Int. j. inf. tecnol.

https://doi.org/10.1007/s41870-023-01204-y

ORIGINAL RESEARCH

MVM‑LBP :  mean−variance−median based LBP for face

recognition
Nitin Arora1 · G. Sucharitha2 · Subhash C. Sharma1 

Received: 21 September 2022 / Accepted: 18 February 2023

© The Author(s), under exclusive licence to Bharati Vidyapeeth’s Institute of Computer Applications and Management 2023

Abstract  This paper proposes a novel descriptor called 1 Introduction

Mean-Variance-Median based Local binary pattern (MVM-
LBP). The Median binary pattern (MBP) calculates the dif- CBIR is a system that can search similar images from a col-
ference between median of all the pixels in a 3 × 3 pixel lection of images for a specific image (query image). Due
window and respective eight neighbours. The main draw- to easy access to digital devices such as cellphones, digital
back of MBP is that two different pixel windows can have cameras, tablets, and other similar digital devices, the num-
same MBP code which is not appropriate in face recognition ber of digital images are rising by the minute. It takes a long
systems. To improve the performance of the system, the pro- time to manually search through a big pool of images for a
posed descriptor (MVM-LBP) involves two more statistics query image. Text-based image retrieval (TBIR) [1], given
mean and variance and uses the mean, variance, and median keywords or hand annotation, takes time, and annotations
values of all the pixels in a 3 × 3 pixel window to get a might vary from person to person and this is one of the diffi-
more robust feature vector. To check the excellence of the culties of TBIR. An image’s content can also be used to find
proposed descriptor, the proposed method is applied on two other images. Color, texture, and shape are three different
publicly available facial data sets AT&T, and faces94. The types of information in every image. The CBIR system [2]
outcomes show that the proposed descriptor gives encourag- allows search for images based on image content. An over-
ing results. view of a CBIR system is seen in its entirety in Fig. 1.
An image database and a query image (an image which
Keywords  Local binary pattern · Median binary pattern · we want to search) are the two inputs to the CBIR system.
Face recognition · CBIR · Feature extraction Image pre-processing is the first step in every CBIR system.
All of the images are scaled to the same size and converted
to gray-scale images in this phase. The feature extraction
step follows, which entails extracting features using a local
pattern descriptor. The retrieved features are converted into
G. Sucharitha and Subhash C. Sharma contributed equally to this histograms in the third step, and the features are matched
work.
using similarity measure matrices in the fourth step. Finally,
* Nitin Arora the best feature match is used to generate the output. A
nitinarora.iitr@gmail.com detailed survey on CBIR is presented in [3–5].
G. Sucharitha The performance of the CBIR systems is dependent
sucharithasu@gmail.com upon the unique feature vectors. The more unique the fea-
Subhash C. Sharma ture vector, the better the system. In MBP [6], difference
scs60fpt@gmail.com between the median of all the pixels in a 3 × 3 pixel win-
1
Electronics and Computer Discipline, Indian Institute dow and respective eight neighbours is used. The main
of Technology, Roorkee, India draw back of MBP is that two different pixel windows
2
Electronics and Communication Engineering, Institute can have same MBP code which is not appropriate in face
of Aeronautical Engineering, Dundigal, Hyderabad, India recognition systems. Fig.  2 show the two examples of

13
Vol.:(0123456789)
 Int. j. inf. tecnol.

1.1 Contribution

This paper proposed a novel descriptor called MVM-LBP.

The descriptor is based on mean, variance, and median of
all the pixels in a 3 × 3 window. The key contributions of
the work are:

1. A novel descriptor MVM-LBP is proposed using mean,

variance, and the Medina of all the pixels in a 3 × 3 pixel
window.
2. The performance of the suggested descriptor is exam-
ined on the two publicly available face images databases
AT&T face images database  [7], and faces94 data-
base [8].
3. The d1 distance-based metric is utilized to compute the
similarity score of the suggested system. The perfor-
mance of the suggested descriptor is compared with
state-of-the-art descriptors LBP  [9], Neighborhood
intensit based LBP (NI-LBP) [10], MBP [6], Center
symmetric based LBP (CSLBP) [11], 6x6 multi-block
based LBP (6x6 MB-LBP) [12], and Logically con-
nected-LBP (LC-LBP) [13].
Fig. 1  Overview of a CBIR system
1.2 Organization

The paper is divided into parts for easy understanding. The

related work is discussed in Section 2. The Section 3 pre-
sents some of the local patterns. The proposed descriptors,
similarity measure, and evaluative measure are all depicted
in Section 4. Section 5 details the databases that are used,
as well as the costs of computation and the outcomes. Sec-
tion 6 concludes with closing remarks and recommendations
for further study.

