Directional Order Topologies based Multi-Scale

Ternary Pattern for Texture Classification.

E.Rachdi
Joao Pessoa, Brasil
Email: [email protected]

Abstract—The local ternary pattern (LTP) is one of theim- performance. Second challenge of LTP is when the threshold
portant variants of the LBP method which has been pro-posed value is exactly the same as the difference between the central
to overcome the disadvantages of the latter, despite theimpres- pixel and the adjacent pixels, in this case the generated code
sive performance of the LTP operator, it suffers fromcertain
limitations. In this paper, a new texture classificationdescriptor of LTP will be zero. Third challenge is that LTP, suffers from
named Directional Order Topologies based Multiscale Ternary certain limitations, such as high dimensionality, results in a
Pattern (DOT-MTP) is proposed. In the DOT-MTP descriptor, longer computation time and may affect the accuracy of the
a directional topology of the order with twodynamic thresholds classification.To work around these LTP descriptor threshold
has been introduced and logical connectionsinspired by the Venn issues, it recently had a notable extension of LTP known as
diagram or (logic diagram),have been incorporated to reducethe
multiscale problem. Experiments are carried out on the well- the Ternary Enhanced Local Model (ILTP) cite ILTP. But
known Brodatz, KTH-TIPS2b, MondialMarmi, TC-00, TC-01, the large dimension problem observed in LTP remains the
CUReT, KTH-TIPS, Jerry Wu , BonnBTF, XUH, Kylberg, USP- same. obstacle. Other methods have also been proposed to
Tex and visTex databases . Compared to many advanced methods, improve the discrimination performance of LBP. However,
the proposeddescriptor achieves superior texture classification there are still some issues to be addressed to overcome the
performance
Keywords: Texture recognition, texture descriptors, LBP, di- dearth of LBP descriptors. Most variants of LBP are affected
rectional order topologies. by this lack attached to classical methods of information
extraction. as For example, El khadir et al proposed (O3S-
I. I NTRODUCTION MTP)[?] with a star oriented sampling structure based on a
In the field of texture analysis, the classification of textures multiscale ternary pattern for texture classification it inherits
is increasingly seen as a serious problem. It plays an important some drawbacks due to its very large size for example,Heikkila
role in many applications, such as image retrieval [?], image et al [?] proposed a centrally symmetrical LBP (CS-LBP)[?]
classification [?], texture classification [?], [?], object recogni- by comparing pairs of centrally symmetric pixels instead of
tion [?], image matching [?], remote detection [?], facial clas- comparing neighbors with central pixels. Liao et al presented
sification [?].owever, textures in the real world vary in rotation, the dominant LBP (DLBP) cite DLBP, in which the dominant
illumination, scale, and affine varieties as imaging conditions models were experimentally chosen among all the rotation
change.Extracting powerful features for texture categorization invariant models.Additionally, they may not be rugged for
is still a gauntlet in texture analysis.he difficulty inherent in rotation, lighting, scale variation and noise, etc. To overcome
extracting efficient texture features is that it is difficult to the above problems, a new descriptor named Directional Order
define a high compromise between classification precision and Topologies based Multi-scale Ternary Pattern (DOT-MTP), is
computational complexity.ver the years numerous approaches proposed for texture classification. In DOT-MTP the extraction
for texture classificationhad been proposed in the literature. of information was carried out according to a directional order
Later, more advanced approacheswere developed for texture topologies (An order topology is a specific topology that can
classification. Typically, Local binary pattern (LBP)[?] fs a be defined in one or more fully ordered sets) connected by
