Noname manuscript No.

(will be inserted by the editor)

Directional Neighborhood Topologies based Multi-scale

Quinary Pattern for Texture Classification
E. Rachdi · Y. El merabet · M. Abdar · M.E. Basiri · D. Zhang · R.
Messoussi

Received: date / Accepted: date

Abstract This paper ideates a new computationally ture representation than other local feature descriptors.
simple and effective local image feature descriptor, re- The effectiveness of DNT-QP is evaluated on sixteen
ferred to as Directional Neighborhood Topologies based challenging representative well-known texture datasets
Multi-scale Quinary Pattern (DNT-MQP) for texture and it is found that it maintains a high level of perfor-
description and classification. The essence of DNT-MQP mance stability where the achieved performances are
is to encode the structure of local neighborhood by an- competitive or better than several recent most promis-
alyzing the differential excitation and directional infor- ing state-of-the-art texture descriptors. Furthermore,
mation using various directional neighborhood topolo- statistical hypothesis testing through Wilcoxon signed
gies and new pattern encoding scheme. We first de- rank test is applied to prove the statistical significance
signed four different versions of single scale DNT-QP of the accuracy improvement obtained in all the datasets.
features based on four directional neighborhood topolo-
Keywords Texture classification · Texture recog-
gies based sampling sets which are then combined to-
nition · Texture descriptors · LBP · Local feature
gether to build the effective multiscale DNT-MQP model.
extraction.
Unlike some existing parametric methods that employ
static thresholds to perform thresholding, the construc-
tion process of DNT-MQP includes an automatic mech- 1 Introduction
anism for dynamic thresholds estimation. Thanks to a
richer local description ensured by exploiting comple- The surface of objects and materials such as crops in
mentary information resulting from single scale DNT- a field, natural scenes, human faces, palmprint, human
QP operators’ combination, DNT-MQP descriptor has skin and many others have their own distinctive tex-
high capability to elicit stable and discriminative fea- ture. The texture analysis, which permit to provide con-
structive information about the structural and spatial
E. Rachdi · Y. El merabet · R. Messoussi
Laboratoire LASTID, Département de Physique, Faculté
arrangement of surfaces, has been extensively studied in
des Sciences, Université Ibn Tofail, BP 133, 14000 Kénitra, the past few years and received considerable attention
Maroc. E-mail: [email protected], y.el-merabet@univ- in many various research subjects such as pattern recog-
ibntofail.ac.ma, [email protected] nition, computer vision and image processing. Texture
M. Abdar classification is suitable for many potential applications
Institute for Intelligent Systems Research and Inno- for instance face classification and recognition, back-
vation (IISRI), Deakin University, Australia. E-mail:
[email protected]
ground subtraction, iris recognition, palmprint recog-
nition, pedestrian detection, biomedical image analysis
M.E. Basiri
Department of Computer Engineering, Shahrekord Univer-
and image retrieval.
sity, Shahrekord 64165478, Iran. E-mail: [email protected] In the literature, there are many approaches available
D. Zhang
for texture analysis with excellent surveys given in [1].
School of Science and Engineering, The Chinese University One can cite human perception-based features [13], ran-
of Hong Kong - Shenzhen, 407605 Shenzhen, China. E-mail: dom features [11], filter-based techniques like Gaussian
[email protected] Markov random fields [9], Gabor [2] and wavelet [10],
2 E. Rachdi et al.

co-occurrence matrix-based approaches [3], ranklet trans- computes two micro patterns from the local relation-
form based approaches [12], fractal analysis based ap- ships by encoding relations amongst the neighbors in
proaches [16], Texton dictionary-based [14], etc. The quadruple space. In another study in [42] presented
method proposed in [5] uses Fourier descriptors to ex- the center symmetric quadruple pattern (CSQP) de-
tract texture feature in spectrum domain. Even though scriptor for facial image retrieval and recognition. As
these above approaches, which take advantage of the for LQPAT, CSQP encodes the facial asymmetry in
merits of spectral and statistical features, ensure im- quadruple space. CSQP computes an eight-bit pattern
proving the ability of texture representation and mod- from 16 pixels in the local neighborhood. Issam et al.
eling, their demerits lie in the fact that they increase [46] proposed local directional ternary pattern (LDTP)
the computation cost in extracting features. Therefore, for texture classification. LDTP, which exploits both
local feature extraction methods have been proposed LDP’s and LTP’s concepts, encodes both contrast in-
and have additionally and impressively been employed formation and directional pattern features in a compact
in texture analysis field. The main advantages of the way based on local derivative variations. Kas et al. [45]
local hand-crafted descriptors are the simplicity in de- proposed mixed neighborhood topology cross decoded
sign and not dependent upon a large volume of training patterns (MNTCDP) for face recognition. MNTCDP
data [36]. adopts 5×5 block which allows combining radii (2) and
Among the local feature extraction approaches, local bi- angles (4) compared to 3×3 block supporting only an-
nary pattern (LBP) developed by Ojala et al. [4], have gle variation.
brightened up as one of the most eminent texture de- The best position of hand-crafted descriptors has been
scriptor and has gained more attention over the past dominated by LBP-like methods for more than a decade.
decades. LBP as an effective texture descriptor is, in The need to design an effective local texture descrip-
fact, highly appreciated by researchers due to its dis- tor with high discrimination capability is no longer to
tinctive advantages including its simplicity, its well in- be demonstrated. Indeed, the development of methods
variability to monotonic gray level changes and its suit- based on local texture descriptors in pattern recogni-
ability for real time applications due to its low comput- tion continue to be designed still today, e.g., median lo-
ing cost. The LBP method, even if originally designed cal ternary patterns (MLTP) [41], averaged local binary
for texture modeling and classification, it showed con- patterns (ALBP) [47], repulsive-and-attractive local bi-
siderable performance in a wide range of applications nary gradient contours (RALBGC) [43], local concave-
like medical and biomedical images analysis, motion de- and-convex micro-structure (LCCMSP) [40], attractive-
tection, image retrieval, face and facial description and and-repulsive center-symmetric local binary patterns
recognition, background subtraction, etc. Despite these (ARCS-LBP) [51], multi-direction local binary pattern
merits, the basic LBP descriptor has some limitations (MDLBP) [52], quaternionic local angular binary pat-
[40] : (1) the output is very sensitive even if to small tern (QLABP) [49], selectively dominant local binary
change in the input; (2) it is not invariant to image ro- patterns (SDLBP) [50], chess pattern (Chess-pat) [48],
tation; (3) based on the employed thresholding scheme, synchronized rotation local ternary pattern [35], Pat-
it is highly susceptible to noise interference; etc. In or- tern of Local Gravitational Force (PLGF) [36] and so
der to deal with these limitations and thus enhance the on.
classification performance of LBP, a great number of Even though LBP and its modifications and extensions
improved LBP algorithms have been proposed in re- achieve satisfactory performance, still an alternative tech-
cent years [17, 54, 18, 19, 45]. The authors in [20] pro- nique to enhance the discriminative power in an image
vided an exhaustive investigation and comprehensive for effective texture modeling and representation is es-
experiments assessing the performance of a great num- sential. In this paper, aiming at the further enhance-
ber of old and recent state-of-the-art texture descriptors ment of texture classification performance and keep-
in face recognition problem. Shiv et al. [44] designed ing the simplicity and effectiveness of the traditional
local wavelet pattern (LWP) descriptor for medical im- LBP and addressing its feebleness, we design a con-
age retrieval. LWP performs local wavelet decomposi- ceptually simple and yet robust model of LBP, named
tion over local neighborhood of pixels to encode the directional neighborhood topologies based multi-scale
relation among the neighboring pixel. The LWP pat- quinary pattern (DNT-MQP). The designed texture op-
tern for the central pixel is computed by comparing erator has basically the following merits: As it will be
its local wavelet decomposed value with those of the shown further, DNT-MQP has low computational com-
neighboring pixels. Chakraborty et al. [38] introduced plexity and by offering highly-desirable features improves
the local quadruple pattern (LQPAT) descriptor for us- both the discriminative capability of LBP-like methods
ing in facial image retrieval and recognition. LQPAT and their invariance to monotonic illumination changes
Directional Neighborhood Topologies based Multi-scale Quinary Pattern for Texture Classification 3

as well as their robustness to small variations, due to in Eq. 1.

image noise. The main contributions of this paper in-
clude the following: 
a3 a2 a1


