Directional Order Topologies based Multi-Scale Ternary Pattern for Texture

Classification

E. Rachdi1 , Y. El merabet1,, R. Messoussi1,

a Laboratoire LASTID, Département de Physique, Faculté des Sciences, Université Ibn Tofail, BP 133, 14000 Kénitra, Maroc

Abstract

In recent decades, the LBP method has received more attention from researchers, due to its distinctive
advantages notammet its simplicity, its good invariability to monotonic gray level changes and iit is suitable
for real-time applications with low computational costs.Despite these advantages, the base LBP descriptor
has some limitations for this purpose, many LBP based descriptors have been proposed to improve its
performance.The local ternary pattern (LTP) is one of the variants important aspects of LBP which has
been proposed to overcome the disadvantages of LBP. However, despite the impressive performance of LTP,
it also suffers from some limitations, such as high dimensionality, which results in longer computation time
and can affect classification accuracy, and a second challenge of LTP is related to difficulty of defining the
fixed threshold for a specific application in order to improve these performances.In this paper, a new texture
classification descriptor named Directional Order Topologies based on the Ternary Multiscale Model (DOT-
MTP) is proposed. In the proposed DOT-MTP descriptor, a neighborhood directional topology is built
into the original LTP. Then, to overcome the difficulty of the fixed threshold, two dynamic thresholds were
introduced and in addition a logical connection inspired by the Venn diagram was integrated to reduce the
multiscale problem seen in the DNT-MQP descriptor.This reduction has had an impact. positive on the
computational complexity of the descriptor and the dimensionality of the resulting histogram. The proposed
DOT-MTP is evaluated using 13 textured data sets and compared to some recent and well-known descriptors.
Experimental results show that DOT-MTP outperformed other descriptors exceeds in classification accuracy
and stability
Keywords: LBP, CS-LBP, LTP, ILTP, O3S-MTP, Feature extraction, Texture classification

1. Introduction

In the field of texture analysis, the classification of textures is increasingly seen as a serious problem.
It plays an important role in many applications, such as image retrieval [? ], image classification [? ],
texture classification [? ? ], object recognition [? ], image matching [? ], remote detection [? ], facial
5 classification [? ].owever, textures in the real world vary in rotation, illumination, scale, and affine varieties as
imaging conditions change.Extracting powerful features for texture categorization is still a gauntlet in texture
analysis.he difficulty inherent in extracting efficient texture features is that it is difficult to define a high
compromise between classification precision and computational complexity.ver the years numerous approaches
for texture classificationhad been proposed in the literature. Later, more advanced approacheswere developed
10 for texture classification. Typically, Local binary pattern (LBP)[? ] fs a simpler and more effective method

∗ Corresponding author

Preprint submitted to Engineering Applications of Artificial Intelligence January 27, 2021

 for texture classification. Due to its superior performance, LBP bypasses almost all other methods, it has been
widely applied to texture and material recognition.Based on the original LBP, several varieties of LBP variants
have been proposed to improve its robustness, distinctiveness and applicability. To improve discrimination.
Local Ternary Mode (LTP) [? ] has been introduced to overcome the sensitivity of LBP to central pixel noise.
15 In LTP, by assigning a threshold to the central pixel, the descriptor is quantized into three levels (-1, 0, and
1) and decomposed into two upper and lower descriptors.However, one of the main challenges of LTP is that
it is difficult to set the correct threshold for a specific application in order to improve performance. Second
challenge of LTP is when the threshold value is exactly the same as the difference between the central pixel
and the adjacent pixels, in this case the generated code of LTP will be zero. Third challenge is that LTP,
20 suffers from certain limitations, such as high dimensionality, results in a longer computation time and may
affect the accuracy of the classification.To work around these LTP descriptor threshold issues, it recently had
a notable extension of LTP known as the Ternary Enhanced Local Model (ILTP) cite ILTP. But the large
dimension problem observed in LTP remains the same. obstacle. Other methods have also been proposed
to improve the discrimination performance of LBP. However, there are still some issues to be addressed
25 to overcome the dearth of LBP descriptors. Most variants of LBP are affected by this lack attached to
classical methods of information extraction. as For example, El khadir et al proposed (O3S-MTP)[? ] with
a star oriented sampling structure based on a multiscale ternary pattern for texture classification it inherits
some drawbacks due to its very large size for example,Heikkila et al [? ] proposed a centrally symmetrical
LBP (CS-LBP)[? ] by comparing pairs of centrally symmetric pixels instead of comparing neighbors with
30 central pixels. Liao et al presented the dominant LBP (DLBP) cite DLBP, in which the dominant models
were experimentally chosen among all the rotation invariant models.Additionally, they may not be rugged
for rotation, lighting, scale variation and noise, etc. To overcome the above problems, a new descriptor
named Directional Order Topologies based Multi-scale Ternary Pattern (DOT-MTP), is proposed for texture
classification. In DOT-MTP the extraction of information was carried out according to a directional order
35 topologies (An order topology is a specific topology that can be defined in one or more fully ordered sets)
connected by logical operators (and) inspired of the Venn diagram mentioned below see (fig :1 and 2) for
more information is to reduce the dimension.