simpler and more effective method for texture classification. logical operators (and) inspired of the Venn diagram or logic
Due to its superior performance, LBP bypasses almost all other diagram, mentioned below see (fig :2) for more information
methods, it has been widely applied to texture and material is to reduce the dimension. mds August 26, 2021
recognition.Based on the original LBP, several varieties of
LBP variants have been proposed to improve its robustness, A. Local binary pattern (LBP) and those extensions
distinctiveness and applicability. To improve discrimination. The local binary pattern presented by Ojala et al [?] for
Local Ternary Mode (LTP) [?] has been introduced to over- textures representation, feature extraction, and then extended
come the sensitivity of LBP to central pixel noise. In LTP, its applicationsto finally include various domains is mainly
by assigning a threshold to the central pixel, the descriptor related to its simplicity, speed of processing and ease of im-
is quantized into three levels (-1, 0, and 1) and decomposed plementation.the LBP code is basically computed as follows:
into two upper and lower descriptors.However, one of the P−1
main challenges of LTP is that it is difficult to set the
X
LBPR,P = δ (gc − gp ) 2p (1)
correct threshold for a specific application in order to improve p=0
 where gc is the gray value of the central pixel,gp is the Represents the mean value of the whole neighbourhood
grayvalue of its neighbors evenly distributed on a circle of R 3×3. h(ILTPS ) and h(ILT PI ) are respectively The two
radius , and P is the number of its neighbors. ,and the function histograms of the descriptors (ILT PS ) and (ILT PI ) which
δ(, ) is defined as follows: are concatenated to form the future histogram h(ILTP) defined
( as follows:
1 si x≥ y
δ (x, y) = (2)
0 otherwise h(ILT P ) = h(ILT PI )||h(ILT PS ) (7)
B. central symmetrical local binary pattern (CS-LBP)
This defines a set of 2 × 29 possible ternary model codes for
Heikkilä et al proposed the central symmetric local binary
the two local ternary patterns (ILT PI and LT PI )
pattern(CS-LBP) [?], inspired by the local binary model
(LBP). Its principle is based on the calculation of the differ-
ence between the values of the pixels that are symmetrical with Although the texture descriptor, the central symmetrical
respect to the central pixel (two by two); i.e. the pixel values local binary pattern (CS-LBP) , has achieved remarkable pre-
are not compared to the central pixel, but to the pixels that cision in classification of textures, it inherits some drawbacks
are relatively symmetrical with respect to the central pixel. from the Local binary pattern (LBP). LBP is sensitive to
The operator (CS-LBP) generates only 16 different binary noise, which reduces its discriminating property. Likewise,
models, which is different from the local binary model (LBP), the Improved local ternary patterns (ILTP) is proposed to be
which generates 256 different binary models, and has several more robust to noise than LBP, however, the latter’s weakness
advantages such as tolerance to lighting changes, robustness may appear with both ILTP and LBP. In this article, a new
over flat image areas, and computational efficiency.lthough the modeling is completed to propose a new descriptor based
texture descriptor, (CS-LBC), has achieved remarkable preci- on the directional order topologies based multi-scale ternary
sion in classification of textures, it inherits some drawbacks pattern (DOT-MTP) to extract the potential texture information
from the Local Binary Model (LBP). LBP is sensitive to of each pixel of a given image, and explore new concepts that
noise, which reduces its discriminating .Figure 4.6 compares include both LBP, CS-LBP and ILTP concepts in order to
the computation of LBP and CS-LBP with eight neighbouring overcome these drawbacks.
pixels the CS-LBP code is basically computed as follows: - First step:Average of neighboring pixels