2 Related work

A with the easy access of digital cameras, the facial images

are increasing rapidly. With this face recognition is one of
the big challenges. Many researchers have been working on
face recognition problem and proposed many solution. In
this section we highlighted some of the existing descriptors
available in literature.
In [14] introduced the grey level co-occurrence matrix
(GLCM) for image categorization. This matrix extracts
Fig. 2  Two different 3 × 3 pixel window but same MBP code
characteristics based on pixel pair co-occurrence. The
GLCM was created as a generalized co-occurrence matrix
to extract some significant spatial features from the distribu-
different pixel arrangements but having the same MBP tion of local maxima  [15,  16] presented a texture feature
codes. This paper’s primary focus is to propose a more computation technique that uses the Prewitt edge detector to
robust descriptor that generates unique codes and hence compute edge images and extracts the co-occurrence matrix
the feature vector. The proposed descriptor takes advan- for those edge images instead of the original images. Statis-
tage of the variance in calculating more unique codes. tical characteristics were used to extract features from the

13
Int. j. inf. tecnol.

co-occurrence matrix. Researchers also used the transforma- early detection and classification of lung diseases based on
tion domain to extract features. [17] utilized wavelet packets lung X-ray images. Three different approaches, Surround-
to extract features and apply them to image classification. ing information retrieval (SIR), Minimum edge retrieval
The methods, such as the k-d tree, co-occurrence matrix, (MER) and Integrated feature retrieval (IFR), for image
etc., are more computationally extensive. The local binary retrieval system are proposed in [27]. SIR extracts the fea-
pattern (LBP) [9] is offered to solve this computational com- tures related to the similarities of the neighborhood intensity
plexity. Ojala et al. [9] invented the LBP for texture analy- values. MER extracts the features related to the similari-
sis. Later LBP is used in many applications, such as texture ties of the neighborhood intensity values. IFR combines the
classification, facial recognition, object detectio, and image properties of feature extraction from SIR and MER. A CBIR
retrieval. In [18], an orthogonal difference - local binary system using local and global features for large dataset of
pattern (OD-LBP) has been introduced. OD-LBP consid- images is proposed in [28]. In this MapReduce paradigm
ered only the orthogonal positions in the 3 × 3 pixel window. with different modes is used to retrieve a queried image.
In OD-LBP, initially, a 24 bits binary pattern is generated, Another query based image management system (QBIMS)
which is divided into three binary patterns of 8 bits each, and is proposed in [29]. In QBIMS, image features are based
to reduce the size of the feature vector, PCA is used. Color on image energy, image entropy, image contrast, horizontal
ZigZag Binary Pattern (CZZBP) and Color Median Block edge, vertical edge, centre point, mean and median.
ZigZag Binary Pattern (CMBZZBP) color face descriptor To increase the efficiency of the CBIR system for face
have been presented in [19]. The ZigZag pattern is used recognition, many researchers have proposed some more
to produce the binary pattern in both the descriptors. In variants of LBP. These includes Extended local binary
CZZBP, the features are extracted for three different colors, pattern (ELBP) [10], Local Neighbourhood Intensity Pat-
R, G, and B, and combined to produce a feature vector. In tern (LNIP) [30], Improved LBP (ILBP) [31], Enhanced
CMBZZBP, 9 × 9 pixel is utilized. This window is divided Local Binary Patterns (E-LBP) [32], Local directional edge
into nine blocks, each of size 3 × 3. The median of each binary pattern (LDEBP) [33], directional local ternary co-
block is computed to produce a window of size 3 × 3. To occurrence pattern (DLTCoP) [34], Graph-Based Structure
reduce the size of the feature vectors, PCA is used, and clas- Binary Pattern (GBSBP) [35], Multi-scale neighbourhood
sification is done using SVM and NN [20] presented a neigh- based-tree binary pattern (MNB-TBP) [36], Most signifi-
bourhood and center difference-based-LBP (NCDB-LBP). cant bits based local binary pattern (m-LBP) [37], Local
NCDB-LBP is based on the difference between neighbour- directional peak valley binary pattern has been presented
hood and center pixel intensities. NCDB-LBP is used in in [38] etc. This research paper presents a novel LBP variant
both the direction clockwise and anti-clockwise. PCA and called Mean-Variance and Median based LBP (MVM-LBP)
FLDA are used to reduce the size of the feature vector, and for face recognition. All the above discussed techniques are
for classification, SVM and NN are used. In [21], a local summarized in Table 1.
directional order pattern (LDOP) has been presented. LDOP
uses multi-scale neighbourhood to improve the robustness
of the descriptor. In LDOP multi-radius pixels, find the
3 Local patterns
relationship between the central pixel and neighbourhood
pixels at different scales. In [22], Local tri-directional pat-
This section briefly describes the existing local binary pat-
terns (LTriDP) is proposed. The traditional LBP method
tern and Median binary pattern.
uses the difference information between central pixel and
neighbourhood pixels. LTriDP is based on the intensities
of a pixel in three directions. Local neighbourhood differ- 3.1 Local binary pattern (LBP)
ence pattern (LNDP) is presented in [23]. LNDP is based
on the mutual relationship of neighbouring pixels. LNDP is Ojala et al. [9] proposed LBP for texture analysis. LBP was
combined with LBP to form a more robust feature descrip- later utilized successfully in several applications. In LBP, the
tor. In [24], the Mean distance local binary pattern (Mean 3 × 3 pixel window’s center pixel is compared to all of its
distance LBP) is proposed. Mean distance LBP is based on neighbors. After comparison, if the difference between the
the Euclidean distance of the neighbouring pixels from the neighbouring pixel and central pixel is greater than or equal
central pixels. In [25], the fuzzy theory was used to detect to zero then the corresponding neighbouring pixel desig-
faces in color images. The intuitionistic fuzzy set theory is nated as 1 else 0. By assigning weights to each binary value,
used to describe local texture patterns in images. an 8-bit binary pattern may be generated, which can then
In [26], authors proposed content-based medical image be translated into decimals. The entire computation for the
retrieval system. In this, the authors proposed CBIR system LBP code is shown in Fig. 3a.The LBP descriptor is defined
for the retrieval of medical images (CBMIR) for enabling the mathematically in Eqs. (1) and (2).