2. PROPOSED DESCRIPTOR

Texture extraction is a major task in computer vision and pattern recognition. In recent years, the rapid
40 development of computer and technological fields has considerably accelerated the method of classifying
textures. In this article, we summarize some representative methods of existing textures. Ojala et al.
originally offered the local binary model (LBP)[? ]. This is a non-parameterizable model that extracts
local structure information from the texture image, based on the difference between the central pixel and
adjacent pixels. Due to its simplicity and ease of use, LBP has grown rapidly over the past decades. However,
45 traditional LBP has two main drawbacks: 1) It is sensitive to image rotation and noise. 2) It cannot adapt to
changes in scale and lighting. To alleviate these problems, many variations of the LBP functionality have been
developed. The LBP has been offered for 20 years [? ? ]. They are divided into four main categories: rotation
invariant functions, noise tolerance functions, scale invariant functions and robust brightness functions

2
 2.1. Local binary pattern (LBP) and those extensions

The local binary pattern presented by Ojala et al [? ] for textures representation, feature extraction, and
then extended its applicationsto finally include various domains is mainly related to its simplicity, speed of
processing and ease of implementation.the LBP code is basically computed as follows:
P−1
X
LBPR,P = δ (gc − gp ) 2p (1)
p=0

50 where gc is the gray value of the central pixel,gp is the grayvalue of its neighbors evenly distributed on a
circle of R radius , and P is the number of its neighbors. ,and the function δ(, ) is defined as follows:

1

si x≥ y
δ (x, y) = (2)
0

otherwise

2.2. central symmetrical local binary pattern (CS-LBP)

Heikkilä et al proposed the central symmetric local binary pattern(CS-LBP) [? ], inspired by the local
binary model (LBP). Its principle is based on the calculation of the difference between the values of the pixels
55 that are symmetrical with respect to the central pixel (two by two); i.e. the pixel values are not compared
to the central pixel, but to the pixels that are relatively symmetrical with respect to the central pixel. The
operator (CS-LBP) generates only 16 different binary models, which is different from the local binary pattern
(LBP), which generates 256 different binary models, and has several advantages such as tolerance to lighting
changes, robustness over flat image areas, and computational efficiency.lthough the texture descriptor, (CS-
60 LBC), has achieved remarkable precision in classification of textures, it inherits some drawbacks from the
Local binary pattern (LBP). LBP is sensitive to noise, which reduces its discriminating .Figure 4.6 compares
the computation of LBP and CS-LBP with eight neighbouring pixels the CS-LBP code is basically computed
as follows:

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

2.3. Improved Local Tertiary Patterns (ILTP)

65 The Improved local ternary patterns (ILTP) [? ] is an extension 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. classification. The two descriptors ILT PI and
70 ILT PS are encoded and concatenated according to the following equations:

P
X −1
ILT PI (xc , yc ) = 2p × ψ(gc − m − τ ) + ψ(m − gp − τ ) × 2p (4)
p=0

P
X −1
ILT Ps (xc , yc ) = 2p × ψ(m − gc − τ ) + ψ(gp + S − τ ) × 2p (5)
p=0

3
 Where
P −1
1 X
m= (gc + gP ) (6)
9 p=0

Represents the mean value of the whole neighbourhood 3×3. h(ILTPS ) and h(ILT PI ) are respectively The
two histograms of the descriptors (ILT PS ) and (ILT PI ) which are concatenated to form the future histogram
h(ILTP) defined as follows:
h(ILT P ) = h(ILT PI )||h(ILT PS ) (7)

This defines a set of 2 × 29 possible ternary model codes for the two local ternary models (ILT PI and LT PI

Although the texture descriptor, the central symmetrical local binary pattern (CS-LBP) , has achieved
remarkable precision in classification of textures, it inherits some drawbacks from the Local binary pattern
75 (LBP). LBP is sensitive to noise, which reduces its discriminating property. Likewise, the Improved local
ternary patterns (ILTP) is proposed to be more robust to noise than LBP, however, the latter’s weakness may
appear with both ILTP and LBP. In this article, a new modeling is completed to propose a new descriptor
based on the directional order topologies based multi-scale ternary pattern (DOT-MTP) to extract the
potential texture information of each pixel of a given image, and explore new concepts that include both
80 LBP, CS-LBP and ILTP concepts in order to overcome these drawbacks.

3. presentation of the proposed descriptor

Figure 1: The adjacent neighbor relationship for each of the 8 neighbors. (top) the two sets of neighboring diagonal and anti
diagonal pixels -vetical and horisontal (in the middle) the four sets of directional neighboring pixels in the center and (at the
bottom) the four sets of neighboring block pixels and their complementary pixels

Then, the νˆ0 , νˆ1 , and νk where k=0,...,3 respectively Shown in red in (c),(d), (e), (f), (g) and (h) in (fig:1)
, are used to construct the averages of the pixels according to the directions in the directions kπ/4 where
k=0,...,3 , νˆ0 and νˆ1 the mean respectively of (ν0 ,ν2 ) resp (ν1 ,ν3 ) (see the eq: 4, 5 and 6),while the averages of
85 the pixel values of the four blocks µ0 , µ1 , µ2 and µ3 including the central pixel Ic as well as their complements
µ¯0 , µ¯1 , µ¯2 and µ¯3 , in the local image I3×3 , they are respectively represented in red resp in blue in ( (i), (j),
(k) and (h) in fig: 1) and their means are calculated as mentioned in the equations (.....) the average values
of the pixels of the local image m indicated in (b) in (fig: 1) and the average values of the pixels of the global
image noted m̄, the local median noted Imed are all the tools that were used to mathematically construct
90 the new descriptor indicated in the following steps:

First step:Average of neighboring pixels

In the directions kπ/4 or k=0,...,3 the following averages are constructed :
1
νk = (Ic + Ik + Ik+4 ) (8)
3
where k ∈ {0, 1, 2, 3} and then the pixel averages of the 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)

