See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/348495185

Local Tetra-Directional Pattern–A New Texture Descriptor for Content-Based

Image Retrieval

Article in Pattern Recognition and Image Analysis · October 2020

DOI: 10.1134/S1054661820040057

CITATIONS READS
4 161

2 authors:

Anterpreet Kaur Bedi Ramesh Kumar Sunkaria

Indian Institute of Technology Ropar National Institute of Technology Jalandhar
16 PUBLICATIONS 26 CITATIONS 83 PUBLICATIONS 514 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Enhanced local Tetra pattern View project

All content following this page was uploaded by Anterpreet Kaur Bedi on 18 January 2021.

The user has requested enhancement of the downloaded file.

 MATHEMATICAL THEORY OF IMAGES AND SIGNALS REPRESENTING,
PROCESSING, ANALYSIS, RECOGNITION, AND UNDERSTANDING

Local Tetra-Directional Pattern–A New Texture Descriptor

for Content-Based Image Retrieval
Anterpreet Kaur Bedia,* and Ramesh Kumar Sunkariaa,**
a
Department of Electronics and Communication Engineering, Dr. B. R. Ambedkar National Institute of Technology,
Jalandhar, Punjab, 144011 India
*e-mail: [email protected]
**e-mail: [email protected]

Abstract—In this present work, a new technique for content-based image retrieval is introduced using local
tetra-directional pattern. In conventional local binary pattern (LBP), each pixel of an image is changed into
a specific binary pattern in accordance with their relationship with neighbouring pixels. Texture feature
descriptor introduced in this work differs from local binary pattern as it exploits local intensity of pixels in four
directions in the neighbourhood. Also, colour feature and gray level co-occurrence matrix have been applied
in this work. Median of images have also been taken under consideration to keep the edge information pre-
served. The proposed technique has been validated experimentally by conducting experiments on two differ-
ent sets of data, viz., Corel-1K and AT&T. Performance was measured using two well-known parameters,
precision and recall, and further comparison was carried with some state-of-the-art local patterns. Compar-
ison of results show substantial improvement in the proposed technique over existing methods.

Keywords: content based image retrieval, local tetra-directional pattern, feature vector, gray level co-occur-
rence matrix, local binary pattern
DOI: 10.1134/S1054661820040057

1. INTRODUCTION retrieval techniques. Also, it performs more efficiently

and at a faster rate.
With the fast development in technology, there is
an expansive availability of online and offline images Various CBIR techniques have been formulated so
in several fields. Coping with such vast databases by far. Some of these include NeTra [3], QBIC [4], SIM-
human annotation requires large amount of efforts. PLIcity [5], MetaSEEK [6], VisualSeek [7], and
This creates a demand for developing an effective DRAWSEARCH [8]. A descriptive survey of several
technique in order to assist in searching for a desired content based image retrieval techniques has been
image automatically from ever increasing databases. provided in [2, 9, 10]. Several low-level primitive fea-
Image retrieval is an active research area in the field of tures are used in CBIR techniques. These include
image processing, which is used to retrieve images color, shape, texture, spatial relationship etc. [10, 11].
similar to a particular image, known as query, from a Color is considered as one of the most significantly
large dataset. The process of image retrieval can be used feature descriptor. It can be treated as distribu-
executed in two ways: text-based image retrieval and tion of intensity in distinct color channels. Taking
content-based image retrieval (CBIR). In order to human perception into consideration, the color spaces
retrieve images using text, features like keywords, text that are much in use include RGB, LAB, HSV, and
and metadata are used. Since text-based image YCrCb. HSV color space is closer to human percep-
retrieval demands manual indexing of images, hence it tion of color and is also invariant to illumination.
becomes time exhausting. Thus, this traditional Thus, it is preferable over other color spaces. Color
method becomes inefficient and incomplete in case of features can be extracted using various statistical mea-
larger database. CBIR was initially introduced by Kato sures like color histograms, color moments, color
in 1992 [1]. CBIR extracts low-level visual features for moments, color-covariance matrix, color correlo-
indexing images for retrieval. These extracted features grams etc. [12, 13]. The most commonly used color
are further evaluated for similarity with the query in feature is histogram, presented by Swain and Ballard
order to retrieve relevant images [2]. Being closer to [14]. Color histogram generates a feature vector by
human perception of visual data, CBIR is considered estimating the count of occurrence of each intensity in
to be more advantageous than the earlier text-based distinct channels. The spatial relationship among
image intensities is represented by color correlogram
[15]. Another feature, known as color coherence vec-
Received March 5, 2020; revised March 5, 2020; tor [16], uses the coherence and incoherence proper-
accepted May 27, 2020 ties of image pixel color in order to create a feature

ISSN 1054-6618, Pattern Recognition and Image Analysis, 2020, Vol. 30, No. 4, pp. 578–592. © Pleiades Publishing, Ltd., 2020.
 LOCAL TETRA-DIRECTIONAL PATTERN–A NEW TEXTURE DESCRIPTOR 579