13
 Int. j. inf. tecnol.

Table 1  Summary of existing Descriptor name Area of application References

descriptors in literature
Local binary pattern (LBP) Texture analysis, [9]
Object detection,
Image retrieval etc.
Orthogonal difference - LBP Face recognition [18]
Color zigzag binary pattern Face recognition [19]
Color median block zigzag Face recognition [19]
binary pattern
Neighbourhood and center Face recognition [20]
difference-based- LBP
Local directional order pattern Face recognition [21]
Local tri-directional patterns Texture descriptor [22]
for image retrieval
Local neighbourhood difference Natural and texture [23]
pattern image retrieval
Mean distance local binary pattern Texture descriptor [39]
for image retrieval
Extended local binary pattern Face recognition [10]
Local neighbourhood intensity pattern Texture descriptor [30]
for image retrieval
Improved local binary pattern Texture classification [31]
Enhanced local binary patterns Face recognition [32]
Local directional edge binary pattern Biomedical image retrieval [33]
Directional local ternary co-occurrence Natural image retrieval [34]
pattern
Graph-based structure binary pattern Face recognition [35]
Multi-scale neighbourhood based-tree Face recognition [36]
binary pattern
Most significant bits based LBP Natural image retrieval [37]
Local directional peak valley binary Texture descriptor [38]
pattern for image retrieval

N-1
∑ (
k VR,n − Vc × 2n (1)
( ) )
LBPN,R xc =
n=0

{
1, if y ≥ 0
k(y) =
0, Otherwise
for n = (0, 1, ........, N − 1) (2)

Where N is the number of neighboring pixels at radius R,

VR,n specifies the locations of the individual pixel, and Vc
specifies the location of the central pixel.
After obtaining the LBP features of the image of size P × Q
,the histogram can be calculated using Eqs. (3) and (4).
P Q
∑ ∑
Hist(l)∣LBP = s(LBP(xc ), l) (3)
p=1 q=1

Fig.  3  a LBP code demonstration, b MBP code demonstration where l ∈ [0, (28 − 1)].

13
Int. j. inf. tecnol.

translated to decimals by assigning weights to each binary

{
1, if a = b
s(a, b) =
0, Otherwise (4) value, and the MBP code can then be generated by summing
decimal values. MBP generates histograms with a size of
In Fig. 3a, 3 × 3-pixel windows are evaluated first, followed 256.
by comparing all eight neighbors to the center pixel. After
comparison, pixels with intensity value higher than or equal
to the central pixel are designated as 1 while those with 4 Proposed descriptor and system framework
intensity value less than or equal to the central pixel are des-
ignated as 0. This generates an 8-bit binary pattern, which This section describes the proposed descriptor with an
may be translated to decimals by assigning weights to each example, and also covers the step-by-step algorithm for
binary value, and the LBP code can then be generated by calculating MVM-LBP codes, the metric used for distance
summing decimal values. LBP generates histograms with similarity between feature vectors, and the evaluation
a size of 256. measures (precision, recall and f-score).

