BGC1 nLBPd ELGS nLBPd CELDP SLGS nLBPd LDTP LESTP
nLBPd SMEPOP LMEBP LBP LBP LQPAT DCLBP LGS nLBPd
LGS DCLBP DCLBP DRLBP LQCH LDTP LBP AELTP BGC1
LDTP LESTP BGC1 DCLBP AELTP ELGS LGS BGC1 LMEBP
DRLBP RDLBP BGC2 LESTP BGC1 LBP DSLGS nLBPd LBP
LBP ELGS LBP LECTP nLBPd LGS SLGS LBP DCLBP
AELTP LMEBP dLBPα AELTP RDLBP SMEPOP RDLBP DRLBP LGS
LQPAT MRELBP nLBPd ELGS LGS AELBP LQPAT LESTP LECTP
BGC2 LQPAT QBP QBP LQPAT RDLBP LECTP DCLBP DSLGS
dLBPα CSQP LQCH LMEBP DSLGS DCLBP SMEPOP AELBP SLGS
LESTP LGS RDLBP RDLBP SLGS BGC2 MRELBP CELDP DRLBP
LMEBP AELTP LQPAT SMEPOP LECTP CSQP CSQP RDLBP RDLBP
RDLBP BGC3 LECTP CELDP SMEPOP MRELBP BGC2 SMEPOP LQCH
SMEPOP DSLGS SMEPOP XCSLBP BGC3 BGC3 BGC3 LQPAT SMEPOP
AELBP SLGS LGS DSLGS LMEBP MMEPOP AELBP BGC2 BGC2
CSQP LECTP BGC3 SLGS DRLBP DRLBP CELDP LQCH BGC3
DCLBP BGC2 AELBP BGC3 AELBP CELDP DRLBP LECTP LQPAT
LECTP AELBP AELTP LGS LESTP LESTP MMEPOP LMEBP CELDP
LQCH MMEPOP DRLBP LQPAT QBP LMEBP LQCH BGC3 AELBP
BGC3 LDNP MRELBP BGC2 BGC2 LECTP LESTP CSQP CSQP
MMEPOP LQCH DSLGS MMEPOP CSQP LDNP LMEBP dLBPα MMEPOP
XCSLBP CELDP SLGS CSQP MMEPOP LQCH XCSLBP LDNP QBP
LDNP XCSLBP MMEPOP AELBP XCSLBP QBP QBP MMEPOP LDNP
CELDP QBP LDNP dLBPα LDNP dLBPα LDNP LDENP XCSLBP
QBP dLBPα XCSLBP LDNP dLBPα LDENP dLBPα QBP dLBPα
LDENP LDENP LDENP LDENP LDENP XCSLBP LDENP XCSLBP LDENP
TABLE IV
R ANKING RESULTS ON EACH TESTED DATASET. T HE PROPOSED METHODS ARE HIGHLIGHTED IN BOLD AND SHADOW.

multi-scale histogram feature vector referred to as attractive- R EFERENCES

and-repulsive decoded gradient contours (ARDGC) model. [1] T. Ojala, M. Pietikainen, T. Maenpaa. Multiresolution gray-scale and
The proposed ARDGC model was evaluated comprehensively rotation invariant texture classification with local binary patterns.IEEE
on nine challenging texture datasets against a total of 30 recent Trans. Pattern Anal. Mach. Intell., 24(7), 971-987, 2002.
[2] A. Fernández, M.X. Álvarez, F. Bianconi. Image classification with
state-of-the-art methods. Experimental results showed that the binary gradient contours. Optics and Lasers in Engineering, 49(9), 2011,
proposed descriptor has high discriminative power to classify 1177-1184.
texture images with better classification accuracy than the [3] S. Sayeed, I. Yusof, H. K. Bashier, J. Hossen. Plant identification based
on leaf shape and texture pattern using local graph structure (LGS).
evaluated methods. As future work, we wish to investigate Aust. J. Basic & Appl. Sci, 2013, vol. 7, no 11, p. 29-35.
high–level challenging applications such as object tracking, [4] M. F. A. Abdullah, M. S. Sayeed, K. S. Muthu, H. K. Bashier, A. Azman,
face detection and object recognition. S. Z. Ibrahim. Face recognition with symmetric local graph structure
(slgs). Expert Systems with Applications, 41(14), 2014, 6131-6137.
 [5] K. Yılmaz, Ö. F. Ertuğrulb, R. Tekinc. Two novel local binary pattern
descriptors for texture analysis. Applied Soft Computing, vol. 34, 2015,
pp. 728-35.
[6] S. Chakraborty, S. K. Singh, P. Chakraborty. Local quadruple pattern: A
novel descriptor for facial image recognition and retrieval. Computers
& Electrical Engineering, 2017, 62, 92-104.