vector. Motif co-occurrence matrix was further pro- local binary co-occurrence pattern was introduced by
posed in which, a 3D matrix was constructed corre- Verma et al. [31] in which local information was
sponding to the local statistics of image [17]. attained by extracting center symmetric LBP. Later,
Verma et al. [32] proposed local extrema co-occur-
Texture is another widely used feature for retrieval
rence pattern (LECoP). In this, local directional
of images. Texture is well able to distinguish images
information was extracted for retrieval of images.
from one another [18]. It varies with local intensities of
Dubey et al. developed an efficient Local diagonal
image, thus providing with its surface characteristics.
extrema pattern (LDEP) which described relationship
Various properties like coarseness, smoothness and
among the diagonal neighbors of center pixel of the
regularity are used to identify these characteristics.
image [33]. Further, a novel texture feature, known as
Texture feature extraction is carried out extensively
local tri-directional pattern (LTriDP) was proposed,
using the signal processing methods. These methods
which exploits the local intensities of neighborhood
make use of mathematical transformations in the
pixels in three directions [34]. Subsequently, Verma
image, thus resulting in image features in the form of
et al. proposed local neighborhood difference pattern
coefficients. Gabor filters [19], discrete wavelet trans-
(LNDP) [35], which creates a pattern by estimating
form [20], etc. are some of the important transforms
differences between two immediate neighborhood
used for extraction of texture features. Another way to
pixels in horizontal or vertical directions. Recently,
extract texture feature involves model-based methods.
local neighborhood intensity pattern (LNIP) for
In these methods, an assumption model for an image
image retrieval was introduced by Banerjee et al. [36]
is initially designed. Estimation of the model parame-
which is more advantageous in terms of resistance to
ters is carried out, which are used as texture descriptor
illumination.
features. One such popularly used method is Markov
random field [21]. Structural analysis is another It is observed that there is no single optimal repre-
method for texture feature extraction. It is performed sentation of an image for retrieval purposes. This is on
in the case where texture primitives consist of regular account of the variant conditions in which photo-
patterns and are large enough to be described and seg- graphs might be captured, viz. view angle, illumina-
mented on individual basis [22]. tion changes, etc. In this paper, the authors have pre-
sented a novel pattern for the purpose of image
In this method, texture features are extracted using retrieval based on color and texture. The method
primitives and placement rules. The analysis of spatial makes use of HSV color space for calculation of hue,
distribution of gray levels is performed using statistical saturation and intensity of the color image. Herein,
methods. Some of the widely used methods include the pattern establishes a relationship of the pixel under
co-occurrence matrix, Tamura feature extraction [23], observation with its surrounding pixels by exploiting
Wold features [24], etc. Haralick et al. introduced the their mutual relationships based on four significant
concept of gray-level co-occurrence matrix (GLCM) directions. A magnitude pattern is also taken into con-
[25] in order to extract statistical features of an image. sideration using the same four directions, and both
GLCM matrix extracts texture features by identifying their histograms are combined to form a feature vec-
spatial correlation of image pixels. In another work, tor. The performance of the proposed method is vali-
Zhang et al. [26] estimated the GLCM of edge images dated using two databases. The remaining paper is
which were formed using Prewitt edge detector in four organized as follows: Section 2 gives an introduction
directions, thus extracting statistical features to form to color space, GLCM and local patterns. Framework
co-occurrence matrices for retrieval purposes. Partio of the proposed method, algorithm and similarity
et al. used GLCM with statistical features for the measure is explained in Section 3. Section 4 presents
retrieval of rock texture images [27]. the experimental results and discussions. Finally, the
An efficient method for the purpose of feature conclusion of the work is achieved in Section 5.
extraction for CBIR includes local binary pattern
(LBP), introduced by Ojala et al. [28]. LBP is a fast,
simple and robust texture descriptor which can cap- 2. COLOR AND TEXTURE DESCRIPTORS
ture local texture features unaffected by illumination. Color Space
LBP was further modified into a uniform and rotation
invariant LBP in order to limit the number of patterns. Broadly, images are classified into three catego-
A number of LBP variants have been developed for ries–binary, gray scale and color image. Binary images
different applications. Murala et al. [29] introduced comprise of only two intensity levels, viz. black and
local tetra pattern (LTrP) that retrieves images by tap- white. Gray scale images consist of a set of gray tone
ping first order derivatives in both horizontal and ver- intensity measures in one band, whereas, color images
tical directions. Further, local ternary co-occurrence compose of multiple bands, each band comprising of a
pattern was developed for retrieval of biomedical range of intensities. The most commonly used color
images [30]. In this pattern, co-occurrence of similar space is RGB color space. In this, the images consist
ternary edges was encoded based on gray level intensi- of three bands labelled as red, green, and blue.
ties in the pixel window. A robust center symmetric Another color space that is in wide use is HSV, which

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

580 BEDI, SUNKARIA

(a) (b)
1 3 0 1 1 pixel 0 1 2 3
pair

0 1 2 3 2 0 1 3 0 1 90
135 45
2 2 1 0 3 1 1 1 2 2
1 3 0 0 1 2 0 2 2 2 0

2 1 2 2 3 3 2 0 1 0

Fig. 1. Example of gray level co-occurrence matrix: (a) original matrix, (b) GLCM matrix.

has three constituents known as hue, saturation and Local Binary Patterns
value. Local binary pattern (LBP), proposed by Ojala
Hue component is related to color, by referring to et al. [28], draws out local information from an image
which pure color it belongs to. It is defined as an angle by using neighbourhood pixel block, while consider-
and ranges from 0° to 360°, with each degree occupy- ing center pixel intensity as the threshold value. The
ing different colors. Saturation describes the level of neighborhood pixels are compared to the center pixel.
brightness and lightness of color component. Its value If the neighborhood pixel intensity is greater than that
ranges between 0 and 1, with the color intensity of the center pixel, the value is set to binary one, else
increasing simultaneously with saturation value. The zero. The resulting zeros and ones are then put
last component, i.e., value defines the intensity of the together to create a binary number in order to generate
color that can be extracted from the color information an LBP code for the center pixel. Mathematically, the
obtained from the image. The range of value compo- LBP operator for p surrounding pixels and radius r can
nent varies from 0 to 1, where 0 refers to completely be expressed as:
dark and 1 to completely bright intensity level. HSV p
color space is mostly preferred by researchers because
of its close proximity to human perception of color and
LBPp,r =  P (I
i =1
1 i − I c ) 2i −1, (1)
invariance [37]. In this present work, images have been
1 t ≥ 0
transformed from RGB to HSV color space for feature P1 ( t ) =  (2)
extraction. 0 else.
Here, Ii and Ic refer to the intensity values of the neigh-
bourhood and the center pixel, respectively. The LBP
Gray Level Co-occurrence Matrix map so created is thereafter converted into histogram
with the help of following equations:
One of the popular statistical approaches for
extraction of image features include gray level co- M N
occurrence matrix (GLCM) [25]. The matrix relates
to the spatial arrangement of co-occurring gray valued
His ( L ) LBP =  P ( LBP ( m, n) , L) ;
m =1 n =1
2
(3)
L ∈ [0,(2 − 1)],
pairs of pixels located at a certain distance in a partic- p