p/2−1
X
CS − LBPP,R,T = δ(gi − gi−(p/2) − T ) × 22 (3)
i=1

C. Improved Local Tertiary Patterns (ILTP)

The Improved local ternary patterns (ILTP) [?] is an exten-
sion of LTP, where each pixel in the neighborhood is limited
to the mean value of gray. Like LTP, the display is divided
into ILT PI lower and ILT PS upper. However, although
the ILTP descriptor has excellent performance and overcomes
the drawback of setting a fixed threshold, it also has some
limitations, such as high dimensionality, which will result in
longer computation time and may affect accuracy. classifica-
tion. The two descriptors ILT PI and ILT PS are encoded and
concatenated according to the following equations:
Fig. 1: The adjacent neighbor relationship for each of the 8 neighbors.
(top) the two sets of neighboring diagonal and anti diagonal pixels
P
X −1 -vetical and horisontal (in the middle) the four sets of directional
ILT PI (xc , yc ) = 2p ×ψ(gc −m−τ )+ ψ(m−gp −τ )×2p neighboring pixels in the center and (at the bottom) the four sets of
p=0
neighboring block pixels and their complementary pixels
(4)
In the directions kπ/4 or k=0,...,3 the following averages
P −1
X are constructed :
ILT Ps (xc , yc ) = 2p ×ψ(m−gc −τ )+ ψ(gp +S −τ )×2p
p=0 1
(5) νk = (Ic + Ik + Ik+4 ) (8)
3
Where
P −1
1 X
where k ∈ {0, 1, 2, 3}, and then the pixel averages of the
m = (gc + gP ) (6)
9 p=0
directions (0,π/2) and (π/4 ,3 × π/4) are calculated as shown
in fig:1 in (c) and (d) (see equations 5 and 6) to solve the ((Zmin ))= 7 which construct two subspaces SU and SL of
coplexity of the clacul in pose: R7 , which are the basis of the proposed method. Indeed to
3 reduce the size of the proposed operator and capture more
1 X
discriminating information, we use an approach which helps
x0 = νˆ0 = (2 × Ic + I2k ), x1 = ν1 , x2 = ν2 , xk = IK
6 us to build the new combined descriptor of several operators
k=0
(9) which are linked together by the logical operator ”and” (&)
Where k ∈ {0, 1, 2, 3}, Inspired by a Venn diagram see figure: 2.from two ordered
3 sets ( SU and SL ) provided with the topology of the order.
1 X
The mathematical description of this method is as follows:
y0 = νˆ1 = (2 × Ic + I2k ), y1 = ν1 , y2 = ν3 , xk = IK
6 we end four sets X, Y ,Zmax and Zmin in the following
k=0
(10) way:
Where k ∈ {4, 5, 6, 7},
8 X = {x0 , x1 , , x3 , x4 , x5 , x6 } (21)
1 X
z̄0 = z0 = m̄ = mean(IM ×N , z̄1 = z1 = m = (Ic + Ik )
9
k=0
(11) Y = {y0 , y1 , y2 , y3 , y4 , y5 , y6 } (22)
2 7
1 X 1 X
µ0 = (2 × Ic + Ik ); µ¯0 = ( Ik ), (12)
4 5 Zmax = {z̄0 , z̄1 , z̄2 , z̄3 , z̄4 , z̄5 , z̄6 } (23)
k=0 k=3

z̄2 = z2 = Imed (13)

4 Zmin = {z0 , z1 , z2 , z3 , z4 , z5 , z6 } (24)
1 X 1 X
µ1 = (2 × Ic + Ik ); µ¯1 = ( Ik ) (14)
4 5
k=2 k∈{0,1,5,6,7}
Smax = {(x ≥ Ic − τ1 )&(y ≥ Ic + τ1 )&(z̄ ≥ Ic + τ2 )}
z̄3 = max(µ0 , µ̄0 ), z̄4 = max(µ1 , µ̄1 ) (15)
(25)
z3 = min(µ0 , µ̄0 ), z4 = min(µ1 , µ̄1 ) (16)
6 Smin = {(x ≤ Ic − τ1 )&(y ≤ Ic + τ1 )&(z ≤ Ic + τ2 )}
1 X 1 X
µ2 = (2 × Ic + Ik ), µ¯2 = ( Ik ) (17) (26)
4 5
k=4 k∈{0,1,2,3,7} where the (xp ), (yp ), (z̄p ) and (zp ) are respectively the
8 5 elements of the sets X, Y, Zmax and Zmin for p = 0,
1 X 1 X ..., P-1 mentioned in equations 21, 22,23 and 24 above.
µ3 = (2 × Ic + Ik ), µ¯3 = ( Ik ) (18)
4 5 h((DOT − M T P )max ) and h((DOT − M T P )min ) are
k=6 k=1
respectively The two histograms of the descriptors (DOT −
z̄5 = max(µ2 , µ̄2 ), z̄6 = max(µ3 , µ̄3 ) (19)
M T P )max and DOT − M T P )min which are concatenated
z5 = min(µ2 , µ̄2 ), z6 = min(µ3 , µ̄3 ) (20) to form the future histogram h(DOT-MTP) This defines a set
of 27 possible ternary pattern codes for each local ternary
-- Second step: construction of the descriptor pattern (DOT − M T P )max and (DOT − M T P )min) ,
K(DOT −M T P )max = 27 and K(DOT −M T P )min = 27 .