3.2 Median binary pattern (MBP) 4.1 Mean‑variance‑median based local binary pattern

(MVM‑LBP)
In [6], MBP descriptor has been suggested for texture analy-
sis. In this, all the eight neighbors of a central pixel in a 3 × The proposed MVM-LBP descriptor is based on the mean,
3 pixel window are compared with the median value of the variance, and median of all the pixels in a 3 × 3 pixel win-
whole window. After comparison, if the difference between dow. In MBP, all the eight neighbors are compared with
the neighbouring pixel and central pixel is greater than or the median of the entire 3 × 3 window. But, in MVM-LBP,
equal to zero then the corresponding neighbouring pixel all the eight neighbors are compared with an average of the
designated as 1 else 0. By assigning weights to each binary mean, square root of the variance, and median of all the
value, an 8-bit binary pattern may be generated, which can pixels in a 3 × 3 pixel window. Mathematically MVM-LBP
then be translated into decimals. The entire computation for is defined in Eq. (8) and (9).
the MBP code is shown in Fig. 3b. The MBP descriptor is
N−1
defined mathematically in Eqs. (5) and (6). ∑
MVM − LBPN,R (xc ) = k(VR,n − Vmvm ) × 2e (8)
N−1 e=0
∑
MBPN,R (xc ) = k(VR,n − Vmedian ) × 2e (5)
e=0 {
1, if y >= 0
k(y) =
0, Otherwise
for e = (0, 1, ........, N − 1) (9)
{
1, if y >= 0
k(y) =
0, Otherwise
for e = (0, 1, ........, N − 1) (6) Where N is the number of neighboring pixels at radius R ,
VR,n specifies the locations of the individual pixel, and VMVM
Where N is the number of neighboring pixels at radius R, is calculated using Eq. (10) for all the pixels in a 3 × 3 pixel
VR,n specifies the locations of the individual pixel, and Vmedian window.
specifies the pixels’ median in a 3 × 3 pixel window. ( )
After obtaining the MBP features of the image of size (Vmean + Vsd. + Vmedian )
Vmvm = floor (10)
P × Q , the histogram can be calculated using Eqs. (7) and 3
(4).
Where Vmean and Vmedian specifies the mean and median val-
P Q
∑ ∑ ues of all the pixels in a 3 × 3 pixel window. Vsd. is calculated
Hist(l)∣MBP = s(MBP(xc ), l); (7) using Eq. (11)
p=1 q=1
�
∑N
Where l ∈ [0, (28 − 1)]. (Vi − Vmean )2
Vsd. = i=0 (11)
In Fig. 3b, 3 × 3-pixel windows are evaluated first, fol- N
lowed by comparing all eight neighbors to the median of all
the nine pixels in the 3 × 3 pixel window. After compari-
son, pixels with intensity value higher than or equal to the After comparison, if the difference between the neigh-
central pixel are designated as 1 while those with intensity bouring pixel and central pixel is greater than or equal to
value less than or equal to the central pixel are designated zero then the corresponding neighbouring pixel designated
as 0. This generates an 8-bit binary pattern, which may be as 1 else 0. By assigning weights to each binary value, an

13
 Int. j. inf. tecnol.

Fig. 4  MVM-LBP code calculation for a pixel window of size 3 × 3

8-bit binary pattern may be generated, which can then be

translated into decimals. The entire computation for the Fig. 6  Transformed images corresponding to input images using pro-
MVM-LBP code is shown in Fig. 4. After obtaining the posed descriptor
MVM-LBP features of the image of size P × Q , the histo-
gram can be calculated using Eqs. (12) and (4). The size
of the histogram produced by MVM-LBP is 256. A block 4.2 Similarity measure
diagram of the proposed system is shown in the Fig. 5.
P Q
After feature extraction, feature matching is another impor-
∑ ∑ tant step in CBIR systems. Many distance similarity met-
Hist(l) = s(MVM − LBP(xc ), l); (12)
p=1 q=1 ric like, Euclidean distance, Manhattan distance, Canberra
distance, d1-distance, Mean-squared distance, etc. [40], are
Where l ∈ [0, (28 − 1)]. existing in the literature. In this paper experiments are per-
In Fig. 4, 3 × 3-pixel windows are evaluated first, fol- formed with all the above mentioned similarity metrics. Out
lowed by comparing all eight neighbors to the average of of these d1-distance similarity metric (Eq. 13) given the best
the mean, square root of the variance, and median of all the results. Results with all the distance metrics are presented in
nine pixels in a 3 × 3 pixel window. After comparison, pixels Table 5. Based on the smallest distances, identical images
with intensity value higher than or equal to the central pixel are retrieved. In our experiment, each image in the database
are designated as 1 while those with intensity value less has been treated as a query image once.
than or equal to the central pixel are designated as 0. This
N Fdb (i) − Fqj (i)
generates an 8-bit binary pattern, which may be translated to ∑
decimals by assigning weights to each binary value, and the
d1 distance = ∣ ∣ (13)
i=1
1 + Fdb (i) + Fqj (i)
MVM-LBP code can then be generated by summing decimal
values. MVM-LBP generates histograms with a size of 256. In all above equations left side is the respective distance
Fig. 6 shows the transformed images using the proposed between the feature vectors of database image db and jt h
descriptor. From the figure, the proposed descriptor is using query image qj . N represents the length of the feature vector.
boundary lines and main features from the image for features Fdb (i) and Fqj represent the feature vector of the database
extraction. image and jth query image.