ular direction. The matrix elements consist of the

1 p = q
P2 ( p, q ) = 
number of times a pixel pair occurs, and its size
depends upon the maximum intensity value available (4)
0 else,
in the image.
where M × N define the image size. An example of
An illustration of GLCM computation is shown in LBP is illustrated in Fig. 2.
Fig. 1. Figure 1a shows an original matrix and its cal-
culated GLCM is depicted in Fig. 1b. First row and
first column of the GLCM matrix indicate pixel values Local Tetra-Directional Pattern
that are present in the original matrix. Co-occurrence Local tetra-directional pattern is a variant of LBP.
for each pair of intensity values, for example, (0, 0), Instead of taking uniform relationship with all neigh-
(0, 1), …, (3, 3) is calculated. As indicated in Fig. 1a, borhood pixels into consideration, LTrDP considers
pair (3, 0) indicating occurrence of value “3” with “0” relationship based on different directions. Each center
at a distance of one in horizontal direction occurs pixel consists of neighboring pixels in a certain radius.
twice. Thus, 2 is entered in GLCM matrix at position Closest neighbor consists of 8 pixels distributed all
(3, 0). Rest of the elements are computed in the similar around the center pixel. Since closest neighbouring
ways in order to create the matrix. pixels are less in count and results in more related

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

 LOCAL TETRA-DIRECTIONAL PATTERN–A NEW TEXTURE DESCRIPTOR 581

1 3 5 0 0 1
Thresholding
8 4 7 1 1 Binary: 00111101
Decimal: 61
3 5 9 0 1 1

Fig. 2. Example of basic LBP.

(a) (b) (c)

I15 I14 I13 I12 I11 9 8 7 6 2 9 8 7 6 2
I16 I4 I3 I2 I10 2 7 4 9 8 2 7 4 9 8
I17 I5 Ic I1 I9 1 4 3 8 2 1 4 3 8 2
I18 I6 I7 I8 I24 4 6 4 9 9 4 6 4 9 9
I19 I20 I21 I22 I26 1 8 9 8 5 1 8 9 8 5
M0 = (8 − 3)2 + (9 − 3)2 + (4 − 3)2 + (7 − 3)2 = 8.83

(d) (e) (f) (g)

9 8 7 6 2 9 8 7 6 2 9 8 7 6 2 9 8 7 6 2
2 7 4 9 8 2 7 4 9 8 2 7 4 9 8 2 7 4 9 8
1 4 3 8 2 1 4 3 8 2 1 4 3 8 2 1 4 3 8 2
4 6 4 9 9 4 6 4 9 9 4 6 4 9 9 4 6 4 9 9
1 8 9 8 5 1 8 9 8 5 1 8 9 8 5 1 8 9 8 5
1001 → 2 1111 → 0 0000 → 0 1000 → 1
(2 − 8)2 + (8 − 8)2 + (9 − 8)2 + (4 − 8)2 = 7.41
→ 0 (8 − 9) + (2 − 9) + (6 − 9) + (7 − 9) = 7.98 → 0 (9 − 4) + (6 − 4) + (7 − 4) + (8 − 4) = 7.35 → 0 (4 − 7) + (7 − 7) + (8 − 7) + (9 − 7) = 8.74 → 0
2 2 2 2 2 2 2 2 2 2 2 2
M0 = 8.83 M0 = 8.88 M0 = 8.83 M0 = 8.83
(h) (i) (j) (k)
9 8 7 6 2 9 8 7 6 2 9 8 7 6 2 9 8 7 6 2
2 7 4 9 8 2 7 4 9 8 2 7 4 9 8 2 7 4 9 8
1 4 3 8 2 1 4 3 8 2 1 4 3 8 2 1 4 3 8 2
4 6 4 9 9 4 6 4 9 9 4 6 4 9 9 4 6 4 9 9
1 8 9 8 5 1 8 9 8 5 1 8 9 8 5 1 8 9 8 5
(3 − 4)2 + (4 − 4)2 + (7 − 4)2 + (2 − 4)2
→ 0 (4 − 6) + (3 − 6) + (4 − 6) + (1 − 6) = 6.48 → 0 (9 − 4) + (8 − 4) + (3 − 4) + (4 − 4) = 6.48 → 0 (9 − 9) + (2 − 9) + (8 − 9) + (3 − 9) = 9.27 → 1
= 8.74 2 2 2 2 2 2 2 2 2 2 2 2
M0 = 8.83 M0 = 8.83 M0 = 8.83 M0 = 8.83

Tetra-direction pattern: 20012013 Magnitude pattern: 00000001

00010010 10001000 00000001

Fig. 3. An illustrated example for local tera-directional pattern.

information, hence we consider 8-neighborhood pix- compute the difference between the neighborhood
els for pattern creation. Each neighborhood pixel is pixel and its four adjacent neighbouring pixels in four
considered, one at a time, and compared with four directions, viz., 0°, 45°, 90°, and 135°. Thus, for each
most adjacent pixels. These four pixels are located at neighborhood pixel, we have
0°, 45°, 90°, and 135°. An example for creating LTrDP
is illustrated in Fig. 3 and explained as follows.
Consider a center pixel Ic. The window chosen D1 = I i, j +1 − I i, j , D2 = I i −1, j +1 − I i, j ,
(5)
consists of closest 8 neighborhood pixels. Firstly, we D3 = I i −1, j − I i, j , D4 = I i −1, j −1 − I i, j ,