I3×3
m,n =  a4 ac a0  (1)
– An automatic mechanism for dynamic thresholds
a5 a6 a7
estimation for quinary pattern creation process is
proposed.
– A family of single scale descriptors named direc- ap is the gray levels of the peripheral pixels (p ∈ {0, 1,
tional neighborhood topologies based single-scale quinary. . . , P-1}). P=8 is the number of local neighbors.
pattern (DNT-QP) is developed based on several di-
rectional neighborhood topologies (DNT) which are
more effective for image texture understanding and
analysis that large number of existing methods.
2.1 Traditional Local Binary Patterns
– We further extend the obtained single scale DNT-
QP descriptors to incorporate multiscale by con-
LBP method [4] characterizes the local spatial struc-
catenating them into a single vector feature to build
ture and the local contrast of each 3×3 local region in
the effective directional neighborhood topologies based
the image by thresholding the intensity of surround-
multi-scale quinary pattern (DNT-MQP) descriptor
ing pixels ap ; p ∈ {0, 1, ..., 7} with the intensity of the
which should be more robust and stable.
central pixel ac . The value of ap is turned into binary
– For performance evaluation, we restrict ourselves to
value 1; otherwise, it is turned into 0. A code of 8 bits
texture classification as basic application of LBP.
is obtained which is transformed into decimal number
Extensive experiments on sixteen challenging pub-
to get the LBP code. Formally, for a 3×3 local region,
licly available texture database is performed.
the definition of the kernel function of LBP operator is
– We provide a fair and systematic comparison and
shown in the equation below:
found that the designed texture operator shows su-
perior performance to 23 recent powerful state-of-
the-art texture descriptors.
P−1
X
3×3
The rest of the paper is organized as follows. Section 2 fLBP (Im,n ) = ϑ(ap − ac ) × 2p (2)
p=0
briefly introduces some typical existing texture descrip-
tors. Section 3 presents the designed DNT-MQP de-
scriptor. Comprehensive experimental results and com- where ϑ(·) is the Heaviside step function (cf. Eq.
parative evaluation are given in Section 4. Section 5 (3)).
concludes the study and presents some solid future re-
search directions. 
1, if x >= 0
ϑ(x) = (3)
0, otherwise

2 Review of the Existing Representative Figure 1 shows the standard steps illustrating the
Texture Methods process of the LBP feature extraction as well as the
quantification of the difference between two texture im-
In this part of study, we briefly introduce some of rep- ages using LBP like features.
resentative texture methods reported in the literature,
our paper’s core idea, contributions and organization.
Given that the designed local texture descriptor is based
on local kernel functions, it is appropriate to define 2.2 Local Quinary Patterns (LQP)
the spatial arrangement of the local structure used in
this paper. In this work, the choice was focused on In local quinary patterns (LQP) [13], the difference of
the use of the 3×3 sized block given its easiness-of- the gray level value between the central pixel ac and its
implementation consideration and which is by far the neighboring pixels ap is encoded according to five levels
most required neighborhood especially in real-time ap- (i.e., 2, -1, 0, 1 and 2) calculated using two thresholds
plications. The mathematical presentation of the set of τ1 and τ2 . The LQP is thus closely relevant to the LTP
gray-scale values of a 3×3 grayscale image patch I3×3 m,n [21], the only difference being that the number of coding
around the central pixel ac of coordinates (m,n) is given levels is five in the LQP while in the LTP it is three.
4 E. Rachdi et al.

Texture Image 1 Texture Image 2

(6)
Texture
Textureimages
imagestoto
be
becompared
compared

7 
X 
77

79
82

70
61
75 3×3
3×3square
square
35 25 88
LQP+1 (I3×3
m,n ) = ϑ(ap −ac −τ1 )×ϑ(ac −ap +τ2 ) ×2p
neighborhood 69 55 42
neighborhood p=0
69 64 79 80 80 91

(7)
1 1 0 0 0 1
Pixel
PixelIntensity
Intensity
1 1 comparison 1 0
comparison
0 0 1 1 1 1

7
X
Binary code: 11011001
Decimal:217
Computation
Computationofof
LBP-Code
Binary code: 00101111
Decimal:47 LQP+2 (I3×3
m,n ) = ϑ(ap − ac − τ2 ) × 2p (8)
LBP-Code
p=0

The four descriptor histograms generated by the

four LQP−2 , LQP−1 , LQP+1 and LQP+2 operators are
LBP image 1 LBP image 2 finally concatenated to form the final hLQP future vec-
LBP-histograms tor as illustrated in Eq. 9. LQP generates 4×28 possible
different patterns.
Frequency

hLQP = hLQP−2 , hLQP−1 , hLQP+1 , hLQP+2 (9)

Texture
Texturedifference
difference
quantification
quantification
2.3 Improved Local Quinary Patterns (ILQP)

Fig. 1: The standard steps for quantification of the Armi et al. [32] proposed the improved local quinary
difference between two texture images using LBP like patterns (ILQP) to overcome some disadvantages of
features. LQP and thus improve its performance. The definition
of the local quinary pattern in ILQP is the same as in
the LQP operator. By contrast, the thresholds τ1 and
τ2 are dynamically defined. τ1 and τ2 are computed as:
The quinary model is divided into four binary models
according to the following rule:
τ1 = median(|lmadmedian(lmad)|); (10)


 +2 ap > ac + τ2 , where lmad = {LocalM ADk |k = 1, 2, ..., M × N },
LocalM AD = median(|Gmedian(G)|) and G = {ap |p =

+1 ac + τ1 6 ap < ac + τ2 ,



0; 1; ...; P 1}
ϕLQP (ap , ac , τ1 , τ2 ) = 0 ac − τ1 6 ap < ac + τ1 ,

−1 a c − τ2 6 ap < ac − τ1 ,



 N X M
1

−2 Otherwise ,
 X
τ2 = LSVi,j ; (11)
M × N i=1 j=1
(4)

where τ1 and τ2 are two user-specified parameters. where LSVPP −1is the local significant value given by
1
The LQP operator, by using the LTP’s concept, splits LSVc = P p=0 (|ac − ap |). The quinary pattern can
each quinary pattern into four parts: LQP−2 , LQP−1 , be divided into the four local binary patterns as follows:
LQP+1 and LQP+2 which can be computed as fol-
lows(cf. Eqs. 12, 13, 14 and 15): 7
X
ILQP−2 (I3×3
m,n ) = ϑ(ac − ap − τ2 ) × 2p (12)
p=0
7
X
LQP−2 (I3×3
m,n ) = ϑ(ac − ap − τ2 ) × 2p (5)
p=0 7 
X 
ILQP−1,−2 (I3×3
m,n ) = ϑ(ap − ac + τ2 ) × ϑ(ac − ap − τ1 )
p=0
7 
X  
LQP−1 (I3×3
m,n ) = ϑ(ap −ac +τ2 )×ϑ(ac −ap −τ1 ) ×2p | ϑ(ac − ap − τ2 ) × 2p
p=0
Directional Neighborhood Topologies based Multi-scale Quinary Pattern for Texture Classification 5

(13) Sk;k∈{0,1,2,3} of directional center-symmetric neigh-

boring pixels with respect to directions (kπ)
4 includ-
ing the central pixel ac . Mathematical definition of
7 
X  Sk is given by Eq. 16.
ILQP+1,+2 (I3×3
m,n ) = ϑ(ac − ap + τ2 ) × ϑ(ap − ac − τ1 )
p=0 (kπ)
 Direction : Sk = {ak , ac , ak+4 }; (16)
4
| ϑ(ap − ac − τ2 ) × 2p
On the other side, from the second row of Figure 2,
(14) it can be inferred that each peripheral pixel is sam-
pled with two of its sequential peripheral neighbor-
ing pixels in each direction. We consider four new
7
X sets Sek;k∈{0,1,2,3} of directional sequential periph-
ILQP+2 (I3×3
m,n ) = ϑ(ap − ac − τ2 ) × 2p (15) eral neighbors with respect to directions (2k+1)π .
4
p=0
Mathematical definition of Sek is given by Eq. 17.