KDOT −M T P = K(DOT −M T P )max + K(DOT −M T P )min

(27)
- First step:Average of neighboring pixels

In this research phase, we carried out extensive tests on

13 representative texture databases,widely used to prove the
efficiency and performance stability of the proposed DOT-
MTP operator.In addition, DOT-MTP was compared with
many newer and most promising advanced texture descriptors
Fig. 2: Venn diagrams which connect three sets by the operator ”and” and several CNN-based features to highlight its performance
to build a single ordered set (≥) improvements.Table 1 summarizes the evaluation methods.
Please note that some of these methods are implemented using
In order to encode the central pixel, we propose a dominant the original code available,for the other methods, we have
intensity order metric, for this we choose four sets X,Y ,Zmax used our own implementation and made improvements based
and Zmin composed of (pixels and pixel averages), such that on the corresponding article to achieve the same result as
cardinal (X) = cardinal ( (Y )) = cardinal ((Zmax )) =cardinal shown in the published article..to get the same output results as
 Table 2 that the designed DOT-MQP operator is clearly the
best performing descriptor compared to the state-of-the-art
methods evaluated, which confirm all the analysis of results
extracted from Table 3. Remarkably, the normalised number
of victories obtained by DOT-MTP is 0.88034 compared to
0.82479 with DNT-MQP (2nd superior), 0.68 with RALBGC
(3rd superior), 0.64103 with LETRIST (4th superior), etc. In
particular, considering the classification performance of DNT-
MQP (the best second descriptor) as a reference as well as
LETRIST (the best fourthdescriptor), the DOT-MTP texture
operator provides a strong improvement compared to the 13
texture data sets tested.
Fig. 3: The adjacent neighbor relationship for each of the 8 neighbors. b) Experiment2: Good results and stable performance
(top) the two sets of neighboring diagonal and anti diagonal pixels
-vetical and horisontal (in the middle) the four sets of directional Table 2 lists the average classification score of each test
neighboring pixels in the center and (at the bottom) the four sets of method and each texture database, as well as the overall
neighboring block pixels and their complementary pixels average performance (GAP) and standard deviation average
(Std average) of each method on all data sets. Table 3 shows
the ranking results based on the average accuracy of each data
given in the publishedpaper. The experiments herein follow the set. Based on these results reported in Tables 2 and 3, and
standard evaluationprotocol for each tested dataset (i.e.split- Figures 4 and 5 above we can easily deduce the following
sample validation)where 50% of the samples are randomly observations:
selected to be usedas the training set, and the remaining 50% Comparison results (
of the samples areregarded as the testing set. The samples The DOT-MTP descriptor provided as an alternative to the
of the test set arethen classified through the parameter-free LBP and ILTP texture operators retains the advantages of
nearest-neighborrule (1-NN) with L1-city block distance. We monotonic lighting changes and low complexity. In addition,
repeat eachexperiment 100 times to remove any bias related DOT-MTP also has other more important advantages, which