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

582 BEDI, SUNKARIA

where (i, j) denote the location of the neighborhood LTrDP2 ( I c ) = {P4 ( f1 ) , P4 ( f2 ) , …, P4 ( f8 )} , (10)
pixel in question. We thus have four differences, D1,
D2, D3, and D4 for each neighborhood pixel. Further, a 1, x=2
pattern is configured based on all the four differences. P4 ( x ) =  (11)
0, else,
f ( D1, D2, D3, D4 ) = {#(Dk < 0)} mod 4
∀k = 1, 2,3, 4,
(6) LTrDP3 ( I c ) = {P5 ( f1 ) , P5 ( f2 ) , …, P5 ( f8 )} , (12)
where #(Dk < 0) represents total count of Dk with neg- 1, x =3
ative values, for all k = 1, 2, 3, 4. #(Dk < 0) results in P5 ( x ) =  (13)
0, else.
values ranging from 0 to 4. In order to calculate value
of each pattern, a mod of #(Dk < 0) is calculated with These three patterns so obtained are further con-
4. The values are calculated according to #(Dk < 0), verted into pattern map using the following equation:
e.g., if all Dk are negative for k = 1, 2, 3, 4, then 7
#(Dk < 0) is 4 and # ( Dk < 0) mod 4 results in zero. In LTrDP ( I c ) i =1,2,3 = 2 l
× LTrDPi ( I c ) ( l + 1) . (14)
this manner, # ( Dk < 0) mod 4 is assigned values 0, 1, l =0
2, and 3. For every neighborhood pixel i = 1, 2, 3, …, After attaining the pattern map, feature is extracted
8, pattern values fi ( D1, D2, D3, D4 ) is computed result- by computing histograms for all three binary patterns
ing in tetra-directional pattern using Eqs. (3) and (4).
LtrDP ( I c ) = { f1, f2, …, f8}. (7) The tetra-directional pattern extracts most of the
In this manner, we achieve a quaternary pattern for local information around the center pixel, but it has
each center pixel, which are further converted into been observed that magnitude pattern can also be
three binary patterns helpful in creating more illustrative feature vector [29].
Thus, we exploit magnitude information in order to
LTrDP1 ( I c ) = {P3 ( f1 ) , P3 ( f2 ) , …, P3 ( f8 )} , (8) create a magnitude pattern. It makes use of the magni-
tude values of each pixel present in the window and its
1, x =1
P3 ( x ) =  (9) neighborhood pixels in four directions, viz., 0°, 45°,
0, else, 90°, and 135°. Magnitude pattern is created as follows:

Mi = ( I a,b +1 − I a,b )2 + ( I a −1,b +1 − I a,b )2 + ( I a −1,b − I a,b )2 + ( I a −1,b −1 − I a,b )2, (15)

where i = 0, 1, 2, …, 8 denotes each pixel located at described by Ic and the neighborhood pixels are
location (a, b) in the window, with M0 referring to the described by the shaded region, denoted from I1 to I8.
value calculated for the center pixel. Binary value is An example window is taken in Fig. 3b. Magnitude
further assigned to each neighborhood pixel based on term M0 for the center pixel is calculated in Fig. 3c
comparison as shown:
using Eq. (15). In Fig. 3d, pixel I1 is compared with its
1, M i ≥ M 0 neighborhood pixels located at 0°, 45°, 90°, and 135°,
Magi ( M i , M 0 ) =  ∀i = 1,2,…,8, (16)
0, else i.e., I9, I10, I2, and I3, respectively and “0” and “1” is
assigned for all four comparisons. Since I1 > I9, I1 =
LtrDPmag ( I c ) = {Mag1, Mag2, …, Mag8} , (17) I10, I1 < I2 and I1 > I3, hence the pattern for I1 is 1001.
7 Thus, according to Eq. (6), the pattern value for I1 is 2.
LTrDP ( I c ) mag = 2
l =0
l
× LTrDPmag ( I c ) ( l + 1) . (18) Further, magnitude value M1 for the pixel is calculated
using Eq. (15) and compared with that of the center
Further, the histogram for this magnitude pattern pixel. Since M1 < M0, hence, “0” pattern value is
is obtained using Eqs. (3), (4). All the four histograms assigned for this. In the similar manner, pattern values
obtained are then concatenated and a joint histogram
is obtained as the final feature vector for rest of the neighborhood pixels is calculated from
Figs. 3e–3k, and one quaternary pattern and one
Hist = His LtrDP1 , His LtrDP2 , His LtrDP3 , His LtrDPmag  . magnitude pattern is formed. The quaternary pattern
(19) is further divided into three patterns based on
Figure 3 illustrates the creation of LTrDP with the Eqs. (8)–(13). All the four binary patterns are further
help of an example. Figure 3a describes the window converted into histograms and are concatenated
chosen for creating the pattern. The center pixel is together to result in a feature vector.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

 LOCAL TETRA-DIRECTIONAL PATTERN–A NEW TEXTURE DESCRIPTOR 583

72 bin quantized Query image

Hue feature vector
histogram
Query
image
RGB to 20 bin Histogram
HSV Saturation quantized Feature
concatenation
color histogram vector
space

Image
database
Value Median

Local tetra-
Centre pixel directional
magnitude pattern
Feature
vector
Gray level database
co-occurrence Resize matrix
matric to vector

Results Similarity
match

Fig. 4. Proposed system framework.

3. PROPOSED SYSTEM FRAMEWORK Input: query image; output: retrieved similar

Figure 4 presents the flowchart of the proposed images.
method. The algorithm for the same is also explained 1. Enter the query image.
further. The algorithm is outlined in two parts. First 2. Construct its feature vector in the manner
division of the algorithm explains the steps for con- depicted in Section 1.
struction of feature vector for the image database. The
second division describes the retrieval process for the 3. Calculate the similarity distance measure, and
system. hence the similarity indices, in order to compare the
feature vector of the query image with that of each
image available in the database.
Algorithm 4. Sort the similarity indices and extract images
Part 1. Feature vector formation. corresponding to least similarity indices as the best
Input: image; output: feature vector. match vectors as final results.
1. Upload the image and transform it from RGB
into HSV color space. Similarity Measure
2. Quantize the hue (H) and saturation (S) compo- Similarity measure is considered as a significant
nents into 72 and 20 bins, respectively. Create the his- element of content-based image retrieval systems. The
tograms for both. function of similarity measurement is to compute the
3. Calculate the local median values for every pixel level of similarity between the query image and the
of the value (V) component. images from the database. Similarity measure com-
4. Apply the proposed pattern on the median image putes the amount of differences between the query
obtained in step 3. image and other images using a distance metric. The
5. Construct GLCM of the pattern map obtained in proposed method makes use of d1 distance measure
step 4 and change it into vector form. since it gives the best results for similarity matching
6. Combine the value of the center pixel, histo- [29, 31, 32, 38].
grams from step 2 and GLCM vector obtained from For a query image Q, its extracted feature vector
step 5 in order to achieve the final histogram as the can be represented as fQ = ( fQ1, fQ2 ,…, fQL ), where L is
feature vector. defined as the length of the feature vector. In a similar
Part 2. Image retrieval. manner, each image from the given database |DB| can