3 Proposed method (2k + 1)π

Direction : Sek = {a2k , a2k+1 , a2k+2 }
4
3.1 Directional neighborhood topologies based (17)
multi-scale quinary patten
It appears from the literature that the average gray
Considering the fact that a texture image is defined level as well as the median of the grey-scale val-
as the local spatial variations in pixel orientation and ues are widely accepted statistical parameters for
intensities, we design directional neighborhood topolo- texture analysis. In view of this and aiming at en-
gies based multi-scale quinary patten (DNT-MQP) to hancing the thresholding range tolerance and thus
describe the spatial variations in pixel intensities and finding a code which is insensitive to noise and more
orientation in a local neighborhood in the image. The robust to illumination changes, several mean and
essence of DNT-MQP is to perform local sampling and median values (denoted as aµ , aµe , aSk , aSek , a% and
pattern encoding in the most informative directions con- a%e) are incorporated as virtual pixels in the mod-
tained within texture images. The construction process eling of the proposed texture model. aµ , aµe are re-
of the proposed DNT-MQP descriptor involves the fol- spectively the average local and global gray levels of
lowing stages: the whole 3 × 3 square neighborhood and the whole
image IM ×N , aSk and aSek are the average direc-
– STAGE #1 (Neighborhood topology): DNT-MQP
tional gray levels according to both sets Sk and Sek
descriptor considers a unit distance radius since clos-
and a% and a%e are the median of the grey-scale val-
est neighboring pixels maintains more discriminat-
ues of the 3 × 3 square region and the whole image
ing information for local texture descriptors keep-
IM ×N , respectively (cf. Eqs. 18, 19, 20 and 21).
ing thus low computational complexity. Thus, the
whole 3 × 3 grayscale image patch is adopted to
design the DNT-MQP method which intends to ex- P −1
1 X
plore the mutual information with respect to the ad- aµ = (ac + ap ) (18)
9 p=0
jacent neighbors. Consider the center pixel ac and
its 8 neighborhood pixels {a0 ,a1 ,...a7 }. On the ba-
sis of the above consideration, directional neighbor- M −1 N −1
1 X X
hood topologies adopted by DNT-MQP and which aµe = a(i,j) (19)
M × N i=0 j=0
are expected to better describe salient local texture
structure are conducted as shown in Figure 2. The
neighbors of a reference pixel ak are categorized ac- (kπ) (ak + ac + ak+4 )
cording to their angular position relative to that Direction : aSk = (20)
4 3
pixel. On the one side, from the first row of Fig-
ure 2, it can be inferred that the central pixel ac is
sampled each time with two pixels in each of four (2k + 1)π (a2k + 2a2k+1 + a2k+2 )
directions (kπ)
4 ; k ∈ {0,1,2,3}. We consider four sets
Direction
4
: aSek =
4
6 E. Rachdi et al.

(21) like coding scheme. The employed indicator func-

tion δ0 (·, ·) that converts the couple relationship in
The pixel samples of both Sk and Sek sets as well quinary form is defined as follows (cf. Eq. 24):
as the considered virtual pixels are then arranged
into two directional neighborhood topologies based
2, if (ax , ay ) ∈ SRU

Sampling Sets, denoted as SS1 and SS2 (cf. Eqs.  τ2
U

 1, if (ax , ay ) ∈ SRτ1

22 and 23). The sampling set SS1 is constructed by

considering the four couples of directional center- δ0 (ax , ay ) = −1, if (ax , ay ) ∈ SRL τ1 (24)
L
−2, ∈

symmetric pixels {(ak , ak+4 ); k ∈ {0, 1, 2, 3}}, the


 if (ax , a y ) SRτ 2

two couples of the average directional gray levels 0, otherwise
{(aS2k ,aS2k+1 ); k ∈ {0, 1}} and the couple formed Where τ1 and τ2 (τ2 > τ1 ) are two positive user-
by the average local and global gray levels (aµ and specified parameters (i.e., thresholds) which are in-
aµe ). In contrast, the sampling set SS2 is constructed troduced to alleviate the effects of external factors
by considering the four couples of pixels {(a2k+1 , aSek ); such as noise which destabilizes the patterns. SRU τ1 ,
k ∈ {0, 1, 2, 3}}, the couple formed by the average U L L
SRτ2 , SRτ1 and SRτ2 are four Sets of pixels Rela-
local and global gray levels (aµ , aµe ) and the cou-
tionship expressed in the following ways:
ple formed by the median of local and global the
grey-scale values (a% , a%e). Mathematical definitions
of both sampling sets SS1 and SS2 are given by Eqs. SRU
τ1 = {(ax , ay ) ∈ SS1 | (ax > ac +τ1 ) & (ay > ac −τ1 )}
22 and 23.
(25)

SS1 = {(aµ , aµe ), (aS0 , aS1 ), (aS2 , aS3 ), {(ak , ak+4 )}}
(22) SRL
τ1 = {(ax , ay ) ∈ SS1 | (ax 6 ac −τ1 ) & (ay 6 ac +τ1 )}
(26)

SS2 = {(a% , a%e), (aµ , aµe ), {(a2k+1 , aSek )}} (23)

SRU
τ2 = {(ax , ay ) ∈ SS2 | (ax > ac +τ2 ) & (ay > ac −τ2 )}
where k ∈ {0, 1, 2, 3}.
(27)

a3 a2 a1 a3 a2 a1 a3 a2 a1 a3 a2 a1

a4 ac a0 a4 ac a0 a4 ac a0 a4 ac a0 SRL
τ2 = {(ax , ay ) ∈ SS2 | (ax 6 ac −τ2 ) & (ay 6 ac +τ2 )}

a5 a6 a7 a5 a6 a7 a5 a6 a7 a5 a6 a7 (28)
Direction 3 π Direction π Direction π Direction 0 It is acknowledged that the related information of
4 2 4
a3 a2 a1 a3 a2 a1 a3 a2 a1 a3 a2 a1
the locality structure can be provided using posi-
tive and negative responses. The proposed descrip-
a4 ac a0 a4 ac a0 a4 ac a0 a4 ac a0
tor splits then, by using the LQP’s concept, each
a5 a6 a7 a5 a6 a7 a5 a6 a7 a5 a6 a7 quinary pattern into four distinct parts: two nega-
Direction π Direction 3 π Direction 5 π Direction 7 π tive (i.e., lower) and two positive (i.e., lower) parts
4 4 4 4 to generate four binary codes. With this convention,
the four thresholding functions that permit to ob-
Fig. 2: The adjacent neighbor relation for each of the tain the two lower and the two upper codes can be
8 neighbors. (Top) the four sets of directional expressed in the following ways (cf. Eqs. 29, 30, 31
center-symmetric neighboring pixels and (Down) the and 32):
four sets of directional peripheral neighboring pixels.
1, if (ax , ay ) ∈ SRU

τ1
δ+1 (ax , ay ) = (29)
0, otherwise
– STAGE #2 (Pattern encoding): The local texture
relationship between each couple of pixels within
1, if (ax , ay ) ∈ SRL

both sampling sets SS1 and SS2 and the central τ1
δ−1 (ax , ay ) = (30)
pixel, is encoded using a threshold values based LQP 0, otherwise
Directional Neighborhood Topologies based Multi-scale Quinary Pattern for Texture Classification 7

1, if (ax , ay ) ∈ SRU
 X
τ2 hESRL (k1 ) = b SRL (I3×3 ), k1 )
δ(E (38)
δ+2 (ax , ay ) = (31) τ1 τ 1
m,n
0, otherwise I3×3
m,n

1, if (ax , ay ) ∈ SRL
 X
τ2 hESRU (k2 ) = b SRU (I3×3 ), k2 )
δ(E (39)
δ−2 (ax , ay ) = (32) τ m,n
0, otherwise τ2
I3×3
2
m,n

The local information is encoded using two upper

and two lower directional binary dual-cross encoders hESRL (k2 ) =
X
b SRL (I3×3 ), k2 )
δ(E (40)
m,n
noted as ESRUτ , ESRUτ , ESRLτ and ESRLτ which are τ2 τ 2
1 2 1 2 I3×3
m,n
based on thresholding functions δ+1 , δ−1 , δ+2 and
δ−2 and associated to the four sets of pixels rela- where k1 ∈ [0, 27 ] is the number of ESRUτ and ESRLτ
1 1
tionship SRU U L L
τ1 , SRτ2 , SRτ1 and SRτ2 , respectively. patterns and k2 ∈ [0, 26 ] is the number of ESRUτ and
2
Formally, the codes produced by the four encoders
ESRL patterns. δ(·,·)
b denotes the Kronecker delta
ESRUτ , ESRUτ , ESRLτ and ESRLτ are computed as: τ2
1 2 1 2 function defined as below:

ESRUτ (I3×3 6
e) × 2 +
m,n ) = δ+1 (aµ , aµ 
1 1, if α = β
1 3 δ(α,
b β) = (41)
X
5−k
X
k
0, otherwise
δ+1 (aS2k , aS2k+1 ) × 2 + δ+1 (ak , ak+4 ) × 2
k=0 k=0 – STAGE #4 (Multi-scale scheme): At present, the
(33) histograms from single scale analysis realized with
the four directional binary dual-cross encoders ESRUτ ,
1
3×3
ESRLτ (Im,n ) = δ−1 (aµ , aµe ) × 2 +6 E SRL
τ1
, ESRUτ and ESRLτ cannot meet the require-
2 2
1
ments for ultimate recognition score.
1 3
X X Taking into consideration that various features have
δ−1 (aS2k , aS2k+1 ) × 25−k + δ−1 (ak , ak+4 ) × 2k
k=0 k=0
various capabilities to demonstrate images and to
(34) make them more robust to scale variations, the richer
detailed texture information can be captured by multi-
scale fusion operation that consists of performing
ESRUτ (I3×3
m,n ) = δ+2 (aµ , aµe ) × 25 + a linear combination of these features. In current
2