3
1 X
νˆ0 = (2 × Ic + I2k ) (9)
6
k=0

4
 2
1 X
νˆ1 = (2 × Ic + I2k+1 ) (10)
6
k=0

2 7
1 X 1 X
µ0 = (2 × Ic + Ik ); µ¯0 = ( Ik ) (11)
4 5
k=0 k=3

4
1 X 1 X
µ1 = (2 × Ic + Ik ); µ¯1 = ( Ik ) (12)
4 5
k=2 k∈{0,1,5,6,7}

6
1 X 1 X
µ2 = (2 × Ic + Ik ); µ¯2 = ( Ik ) (13)
4 5
k=4 k∈{0,1,2,3,7}

8 5
1 X 1 X
µ3 = (2 × Ic + Ik ); µ¯3 = ( Ik ) (14)
4 5
k=6 k=1

8
1 X
m= (Ic + Ik ); m̄ = mean(IM ×N ) (15)
9
k=0

-Second step: construction of the descriptor

95

In order to encode the central pixel, we propose a dominant intensity order metric, for this we choose four
sets X,Y ,Zmax and Zmin composed of (pixels and pixel averages), such that cardinal (X) = cardinal ( (Y)) =
cardinal ((Zmax )) =cardinal ((Zmin ))= 7 which construct two subspaces SU and SL of R7 , which are the basis
of the proposed method. Indeed to reduce the size of the proposed operator and capture more discriminating
100 information, we use an approach which helps us to build the new combined descriptor of several operators
which are linked together by the logical operator ”and” (&) Inspired by a Venn diagram see figure: 2.from
two ordered sets ( SU and SL ) provided with the topology of the order. The mathematical description of
this method is as follows:
;
Figure 2: Venn diagrams which connect three sets by the operator ”and” to build a single set

we end four sets X, Y ,Zmax and Zmin in the following way:

X = {(x0 = νˆ0 ), (x1 = ν0 ), (x2 = ν2 ), (x3 = I0 ), (x4 = I1 ), (x5 = I2 ), (x6 = I3 )} (16)

Y = {(y0 = νˆ1 ), (y1 = ν1 ), (y2 = ν3 ), (y3 = I4 ), (y4 = I5 ), (y5 = I6 ), (y6 = I7 )} (17)

Zmax = {(z̄0 = m̄), (z̄1 = m), (z̄2 = Imed), (z̄3 = max(µ0 , µ̄0 )), (z̄4 = max(µ1 , µ̄1 )), (z5 = max(µ2 , µ̄2 )), (z6 = max(µ3 , µ̄3 ))}
(18)

Zmin = {(z0 = m̄), (z1 = m), (z2 = Imed), (z3 = min(µ0 , µ̄0 )), (z4 = min(µ1 , µ̄1 )), (z5 = min(µ2 , µ̄2 )), (z6 = min(µ3 , µ̄3 ))}
(19)

5
 Smax = {(x, y, z̄) ∈ X × Y × Zmax |(x ≥ Ic − τ1 )&(y ≥ Ic + τ1 )&(z̄ ≥ Ic + τ2 )} (20)

Smin = {(x, y, z) ∈ X × Y × Zmin |(x ≤ Ic − τ1 )&(y ≤ Ic + τ1 )&(z ≤ Ic + τ2 )} (21)

The reduction of the subsets which form the binary patterns via logical connectors “and”, the presence of two
dynamic thresholds (τ1 ) and (τ2 ) see in the paper ( DNT-MQP ) and the construction of two subspaces Smax
and Smin lead to a descriptor a of local ternary values quantized in (+1) and (-1) on a multiscale scheme
based on directional neighborhood topologies called MNT-MTP coded as follows



 1 if (x,y,z̄) ∈ Smax


δτ1 ,τ2 (x, y, z̄, z) = 0 otherwise (22)




−1 if (x,y,z) ∈ S
min

So we define two symbols δ1 and δ2 such that:


1

if δτ1 ,τ2 (x, y, z̄, z) = 1
δ1 (x, y, z̄) = (23)
0

otherwise ,

1 if δτ1 ,τ2 (x, y, z̄, z) = −1

δ2 (x, y, z) = (24)
0

otherwise ,
With this convention, the kernel functions which define the two distributions the ternary models ψmax and
ψmin can be expressed as follows:
X
ψmax (M N T − M T P ) = (M N T − M T P )max = δ1 (xp , yp , z̄p ) 2p (25)
(xp ,yp ,z̄p )∈X×Y×Zmax

X
ψmin (M N T − M T P ) = (M N T − M T P )min = δ2 (xp , yp , zp ) 2p (26)
(xp ,yp ,zp )∈X×Y×Zmin

where the (xp ), (yp ), (z̄p ) and (zp ) are respectively the components of the vectors X, Y, Zmax and Zmin for
p = 0, ..., P-1 mentioned in equations 11,12, and 13 above.
h((DOT − M T P )max ) and h((DOT − M T P )min ) are respectively The two histograms of the descriptors
(DOT − M T P )max and DOT − M T P )min which are concatenated to form the future histogram h(MNT-
MTP) defined as follows:

h(DOT − M T P ) = h((DOT − M T P )max )||h((DOT − M T Pmin ) (27)

105 This defines a set of 27 possible ternary pattern codes for each local ternary 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 (28)

Figure 3: Calculation of the DOT-MTP for a local 3 3 neighborhood of a Brodatz (D12-005) image

6
 4. Experemental resultts and discussion

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
110 was compared with many newer and most promising advanced texture descriptors and several CNN-based
features to highlight its performance improvements.Table 1 summarizes the evaluation methods. Please note
that some of these methods are implemented using the original code available,for the other methods, we
have used our own implementation and made improvements based on the corresponding article to achieve
the same result as shown in the published article..to get the same output results as given in the publishedpa-
115 per. The experiments herein follow the standard evaluationprotocol for each tested dataset (i.e.split-sample
validation)where 50% of the samples are randomly selected to be usedas the training set, and the remaining
50% of the samples areregarded as the testing set. The samples of the test set arethen classified through
the parameter-free nearest-neighborrule (1-NN) with L1-city block distance. We repeat eachexperiment 100
times to remove any bias related to the divi-sion of the dataset and the averaged results are consideredas
120 estimated accuracies. In what follows, the texture datasets considered in the experiments are first presented
and the obtained experimental results are then discussed.