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

584 BEDI, SUNKARIA

Table 1. LBP variant abbreviations

LeCoP [32] Local extrema co-occurrence pattern
Tri-directional [34] Local tri-directional pattern
LNDP [35] Local neighbourhood difference pattern
LNIP [36] Local neighbourhood intensity pattern
PM (proposed method) Local tetra-directional pattern

be represented by its feature vector as fDB j = N qc

( fDB , fDB
j1
,…, fDB jL ) , j = 1, 2, …, |DB|. The distance
j2
Ravg ( b, n) = 1
N qc  R (q, n) ,
q =1
(24)
measure d1 can be calculated using the following for-
mula: where b refers to the category number in the database.
Further, we compute total precision and total recall for
L
fDB ji − fQi the whole dataset using the following two equations:
D (Q,DB j ) = 1+ f + fQi
, (20) Nc
Ptotal ( n) = 1 P (b, n) ,
i =1 DB ji
avg (25)
where fDB ji refers to the ith feature of the jth image Nc b =1

from the database |DB|. Nc

Rtotal ( n) = 1
Nc  R (b, n) ,
b =1
(26)
4. EXPERIMENTAL RESULTS AND ANALYSIS where Nc denotes the total number of categories that
To affirm the quality of the proposed method, it exist in the dataset. Total recall is also referred to as
has been examined on two benchmark databases. The average recall rate (ARR). The performance of the
performance is graded on the basis of two measures, proposed method is validated by comparing it with
viz., precision and recall [39]. These parameters are recent methods listed in Table 1. In each experiment,
computed based on the count of relevant images that retrieval is performed taking into consideration every
are retrieved from the given database. Relevant images image from the database as the query image. Specific
are those images that belong to the category similar to number of images is retrieved for every query image.
that of the query image. The remaining images are Precision and recall are computed and final results are
regarded as non-relevant. Precision and recall are then analyzed. The results evaluated on two datasets
effective evaluation parameters for CBIR. High values are explained below.
of precision and recall suggest better results. Precision
is described as the ratio of total count of relevant
Database-1 (Corel-1K)
images retrieved to the total number of images
retrieved from the database. Recall is defined as the The first database considered for experimentation
ratio of the total count of relevant images to the total is Corel-1K [40]. This dataset comprises of 1000 natu-
number of images that are present in the database. ral images that are divided into ten distinct categories,
Given a query image Q, total number of images viz., Africans, Beaches, Buildings, Buses, Dinosaurs,
being retrieved be n, precision and recall values can be Elephants, Flowers, Horses, Mountains, and Food.
calculated from the following equations: Each category comprises of 100 images in it. Each
image is either of size 256 × 384 or 384 × 256. Some
No. of relevant images retrieved of the images form the dataset are shown in Fig. 5.
P ( q, n) = , (21) For this database, 10 images were retrieved initially
n
and the count is increased by 10 in every experiment.
No. of relevant images retrieved Maximum of 100 images were retrieved and hence-
R ( q, n) = , (22) forth, results were analyzed. The experimentation was
N qc
carried out on each image of the dataset. Figure 6
where N qc defines the total count of relevant images shows one query image from each category along with
available in the dataset which is defined by the number some of its retrieved images. From the retrieval results,
of images present in each category of the database. For it was observed that similar images have been retrieved
each category, we can calculate average precision and for each category. Precision and recall values for every
recall using the following equations: experiment were calculated and analyzed. Average
precision and recall percentage for different number of
N qc retrieved images have been described in Tables 2 and
Pavg ( b, n) = 1
N qc  P (q, n) ,
q =1
(23) 3, respectively. Average precision rate decreases
whereas average recall rate increases with the increase

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

 LOCAL TETRA-DIRECTIONAL PATTERN–A NEW TEXTURE DESCRIPTOR 585

Fig. 5. Sample images from Corel-1K dataset.

in the number of images that are to be retrieved. It can case of LNDP. Thus, it can be seen that the proposed
be observed from the table that the proposed method method performs well for retrieving natural images
is more benevolent than the other existing methods in from a huge database.
terms of both precision and recall. Figures 7a, 7b
depict the graphical results for precision and recall for
Corel-1K database. It can be observed that the pro- Database-2 (AT&T)
posed method outperforms the other methods. For
retrieval of 10 images, the average retrieval rate for the Another experimentation for CBIR is carried out
proposed method is improved by 0.805% in case of using AT&T face database [41]. The database com-
LeCoP, 7.92% for LNIP, 12.13% for Tri-Directional
Pattern, and 16.20% in case of LNDP. While retriev- prises of 400 facial images of size 92 × 112. These
ing 100 images, the proposed method shows an 400 images are further divided into 40 categories with
improvement of 10.48% for LeCoP, 16.90% for LNIP, each category consisting of ten images. Figure 8 shows
21.21% for Tri-Directional Pattern, and 19.31% in a few sample images from the AT&T dataset.