X3 study, a novel hybrid histogram is generated by com-

+ δ+2 (a% , a%e) × 24 + δ+2 (a2k+1 , aSek ) × 2k bining the information obtained by ESRUτ , ESRLτ ,
1 1
k=0 ESRUτ and ESRLτ encoders into a single vector fea-
(35) 2 2
ture. Note that this new hybrid texture description
model is expected to be more effective as it permits
ESRLτ (I3×3
m,n ) = δ (a
−2 µ , a µ
e ) × 25
+ δ (a
−2 % , a%e) × 24
+ to reduce the noise sensitivity and improves the dis-
2
3
crimination capability of ESRUτ , ESRLτ , ESRUτ and
1 1 2
ESRLτ operators through their complementary infor-
X
k
δ−2 (a2k+1 , aSek ) × 2 2
k=0 mations. The constructed histogram feature vector
(36) of multi-scale analysis is illustrated as follows:

– STAGE# 3 (Features extraction): After encoding D E

each pixel in the input texture image using the four hDNT-MQP = hESRU , hESRL , hESRU , hESRL (42)
τ1 τ1 τ2 τ2
directional binary dual-cross encoders ESRUτ , ESRLτ ,
1 1
ESRUτ and ESRLτ , four code maps are produced. The
2 2
histograms used as features representing the texture 3.2 Proposed dynamic thresholds for DNT-MQP
are generated from the obtained four code maps by
the following equations: Evidently, a parametric method evaluated with its user-
specified parameters optimized over each tested dataset
X leads to achieving satisfactory classification results. In
hESRU (k1 ) = b SRU (I3×3 ), k1 )
δ(E (37) this paper, to ensure a meaningful and reasonable com-
τ 1
m,n
τ1
3×3
Im,n parison with parameter-free state-of-the-arts methods,
8 E. Rachdi et al.

we propose to define locally and dynamically both pa- Algorithm 1: Computing DNT-MQP descrip-
rameters τ1 and τ2 for quinary pattern creation pro- tor
cess of our method. For that, we consider a local image Require: I ← input grayscale image IM ×N .
patch of size 3 × 3 and the neighbor to center difference Output: hDNT-MQP ← the multi-scale histogram feature.
denoted as df3×3 (cf. Eq. 43) is first calculated. Then, 1: Calculate the average global gray levels aµe of the whole
mean−
the mean of all negative difference values df3×3 and image IM ×N using Eq. 19
mean+ 2: Calculate the median of the grey-scale values a%e of the
the mean of all positive difference values df3×3 are
whole image IM ×N .
calculated from the df3×3 set using Eq. 44.
3: for Each image pixel ac of IM ×N do
4: Consider a local square window I3×3
m,n of dimension
3 × 3 around ac .
df3×3 = [a1 − ac , a2 − ac , ..., a7 − ac ] (43)
5: Calculate the neighbor to center difference df3×3 and
then calculate both dynamic thresholds τ1 and τ2
using Eq. 45.
6: Calculate the average directional gray levels aSk and
pv nv
mean + 1 X + mean − 1 X − aSek according to both sets Sk and Sek (cf. Eqs 16 and
df3×3 = dfk df3×3 = |dfk | 17) using Eqs. 20 and 21.
pv nv
k=1 k=1 7: Calculate (using Eqs. 33, 34, 35 and 36, respectively):
(44) – ESRU (I3×3
m,n ) ← the upper directional binary code
τ1
based on thresholding function δ+1 and
where dfk+ and dfk− are, respectively, the positive associated to the set of pixels relationship SRUτ1 .
and negative difference values in the df3×3 set, pv is – ESRL (I3×3
m,n ) ← the lower directional binary code
τ1
the number of dfk+ elements and nv is the number of based on thresholding function δ−1 and
dfk− elements (pv + nv = P). Finally, both parameters associated to the set of pixels relationship SRLτ1 .

τ1 and τ2 are measured using the following equation (cf. – ESRU (I3×3
m,n ) ← the upper directional binary code
τ2

Eq. 45): based on thresholding function δ+2 and

associated to the set of pixels relationship SRUτ2 .
– ESRL (I3×3
m,n ) ← the lower directional binary code
τ2

mean +
mean −
mean +
mean − based on thresholding function δ−2 and
|df3×3 − df3×3 | df3×3 + df3×3 associated to the set of pixels relationship SRL
τ1 = τ = τ2 .
+ −
mean , df mean ) 2 mean , df mean− )
+
max(df3×3 3×3 min(df3×3 3×3 8: end for
(45) 9: Calculate (using Eqs. 37, 38, 39 and 40, respectively):
– hESRU ← histogram feature of ESRU code map.
τ1 τ1
The pseudo-code of the proposed DNT-MQP de- – hESRL ← histogram feature of ESRL code map.
τ1 τ1
scriptor is illustrated in algorithm 1. – hESRU ← histogram feature of ESRU code map.
τ2 τ2

– hESRL ← histogram feature of ESRL code map.

τ2 τ2

10: Calculate the multi-scale histogram feature hDNT-MQP

3.3 Advantages of DNT-MQP and discussion using Eq. 42.

As a LBP and LQP texture operators variant, the de- 11: return hDNT-MQP
signed DNT-MQP descriptor preserves the same mer-
its concerning the invariance with the monotonic light-
ing change as well as the low complexity. Furthermore, It is easy to observe that DNT-MQP carries more infor-
DNT-MQP presents other merits that are discussed mation than those of the other descriptors which per-
herein below. mit to DNT-MQP to have better discriminative abil-
Compared to some recent state-of-the-art texture ity compared to LQPAT, LCCMSP and ARCSLBP.
descriptors like LQPAT [38], LCCMSP [40] and ARC- Moreover, the histograms of both images using DNT-
SLBP [51], which suffer from an inborn defect of LBP, MQP are considered as not similar since the distance
DNT-MQP descriptor can better describes local tex- (using L1-city block distance) between the two sam-
ture characteristics of the image with lower computa- ples, which is 1.1298 is higher compared to LQPAT,
tional complexity. In order to visually show the effec- LCCMSP and ARCSLBP where thier measured dis-
tiveness of the coding strategy of DNT-MQP, we illus- tances which are respectively 0.6173, 0.6924 and 0.7119,
trate in Figure 3 examples of histogram based-matching can be considered similar (lower than those obtained
of two sample texture images, selected from USPTex by DNT-MQP). To further highlight its effectiveness,
database and belonging to two distinct texture classes DNT-MQP is applied on various challenging represen-
using DNT-MQP, LQPAT, LCCMSP and ARCSLBP. tative texture datasets in the Section 4.
Directional Neighborhood Topologies based Multi-scale Quinary Pattern for Texture Classification 9

ing set, and the remaining 50% of the samples are re-
garded as the testing set. The samples of the test set
are then classified through the parameter-free nearest-
neighbor rule (1-NN) with L1-city block distance. We
repeat each experiment 100 times to remove any bias
related to the division of the dataset and the averaged
results are considered as estimated accuracies. In what
follows, the texture datasets considered in the experi-
ments are first presented and the obtained experimental
results are then discussed.