4.1. Texture dataset

In an effort to verify the efficiency and performance stability of the DNT-MTP, we performed extensive
testing on 13 well-known texture databases, including Brodatz, KTH-TIPS2b,, MondialMarmi, TC-00, TC-
125 01, CUReT, KTH-TIPS, Jerry Wu, BonnBTF, XUH, Kylberg, USPTex and visTex databases (the same
datasets used in [52], [63]),. These well-known data sets for texture classification are selected to cover various
characteristics of sample size, number of categories, and category uniformity related to perspective and scale.
More information about each data set is summarized in the table:2

No. Name C TS Challenge

1 Brodatz 13 208 Images-are not acquired under controlled conditions or corrected in any way and there is a lack of intraclass varia
2 KTH-TIP2b 11 176 Image acquired under illumination, rotation, pose scale changes.
3 MondialMarmi 12 768 Images acquired under controlled illumination conditions. The dataset includes hardware and software-rotated im
4 TC-00 24 480 Images acquired under controlled and variable conditions of illumination, rotation and spatial resolution.
5 TC-01 24 2112 Some properties as TC-00
6 CUReT 61 5612 Images present wide variety of geometric and photometric properties as intra-class variations in rotation, illumina
and viewing angle.
7 KTH-TIPS 10 40 Images captured under nine scales, nine illumination conditions and three poses.
8 JerryWu 39 156 Images characterize the natural color textures acquired under natural illumination conditions.
9 BonnBTF 10 160 Images acquired under different combinations of illumination direction, imaging direction and surface rotation.
10 XUH 25 1000 Images acquired under varying viewing and illumination angle.
11 Kylberg 28 4480 Images acquired under different rotation angles, different viewpoints, and different illuminations.
12 USPtex 191 2292 Real-word texture images acquired in the wild which are obtained by photographing natural objects and landscap
different conditions for lighting, geometrical positioning, noise, etc.
13 visTex 167 2672 Images acquired under different .

Table 1: Image datasets considered in the experiments.

4.2. Comparative Assessment of Performance

130 a)Experiment 1: Results and statistical significance

The main idea of this experiment is to use the ranking technology based on the Wilcoxon signature ranking
test proposed in [49] to statistically verify the ranking results obtained by the design method of the high-
level descriptors evaluated. This technique is applied to all pairwise combinations of 27 evaluation texture
descriptors considered out of the 13 texture data sets used in this experiment. Table 2 gathers the ranking

7
135 results obtained according to the standardised number of wins (number of wins / (number of data sets tested
× (number of descriptors tested-1)).by each method evaluated on all data sets tested. It is easy to see from the
results presented in Table 2 that the designed DNT-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 DNT-MTP is 0.86, compared to 0.80
140 with DNT-MQP (2nd superior), 0.68 with LCCMSP (3rd superior), 0.77 with LETRIST (7th superior),
etc. In particular, considering the classification performance of DNT-MQP (the best second descriptor) as a
reference as well as LETRIST (the best third descriptor), the DNT-MTP texture operator provides a strong
improvement compared to the 13 texture data sets tested.
b) Experiment2: Good results and stable performance
145 Table 2 lists the average classification score of each test method and each texture database, as well as
the overall 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 set. Based on these
results reported in Tables 2 and 3, and Figures 4 and 5 above we can easily deduce the following observations:

Ranking Texture descriptor and Ref Victories/ Dimension Year

1- comparisons
NN
1 Directional order topologies based multi-scale Tinary pattern : DOT-MTP proposed 0.85989 256 2021
2 Directional neighborhood topologies based multi-scale quinary pattern :DNT-MQP [? ] 0.80220 384 2020
3 Local concave-and-convex micro-structure pattern: LCCMSB[? ] 0.77198 2046 2018
4 Attractive-and-repulsive center-symmetric LBP: ARCSLBP [? ] 0.70604 256 2018
5 Mixed neighborhood topoloycross decodedpatterns : MNTCDP [? ] 0.70330 2048 2018
6 Repulsive-and-attractive local gradient contour: RALBGC [? ] 0.65659 1022 2017
7 LOcally encoded transform featurehistogram : LETRIST [? ] 0.60440 413 2017
8 Oriented star sampling structure based multi-scale ternary pattern O3S-MTP [? ] 0.59615 4096 2020
9 Improved local ternary pattern: ILTP [? ] 0.59615 1024 2010
10 Local directioal ternary pattern : LDTP(2) [? ] 0.59341 1022 2018
11 Improved local quinary patterns: ILQP [? ] 0.58791 1024 2019
12 Multi kernel based local binary pattern:KLBP [? ] 0.57967 1280 2019
13 local Ternary Pattern: LTP [? ] 0.57967 512 2007
14 Local directional edge binary patterns: LDEBP [? ] 0.49176 64 2019
15 Local extreme complete trio pattern: LECTP [? ] 0.43407 768 2016
16 Medianrobustextended local binary pattern: MRELBP [? ] 0.41209 200 2016
17 local extreme sign trio patterns: LESTP [? ] 0.41209 512 2016
18 Local neighborhood difference pattern: LNDP [? ] 0.38736 256 2019
19 Local optimal-oriented pattern : LOOP [? ] 0.34615 256 2018
20 Local directional ZigZag patterns:LDZP [? ] 0.32418 354 2019
21 Local neighborhood intensity pattern: LNIP [? ] 0.31868 512 2019
22 Local quadruple pattern: LQPAT [? ] 0.30220 512 2017
23 Cascaded asymmertric local pattern: CALP [? ] 0.27473 192 2019
24 Multi-direction local binary pattern: MDLBP [? ] 0.24725 256 2019
25 Center symmetric quadruple pattern: CSQP [? ] 0.21703 256 2017
26 Directional codin: DC [? ] 0.12363 225 2018
27 Elliptical upperand lower localternary pattern: EULLTP [? ] 0.09341 32 2019
28 Local diagonal extrema number pattern: LDENP [? ] 0.02473 15 2017
29 Local directional value: LDV [? ] 0.01648 256 2019

Table 2: Summary of the texture descriptors used for the evaluation and results of the comparison with the proposed descriptor.