Table 2. Precision (%) for Corel-1K dataset

No. of images
LeCoP Tri-directional LNDP LNIP PM
retrieved (n)
10 74.480 66.957 64.610 69.570 75.08
20 66.855 59.830 58.085 62.347 68.62
30 62.243 55.540 54.403 57.945 64.433
40 58.725 52.308 51.515 54.279 61.318
50 55.406 49.596 48.988 51.369 58.59
60 52.622 47.213 46.903 48.879 56.09
70 50.069 45.036 45.050 46.561 53.803
80 47.546 43.02 43.252 44.522 51.554
90 45.216 41.033 41.444 42.437 49.467
100 42.913 39.116 39.736 40.557 47.412

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

586 BEDI, SUNKARIA

Retrieved images

Fig. 6. Query image and retrieved images from Corel-1K dataset.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

 LOCAL TETRA-DIRECTIONAL PATTERN–A NEW TEXTURE DESCRIPTOR 587

Table 3. Recall (%) for Corel-1K dataset

No. of images
LeCoP Tri-directional LNDP LNIP PM
retrieved (n)
10 7.448 6.696 6.461 6.9570 7.508
20 13.371 11.966 11.617 12.470 13.724
30 18.673 16.662 16.321 17.383 19.33
40 23.490 20.923 20.606 21.712 24.527
50 27.703 24.798 24.494 25.685 29.295
60 31.573 28.328 28.142 29.327 33.654
70 35.048 31.525 31.535 32.593 37.662
80 38.037 34.416 34.602 35.618 41.243
90 40.694 36.930 37.300 38.193 44.52
100 42.913 39.116 39.736 40.557 47.412

Table 4. Precision (%) for AT&T dataset

No. of images
LeCoP Tri-directional LNDP LNIP PM
retrieved (n)
1 100 100 100 100 100
2 72.000 85.000 88.000 81.750 88.875
3 59.418 75.000 77.500 71.167 77.833
4 50.812 68.125 70.438 63.125 71.375
5 45.600 62.600 64.550 58.000 64.85
6 42.292 56.958 59.708 53.250 59.875
7 39.179 53.000 55.209 50.036 55.714
8 36.656 49.375 51.719 46.563 52.281
9 34.583 45.9722 48.556 43.722 49.306
10 32.800 43.025 45.325 41.275 45.8

Table 5. Recall (%) for AT&T dataset

No. of images
LeCoP Tri-directional LNDP LNIP PM
retrieved (n)
1 10 10 10 10 10
2 14.400 17.000 17.600 16.350 17.775
3 17.825 22.500 23.250 21.350 23.36
4 20.325 27.250 28.175 25.250 28.55
5 22.800 31.300 32.275 29.000 32.425
6 25.375 34.175 35.825 31.950 36.069
7 27.425 37.100 38.647 35.025 39.015
8 29.325 39.500 41.375 37.250 41.825
9 31.125 41.375 43.700 39.350 44.375
10 32.800 43.025 45.325 41.275 45.8

In this experiment, initially, a single image was was carried out for every image present in the data-
retrieved and the count of retrieval was increased by 1. base. Some of the query images and their correspond-
Ten images were retrieved maximum for each category ing retrieved images are represented in Fig. 9. Preci-
and results obtained were analysed. The experiment sion and recall values were evaluated for each experi-

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

588 BEDI, SUNKARIA

(a) LeCoP
80
Tri-directional
70 LNDP
LNIP
60
Precision, % PM
50

40

30

20

10

0
10 20 30 40 50 60 70 80 90 100
No. of images retrieved
50
(b)
10
LeCoP
10
Tri-directional
10
LNDP
Recall , %

10 LNIP
10 PM
10
10
10
10
0
10 20 30 40 50 60 70 80 90 100
No. of images retrieved

Fig. 7. (a) Precision (%) and (b) recall (%) for Corel-1K dataset.

ment. Tables 4 and 5 depict the average precision and pattern, the proposed method shows an improvement
recall percentage for different number of retrieved by 6.45% and for LNDP, it shows an improvement of
images for all the techniques. From the results, it was 1.048%. This indicates the method is good for detec-
observed that the proposed method results in the bet- tion and identification of faces in comparison to the
ter precision and retrieval rates in comparison to other existing methods.
methods. Graphs for precision and recall are shown in
Figs. 10a, 10b. From the graphs, the performance of
the proposed method can be compared with the other Feature Vector Length
existing methods. For each method, the results show Table 4 calculates and describes the feature vector
100% in the case of retrieving only one image, thus lengths for every method. It can be seen that the pro-
indicating that every query image is retrieved success- posed method has the highest feature vector length out
fully for each of the methods. Further, when 2 images of all the methods. Further, the computational time
are being retrieved, the proposed method shows an for each method depends on its corresponding feature
improvement in the average retrieval rate by 23.437% vector length. Higher length of feature vector corre-
for LeCoP, 8.715% for LNIP, 4.558% for Tri-direc- sponds to higher amount of time taken for retrieving a
tional method, and 0.99% in case of retrieval by single image. Hence, compared to other existing
LNDP. While retrieving 10 images, the proposed methods, the proposed method results in highest
method is improved by 39.634% in case of LeCoP and computational time on account of highest feature vec-
10.96% for LNIP pattern. In case of Tri-directional tor length.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

 LOCAL TETRA-DIRECTIONAL PATTERN–A NEW TEXTURE DESCRIPTOR 589

Fig. 8. Sample images for AT&T.

Retrieved images

Fig. 9. Query image and retrieved images from AT&T dataset.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

590 BEDI, SUNKARIA

100 LeCoP
(a)
90 Tri-directional
80 LNDP
LNIP
70

Precision, %
PM
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10
No. of images retrieved
50
LeCoP (b)
45 Tri-directional
40 LNDP
35 LNIP
PM
Recall, %

30
25
20
15
10
5
0
1 2 3 4 5 6 7 8 9 10
No. of images retrieved

Fig. 10. (a) Precision (%) and (b) recall (%) for AT&T dataset.

5. CONCLUSIONS 1. The average precision/average recall have sig-

nificantly improved from 42.913, 39.116, 39.736, and
In the present paper, a new feature descriptor, 40.557–47.412% when compared with LeCOP, Tri-
known as local tetra-directional pattern, is proposed
for CBIR. The descriptor combines the properties of Directional, LNDP, and LNIP, respectively on Data-
basic LBP and its directional measures to create a pat- base 1.
tern. It is further combined with color features and co- 2. The average precision/average recall have sig-
occurrence matrix in order to create the final feature nificantly improved from 32.8, 43.025, 45.325, and
vector for retrieval process. 41.275–45.8% when compared with LeCOP, tri-
The execution of the proposed method has been directional, LNDP, and LNIP, respectively on Data-
compared with the existing methods, viz., LeCOP, tri- base 2.
driectional, LNDP, and LNIP and has been detailed
below. 3. From the above analysis, it can be concluded
that the proposed method is able to retrieve coloured
natural images in a better way than the rest of the
Table 6. Feature vector lengths of various methods. methods. Also, the method is good for retrieving faces
Feature vector length irrespective of change in emotions or angles.