to the divi-sion of the dataset and the averaged results are will be discussed below. Compared to some modern advanced
consideredas estimated accuracies. In what follows, the texture texture descriptors listed in Table II, which have inherent flaws
datasets considered in the experiments are first presented and in LBP descriptors, DOT-MTP descriptors can describe the
the obtained experimental results are then discussed. texture characteristics of local images with less computational
complexity and small size in size. because of topology direc-
D. Texture dataset tional, thus building this newly proposed operator.
In an effort to verify the efficiency and performance stability - whether in Tables II and III or in Figure 4 and 5 It is
of the DOT-MTP, we performed extensive testing on 13 well- very clear to observe that descriptors such as CALP, DC
known texture databases, including Brodatz, KTH-TIPS2b,, and EULLTP, and LDV, show the poorest performance on
MondialMarmi, TC-00, TC-01, CUReT, KTH-TIPS, Jerry Wu, almost any dataset used. and on the ranking (Victories /
BonnBTF, XUH, Kylberg, USPTex and visTex databases (the comparisons) their measurements are respectively less than
same datasets used in [52], [63]),. These well-known data (0.08120 and 0.00855) compared to the other descriptors
sets for texture classification are selected to cover various tested. - Indeed, given the classification among the descriptors
characteristics of sample size, number of categories, and tested in each data set used, DNOT-MTP stands out as the
category uniformity related to perspective and scale. More best stable descriptor for the ten textured data sets: Brodatz,
information about each data set is summarized in the table:1 KTH-TIP2b, MondialMarmi, TC-00, KTH -TIPS, BonnBTF,
XUH, Kylberg and visTex, in databases (Brodatz, KTH-TIP
E. Comparative Assessment of Performance and BonnBTF) DOT-MTP reaches the max rank (100%),
a)Experiment 1: Results and statistical significance while it is close to 100% in both databases (TC00 (99.95%)
The main idea of this experiment is to use the ranking tech- and (XUH (99.89%))). In addition, it is part of the second
nology based on the Wilcoxon signature ranking test proposed best method in the database (TC-01 (99.09%)), In addition,
in [?] to statistically verify the ranking results obtained by it is also that it has reached the third best classification of
the design method of the high-level descriptors evaluated. the database (CUReT (95.20%), slightly different from the
This technique is applied to all pairwise combinations of 19 first descriptor LETRIST (99.56%), therefore, the proposed
evaluation texture descriptors considered out of the 13 texture descriptor DNT -MTP reached the second best ranking in
data sets used in this experiment. Table 2 gathers the ranking the database (JerryWu (99.77%)) with a slight difference
results obtained according to the standardised number of wins compared to the first LETRIST descriptor (100%), and at the
(number of wins / (number of data sets tested × (number of end the DOT-MQP descriptor reached the first rank of the
descriptors tested-1)).by each method evaluated on all data database (USPtex (91.4 7%)), while the proposed descriptor,
sets tested. It is easy to see from the results presented in the DNT-MTP descriptor ranks second (USPtex (90, 52%))
 No. Name C TS Challenge