4.1 Texture datasets

For the sake of verification of the effectiveness and per-

formance stability of DNT-MQP, we carried out exten-
sive tests on sixteen well-known texture databases in-
cluding Jerry Wu, CUReT, KTH-TIPS, USPTex, Bro-
datz, KTH-TIPS2b, Bonn BTF, MondialMarmi, Ou-
TeX TC 00000, OuTeX TC 00001, OuTeX TC 00013,
Kylberg, XUH, 2D HeLa databases (the same datasets
used in [51, 43]), NewBarkTex1 and MBT2 databases.
These well-known datasets for texture classification (recog-
nition) were selected to cover various characteristics in
terms of number of samples, number of classes and class
homogeneity concerning perspective and scale.More in-
formation about each dataset is summarized in Table
Fig. 3: Comparison of DNT-MQP, LCCMSP, 2.
ARCSLBP and LQPAT features on two sample
texture images representing two distinct texture
classes from USPTex dataset, left and right columns 4.2 Comparative Assessment of Performance
present, respectively, the histograms of the images (a)
and (b) using the different descriptors. 4.2.1 Experiment #1: Investigation on Performance
Stability

Table 3 depicts the average classification scores (i.e.,

4 Experimental results and discussion over the 100 subdivisions) of each tested method and
for each texture database as well as the global average
In this phase of study, we show the effectiveness and val- performance (GAP) and the mean of standard devia-
idate the performance stability of the proposed DNT- tion (mean Std) of each method over all the datasets.
MQP operator with comprehensive tests on sixteen rep- Table 4 reports the average accuracy based ranking re-
resentative widely used texture databases. Furthermore, sults for each dataset. According to the results reported
DNT-MQP is compared with 23 recent most promising in both Tables 3 and 4, we can readily make the follow-
state-of-the-art texture descriptors to highlight its pro- ing observations:
vided performance improvement. The evaluated meth-
ods are summarized in Table 1. Note that, some of these – It is easy to observe that descriptors like LDENP,
methods are implemented using the original available LDV, DC and EU LLTP present the lowest perfor-
codes while for the others we have used our own im- mance on almost all the used dataset. Indeed, these
plementation according to their respective papers and methods are often found at the bottom of the rank-
which are fine tuned to get the same output results as ing as they achieve classification performance which
given in the published paper. The experiments herein 1
https://www-lisic.univ-littoral.fr/ poreb-
follow the standard evaluation protocol for each tested ski/BarkTex image test suite.html
dataset (i.e., split-sample validation) where 50% of the 2
https://multibandtexture.recherche.usherbrooke.ca/multi-
samples are randomly selected to be used as the train- band.html
10 E. Rachdi et al.

No. Name Acronym Dim. Ref. Year

1 Median robust extended local binary pattern MRELBP 200 [37] 2016
2 Statistical binary patterns SBP 400 [24] 2016
3 Local quadruple pattern LQPAT 29 [38] 2017
4 Locally encoded transform feature histogram LETRIST 413 [53] 2017
5 Local neighborhood difference pattern LNDP 28 [39] 2018
6 Centre symmetric quadruple pattern CSQP 28 [42] 2018
7 Local directional ZigZag pattern LDZP 354 [23] 2018
8 Mixed neighborhood topology cross decoded patterns MNTCDP 211 [45] 2018
9 Local directional ternary pattern LDTP 210 -2 [46] 2018
10 Local neighborhood intensity pattern LNIP 29 [22] 2018
11 Local diagonal extrema number pattern LDENP 24 -1 [27] 2018
12 Local concave-and-convex micro-structure patterns LCCMSP 21 1-2 [40] 2018
13 Local optimal-oriented pattern LOOP 28 [26] 2018
14 Repulsive-and-attractive local binary gradient contours RALBGC 210 -2 [43] 2018
15 Local directional edge binary patterns LDEBP 26 [25] 2019
16 Directional coding DC 225 [28] 2019
17 Local directional value LDV 29 [29] 2019
18 Attractive-and-repulsive center-symmetric local binary patterns ARCSLBP 28 [51] 2019
19 Multi kernel based local binary pattern KLBP 1280 [30] 2019
20 Cascaded asymmetric local pattern CALP 192 [31] 2019
21 Improved local quinary patterns ILQP 210 [32] 2019
22 Elliptical upper and lower local ternary pattern EU LLT P 25 [33] 2019
23 Multi-direction local binary pattern MDLBP 28 [52] 2019

Table 1: Summary of texture descriptors tested and compared with our proposed descriptors.

No. Name Classes Samples Total samples Sample resolution Image format Predefined
per class (pixels) train/test sets?
1 USPtex 191 12 2292 128×128 Colour (PNG) No
2 Brodatz 13 16 208 256×256 Monochrome (TIFF) No
3 KTH-TIPS2b 11 16 176 100×100 Monochrome (PNG) No
4 MondialMarmi 12 64 768 136×136 Monochrome (BMP) No
5 OuTeX TC 00000 24 20 480 128×128 Monochrome (RAS) Yes
6 OuTeX TC 00001 24 88 2112 64×64 Monochrome (RAS) Yes
7 OuTeX TC 00013 68 20 1360 128×128 Colour (BMP) No
8 CUReT 61 92 5612 200×200 Monochrome (PNG) No
9 KTH-TIPS 10 4 40 100×100 Monochrome (PNG) No
10 NewBarkTex 06 68 1632 64×64 Colour (BMP) No
11 Jerry Wu 39 4 156 256×256 Monochrome (BMP) No
12 Bonn BTF 10 16 160 200×200 and 64×64 Monochrome (JPEG) No
13 2D HeLa 10 20 200 Variable Monochrome (TIFF) No
14 XUH 25 40 1000 1280×960 Monochrome (PNG) No
15 Kylberg 28 160 4480 576×576 Monochrome (BMP) No
16 MBT 154 16 2464 160×160 Colour (TIFF) No

Table 2: Image datasets used in this study.

are mostly poor or at least lowest than the other including USPtex, MondialMarmi, KTH-TIPS2b, Ou-
evaluated descriptors. TeX TC 00013, NewBarkTex, 2D HeLa and MBT
– Results indicate that, except operators like LDENP, (datasets 1, 3, 4, 7, 10, 13, and 16 in Tables 3 and 4).
LDV, DC and EU LLTP which perform poorly, all The overall classification accuracy achieved on Ou-
the other evaluated methods show promising results TeX TC 00013, NewBarkTex, 2D HeLa and MBT
on Brodatz database (dataset 2 in Table 3) where are below 90%. It is worth mentioning that in this
their achieved score are above 96%. Descriptors like study the 1-NN classifier was used for classifica-
LDZP, KLBP, ILQP, RALBGC, LCCMSP, ARCS- tion but more complicated machine learning algo-
LBP , LDTP, MNTCDP as well as the designed rithms such as support vector machine (SVM) and
DNT-MQP operator manage to differentiate all classes extended nearest neighbor (ENN) may improve the
perfectly (score equal to 100%) on Brodatz database, overall performance.
leaving then essentially no room for improvement. – It is evident from the results reported in Table 3
This remark is also true for KTH-TIPS database that none of the evaluated methods ensure obtain-
(dataset 9 in Table 3) where DNT-MQP as well as ing satisfactory classification results over all the six-
several evaluated state-of-the-art descriptors man- teen tested datasets. Considering for example the
age to differentiate all classes perfectly. DC descriptor, it performs nicely on Brodatz, KTH-
– It can be found that there is a notable performance TIPS (it gets a score of 100%) and Kylberg datasets
drop for all the tested operators on several datasets but shows poor performance on the others tested
 Directional Neighborhood Topologies based Multi-scale Quinary Pattern for Texture Classification 11