Figure 4: curve also representing the overall average performance (GAP) of the tested descriptor including the descriptor offers

Figure 5: curve representing the classification (red line) and dimension (blue line) of the descriptors tested as well as the
performance of the proposed descriptor

The DOT-MTP descriptor provided as an alternative to the LBP and ILTP texture operators retains
150 the advantages of monotonic lighting changes and low complexity. In addition, DOT-MTP also has other
more important advantages, which will be discussed below. Compared to some modern advanced texture

8
 Texture descriptor 1 2 3 4 5 6 7 8 9 10 11 12 13 GAP mean Std
DOT-MTP 100.00 95.26 93.52 99.95 99.09 95.20 100.00 99.77 100.00 99.89 97.47 90.52 78.71 96.11 *
DNT-MQP 100.00 95.12 91.88 99.83 99.09 95.30 100.00 99.67 99.25 99.85 96.59 91.47 75.99 95.70 *
LETRIST 99.99 90.08 82.89 99.92 98.51 95.56 100.00 100.00 100.00 97.91 94.26 82.24 64.31 92.74 *
MNTCDP 100.00 90.93 89.31 99.78 99.40 89.70 100.00 100.00 100.00 99.52 92.92 85.29 75.81 94.05 *
LCCMSB 100.00 93.32 89.74 99.95 99.40 94.20 100.00 98.51 99.72 99.71 94.30 88.30 74.93 94.77 *
ARCSLBP 100.00 93.23 89.92 99.72 98.88 92.62 100.00 99.60 99.49 99.68 94.35 86.50 72.67 94.36 *
LDZP 99.93 88.53 83.27 97.92 96.43 85.12 99.75 96.51 99.96 92.50 83.99 70.51 68.62 89.46 *
KLBP 100.00 91.24 84.36 99.38 98.48 91.42 100.00 98.00 99.74 99.25 93.24 85.37 72.84 93.33 *
MRELBP 99.83 90.51 84.07 98.99 97.71 85.76 97.85 99.23 98.79 96.94 94.29 79.34 62.81 91.24 *
LOOP 99.42 85.80 73.66 99.44 98.57 82.50 100.00 98.28 99.30 96.10 89.69 77.21 68.62 89.89 *
LDEBP 99.66 87.30 82.55 99.77 99.08 84.44 100.00 97.01 99.97 99.06 92.91 83.83 67.53 91.78 *
LDENP 93.26 65.80 54.03 94.44 82.94 52.97 88.80 88.74 93.33 87.77 69.95 41.24 41.86 73.47 *
LDTP(2) 99.78 90.47 83.53 99.94 99.07 90.34 100.00 98.22 99.97 99.09 91.58 82.63 73.82 92.96 *
DC 97.35 79.78 66.36 92.66 87.68 68.72 100.00 96.46 97.94 91.14 85.91 59.05 54.34 82.87 *
LDV 90.39 58.37 49.69 92.20 83.49 71.97 88.45 80.91 94.36 85.32 73.68 40.20 37.56 72.81 *
O3S-MTP 99.56 91.95 86.03 99.64 98.71 93.34 99.95 99.00 99.46 99.38 94.08 82.33 71.34 93.44 *
LNIP 99.30 82.97 82.92 99.10 97.83 81.08 98.45 97.21 99.95 95.54 87.58 75.54 69.63 89.77 *
CALP 99.40 82.63 75.28 98.99 94.89 92.66 98.15 95.76 95.83 98.72 89.64 75.00 66.05 89.46 *
ILQP 100.00 93.39 90.07 99.63 97.57 91.22 100.00 98.03 98.72 98.33 92.11 87.54 75.39 94 *
EULLTP 99.38 78.63 69.33 96.68 88.98 73.72 93.90 92.58 93.40 94.87 82.10 64.95 53.53 83.23 *
LESTP 99.79 88.80 77.74 97.14 98.54 79.06 100.00 99.18 99.38 95.80 95.23 78.97 67.93 90.58 *
LECTP 100.00 88.59 81.84 97.23 98.21 82.18 99.60 98.69 99.08 98.64 94.50 80.70 68.07 91.33 *
LNDP 99.89 88.72 76.64 99.20 97.47 88.40 100.00 96.81 99.15 97.77 90.04 79.52 71.57 91.6 *
LQPAT 99.16 83.76 75.47 99.23 97.11 89.80 99.60 96.15 99.13 98.03 89.49 76.74 67.71 90.11 *
CSQP 98.56 80.13 64.44 98.77 93.77 85.10 100.00 96.74 99.03 97.53 88.33 69.90 61.75 87.23 *
MDLBP 99.23 77.99 76.33 99.63 95.05 81.65 99.60 94.55 96.45 97.04 88.21 69.47 70.17 8810 *
RALBGC 100.00 93.25 88.93 99.83 99.01 92.51 100.00 97.92 98.36 99.46 93.35 87.08 75.51 94.25 *
ILTP 100.00 91.88 88.54 99.35 97.70 90.57 100.00 98.03 97.31 99.84 96.38 83.94 73.23 93.60 *
LTP 99.99 91.31 86.51 99.18 98.80 91.59 99.90 98.13 94.51 99.60 94.52 87.38 72.91 93.41 *