1 BonnBTF 10 160 Images acquired under different combinations of illumination direction, imaging direction and surface rotation.
2 visTex 167 2672 Images acquired under different .
3 KTH-TIP2b 11 176 Image acquired under illumination, rotation, pose scale changes.
4 KTH-TIPS 10 40 Images captured under nine scales, nine illumination conditions and three poses.
5 TC-00 24 480 Images acquired under controlled and variable conditions of illumination, rotation and spatial resolution.
6 TC-01 24 2112 Some properties as TC-00

TABLE I: Image datasets considered in the experiments.

Ranking Texture descriptor and Ref Victories/ Dimension Year

1-NN comparisons

1 DOT-MTP Proposed 0,82456 256 2021

2 DNT-MQP [?] 0,76316 384 2020
3 LCCMSB [?] 0,73684 2046 2018
4 ARCSLBP [?] 0,64912 256 2018
5 RALBGC [?] 0,64035 1022 2017
6 O3S-MTP [?] 0,56140 4096 2020
7 LETRIST [?] 0,55263 413 2017
8 ILQP [?] 0,52632 1024 2019
9 ILTP [?] 0,45614 1024 2010
10 LTP [?] 0,42982 512 2007
11 LOOP [?] 0,41228 256 2018
12 LNIP [?] 0,37719 512 2019
13 LNDP [?] 0,35965 256 2019
14 LESTP [?] 0,35088 512 2016
15 LECTP [?] 0,30702 768 2016
16 LQPAT [?] 0,27193 512 2017
17 MRELBP [?] 0,25439 200 2016
18 CALP [?] 0,14912 192 2019
19 EULLTP [?] 0,04386 32 2019
20 LDV [?] 0,00877 256 2019

TABLE II: Summary of the texture descriptors used for the evaluation and results of the comparison with the proposed descriptor.

N Texture descriptor 1 2 3 4 5 6 GAP

1 DOT-MTP 99,89 80,26 95,28 100,00 99,43 99,94 95,80

2 DNT-MQP 99,25 78,22 95,12 100,00 99,70 99,97 95,38
3 RALBGC 98,36 75,51 93,25 100,00 99,01 99,83 94,33
4 LETRIST 100,00 64,31 90,08 100,00 98,51 99,92 92,14
5 O3S-MTP 99,46 71,34 91,95 99,95 98,71 99,64 93,51
6 ILTP 97,31 73,23 91,88 100,00 97,70 99,35 93,24
7 ILQP 98,72 75,39 93,39 100,00 97,57 99,63 94,12
8 LTP 94,51 72,91 91,31 99,90 98,80 99,18 92,77
9 LECTP 99,08 68,07 88,59 99,60 98,21 97,23 91,80
10 LESTP 99,38 67,93 88,80 100,00 98,54 97,14 91,96
11 MRELBP 98,79 62,81 90,51 97,85 97,71 98,99 91,11
12 LNDP 99,15 71,57 88,72 100,00 97,47 99,20 92,68
13 LOOP 99,30 68,62 85,80 100,00 98,57 99,44 91,95
14 LNIP 99,95 69,63 82,97 98,45 97,83 99,10 91,32
15 LQPAT 99,13 67,71 83,76 99,60 97,11 99,23 91,09
16 CALP 95,83 66,05 82,63 98,15 94,89 98,99 89,42
17 EULLTP 93,40 53,53 78,63 93,90 88,98 96,68 84,19
18 LDV 94,36 37,56 58,37 88,45 83,49 92,20 75,74
19 LCCMSB 99,72 74,93 93,32 100,00 99,40 99,95 94,55
28 ARCSLBP 99,49 72,67 93,23 100,00 98,88 99,72 94,00

TABLE III: Comparison results (%) of different texture representations on 6 data sets..