[7] Chakraborty, S., Singh, S. K., Chakraborty, P. (2018). Cen-
tre Symmetric Quadruple Pattern: A Novel Descriptor for Fa-
cial Image Recognition and Retrieval. Pattern Recognition Letters.
https://doi.org/10.1016/j.patrec.2017.10.015
[8] I. El khadiri, A. Chahi, Y. El merabet, Y. Ruichek and R. Touahni.
Local Directional Ternary Pattern: A New Texture Descriptor for Texture
Classification. Computer Vision and Image Understanding, 2018.
[9] L. Liu, L. Zhao, Y. Long, G. Kuang, P. Fieguth. Extended local binary
patterns for texture classification. Image and Vision Computing, 2012,
vol. 30, no 2, p. 86-99.
[10] M. Subrahmanyam, R. P. Maheshwari, R. Balasubramanian. Local
Maximum Edge Binary Patterns: A New Descriptor for Image Retrieval
and Object Tracking. Signal Processing, 2012, 92 (6), pp. 1467-1479.
[11] J. Ylioinas, A. Hadid, Y. Guo, M. Pietikäinen. Efficient image appear-
ance description using dense sampling based local binary patterns. 11th
Asian conference on Computer Vision, (pp. 375-388), 2012, Springer-
Verlag.
[12] K. Song, Y. Yan, Y. Zhao, C. Liu. Adjacent evaluation of local binary
pattern for texture classification. Journal of Visual Communication and
Image Representation, 2015, vol. 33, p. 323-339.
[13] S. K. Vipparthi, S. Murala, S. K. Nagar, A. B. Gonde. Local Gabor max-
imum edge position octal patterns for image retrieval. Neurocomputing,
2015, vol. 167, p. 336-345.
[14] S. K. Vipparthi, S. K. Nagar. Local Extreme Complete Trio Pattern for
Multimedia Image Retrieval System. International Journal of Automa-
tion and Computing, 2015.
[15] C. Silva, T. Bouwmans, C. Frélicot. An extended center-symmetric local
binary pattern for background modeling and subtraction in videos. Inter-
national Joint Conference on Computer Vision, Imaging and Computer
Graphics Theory and Applications, 2015.
[16] H. K. Bashier, L. S. Hoe, L. T. Hui, M. F. Azli, P. Y. Han, W. K. Kwee,
M. S Sayeed. Texture classification via extended local graph structure.
Optik, vol.127, pp. 638-643, 2016.
[17] H. Zeng, J. Chen, X. Cui, C. Cai, K. K. Ma. Quad Binary Pattern and
Its Application in Mean-Shift Tracking. Neurocomputing, 2016, 3-10.
[18] M. R. Faraji, X. Qi. Face recognition under varying illuminations using
logarithmic fractal dimension-based complete eight local directional
patterns. Neurocomputing, 199, 2016, 16-30.
[19] F. Bianconi, E. González, A. Fernández. Dominant local binary patterns
for texture classification: Labelled or unlabelled?. Pattern Recogn Lett ,
2015, vol. 65, p. 8-14.
[20] L. Liu, S. Lao, P.W. Fieguth, Y. Guo, X. Wang, M. Pietikäinen. Median
robust extended local binary pattern for texture classification. IEEE
Transactions on Image Processing, 25(3), 1368-1381, 2016.
[21] Y. Zhao, R. G. Wang, M. W. Wang, W. Gao. Local quantization code
histogram for texture classification. Neurocomputing, 207, 354-364,
2016.
[22] R. Mehta, K. Egiazarian. Dominant Rotated Local Binary Patterns
(DRLBP) for texture classification. Pattern Recogn Lett , vol. 71, 2016,
pp. 16-22.
[23] Pillai, A., Soundrapandiyan, R., Satapathy, S., Satapathy, S. C., Jung,
K. H., Krishnan, R. (2018). Local diagonal extrema number pattern: A
new feature descriptor for face recognition. Future Generation Computer
Systems, 81, 297-306.
[24] S. Dong, J. Yang, C. Wang, Y. Chen, D. Sun. A New Finger Vein
Recognition Method Based on the Difference Symmetric Local Graph
Structure (DSLGS). IJSIP, vol.8, No.10 (2015), pp.71-80.
[25] X. Chen, Z. Zhou, J. Zhang, Z. Liu, Q. Huang. Local convex-and-
concave pattern: An effective texture descriptor. Information Sciences,
363, 2016, 120-139.
[26] M. Verma, B. Raman. Local neighborhood difference pattern: A new
feature descriptor for natural and texture image retrieval. Multimedia
Tools and Applications, 1-24, 2017.