Table 3: Comparison results (%) of different texture representations on 13 data sets..

descriptors listed in Table 1, which have inherent flaws in LBP descriptors, DOT-MTP descriptors can
describe the texture characteristics of local images with less computational complexity and small size in size.
because of topology. directional, thus building this newly proposed operator.
155 - whether in Tables 2 and 3 or in Figure 4 and 5 It is very clear to observe that descriptors such as CALP,
MDLBP, CSQP, DC and EULLTP, LDENP and LDV, show the poorest performance on almost all datasets
used. and on the ranking less than 28% (Victories/ comparisons) para ports to other descriptors tested. -
Indeed, given the classification among the descriptors tested in each data set used, DOT-MTP stands out
as the best stable descriptor for the ten textured data sets: Brodatz, KTH-TIP2b, MondialMarmi, TC-00,
160 KTH -TIPS, BonnBTF, XUH, Kylberg and visTex, in databases (Brodatz, KTH-TIP and BonnBTF) DOT-
MTP reaches the max rank (100%), while it is close to 100% in both databases (TC00 (99.95%) and (XUH
(99.89%))). In addition, it is part of the second best method in the database (TC-01 (99.09%)), which is
slightly different from the first descriptor MNTCDP (99.40%). In addition, it is also that it has reached
the third best classification of the database (CUReT (95.20%), slightly different from the first descriptor
165 LETRIST (99.56%), therefore, the proposed descriptor DOT -MTP reached the second best ranking in the
database (JerryWu (99.77%)) with a slight difference compared to the first LETRIST descriptor (100%), and
at the end the DOT-MQP descriptor reached the first rank of the database (USPtex (91.4 7%)), while the
proposed descriptor, the DOT-MTP descriptor ranks second (USPtex (90, 52%))
c) Experiments 3: Comparisons with the characteristics of CNN
170 To thoroughly assess the effectiveness of DOT-MTP at the highest level, we also used the preformed
deep learning models ResNet50, ResNet101, AlexNet, VGG16 and VGG19 to compare its performance with
well-known CNN functions. Some of them have varying layers. It should be noted that the result of the
CNN- evaluated function is considered a function vector and its method of expansion is similar to how the
process function is usually used. It can be noted that DOT-MTP achieves the highest average performance
175 in a manner similar to local descriptors against all CNNs functions tested in 10 of 13 datasets. It should be

9
 pointed out that when DNT-MTP is not in the first place, even if it does not achieve the highest score, it
can still achieve interesting average performance (i.e. its accuracy is similar to that of the previously ranked
CNNs), but the advantage is DOT-MTP is conceptually easier to implement and requires no training. Please
note that CNNs are best suited for places with high intra-class variability, while local descriptors are best
180 suited for homogeneous fine-grained textures with low intra-class variability. In fact, when the images have
similar content and less patterns, the depth model tends to extract redundant characteristic values, which
will make the classifier perform poorly on those characteristics, and different models will provide almost the
same performance. On the other hand, when the image has complex content and different patterns (as in
the USP text dataset), the deep features provide a better representation than the local descriptors [70].

185 5. Conclusion

The proposed descriptor named DOT-MTP, compared and tested respectively with several recent and
advanced LBP-type methods and 13 reference textured data sets, these comparisons and ballasts have shown
that the DOT-MTP descriptor, is conceptually simple and easy to implement, is more decriminating than
all the tested methods ie it captures more information thanks to its specific form its logical connections its
190 directional topology on these neighborhoods which also makes its small dimension compared to improved
variant of the LTP type. Future work will concern the study of the use of other well-known classifiers to
increase the classification performance of the proposed model. In addition, the proposed texture operator
has potential for deployment in high-level complex applications related to texture classification, including
dynamic texture classification, subtraction of backgrounds in complex scenes, recognition objects and video-
195 based face analysis, etc.

References

[1] Yu, J., Qin, Z., Wan, T., Zhang, X.: Feature integration analysis ofbag of features model for image
retrieval. Neurocomputing120,355–364 (2013)

[2] anerji, S., Verma, A., Liu, C.: Cross disciplinary biometric sys-tems. LBP and Color Descriptors for
200 Image Classification, pp.205–225. Springer, Berlin (2012)

[3] Ledoux, A., Losson, O., Macaire, L.: Color local binary patterns:compact descriptors for texture classifi-
cation. J Electron Imaging25(6), 061404 (2016)

[4] Yongsheng, D., et al.: Multi-scale counting and difference repre-sentation for texture classification. Vis.
Comput. (2017).https://doi.org/10.1007/s00371-017-1415-4

205 [5] Lowe, D.G.: Distinctive image features from scale-invariant key-points. Int. J. Comput. Vision60(2),
91–110 (2004

[6] Xuemei, H., Yan, D.: Image matching with an improved descriptorbased on SIFT. In: Proceedings
Volume 10322, Seventh Interna-tional Conference on Electronics and Information Engineering, pp.1-7
(2017).https://doi.org/10.1117/12.2265595

210 [7] Romero, A., Gatta, C., Camps-Valls, G.: Unsupervised deep featureextraction for remote sensing image
classification. IEEE Trans.Geosci. Remote Sens.54(3), 1349–1362 (2016)

10
 [8] Berbar, M.: Three robust features extraction approaches for facialgender classification. Vis. Com-
put.30(1), 19–31 (2014)

[9] Y. Xu, H. Ji, and C. Fermüller, “Viewpoint invariant texture descriptionusing fractal analysis,”Int. J.
215 Comput. Vis., vol. 83, no. 1, pp. 85–100,Jun. 2009.