Fig. 5  Block diagram of the

proposed methodology

13
Int. j. inf. tecnol.

4.3 Algorithm A
Precision(P) = (14)
B
Step-by-step working of the proposed method is presented
in Algorithm 1. In this, image retrieval using the proposed A
descriptor works in three steps. Image pre-processing is the
Recall(R) = (15)
C
initial step. In this step, all the images are loaded and con-
verted into gray-scale images of equal size. Feature vec- Where A is the total number of relevant images retrieved
tor generation is the second step. The feature vectors of the from the dataset, B is Total number of images in the data-
database images and the query image are evaluated using set, and C is Total number of relevant images present in the
Eqs. (8-11), and for histograms Eqs. (12) and (4) are used. dataset.
The distance between the query image and all the images in Further F-score is evaluated using P and R values from
the database is calculated using Eq. (13) in step 3. Finally, Eqs. (14) and (15). F-score is given in Eq. (16).
similar images are retrieved based on the least distance. 2×P×R
F − Score = (16)
P+R
Average recall rate (ARR) and Average precision rate (APR)
Algorithm 1 Features Vector (FV) calculation are calculated using Eq. (17).
using MVM-LBP
N N
Require: Input image I 1∑ 1∑
Ensure: Output F VM V M LBP ARR = R(qi ); ARR = P(qi ) (17)
N i=1 N i=1
1: Ir ⇐ number of rows in I  number of rows
in input image
2: Ic ⇐ number of columns in I  number of
columns in input image
3: i ⇐ 2
4: j ⇐ 2 5 Experiment results and analysis
5: while i = Ir − 1 do
6: while j = Ic − 1 do Different experiments are performed to check the excellence
7: J 0 ⇐ I (i, j )  central pixel position in of the proposed descriptor. This section provides experi-
a 2 × 2 matrix ments performed and the outputs.
8: calculate Vmean , Vmedian , and
Vsqrt(variance)
9: calculate Vmvm using Eq. (10) 5.1 Experiment
10: calculate M V M − LBP using Eqs. (8-
9) To perform the experiments, two publicly available facial
11: j ⇐j+1 databases, AT&T [7], and faces94 [8] database of facial
12: end while images are used. Both the databases are having the same
13: i⇐i+1 challenges of lighting, face emotions and facial features.
14: end while Both the databases are summarized in Table 2.

4.4 Evaluation measure
The performance of the proposed descriptor is compared 5.1.1 Experiment 1: AT&T database of facial images  [7]
to some of the existing state-of-the-art methods. Because
the proposed descriptor takes advantage of mean, variance The first database for our experiment is the AT&T face data-
and median, therefore, it generates more unique codes. To base [7]. This database contains 400 face images of 40 sub-
prove the excellence of the proposed descriptor, the results jects, and each subject is represented by ten different images.
are compared with LBP [9], MBP [6], 6 x 6 MB-LBP [12], The images for certain persons were taken several times,
CSLBP [11], NI-LBP  [10], and LCLBP [13]. with varied lighting, face emotions (open/closed eyes, smil-
In each experiment, the images are retrieved by taking ing / not smiling), and facial features (glasses / no glasses)
every image from the database as a query image. In this being used. Fig. 7a shows some of the sample images of this
paper, the precision and recall values are evaluated using database.
Eqs. (14) and (15).

13
 Int. j. inf. tecnol.

The maximum improvement of 30% is from LBP, and the

least improvement of 2.1% is from MBP is recorded. The
proposed descriptor outperformed by 21.14% from NI-LBP,
12% from CS-LBP, 11.77% from 6 × 6 MB-LBP, and 16%
from LC-LBP. The APR% and the ARR% are presented in
Table 3.

5.1.2 Experiment 2: faces94 facial images database [8]

Faces94 database [8] is the second database used in this

paper. There are 2980 color images of 149 people in this
database [8]. Each subject has 20 images with considerable
expression changes and minor changes in face position with
head turn, tilt, and slant. For this experiment have we con-
sidered only 50 individuals randomly with 20 images each.
Therefore, we have taken 1000 images in total. Fig. 7b shows
some of the sample images of this database.
The performance of the proposed descriptor on retrieving
the different numbers of images from the database is shown
Fig. 7  Sample images from a database 1 [7], and b database 2 [8] in Fig. 9. Initially, two images are retrieved, and the num-
ber of retrieved images is increased by two subsequently.
A maximum twenty number of images have been retrieved
The performance of the proposed descriptor on retrieving from this database. The APR of the proposed descriptor is
the different numbers of images from the database is shown outperformed the other descriptors on retrieving twenty
in Fig. 8. Initially, only one image is retrieved, and the num- images. The maximum increased of 51.22% is from LBP,
ber of retrieved images is increased by one subsequently. A and the least increased of 4.2% is from 6 × 6 MB-LBP. The
maximum ten number of images have been retrieved from proposed descriptor outperformed by 19% from NI-LBP,
this database. The APR of the proposed descriptor is out- 16.27% from MBP, 5.6% from CS-LBP, and 19.9% from
performed the other descriptors on retrieving ten images. LC-LBP. The APR% and the ARR% are shown in Table 4.