c) Experiments 3: Comparisons with the characteristics manner similar to local descriptors against all CNNs functions
of CNN tested in 10 of 13 datasets. It should be pointed out that
when DNT-MTP is not in the first place, even if it does
To thoroughly assess the effectiveness of DOT-MTP at not achieve the highest score, it can still achieve interesting
the highest level, we also used the preformed deep learning average performance (i.e. its accuracy is similar to that of
models ResNet50, ResNet101, AlexNet, VGG16 and VGG19 the previously ranked CNNs), but the advantage is DOT-
to compare its performance with well-known CNN functions. MTP is conceptually easier to implement and requires no
Some of them have varying layers. It should be noted that training. Please note that CNNs are best suited for places
the result of the CNN- evaluated function is considered a with high intra-class variability, while local descriptors are
function vector and its method of expansion is similar to best suited for homogeneous fine-grained textures with low
how the process function is usually used. It can be noted intra-class variability. In fact, when the images have similar
that DNT-MTP achieves the highest average performance in a
content and less patterns, the depth model tends to extract [14] E. Rachdi, et al. ”Directional Neighborhood Topologies Based Multi-
redundant characteristic values, which will make the classifier Scale Quinary Pattern for Texture Classification.” IEEE Access 8 (2020):
212233-212246.
perform poorly on those characteristics, and different models [15] Y. El merabet, Y. Ruichek, A. El idrissi, Attractive-and-repulsive center-
will provide almost the same performance. On the other hand, symmetric local binary patterns for texture classification, Eng. Appl. Artif.
when the image has complex content and different patterns Intell. 78 (2019) 158–172.
[16] T. Song, H. Li, F. Meng, Q. Wu, J. Cai, Letrist: locally encoded
(as in the USP text dataset), the deep features provide a better transform feature histogram for rotation-invariant texture classification,
representation than the local descriptors [70]. IEEE Trans. Circuits Syst. Video Technol. (2017).
[17] I. El khadiri, Ruichek, Y., D. Chetverikov, and R. Touahni. ”O3S-MTP:
II. C ONCLUSION Oriented star sampling structure based multi-scale ternary pattern for
texture classification.” Signal Processing: Image Communication (2020):
The proposed descriptor named DOT-MTP, compared and 115830.
tested respectively with several recent and advanced LBP-type [18] W. Yang, C. Sun, 2011. Face recognition using improved local texture
patterns. 9th World Congress on Intelligent Control and Automation
methods and 13 reference textured data sets, these comparisons (WCICA), IEEE, 2011, 48-51.
and ballasts have shown that the DOT-MTP descriptor, is con- [19] . Armi and S. Fekri-Ershad, “Texture image Classification based onim-
ceptually simple and easy to implement, is more decriminating proved local Quinary patterns,”Multimedia Tools Appl., vol. 78,pp. 1–24,
Feb. 2019.
than all the tested methods ie it captures more information [20] X. Tan, B. Triggs, Enhanced local texture feature sets for face recog-
thanks to its specific form its logical connections its directional nition under difficult lighting conditions, IEEE Trans. Image Process. 19
topology on these neighborhoods which also makes its small (6) (2010) 1635–1650.
[21] S.K. Vipparthi, S.K. Nagar, Local extreme complete trio pattern for
dimension compared to improved variant of the LTP type. multimedia image retrieval system, Int. J. Autom. Comput. 13 (5) (2016)
Future work will concern the study of the use of other well- 457–467.
known classifiers to increase the classification performance [22] L. Liu, S. Lao, P.W. Fieguth, Y. Guo, X. Wang, M. Pietikäinen, Median
robust extended local binary pattern for texture classification, IEEE Trans.
of the proposed model. In addition, the proposed texture Image Process. 25 (3) (2016) 1368–1381.
operator has potential for deployment in high-level complex [23] S.K. Vipparthi, S.K. Nagar, Local extreme complete trio pattern for
applications related to texture classification, including dynamic multimedia image retrieval system, Int. J. Autom. Comput. 13 (5) (2016)