LeCOP 72 + 20 + 256 348 In this paper, directional patterns have been taken
into consideration and weightage has also been given
Tri-directional 256 + 256 + 256 768 to the center pixel. Median has also been taken into
LNDP 256 256 account because of its robustness and preservation of
LNIP 256 + 256 512 edges. The proposed method can be used in various
applications like face detection, fingerprint recogni-
PM 1 + 72 + 20 + 256 + 256 + 256 + 256 1117 tion, etc.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

 LOCAL TETRA-DIRECTIONAL PATTERN–A NEW TEXTURE DESCRIPTOR 591

CONFLICT OF INTEREST 14. M. J. Swain and D. H. Ballard, “Color indexing,” Int.

J. Comput. Vision 7 (1), 11–32 (1991).
The authors declare no conflict of interest neither in
financial nor in any other area. 15. J. Huang, S. R. Kumar, M. Mitra, W. J. Zhu, and
R. Zabih, “Image indexing using color correlograms,”
in Proceedings of IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (San Juan,
REFERENCES 1997).
1. T. Kato, “Database architecture for content-based im- 16. G. Pass, R. Zabih, and J. Miller, “Comparing images
age retrieval,” in Image Storage and Retrieval Systems using color coherence vectors,” in Proceedings of the
(SPIE, San Jose, 1992). Fourth ACM International Conference on Multimedia
(Boston, 1997).
2. R. Datta, D. Joshi, J. Li, and J. Z. Wang, “Image re-
trieval: Ideas, influences, and trends of the new age,” 17. N. Jhanwar, S. Chaudhuri, G. S. Seetharaman, and
ACM Comput. Surv. 40 (2), 5:1–5:60 (2008). B. Zavidovique, “Content based image retrieval using
motif cooccurrence matrix,” Image Vision Comput. 22
3. B. S. Manjunath and W. Y. Ma, “Netra: A toolbox for (14), 1211–1220 (2004).
navigating large image databases,” Multimedia Syst. 7
(3), 184–198 (1999). 18. Z. He, X. You, and Y. Yuan, “Texture image retrieval
based on non-tensor product wavelet filter banks,” Sig-
4. C. W. Niblack, R. Barber, W. Equitz, M. D. Flickner, nal Process. 89 (8), 1501–1510 (2009).
E. H. Glasman, D. Petkovic, P. Yanker, C. Faloutsos,
and G. Taubin, “QBIC project: Querying images by 19. A. Ahmadian and A. Mostafa, “An efficient texture
content, using color, texture, and shape,” in Storage and classification algorithm using Gabor wavelet,” in 25th
Retrieval for Image and Video Databases, SPIE (San Annual International Conference of Engineering in Med-
Jose, 1993). icine and Biology Society (Cancun, 2003).
5. J. Z. Wang, J. Li, and G. Wiederhold, “SIMPLIcity: 20. H. Farsi and S. Mohamadzadeh, “Colour and texture
Semantics-sensitive integrated matching for picture li- feature-based image retrieval by using Hadamard ma-
braries,” IEEE Trans. Pattern Anal. Mach. Intell. 23 trix in discrete wavelet transform,” IET Image Process.
(9), 947–963 (2001). 7 (3), 212–218 (2013).
6. M. Beigi, A. B. Benitez, and S. F. Chang, “MetaSEEK: 21. L. Ainhoa, R. Manmatha, and S. Rüger, “Image re-
A content-based metasearch engine for images,” in trieval using markov random fields and global image
Storage and Retrieval for Image and Video Databases VI features,” in ACM International Conference on Image
(SPIE, San Jose, 1997). and Video Retrieval (New York, 2010).
7. J. R. Smith and S. F. Chang, “Querying by color re- 22. R. Jain, R. Kasturi, and B. G. Schunck, Texture, in Ma-
gions using the VisualSEEk content-based visual query chine Vision (McGraw-Hill, New York, 1995),
system,” Intell. Multimedia Inf. Retr. 7 (3), 23–41 pp. 234–248.
(1997).
23. H. Tamura, S. Mori, and T. Yamawaki, “Textural fea-
8. E. D. Sciascio and M. Mongiello, “DrawSearch: A tool tures corresponding to visual perception,” IEEE Trans.
for interactive content-based image retrieval over the Syst. Man Cybern. 8 (6), 460–473 (1978).
Internet,” in Storage and Retrieval for Image and Video
Databases VII (SPIE, San Jose, 1998). 24. F. Liu and R. W. Picard, “Periodicity, directionality,
and randomness: Wold features for image modeling and
9. P. Aigrain, H. Zhang, and D. Petkovic, “Content- retrieval,” IEEE Trans. Pattern Anal. Mach. Intell. 18
based representation and retrieval of visual media: A (7), 722–733 (1996).
state-of-the-art review,” Multimedia Tools Appl. 3 (3),
179–202 (1996). 25. R. M. Haralick, K. Shanmugam, and I. Dinstein,
“Textural features for image classification,” IEEE
10. Y. Liu, D. Zhang, G. Lu, and W. Y. Ma, “A survey of Trans. Syst. Man. Cybern. 6, 610–621 (1973).
content-based image retrieval with high-level seman-
tics,” Pattern Recognit. 40 (1), 262–282 (2007). 26. J. Zhang, G.-l. Li, and S.-W. He, “Texture-based im-
age retrieval by edge detection matching GLCM,” in
11. B. S. Manjunath, J. R. Ohm, V. V. Vasudevan, and 2008 10th IEEE International Conference on High Per-
A. Yamada, “Color and texture descriptors,” IEEE formance Computing and Communications (Dalian,
Trans. Circuits Syst. Video Technol. 11 (6), 703–715 2008).
(2001).
27. M. Partio, B. Cramariuc, M. Gabbouj, and A. Visa,
12. G. Pass and R. Zabih, “Comparing images using joint “Rock texture retrieval using gray level co-occurrence
histograms,” J. Multimedia Syst. 7 (3), 234–240 matrix,” in Proc. of 5th Nordic Signal Processing Sympo-
(1999). sium (2002).
13. G. Pass and R. Zabih, “Histogram refinement for con- 28. T. Ojala, M. Pieti, and T. Maenpaa, “Multiresolution
tent-based image retrieval,” in Proceedings Third IEEE gray-scale and rotation invariant texture classification
Workshop on Applications of Computer Vision (Sarasota, with local binary patterns,” IEEE Trans. Pattern Anal.
1996). Mach. Intell. 24 (7), 971–987 (2002).