Table 2  Summary of databases Database Image size #classes #images Total Challenges

per class images

AT&T (gray) [7] 92x112 40 10 400 Lighting,

face emotions,
facial features
faces94 (color) [8] 180x200 149 20 2980
(utilized 50) (utilized 1000)

Fig.  8  a APR%, b ARR%, c F-Score, and d Prevision vs Recall, for database 1

13
Int. j. inf. tecnol.

Table 3  APR% and ARR% for # LBP NI-LBP MBP CSLBP 6 x 6MB-LBP LCLBP PM
database 1
Average precision rate (APR)%
  1 100.00 100.00 100.00 100.00 100.00 100.00 100.00
  2 74.38 79.38 85.75 82.50 84.13 80.38 87.50
  3 63.00 67.33 76.08 71.92 73.00 69.08 78.33
  4 55.00 60.75 69.25 64.06 65.56 62.00 71.63
  5 50.40 55.15 63.45 59.20 60.30 56.15 66.05
  6 46.54 50.58 59.00 54.50 54.83 52.13 60.88
  7 43.21 46.64 54.93 50.54 50.86 49.11 56.54
  8 40.72 43.53 51.53 47.09 47.34 45.78 52.38
  9 37.92 40.81 48.33 43.86 44.22 42.69 49.00
  10 35.90 38.55 45.70 41.35 41.78 40.25 46.70
Average retrieval rate (ARR)%
  1 10.00 10.00 10.00 10.00 10.00 10.00 10.00
  2 14.88 15.88 17.15 16.50 16.83 16.08 17.50
  3 18.90 20.20 22.83 21.58 21.90 20.73 23.50
  4 22.00 24.30 27.70 25.63 26.23 24.80 28.65
  5 25.20 27.58 31.73 29.60 30.15 28.08 33.03
  6 27.93 30.35 35.40 32.70 32.90 31.28 36.53
  7 30.25 32.65 38.45 35.38 35.60 34.38 39.58
  8 32.58 34.83 41.23 37.68 37.88 36.63 41.90
  9 34.13 36.73 43.50 39.48 39.80 38.43 44.10
  10 35.90 38.55 45.70 41.35 41.78 40.25 46.70

#No. of output images

The bold entries in the table is represented the outcome of the proposed method

Fig.  9  a APR%, b ARR%, c F-Score and d Prevision vs Recall, for database 2

Table 5 presented the summary of all the ARR obtained vectors generated using the proposed descriptor and state-
using different distance similarity metrics discussed in of-the-art descriptors except the CS-LBP is same. But, the
subsection 4.2. performance of the proposed descriptor is much better than
other techniques. Feature vector size and ARR% of all the
5.2 Computational cost used techniques on faces94 database [8] is presented in the
Table 6. From the table, the performance of the proposed
The performance of the CBIR systems is dependent upon descriptor is significantly improved from the other existing
the size of the feature vector (#FV). The size of the feature descriptors.

13
 Int. j. inf. tecnol.

Table 4  APR% and ARR% for # LBP NI-LBP MBP CSLBP 6 x 6MB-LBP LCLBP PM
database 2
Average precision rate (APR)%
 2 88.25 93.00 95.45 96.30 97.65 93.25 97.95
 4 77.73 85.73 89.83 92.35 94.78 88.38 96.20
 6 70.90 81.62 85.95 89.40 91.97 84.00 94.18
 8 66.29 78.81 82.26 86.88 89.09 80.45 92.21
 10 62.06 75.60 78.55 84.30 86.19 76.88 89.84
 12 58.88 72.86 74.95 81.30 83.05 73.31 87.21
 14 55.99 69.50 71.91 78.48 79.65 69.85 84.16
 16 53.05 66.64 68.76 75.58 76.37 66.68 80.88
 18 50.49 63.80 65.61 72.33 72.91 63.54 77.13
 20 48.04 61.01 62.48 68.77 69.72 60.56 72.65
Average retrieval rate (ARR)%
 2 8.83 9.30 9.55 9.63 9.77 9.40 9.80
 4 15.55 17.15 17.97 18.47 18.96 17.68 19.24
 6 21.27 24.49 25.79 26.82 27.59 25.20 28.26
 8 26.52 31.53 32.91 34.75 35.64 32.18 36.89
 10 31.03 37.80 39.28 42.15 43.10 48.43 44.92
 12 35.33 43.72 44.97 48.78 49.83 43.99 52.33
 14 39.19 48.65 50.34 54.94 55.76 48.90 58.91
 16 42.44 53.31 55.01 60.46 61.10 53.35 64.71
 18 45.45 57.42 59.05 65.10 65.62 57.19 69.42
 20 48.04 61.01 62.48 68.77 69.72 60.56 72.65