Texture Dataset GAP mean

descriptor 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Std
DNT-MQP 92.46 100.00 95.12 93.64 99.97 99.70 86.59 95.52 100.00 86.98 99.88 99.25 71.19 98.11 99.92 86.91 94,07 0.89
SBP2 89.28 99.92 90.89 89.71 98.91 97.65 84.24 96.45 99.45 64.56 98.74 99.91 74.38 98.40 99.71 76.63 91,17 1.11
LETRIST 89.34 99.99 90.08 86.99 99.92 98.51 83.65 98.00 100.00 63.12 100.00 100.00 72.22 97.99 99.84 81.39 91,31 0.98
LCCMSP 88.42 100.00 93.32 90.83 99.97 99.68 84.44 94.69 100.00 84.72 98.18 97.64 68.98 96.46 99.69 85.41 92,65 1.05
MDLBP 69.47 99.23 77.99 76.33 99.63 97.12 74.77 81.65 99.60 65.53 94.55 96.45 45.98 91.16 98.75 72.93 83,82 1.48
LOOP 79.47 99.97 85.81 79.33 99.49 97.83 76.48 90.82 100.00 70.00 96.86 98.78 58.81 93.71 97.75 83.99 88,06 1.21
LDEBP 84.85 99.96 87.30 82.93 99.27 98.19 80.66 89.56 100.00 71.13 98.87 92.96 56.54 94.97 99.66 84.15 88,81 1.20
ARCS-LBP 86.89 99.88 93.23 90.95 99.65 99.31 85.29 94.32 100.00 78.66 99.53 99.17 73.67 96.95 99.80 83.23 92,53 1.06
LDENP 44.99 95.40 65.80 57.87 95.12 81.66 63.61 64.82 88.80 54.10 88.50 88.29 42.79 78.61 87.13 60.40 72,37 2.17
LDTP 83.00 100.00 90.47 86.00 99.76 99.20 80.32 91.81 100.00 72.66 98.23 99.88 62.28 94.30 97.74 82.97 89,91 1.09
LDZP 70.27 100.00 88.47 83.38 98.91 97.84 70.27 85.14 99.75 67.53 96.74 97.14 33.35 88.88 92.54 84.18 84,65 1.36
DC 61.09 99.48 79.78 67.54 92.91 88.13 68.29 81.05 100.00 51.86 94.95 93.36 64.32 93.14 97.92 56.24 80,63 1.4
LDV 45.24 96.21 59.16 56.06 95.08 75.97 60.54 71.05 87.65 54.05 82.64 94.61 25.98 81.41 91.85 59.72 71,08 2.13
KLBP 85.07 100.00 91.24 86.63 99.54 99.19 79.90 91.38 100.00 81.19 97.49 97.34 50.95 94.55 98.60 86.78 89,99 1.13
LNIP 76.48 99.73 82.97 83.05 99.02 98.14 76.10 86.18 98.45 68.02 96.86 77.86 38.42 90.63 96.55 83.99 84,53 1.49
CALP 75.00 99.40 82.63 75.28 98.99 96.41 76.44 92.66 98.15 72.99 95.76 95.83 54.56 90.70 98.72 84.30 86,74 1.54
ILQP 87.54 100.00 93.39 90.07 99.63 99.32 80.12 91.22 100.00 83.85 98.03 98.72 59.94 95.64 98.33 85.80 91,35 1.09
EU LLT P 64.95 99.38 78.63 69.33 96.68 90.92 66.91 73.72 93.90 69.81 92.58 93.40 37.73 88.89 94.87 71.75 80,22 1.76
MRELBP 83.33 99.94 90.51 86.53 98.90 97.10 82.02 93.08 97.85 64.01 99.05 99.26 73.66 97.60 99.78 75.36 89,87 1.29
MNTCDP 85.73 100.00 90.93 91.72 99.93 99.57 82.97 84.32 100.00 71.03 99.36 100.00 63.68 94.99 97.45 78.95 90,04 0.98
LNDP 80.18 99.97 88.72 80.92 99.48 96.85 75.97 91.31 100.00 72.79 97.44 93.85 49.38 93.55 97.89 87.49 87,86 1.27
LQPAT 77.13 99.93 83.76 81.60 99.81 97.67 76.89 90.70 99.60 76.28 97.28 94.94 51.22 91.49 96.92 84.73 87,50 1.34
CSQP 72.61 99.93 80.12 74.87 99.10 96.92 74.90 88.51 100.00 69.63 98.42 95.26 46.30 90.79 96.87 81.84 85,38 1.3
RALBGC 86.94 100.00 93.25 90.58 99.83 99.22 82.69 93.96 100.00 83.43 97.92 98.36 64.88 95.57 99.22 87.40 92,08 1.07

Table 3: Average classification accuracy. The last two columns represent, respectively, the global average
performance (GAP) and the mean standard deviation of each method over all the datasets.

Datasets
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

DNT- DNT- DNT- DNT- DNT- DNT- DNT- LETRIST DNT- DNT- LETRIST LETRIST SBP2 SBP2 DNT- LNDP
MQP MQP MQP MQP MQP MQP MQP MQP MQP MQP
LETRIST LCCMSP ILQP MNTCDP LCCMSP LCCMSP ARCS- SBP2 LCCMSP LCCMSP DNT- MNTCDP ARCS- DNT- LETRIST RALBGC
LBP MQP LBP MQP
SBP2 KLBP LCCMSP ARCS- MNTCDP MNTCDP LCCMSP DNT- LETRIST ILQP ARCS- SBP2 MRELBP LETRIST ARCS- DNT-
LBP MQP LBP LBP MQP
LCCMSP ILQP RALBGC LCCMSP LETRIST ILQP SBP2 LCCMSP LDEBP RALBGC MNTCDP LDTP LETRIST MRELBP MRELBP KLBP
ILQP RALBGC ARCS- RALBGC RALBGC ARCS- LETRIST ARCS- DC KLBP MRELBP MRELBP DNT- ARCS- SBP2 ILQP
LBP LBP LBP MQP LBP
RALBGC LDZP KLBP ILQP LQPAT RALBGC MNTCDP RALBGC KLBP ARCS- LDEBP DNT- LCCMSP LCCMSP LCCMSP LCCMSP
LBP MQP
ARCS- LDTP MNTCDP SBP2 LDTP LDTP RALBGC MRELBP ILQP LQPAT SBP2 ARCS- RALBGC ILQP LDEBP LQPAT
LBP LBP
MNTCDP MNTCDP SBP2 LETRIST ARCS- KLBP MRELBP CALP LNDP CALP CSQP LOOP DC RALBGC RALBGC CALP
LBP
KLBP LETRIST MRELBP KLBP MDLBP LETRIST LDEBP LDTP CSQP LNDP LDTP ILQP MNTCDP MNTCDP MDLBP LDZP
LDEBP LOOP LDTP MRELBP ILQP LDEBP LDTP KLBP RALBGC LDTP LCCMSP RALBGC LDTP LDEBP CALP LDEBP
MRELBP LNDP LETRIST LDTP KLBP LNIP ILQP LNDP LOOP LDEBP ILQP LCCMSP ILQP KLBP KLBP LOOP
LDTP LDEBP LNDP LDZP LOOP LDZP KLBP ILQP ARCS- MNTCDP RALBGC KLBP LOOP LDTP ILQP LNIP
LBP
LNDP MRELBP LDZP LNIP LNDP LOOP LQPAT LOOP LDTP LOOP KLBP LDZP LDEBP LOOP DC ARCS-
LBP
LOOP LQPAT LDEBP LDEBP LDEBP LQPAT LOOP LQPAT MNTCDP EU LLTP LNDP MDLBP CALP LNDP LNDP LDTP
LQPAT CSQP LOOP LQPAT CSQP SBP2 CALP LDEBP LDZP CSQP LQPAT CALP LQPAT DC LOOP CSQP
LNIP SBP2 LQPAT LNDP LNIP MDLBP LNIP CSQP MDLBP LNIP LOOP CSQP KLBP LQPAT LDTP LETRIST
CALP ARCS- LNIP LOOP CALP MRELBP LNDP LNIP LQPAT LDZP LNIP LQPAT LNDP MDLBP MNTCDP MNTCDP
LBP
CSQP LNIP CALP MDLBP SBP2 CSQP CSQP LDZP SBP2 MDLBP LDZP LDV CSQP CSQP LQPAT SBP2
LDZP DC CSQP CALP LDZP LNDP MDLBP MNTCDP LNIP SBP2 CALP LNDP MDLBP CALP CSQP MRELBP
MDLBP CALP DC CSQP MRELBP CALP LDZP MDLBP CALP MRELBP DC EU LLTP LDENP LNIP LNIP MDLBP
EU LLTP EU LLTP EU LLTP EU LLTP EU LLTP EU LLTP DC DC MRELBP LETRIST MDLBP DC LNIP EU LLTP EU LLTP EU LLTP
DC MDLBP MDLBP DC LDENP DC EU LLTP EU LLTP EU LLTP LDENP EU LLTP LDEBP EU LLTP LDZP LDZP LDENP
LDV LDV LDENP LDENP LDV LDENP LDENP LDV LDENP LDV LDENP LDENP LDZP LDV LDV LDV
LDENP LDENP LDV LDV DC LDV LDV LDENP LDV DC LDV LNIP LDV LDENP LDENP DC

Table 4: Ranking results on each tested dataset. The proposed texture operator is highlighted in shadow.

datasets. We can express the same remark for many of around 23.86%) realized with the top 1 descrip-
other evaluated methods. LETRIST descriptor which tor (i.e., the proposed DNT-MQP descriptor). The
is ranked as the 4th best descriptor as will be shown same comment is true for many other methods such
later, it allows achieving good classification results as CSQP, LOOP, LDEBP and so on.
on the majority of the considered datasets but its – It can also be seen in Tables 3 and 4 that DNT-MQP
weaknesses appeared in classifying the images of the realizes a performance that is competitive or better
NewBarkTex dataset as the achieved score drops than all the evaluated state-of-the-art descriptors.
dramatically and is about 63% vs 86.98% (difference Furthermore, according to literature review it can
12 E. Rachdi et al.