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

592 BEDI, SUNKARIA

29. M. Subrahmanyam, R. P. Maheshwari, and R. Balasu- Anterpreet Kaur Bedi was born in
bramanian, “Local tetra patterns: A new feature de- Pathankot, Punjab, on July 7, 1992.
scriptor for content-based image retrieval,” IEEE She received B.E. degree and M.E.
Trans. Image Process. 21 (5), 2874–2886 (2012). degree (Gold Medalist) in Electron-
ics from Department of Electronics
30. M. Subrahmanyam and Q. J. Wu, “Local ternary co- and Communication Engineering,
occurrence patterns: A new feature descriptor for MRI Dr. B. R. Ambedkar National Insti-
and CT image retrieval,” Neurocomputing 119, 399– tute of Technology, Jalandhar, India
412 (2013). in 2014 and 2017, respectively. Cur-
rently, she is pursuing her Ph.D.
31. M. Verma and B. Raman, “Center symmetric local bi- degree in Electronics and Commu-
nary co-occurrence pattern for texture, face and bio- nication Engineering Department,
medical image retrieval,” J. Visual Commun. Image Dr. B. R. Ambedkar National Insti-
tute of Technology, Jalandhar, Punjab, India. Her major
Representation 32, 224–236 (2015). fields of interest are content-based image retrieval and bio-
32. M. Verma, B. Raman, and S. Murala, “Local extrema medical imaging.
co-occurrence pattern for color and texture image re-
trieval,” Neurocomputing 165, 255–269 (2015).
Ramesh Kumar Sunkaria was born in
33. S. R. Dubey, S. K. Singh, and R. K. Singh, “Local di- Kot Khalsa, (Amritsar), on February
agonal extrema pattern: A new and efficient feature de- 20, 1967. He received B.Tech. degree
in Electronics Engg. from Depart-
scriptor for CT image retrieval,” IEEE Signal Process. ment of Electronics Technology,
Lett. 22 (9), 1215–1219 (2015). Guru Nanak Dev University Amrit-
34. M. Verma and B. Raman, “Local tri-directional pat- sar, and M.Tech. (Hons.) Degree in
Electronics Engg. from Guru Nanak
terns: A new texture feature descriptor for image re- Dev Engineering College, Ludhiana
trieval,” Digital Signal Process. 51, 62–72 (2016). (Punjab Technical University, Jaland-
35. M. Verma and B. Raman, “Local neighborhood differ- har) in 1990 and 2004, respectively.
He did his Ph.D. degree from Electri-
ence pattern: A new feature descriptor for natural and cal Engineering Department, Indian
texture image retrieval,” Multimedia Tools Appl. 77 Institute of Technology Roorkee, Roorkee in June 2010.
(10), 11843–11866 (2018). He worked as Engineer in Bharat Heavy Electricals
36. P. Banerjee, A. K. Bhunia, A. Bhattacharyya, P. P. Roy, Limited, Haridwar upto June 2004. Thereafter he joined
and S. Murala, “Local neighborhood intensity pattern— Indian Broadcasting (Engineers) Service, Govt. of India
a new texture feature descriptor for image retrieval,” Ex- and worked as Assistant Director (Engineering) until
August 1996. He directed the installation of numerous radio
pert Syst. Appl. 113, 100–115 (2018). and TV projects all-over India, namely, VLPT Joginder
37. A. Vadivel, S. Sural, and A. K. Majumdar, “An inte- Nagar, VLPT Ahjufort, VLPT Kalakot etc. Afterwards he
grated color and intensity co-occurrence matrix,” Pat- joined academics and at present he is Associate Professor in
tern Recognit. Lett. 28 (8), 974–983 (2007). Department of Electronics & Communication Engineer-
ing, Dr. B. R. Ambedkar National Institute of Technology,
38. S. Murala, R. P. Maheshwari, and R. Balasubramani- where he has been teaching UG/PG courses related to sig-
an, “Directional local extrema patterns: A new descrip- nal processing and communication engineering. Presently,
tor for content based image retrieval,” Int. J. Multime- he is also serving as Head in the Department of Electronics
and Communication Engineering in the institute. His cur-
dia Inf. Retr. 1 (3), 191–203 (2012). rent research interests are in bio-electric signal and image
39. H. Müller, W. Müller, D. M. Squire, S. Marchand- processing, filter banks and non-linear heart variability
Maillet, and T. Pun, “Performance evaluation in con- studies. He has 57 publications in reputed international
tent-based image retrieval: Overview and proposals,” journals and 45 publications in conference proceedings. At
present, seven PhD thesis have been guided and six PhD
Pattern Recognit. Lett. 22 (5), 593–601 (2001). thesis are in progress under his guidance, 34 M.Tech thesis
40. The Corel-1K Database. have already been completed and more than 44 B.Tech
http://wang.ist.psu.edu/docs/related/ projects have been completed. Also, there are two ongoing
projects under his supervision, funded by DST SERB and
41. The AT&T Database of Faces (2002). AICTE. He is member of IEEE Engineering in Medicine
http://www.uk.research.att.com/facedatabase.html and Biology Society and IEEE Computer society.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 30 No. 4 2020

View publication stats