#No. of output images

The bold entries in the table is represented the outcome of the proposed method

Table 5  Performance of the proposed descriptor using different dis- is same as to the state-of-the-art techniques. In this paper,
tance measure techniques on both the databases different similarity measure metrics are used for experi-
Sr. No. Distance measure ARR% ment purpose but d1-distance metric given the best results.
For the proposed descriptor, the concluding points are
AT & T Faces 94
as follows:
1 Euclidean 37.45 62.21
2 Manhattan 42.6 68.7 1. The proposed descriptor takes the advantage of includ-
3 Canberra 45.2 58.34 ing more statistics measures for calculating the patterns.
4 Mean-Squared 37.42 60.18 2. The average precision/average recall has consider-
5 d1 46.7 72.65 ably increased from 35.90%, 38.55%, 45.70%, 41.35%,
41.78%, and 40.25% to 46.70% when compared with
The bold entries in the table is represented the outcome of the pro-
posed method
basic LBP, NI-LBP, MBP, CS-LBP, 6 × 6 MB-LBP, and
LC-LBP, respectively on AT&T database [7].
3. The average precision/average recall has significantly
6 Conclusion increased from 48.04%, 61.01%, 62.48%, 68.77%,
69.72%, and 60.56% to 72.65% when compared with
A novel descriptor MVM-LBP has been proposed in basic LBP, NI-LBP, MBP, CS-LBP, 6 × 6 MB-LBP, and
this paper. This descriptor performs well in precision LC-LBP, respectively on faces94 database [8].
and recalls values tested on two facial databases AT&T
database [7], and the faces94 database [8]. The proposed In the future, we should focus on generating more efficient
MVM-LBP descriptor takes the advantage of using extra descriptors. To reduce the size of the feature vectors, tech-
statistics while calculating MVM-LBP codes. Although, niques like PCA and LDA can also be used.
in comparison the size of the feature vectors in MVM-LBP