[10] . Song, H. Li, F. Meng, Q. Wu, and B. Luo, “Exploring space–frequencyco-occurrences via local quantized
patterns for texture representation,”Pattern Recognit., vol. 48, no. 8, pp. 2621–2632, Aug. 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.

220 [12] M. Heikkilä, M. Pietikäinen, and C. Schmid, “Description of interestregions with center-symmetric local
binary patterns,” inProc. Com-put. Vis., Graph. Image Process., vol. 4338. 2006, pp. 58–69.

[13] S. Liao, M. W. K. Law, and A. C. S. Chung, “Dominant local binarypatterns for texture classifica-
tion,”IEEE Trans. Image Process., vol. 18,no. 5, pp. 1107–1118, May 2009.

[14] E. Rachdi, et al. ”Directional Neighborhood Topologies Based Multi-Scale Quinary Pattern for Texture
225 Classification.” IEEE Access 8 (2020): 212233-212246.

[15] Y. El merabet, Y. Ruichek, Local concave-and-convex micro-structure patterns for texture classification,
Pattern Recognit. 76 (2018) 303–322.

[16] Y. El merabet, Y. Ruichek, A. El idrissi, Attractive-and-repulsive center-symmetric local binary patterns

for texture classification, Eng. Appl. Artif. Intell. 78 (2019) 158–172.

230 [17] M. Kas, Y. El Merabet, Y. Ruichek, and R. Messoussi, “Mixed neighbor-hood topology cross decoded
patterns for image-based face recognition,”Expert Syst. Appl., vol. 114, pp. 119–142, Dec. 2018.

[18] T. Song, H. Li, F. Meng, Q. Wu, J. Cai, Letrist: locally encoded transform feature histogram for
rotation-invariant texture classification, IEEE Trans. Circuits Syst. Video Technol. (2017).

[19] I. El khadiri, Ruichek, Y., D. Chetverikov, and R. Touahni. ”O3S-MTP: Oriented star sampling structure
235 based multi-scale ternary pattern for texture classification.” Signal Processing: Image Communication
(2020): 115830.

[20] W. Yang, C. Sun, 2011. Face recognition using improved local texture patterns. 9th World Congress on
Intelligent Control and Automation (WCICA), IEEE, 2011, 48-51.

[21] I. El Khadiri, A. Chahi, Y. El Merabet, Y. Ruichek, R. Touahni, Local directional ternary pattern: a
240 new texture descriptor for texture classification, Comput. Vis. Image Underst. 169 (2018) 14–27.

[22] . Armi and S. Fekri-Ershad, “Texture image Classification based onimproved local Quinary pat-
terns,”Multimedia Tools Appl., vol. 78,pp. 1–24, Feb. 2019.

[23] T. Tuncer and S. Dogan, “Pyramid and multi kernel based local binarypattern for texture recognition,”J.
Ambient Intell. Humanized Comput.,vol. 11, pp. 1–12, May 2019.

11
245 [24] X. Tan, B. Triggs, Enhanced local texture feature sets for face recognition under difficult lighting con-
ditions, IEEE Trans. Image Process. 19 (6) (2010) 1635–1650.

[25] . Sucharitha and R. K. Senapati, “Biomedical image retrieval by usinglocal directional edge binary
patterns and Zernike moments,”MultimediaTools Appl., vol. 79, pp. 1–18, Nov. 2019.

[26] S.K. Vipparthi, S.K. Nagar, Local extreme complete trio pattern for multimedia image retrieval system,
250 Int. J. Autom. Comput. 13 (5) (2016) 457–467.

[27] L. Liu, S. Lao, P.W. Fieguth, Y. Guo, X. Wang, M. Pietikäinen, Median robust extended local binary
pattern for texture classification, IEEE Trans. Image Process. 25 (3) (2016) 1368–1381.

[28] S.K. Vipparthi, S.K. Nagar, Local extreme complete trio pattern for multimedia image retrieval system,
Int. J. Autom. Comput. 13 (5) (2016) 457–467.

255 [29] . Verma and B. Raman, “Local neighborhood difference pattern: A newfeature descriptor for natural
and texture image retrieval,”MultimediaTools Appl., vol. 77, pp. 1–24, May 2017.

[30] . Chakraborti, B. McCane, S. Mills, and U. Pal, “LOOP descriptor:Local optimal-oriented pattern,”IEEE
Signal Process. Lett., vol. 25, no. 5,pp. 635–639, May 2018.

[31] S. K. Roy, B. Chanda, B. B. Chaudhuri, S. Banerjee, D. K. Ghosh,and S. R. Dubey, “Local directional

260 ZigZag pattern: A rotation invariantdescriptor for texture classification,”Pattern Recognit. Lett., vol.
108,pp. 23–30, Jun. 2018

[32] P . Banerjee, A. K. Bhunia, A. Bhattacharyya, P. P. Roy, and S. Murala,“Local neighborhood intensity

pattern—A new texture feature descriptorfor image retrieval,”Expert Syst. Appl., vol. 113, pp. 100–115,
Dec. 2018.

265 [33] S. Chakraborty, S. K. Singh, and P. Chakraborty, “Local quadruple pattern:A novel descriptor for facial
image recognition and retrieval,”Comput.Electr. Eng., vol. 62, pp. 92–104, Aug. 2017

[34] S. Chakraborty, S. K. Singh, and P. Chakraborty, “Cascaded asymmetriclocal pattern: A novel descriptor
for unconstrained facial image recognitionand retrieval,”Multimedia Tools Appl., vol. 78, pp. 1–20, May
2019

270 [35] T. Song, H. Li, F. Meng, Q. Wu, and J. Cai, “LETRIST: Locally encodedtransform feature histogram
for rotation-invariant texture classification,”IEEE Trans. Circuits Syst. Video Technol., vol. 28, no. 7, pp.
1565–1579,Jul. 2018.