457–467.
texture classification, subtraction of backgrounds in complex [24] . Verma and B. Raman, “Local neighborhood difference pat-
scenes, recognition objects and video-based face analysis, etc. tern: A newfeature descriptor for natural and texture image re-
trieval,”MultimediaTools Appl., vol. 77, pp. 1–24, May 2017.
R EFERENCES [25] . Chakraborti, B. McCane, S. Mills, and U. Pal, “LOOP descriptor:Local
optimal-oriented pattern,”IEEE Signal Process. Lett., vol. 25, no. 5,pp.
[1] Yu, J., Qin, Z., Wan, T., Zhang, X.: Feature integration analysis ofbag of 635–639, May 2018.
features model for image retrieval. Neurocomputing120,355–364 (2013) [26] P . Banerjee, A. K. Bhunia, A. Bhattacharyya, P. P. Roy, and S.
[2] anerji, S., Verma, A., Liu, C.: Cross disciplinary biometric sys-tems. LBP Murala,“Local neighborhood intensity pattern—A new texture feature
and Color Descriptors for Image Classification, pp.205–225. Springer, descriptorfor image retrieval,”Expert Syst. Appl., vol. 113, pp. 100–115,
Berlin (2012) Dec. 2018.
[3] Ledoux, A., Losson, O., Macaire, L.: Color local binary patterns:compact [27] S. Chakraborty, S. K. Singh, and P. Chakraborty, “Local quadru-
descriptors for texture classification. J Electron Imaging25(6), 061404 ple pattern:A novel descriptor for facial image recognition and re-
(2016) trieval,”Comput.Electr. Eng., vol. 62, pp. 92–104, Aug. 2017
[4] Yongsheng, D., et al.: Multi-scale counting and difference [28] S. Chakraborty, S. K. Singh, and P. Chakraborty, “Cascaded asym-
repre-sentation for texture classification. Vis. Comput. metriclocal pattern: A novel descriptor for unconstrained facial image
(2017).https://doi.org/10.1007/s00371-017-1415-4 recognitionand retrieval,”Multimedia Tools Appl., vol. 78, pp. 1–20, May
[5] Lowe, D.G.: Distinctive image features from scale-invariant key-points. 2019
Int. J. Comput. Vision60(2), 91–110 (2004 [29] . Ouslimani, A. Ouslimani, and Z. Ameur, “Rotation-invariant fea-
[6] Xuemei, H., Yan, D.: Image matching with an improved descrip- turesbased on directional coding for texture classification,”Neural Com-
torbased on SIFT. In: Proceedings Volume 10322, Seventh Interna- put.Appl., vol. 31, no. 10, pp. 6393–6400, Oct. 2019.
tional Conference on Electronics and Information Engineering, pp.1-7 [30] L. Kabbai, M. Abdellaoui, and A. Douik, “Image classification by
(2017).https://doi.org/10.1117/12.2265595 com-bining local and global features,”Vis. Comput., vol. 35, no. 5, pp.
[7] Romero, A., Gatta, C., Camps-Valls, G.: Unsupervised deep featureextrac- 679–693,May 2019
tion for remote sensing image classification. IEEE Trans.Geosci. Remote [31] X. Wu and J. Sun, “Face recognition based on multi-scale local direc-
Sens.54(3), 1349–1362 (2016) tionalvalue,”Multimedia Tools and Applications, vol. 79, pp. 1–17, Nov.
[8] Berbar, M.: Three robust features extraction approaches for facialgender 2019.
classification. Vis. Comput.30(1), 19–31 (2014) [32] Y. El merabet and Y. Ruichek. Local Concave-and-Convex Micro-
[9] Y. Xu, H. Ji, and C. Fermüller, “Viewpoint invariant texture descriptionus- Structure Patterns for texture classification. Pattern Recognition. Vol. 76,
ing fractal analysis,”Int. J. Comput. Vis., vol. 83, no. 1, pp. 85–100,Jun. 2018, pp. 303-322.
2009. [33] I. El Khadiri, M. Kas, Y. El Merabet, Y. Ruichek, and R.
[10] . Song, H. Li, F. Meng, Q. Wu, and B. Luo, “Exploring Touahni,“Repulsive-and-attractive local binary gradient contours: New
space–frequencyco-occurrences via local quantized patterns for texture and effi-cient feature descriptors for texture classification,”Inf. Sci., vol.
representation,”Pattern Recognit., vol. 48, no. 8, pp. 2621–2632, Aug. 467,pp. 634–653, Oct. 2018.
2015
[11] T. Ojala, M. Pietikäinen, D. Harwood, A comparative study of texture
measures with classification based on featured distributions, Pattern
Recognit. 29 (1) (1996) 51–59.
[12] M. Heikkilä et al “Description of interestregions with center-symmetric
local binary patterns,” in Proc. Comput. Vis., Graph. Image Process., vol.
4338. 2006, pp. 58–69.
[13] S. Liao, M. W. K. Law, and A. C. S. Chung, “Dominant local binarypat-
terns for texture classification,”IEEE Trans. Image Process., vol. 18,no.
5, pp. 1107–1118, May 2009.