be found that DNT-MQP universally is among one 24 evaluated texture descriptors considered in this ex-
of the best descriptors in terms of overall accuracy periment on the sixteen used texture datasets. Table
as it achieves outstanding classification outcomes on 5 gathers the obtained ranking results according to the
nearly all the selected datasets. The striking perfor- normalized number of victories (number of wins/(number
mance of NT-MQP can be judged by the fact that of tested datasets*(number of tested descriptors - 1)))
it realizes the highest global average performance achieved by each evaluated method on all the tested
(GAP) against all the evaluated methods. datasets. It is easy to observe from the results shown
– Noticeably, considering the ranking between the tested in Table 5 that the designed DNT-MQP operator is
descriptors within each used dataset (cf. Table 4), clearly the best performing descriptor in comparison
DNT-MQP stands out as the best descriptor as it with the evaluated state-of-the-art methods, which cor-
performs consistently and significantly the best for roborate all the result analysis extracted from Tables 3
ten texture datasets: USPtex, Brodatz, KTH-TIPS2b, and 4. Remarkably, the normalized number of victo-
MondialMarmi, OuTeX TC 00000, OuTeX TC 00001, ries achieved by DNT-MQP is 0.87228 vs. 0.75815 with
OuTeX TC 00013, KTH-TIPS, NewBarkTex and LCCMSP (top 2nd ), vs. 0.74185 with ARCS-LBP (top
Kylberg datasets (1-7, 9, 10, and 15, respectively). 3th ), vs. 0.70924 with LETRIST (top 4th ), vs. 0.70652
Furthermore, it is in the top 3 methods on four with RALBGC (top 5th ) vs. 0.65217 with ILQP (top
tested datasets. It is interesting to note that when 6th ), etc. Particularly, if we consider the classification
DNT-MQP does not achieve the highest specific scores performance of ARCS-LBP (the best second descrip-
(i.e., it is not the top texture descriptor), it provides tor) as baseline, DNT-MQP texture operator provides
an interesting competitive average performance com- about 15.05% improvement over the sixteen tested tex-
pared to the score yielded by the top one texture op- ture datasets.
erator. Considering for example Bonn BTF dataset
(number 12) where DNT-MQP has the sixth-highest Ranking 1-NN Texture descriptor Victories/ Dimension
average performance (i.e., ranked at the 6th posi- comparisons
1 DNT-MQP 0.87228 384
tion), it allowed to reach a score of 99.25% which is 2 LCCMSP 0.75815 2046
3 ARCS-LBP 0.74185 256
considered as very satisfactory classification result 4 LETRIST 0.70924 413
5 RALBGC 0.70652 1022
(very close to the score of the descriptors ranked 6 ILQP 0.65217 1024
7 MNTCDP 0.63043 2048
before it: 100% for the top one). 8 SBP2 0.62500 400
9 LDTP 0.57065 1022
10 KLBP 0.56793 1280
Considering the above discussions, it can be noticed 11
12
MRELBP
LDEBP
0.54076
0.49457
200
64
that the designed DNT-MQP operator shows a signif- 13 LOOP 0.42935 256
14 LNDP 0.42663 256
icant performance stability over the evaluated state- 15 LQPAT 0.42120 512
16 CALP 0.34239 192
of-the-art texture methods on almost all the consid- 17 LDZP 0.30435 354
18 CSQP 0.29348 256
19 LNIP 0.27174 512
ered datasets. Indeed, DNT-MQP has lower oscillation 20 MDLBP 0.26902 256
21 DC 0.18207 225
of performance as it keeps a consistent performance 22 EU LLT P 0.12228 32
23 LDENP 0.03804 15
throughout various texture databases while most of the 24 LDV 0.03804 512
literature methods oscillate. The performance stability
can also be confirmed by the fact that DNT-MQP has Table 5: Ranking outcomes based on the normalized
achieved the lowest mean standard deviation, corrob- number of victories achieved by each applied method
orating to its robustness against any bias related to on all the used datasets after applying the
the division of the datasets. These indicate that the Wilcoxon-based ranking test.
combination of single scale DNT-QP features helped to
construct a texture operator that performs better on a
wide selection of various texture datasets.

4.2.2 Experiment #2: Statistical significance of the 4.3 Research Implementation

obtained results in terms of accuracy improvement
A laptop, equipped with a 2.10 GHz Core i7 CPU, 8
The major reason to carry out this experiment is to GB of RAM and having Ubuntu 14.04 trusty operating
further validate statistically the obtained classification system was used in the experiments. We implemented
results via the designed method vs the evaluated state- all the methods using MATLAB R R2013a. The imple-
of-the-art descriptors using the Wilcoxon signed rank mentation of the used methods needed approximately
test based ranking technique proposed in [40]. The tech- 38.3 hours. Figure 4 demonstrates the required time
nique is applied on all the pairwise combinations of the for processing (in minutes) containing time for feature
Directional Neighborhood Topologies based Multi-scale Quinary Pattern for Texture Classification 13

extraction, calculation of distance and also 1-NN clas- References

sification applied to 23140 images in the 16 datasets
1. Liu, L., Chen, J., Fieguth, P., Zhao, G., Chellappa, R.,
used in this study. According to Figure 4, it can be seen
Pietikäinen, M. (2019). From BoW to CNN: Two decades
that the proposed DNT-MQP method is faster than the of texture representation for texture classification. Inter-
five best performing descriptors ranked after it which national Journal of Computer Vision, 127(1), 74-109.
confirms that DNT-MQP makes an admissible tradeoff 2. B.S. Manjunath and W. Y. Ma. Texture features for
browsing and retrieval of image data. IEEE Trans. Pat-
between speed and classification performance. tern Anal. Mach. Intell., vol. 18, no. 8, pp. 837–842, Aug.
1996
3. L.S. Davis. Polarograms: A new tool for image texture
PROCESSING TIME (IN MINUTES) analysis. Pattern Recognition 13(3), (1981), pp. 219-223.
4. T. Ojala, M. Pietikainen, and T. Maenpaa. Multireso-

433.5
420.4
407.9
lution gray-scale and rotation invariant texture classifi-
cation with local binary patterns. IEEE Trans. Pattern
TIME (IN MINUTES)

Anal. Mach. Intell., vol. 24, no. 7, pp. 971-987, 2002.

5. J.Duvernoy, Optical digital processing of directional ter-
218.9
rain textures invariant under translation, rotation, and
129.3
116.3

change of scale, Appl. Opt. 23(6), (1984), pp. 828-837.

82.3

6. R.K. Goyal, W.L. Goh, D.P. Mitaletal. Scale and rotation

52.4
47.5

67
28.7
27.5
24.1

45
21.7
20.8
19.9
17.1
16.5
13.8
12.6
10.3

invariant texture analysis based on structural property.

9.5
8.5

IEEE International Conference on Industrial Electronics,

Control, and Instrumentation, 1-2, 1995, pp. 1290-1294.
7. G. Eichmann, T. Kasparis. Topologically invariant tex-
DATASET ture descriptors. Comput. Vision Graphics Image Pro-
cess. 41(3), (1988), pp. 267-281.
8. R.L. Kashyap, A. Khotanzad. A model-based method
Fig. 4: Processing time (in minutes) of the tested for rotation invariant texture classification. IEEE Trans.
texture descriptors over all considered dataset. Pattern Anal. Mach. Intell. 8(4), (1986), pp. 472–481.
9. F.S. Cohen, Z.G. Fan, M.A. Patel. Classification of ro-
tated and scaled textured images using Gaussian Markov
random field models. IEEE Trans. Pattern Anal. Mach.
Intell. 13(2), (1991), pp. 192–202.
10. R. Porter, N. Canagarajah. Robust rotation invariant
texture classification 1997. IEEE International Confer-
5 Conclusion ence on Acoustics, Speech, and Signal Processing, I-V,
1997, pp. 3157-3160.
11. Liu, L. and Fieguth, P. (2012). Texture classification from
Herein, motivated by several state-of-the-art LBP-like random features. IEEE Transactions on Pattern Analysis
methods, a conceptually simple, easy-to-implement yet and Machine Intelligence, 34(3), 574-586.
highly discriminative texture operator so-called Direc- 12. Bianconi, F., Fernández, A., González, E. and Armesto,
J. (2009). Robust color texture features based on ran-
tional Neighborhood Topologies based Multi-scale Quinary
klets and discrete Fourier transform. Journal of Elec-
Pattern (DNT-MQP) is designed for texture descrip- tronic Imaging, 18(4), 043012-043012.
tion and classification. DNT-MQP operator capture, 13. Nanni, L., Lumini, A., Brahnam, S.: Local binary pat-
thanks to the use of several directional neighborhood terns variants as texture descriptors for medical image
analysis. Artif. Intell. Med. 49(2), 117–125 (2010)
topologies, richer detailed and complementary texture 14. M.Varma, A. Zisserman. A statistical approach to ma-
information. As it has been proved through the experi- terial classification using image patch exemplars. IEEE
mental results, DNT-MQP showed its high ability to Trans. Pattern Anal. Mach. Intell. 31(11), (2009),pp.
differentiate, with high precision, different classes of 2032–2047.
15. Y. Xu, H. Ji, C. Fermuller. A projective invariant for tex-
a great number of publicly available texture datasets. ture. International Conference on Computer Vision and
Furthermore, the discriminative power of DNT-MQP Pattern Recognition, 2006, pp.1932-1939.
is proved against 23 recent state-of-the-art texture de- 16. Y. Xu, H. Ji, C. Fermuller. View point invariant texture
description using fractal analysis. Int .J. Comput. Vision
scriptors, indicating that DNT-MQP is a strong candi- 83 (1), (2009), pp. 85-100.
date for texture modeling. Future work will concern the 17. Guo, Z., Li, Q., You, J., Zhang, D., Liu, W. (2012). Local
investigation of the use of other well known classifiers directional derivative pattern for rotation invariant tex-
to increase classification performance of the proposed ture classification. Neural Computing and Applications,
21(8), 1893-1904.
model. In addition, the proposed texture operator holds 18. Guo, Z., Zhang, L., Zhang, D. (2010). Rotation invariant
potential for deployment across high-level challenging texture classification using LBP variance (LBPV) with
applications related to texture classification including global matching. Pattern recognition, 43(3), 706-719
19. Guo, Z., Zhang, L., Zhang, D. (2010). A completed mod-
dynamic texture classification, background subtraction
eling of local binary pattern operator for texture classi-
in complex scenes, object recognition and video-based fication. IEEE transactions on image processing, 19(6),
face analysis and so on. 1657-1663
14 E. Rachdi et al.