13
Int. j. inf. tecnol.

Table 6  Size of feature vector(#FV) and ARR% 15. Davis Larry S, Johns Steven A, Aggarwal JK (1979) Texture
analysis using generalized co-occurrence matrices. IEEE Trans
Sr. No. Method #FV ARR% Pattern Anal Mach Intell 3:251–259
AT&T [7] Faces 94 [8] 16. Zhang J, Li GL, He SW (2008) Texture-based image retrieval by
edge detection matching glcm. In 2008 10th IEEE International
1 LBP [9] 256 35.9 48.04 Conference on High Performance Computing and Communica-
tions, pages 782–786. IEEE
2 NI-LBP [10] 256 38.55 61.01
17. Laine A, Fan J (1993) Texture classification by wavelet packet sig-
3 MBP [6] 256 45.7 62.48 natures. IEEE Trans Pattern Anal Mach Intell 15(11):1186–1191
4 CSLBP [11] 16 41.35 68.77 18. Karanwal S, Diwakar M (2021) OD-LBP: Orthogonal difference-
5 6x6MB-LBP [12] 256 41.78 69.72 local binary pattern for face recognition. Digit Signal Process
110:102948
6 LCLBP [13] 256 40.25 60.56
19. Karanwal Shekhar, Diwakar Manoj (2021) Two novel color local
7 MVM-LBP(PM) 256 46.7 72.65 descriptors for face recognition. Optik 226:166007
20. Karanwal S, Diwakar M (2021) Neighborhood and center dif-
The bold entries in the table is represented the outcome of the pro- ference-based-LBF for face recognition. Pattern Anal Appl
posed method 24(2):741–761
21. Dubey SR, Mukherjee S (2020) LDOP: local directional
order pattern for robust face retrieval. Multimedia Tools Appl
References 79(9):6363–6382
22. Verma M, Raman B (2016) Local tri-directional patterns: a new
1. Al-halafawy Waleed S, Tawfiq Marwa Z (2014) The relationship texture feature descriptor for image retrieval. Digit Signal Process
between types of image retrieval and cognitive style in developing 51:62–72
visual thinking skills. Life Sci J 11(9):865–879 23. Verma M, Raman B (2018) Local neighborhood difference
2. Gudivada Venkat N, Raghavan Vijay V (1995) Content based pattern: a new feature descriptor for natural and texture image
image retrieval systems. Computer 28(9):18–22 retrieval. Multimedia Tools Appl 77(10):11843–11866
3. Ashish Mohan Yadav BPSS, and (2014) A survey on:‘content 24. Bedi AK, Sunkaria RK (2021) Mean distance local binary pattern:
based image retrieval systems’. Int J Emerg Technol Adv Eng a novel technique for color and texture image retrieval for liver
4(6):22–26 ultrasound images. Multimedia Tools Appl 80(14):20773–20802
4. Liu Y, Zhang D, Lu G, Ma WYing (2007) A survey of content- 25. Chandan S, Ekta W, Preet KK (2018) Color texture description
based image retrieval with high-level semantics. Pattern Recognit with novel local binary patterns for effective image retrieval. Pat-
40(1):262–282 tern Recognit 76:50–68
5. Dharani T, Aroquiaraj IL (2013) A survey on content based image 26. Agrawal S, Chowdhary A, Agarwala S, Mayya V, Kamath SS
retrieval. 2013 International Conference on Pattern Recognition. et al (2022) Content-based medical image retrieval system for
Informatics and mobile engineering. IEEE, IEEE, Cham, pp lung diseases using deep cnns. Int J Inf Technol. https://​doi.​org/​
485–490 10.​1007/​s41870-​022-​01007-7
6. Adel H, Guna S, Bertrand Z (2007) Median binary pattern for 27. Rosline NG et al (2022) Image retrieval system based on multi
textures classification. International Conference Image Analysis feature extraction and its performance assessment. Int J Inf Tech-
and Recognition. Springer, Berlin, pp 387–398 nol 14(2):1161–1173
7. AT&T laboratories cambridge, the AT&T database of faces, avail- 28. Kanaparthi SK (2022) Content based image retrieval on big image
able online:. http://​www.​uk.​resea​rch.​att.​com/​faced​ataba​se.​html data using local and global features. Int J Inf Technol 14(1):49–68
Access July, 2021 29. Khaleel A, Monika S, Madhup S, Abbas RM, Vishal J (2020)
8. Libor Spacek. Faces94 dataset. https://c​ mp.f​ elk.c​ vut.c​ z/~ ​ spacel​ ib/​ An efficient image retrieval tool: query based image management
faces/​faces​94.​html system. Int J Inf Technol 12(1):103–111
9. Ojala T, Pietikäinen M, Harwood D (1996) A comparative study 30. Prithaj B, Kumar BA, Avirup B, Pratim RP, Subrahmanyam M
of texture measures with classification based on featured distribu- (2018) Local neighborhood intensity pattern-a new texture feature
tions. Pattern Recognit 29(1):51–59 descriptor for image retrieval. Expert Syst Appl 113:100–115
10. Liu L, Zhao L, Long Y, Kuang G, Fieguth P (2012) Extended local 31. Hongliang J, Liu Qingshan L, Hanqing TX (2004) Face detection
binary patterns for texture classification. Image Vision Comput using improved LBF under bayesian framework. Third Interna-
30(2):86–99 tional Conference on Image and Graphics (ICIG’04). IEEE, pp
11. Heikkilä M, Pietikäinen M, Schmid C (2009) Description of 306–309
interest regions with local binary patterns. Pattern Recognit 32. Pavel Kál, Antonín V, Ladislav L (2019) Enhanced local binary
42(3):425–436 patterns for automatic face recognition. International conference
12. Shengcai L, Xiangxin Z, Zhen L, Lun Z, Li Stan Z (2007) Learn- on artificial intelligence and soft computing. Springer, pp 27–36
ing multi-scale block local binary patterns for face recognition. 33. Sucharitha G, Kumar SR (2020) Biomedical image retrieval by
International Conference on Biometrics. Springer, Berlin, pp using local directional edge binary patterns and zernike moments.
828–837 Multimedia Tools Appl 79(3):1847–1864
13. Rinku DR, Chandra NS, Ranjan KD (2017) An improved 34. Amit S, Megha A, Bilas PR (2021) Directional local ternary co-
local pattern descriptor for biometrics face encoding: a LC- occurrence pattern for natural image retrieval. Multimedia Tools
LBP approach toward face identification. J Chinese Inst Eng Appl 80(10):15901–15920
40(1):82–92 35. Karanwal S (2021) Graph based structure binary pattern for face
14. Haralick Robert M, Karthikeyan S, Hak DI (1973) Textural analysis. Optik 241:166965
features for image classification. IEEE Trans Syst Man Cybern 36. Karanwal S (2021) Multi-scale neighbourhood based-tree binary
6:610–621 pattern: a new feature descriptor for face recognition. Int J Biom-
etrics 13(2–3):322–342

13
 Int. j. inf. tecnol.

37. Martolia M, Dhanore N, Singh A, Shahare V, Arora N (2020) Springer Nature or its licensor (e.g. a society or other partner) holds
A modified local binary pattern (LBP) for content-based image exclusive rights to this article under a publishing agreement with the
retrieval. Int Adv Sci Technol 29(1):1630–1644 author(s) or other rightsholder(s); author self-archiving of the accepted
38. Srishti G, Pratim RP, Prosad DD, Byung-Gyu K (2020) Retrieval manuscript version of this article is solely governed by the terms of
of colour and texture images using local directional peak valley such publishing agreement and applicable law.
binary pattern. Pattern Anal Appl 23(4):1569–1585
39. Bedi AK, Sunkaria RK (2020) Local tetra-directional pattern-a
new texture descriptor for content-based image retrieval. Pattern
Recognit Image Anal 30(4):578–592
40. Verma M, Raman B, Murala S (2015) Local extrema co-occur-
rence pattern for color and texture image retrieval. Neurocomput-
ing 165:255–269

13