[36] S. Chakraborty, S. K. Singh, and P. Chakraborty, “Centre symmetricquadruple pattern: A novel de-
scriptor for facial image recognition andretrieval,”Pattern Recognit. Lett., vol. 115, pp. 50–58, Nov. 2018

275 [37] . Ouslimani, A. Ouslimani, and Z. Ameur, “Rotation-invariant featuresbased on directional coding for
texture classification,”Neural Comput.Appl., vol. 31, no. 10, pp. 6393–6400, Oct. 2019.

[38] L. Kabbai, M. Abdellaoui, and A. Douik, “Image classification by com-bining local and global fea-
tures,”Vis. Comput., vol. 35, no. 5, pp. 679–693,May 2019

12
 [39] A . Pillai, R. Soundrapandiyan, S. Satapathy, S. C. Satapathy, K.-H. Jung,and R. Krishnan, “Local
280 diagonal extrema number pattern: A new featuredescriptor for face recognition,”Future Gener. Comput.
Syst., vol. 81,pp. 297–306, Apr. 2018

[40] X. Wu and J. Sun, “Face recognition based on multi-scale local directionalvalue,”Multimedia Tools and
Applications, vol. 79, pp. 1–17, Nov. 2019.

[41] Rivera, A. R., Castillo, J. R., & Chae, O. O. (2013). Local directional number pattern for face analysis:
285 Face and expression recognition. IEEE transactions on image processing, 22(5), 1740-1752.

[42] Vipparthi, S. K., Nagar, S. K. Local Extreme Complete Trio Pattern for Multimedia Image Retrieval
System. International Journal of Automation and Computing, 2015.

[43] 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). International Journal of Signal Processing, Image
290 Processing and Pattern Recognition, vol.8, No.10 (2015), pp.71-80.

[44] Xu, Y., Huang, S., Ji, H., & Fermüller, C. (2012). Scale-space texture description on SIFT-like textons.
Computer Vision and Image Understanding, 116(9), 999-1013.

[45] Zhu, Z., You, X., Chen, C. P., Tao, D., Ou, W., Jiang, X., Zou, J. (2015). An adaptive hybrid pattern
for noise-robust texture analysis. Pattern Recognition, 48(8), 2592-2608.

295 [46] Rivera, A. R., Castillo, J. R., Chae, O. (2015). Local directional texture pattern image descriptor.
Pattern Recognition Letters, 51, 94-100.

[47] Dubey, S. R., Singh, S. K., Singh, R. K. . Local Wavelet Pattern: A New Feature Descriptor for Image
Retrieval in Medical CT Databases. IEEE Trans. Image Process, 24(12): 5892-5903, 2015

[48] Dubey, S. R., Singh, S. K., Singh, R. K. . Local Diagonal Extrema Pattern: A New and Efficient Feature
300 Descriptor for CT Image Retrieval. IEEE Signal Processing Letters, 22(9): 1215-1219, 2015

[49] Dubey, S. R., Singh, S. K., Singh, R. K. . Local Bit-plane Decoded Pattern: A Novel Feature Descriptor
for Biomedical Image Retrieval. IEEE J. Biomed. Health Inform, 20(4): 1139-1147, 2016.

[50] Dubey, S. R., Singh, S. K., Singh, R. K. . Multichannel Decoded Local Binary Patterns for Content
Based Image Retrieval. IEEE Trans. Image Process, 25(9): 4018-4032, 2016.

305 [51] Chakraborty, S., Singh, S. K., Chakraborty, P. (2017). Centre Symmetric Quadruple Pattern:
A Novel Descriptor for Facial Image Recognition and Retrieval. Pattern Recognition Letters.
https://doi.org/10.1016/j.patrec.2017.10.015

[52] Local Directional Relation Pattern for Unconstrained and Robust Face Retrieval. arXiv:1709.09518, 2017

[53] Face Retrieval using Frequency Decoded Local Descriptor. arXiv:1709.06508, 2017

310 [54] Pillai, A., Soundrapandiyan, R., Satapathy, S., Satapathy, S. C., Jung, K. H., Krishnan, R. (2017).
Local diagonal extrema number pattern: A new feature descriptor for face recognition. Future Generation
Computer Systems. https://doi.org/10.1016/j.future.2017.09.055

13
 [55] Y. El merabet and Y. Ruichek. Local Concave-and-Convex Micro-Structure Patterns for texture classi-
fication. Pattern Recognition. Vol. 76, 2018, pp. 303-322.

315 [56] Liu, L., Lao, S., Fieguth, P. W., Guo, Y., Wang, X., Pietikäinen, M. (2016). Median robust extended
local binary pattern for texture classification. IEEE Trans. Image Process, 25(3), 1368-1381.

[57] Ding, C., Choi, J., Tao, D., Davis, L. S. (2016). Multi-directional multi-level dual-cross patterns for
robust face recognition. IEEE transactions on pattern analysis and machine intelligence, 38(3), 518-531.

[58] Verma, M., & Raman, B. (2017). Local neighborhood difference pattern: A new feature descriptor for
320 natural and texture image retrieval. Multimedia Tools and Applications, 1-24.

[59] Chakraborty, S., Singh, S. K., & Chakraborty, P. (2017). Local quadruple pattern: A novel descriptor
for facial image recognition and retrieval. Computers & Electrical Engineering.

14