20. Kas, M., El-merabet, Y., Ruichek, Y., Messoussi, R. 38. Chakraborty, S., Singh, S. K., & Chakraborty, P. (2017).
(2019). A comprehensive comparative study of hand- Local quadruple pattern: A novel descriptor for facial im-
crafted methods for face recognition LBP-like and non age recognition and retrieval. Computers & Electrical En-
LBP operators. Multimedia Tools and Applications, 1- gineering.
39. 39. Verma, M., & Raman, B. (2017). Local neighborhood dif-
21. X. Y.Tan , B. Triggs , Enhanced local texture feature ference pattern: A new feature descriptor for natural and
sets for face recognition under difficult lighting condi- texture image retrieval. Multimedia Tools and Applica-
tions, IEEE Trans. Image Process. 19 (6) (2010) 1635- tions, 1-24.
1650. 40. El-merabet, Y., Ruichek, Y. (2018). Local concave-and-
22. Banerjee, P., Bhunia, A. K., Bhattacharyya, A., Roy, P. convex micro-structure patterns for texture classification.
P., Murala, S. (2018). Local Neighborhood Intensity Pat- Pattern Recognition, 76, 303-322.
tern–A new texture feature descriptor for image retrieval. 41. Ji, L., Ren, Y., Pu, X., Liu, G. (2018). Median lo-
Expert Systems with Applications, 113, 100-115. cal ternary patterns optimized with rotation-invariant
23. Roy, S. K., Chanda, B., Chaudhuri, B. B., Banerjee, S., uniform-three mapping for noisy texture classification.
Ghosh, D. K., Dubey, S. R. (2018). Local directional Pattern Recognition, 79, 387-401.
ZigZag pattern: A rotation invariant descriptor for tex- 42. Chakraborty, S., Singh, S. K., Chakraborty, P. (2018).
ture classification. Pattern Recognition Letters, 108, 23- Centre symmetric quadruple pattern: A novel descriptor
30. for facial image recognition and retrieval. Pattern Recog-
24. Nguyen, T. P., Vu, N. S., Manzanera, A. (2016). Sta- nition Letters, 115, 50-58.
tistical binary patterns for rotational invariant texture 43. El Khadiri, I., Kas, M., El Merabet, Y., Ruichek, Y.,
classification. Neurocomputing, 173, 1565-1577. Touahni, R. (2018). Repulsive-and-attractive local binary
25. Sucharitha, G., Senapati, R. K. (2019). Biomedical image gradient contours: New and efficient feature descriptors
retrieval by using local directional edge binary patterns for texture classification. Information Sciences, 467, 634-
and Zernike moments. Multimedia Tools and Applica- 653.
tions, 1-18. 44. Dubey, S. R., Singh, S. K., Singh, R. K. Local Wavelet
26. Chakraborti, T., McCane, B., Mills, S., Pal, U. (2018). Pattern: A New Feature Descriptor for Image Retrieval
LOOP Descriptor: Local Optimal-Oriented Pattern. in Medical CT Databases. IEEE Trans. Image Process,
IEEE Signal Processing Letters, 25(5), 635-639. 24(12): 5892-5903, 2015
27. Pillai, A., Soundrapandiyan, R., Satapathy, S., Satapa- 45. Kas, M., Ruichek, Y., Messoussi, R. (2018). Mixed neigh-
thy, S. C., Jung, K. H., Krishnan, R. (2018). Local diago- borhood topology cross decoded patterns for image-based
nal extrema number pattern: a new feature descriptor for face recognition. Expert Systems with Applications, 114,
face recognition. Future Generation Computer Systems, 119-142.
81, 297-306. 46. El-khadiri, I., Chahi, A., El-merabet, Y., Ruichek, Y.,
28. Ouslimani, F., Ouslimani, A., Ameur, Z. (2019). Touahni, R. (2018). Local directional ternary pattern: A
Rotation-invariant features based on directional coding new texture descriptor for texture classification. Com-
for texture classification. Neural Computing and Appli- puter Vision and Image Understanding, 169, 14-27.
cations, 31(10), 6393-6400. 47. Hassaballah, M., Alshazly, H. A., Ali, A. A. (2019). Ear
29. Wu, X., Sun, J. (2019). Face recognition based on multi- recognition using local binary patterns: A comparative
scale local directional value. Multimedia Tools and Ap- experimental study. Expert Systems with Applications,
plications, 1-17. 118, 182-200.
30. Tuncer, T., Dogan, S. (2019). Pyramid and multi kernel 48. Tuncer, T., Dogan, S., Ataman, V. (2019). A novel and
based local binary pattern for texture recognition. Jour- accurate chess pattern for automated texture classifica-
nal of Ambient Intelligence and Humanized Computing, tion. Physica A: Statistical Mechanics and its Applica-
1-12. tions, 536, 122584.
31. Chakraborty, S., Singh, S. K., Chakraborty, P. (2019). 49. Lan, R., Lu, H., Zhou, Y., Liu, Z., Luo, X. (2019). An
Cascaded asymmetric local pattern: a novel descriptor LBP encoding scheme jointly using quaternionic repre-
for unconstrained facial image recognition and retrieval. sentation and angular information. Neural Computing
Multimedia Tools and Applications, 1-20. and Applications, 1-7.
32. Armi, L., Fekri-Ershad, S. (2019). Texture image Classi- 50. Luo, Q., Fang, X., Sun, Y., Liu, L., Ai, J., Yang, C.,
fication based on improved local Quinary patterns. Mul- Simpson, O. (2019). Surface Defect Classification for Hot-
timedia Tools and Applications, 1-24. Rolled Steel Strips by Selectively Dominant Local Binary
33. Kabbai, L., Abdellaoui, M., Douik, A. (2019). Image clas-
Patterns. IEEE Access, 7, 23488-23499.
sification by combining local and global features. The Vi- 51. El-merabet, Y., Ruichek, Y. (2019). Attractive-and-
sual Computer, 35(5), 679-693. repulsive center-symmetric local binary patterns for tex-
34. Liu, J., Chen, Y., Sun, S. (2019). A novel local texture
ture classification. Engineering Applications of Artificial
feature extraction method called multi-direction local bi-
Intelligence, 78, 158-172.
nary pattern. Multimedia Tools and Applications, 1-16. 52. Liu, J., Chen, Y., Sun, S. (2019). A novel local texture
35. Shih, H. C., Cheng, H. Y., Fu, C. (2020). Image Classifi-
feature extraction method called multi-direction local bi-
cation Using Synchronized Rotation Local Ternary Pat-
nary pattern. Multimedia Tools and Applications, 1-16.
tern. IEEE Sensors Journal.
53. Song, T., Li, H., Meng, F., Wu, Q., Cai, J. (2018).
36. Bhattacharjee, D., Roy, H. (2020). Pattern of Local Grav-
LETRIST: Locally encoded transform feature histogram
itational Force (PLGF): A novel Local Image Descriptor.
for rotation-invariant texture classification. IEEE Trans-
IEEE transactions on pattern analysis and machine in-
actions on circuits and systems for video technology,
telligence.
37. Liu, L., Lao, S., Fieguth, P. W., Guo, Y., Wang, X., 28(7), 1565-1579.
Pietikäinen, M. (2016). Median robust extended local bi- 54. Fernández, A., Álvarez, M.X., Bianconi, F. Image classi-
nary pattern for texture classification. IEEE Transactions fication with binary gradient contours. Opt. Lasers Eng.
on Image Processing, 25(3), 1368-1381. 49(9-10), 1177-1184 (2011).