International Journal of Advanced Science and Technology

Vol. 29, No. 3s, (2020), pp. 1715-1721

ROBUST FACE VERIFICATION USING SPARSE REPRESENTATION

P. Ajay Kumar Reddy1, Dr G.N Kodanda Ramaiah2, Dr M.N Giri Prasad3
1
Research Scholar, Dept. of ECE, JNTU Ananthapuramu/ KEC Kuppam.
2
Professor, Dept. of ECE, Kuppam Engineering College, Kuppam.
3
Professor, Dept. of ECE, JNTU Ananthapuramu.

Abstract
For identity of humans and their emotional status of the human, face carries more information.
Recognition of face becomes interesting and challenging task. Most researchers investigated
extensively on faces because of its huge range of applications. The automatic recognition of face
from different angles, changes in illumination, expressions, noise in the images are several issues of
face recognition systems. In this paper, the research work concentrated mainly on face detection,
feature extraction and classification. In face detection, by successive methods faces are detected.
LBP method was used for feature extraction. For classification, the features of face are related with
the features of faces from the database. Here Sparse Representation is used for classification. The
experimental result indicates that the proposed approach has better efficiency in face verification
under noise conditions.

Keywords: Face Verification, Face Recognition, LBP, PCA, Sparse Representation

Classifier (SRC).

1. Introduction

In last two decades, Facial Recognition Techniques (FRT) has gained a huge interest in the
image and computer vision applications. FRT is extensively used in surveillance systems [16] for
automatically detecting and recognition of the human faces. FRT requires familiarity in the fields
likes computer machine learning, psychology, Artificial Intelligence, image recognition, computer
vision and pattern recognition. Most researches evaluate the Face Recognition system (FRS) in terms
of Recognition rate. The recognition rate of the FRS is affected by variations in face image detection,
extraction of features in a face and face identification.
Face recognition method includes authentication and verification. In Authentication process
a query image is compared with several images of the same person. In identification process the
query image is compared with various other faces of several persons which are present in the
database.

The face verification or recognition process starts with detecting the face or a number of
faces in given image. The major issue with the face detection process are, variation in the facial
expressions, position of face, orientation, illumination and occlusions in a given face image. These
issues have to be addressed first by the face detecting system because they effect the accuracy of face
detection.

Once the face is detected in the given image, the features are extracted from the detected
face. The extracted features are converted into a feature vector which is adequate to represent a given
image.

General block diagram is of face recognition/verification system shown in Fig 1.

ISSN: 2005-4238 IJAST 1715

Copyright ⓒ 2020 SERSC
 International Journal of Advanced Science and Technology
Vol. 29, No. 3s, (2020), pp. 1715-1721

TRAINING PRE DETECTION FEATURE

IMAGES PROCESSIN EXTRACTION
G

OUTPUT CLASSIFICATIO
N

TEST PRE DETECTION FEATURE

IMAGES PROCESSIN EXTRACTION
G

Fig 1 Basic algorithm of face recognition/verification system

II. Face Recognition Techniques

Computer vision offers a high demanding applications and outcomes specifically face
recognition. This area has always become the researchers major focus in image analysis because of
its nature as human-face primary identification method. It is very interesting and becomes such a
challenge to teach a machine to do this task. Numerous robust algorithms have been developed and
claimed to have accurate performance to tackle face recognition problems. These algorithms or
methods are the most successfully and widely used for face detection and recognition applications.

There are three main approaches for face verification [16]. To comprehend the standard of every
procedure, we will order these methodologies into three classes as per the way of treating the face
image. In the first approach, the total face is considered as an input image or query image. This
method is called as global approach. Here the entire face image is projected into subspaces of lesser
dimension. The various global approaches used widely are Eigenface [2], 2D PCA [3], LDA [4],
SVM [5], KICA [6] etc. Though they are faster in implementation they are sensitive to the variations
in poses, expressions and illuminations.

In the second approach, instead of entire face image, the system extracts few features from the
detected face. This method is called local approach. The various local approaches are EGBM [7],
DLA [8], LBP [9], Haar wavelets [10], SIFT [11] etc. the performance of those approaches depends
mainly on the effectiveness of the algorithms.

The third approach is called as hybrid approach, which includes both features of local and global
approaches. This method uses the statistical model, which uses the mathematical formulae to
describe the relationships between various random variables. The model is viewed as factual when
the factors are not deterministic, yet stochastically related. The various hybrid approaches are PCA+
Gabor [12], HMM [13], GWT-PHMM [14], HMM-LBP [15] etc. Though these algorithms are
efficient these methods are difficult to implement than the other approaches.

Local Binary Patterns (LBP)

The LBP is extensively is used for texture and shape analysis [19]. In Face recognition systems
(FRS) LBP is used for detection and recognition of the faces in an image. This algorithm is used to
extract the facial features. Ojala et.al [19] developed the LBP algorithm as texture descriptor. The
LBP technique divides the face image into smaller regions. Each region is divided into eight
neighbourhood pixels with combination of a centre pixel in an 3x3 region. The grey value of the

ISSN: 2005-4238 IJAST 1716

Copyright ⓒ 2020 SERSC
 International Journal of Advanced Science and Technology
Vol. 29, No. 3s, (2020), pp. 1715-1721

centre pixel is compared with the values of its neighbourhood pixel. If the grey value of neighbour
pixel is greater then the centre pixel then it is treated as `1` else it is treated ad `0`. Final the LBP
code is generated by combining these eight values of either ones or zeros. These codes are concise
into histograms. These smaller histograms are combined together to represent the description of face
image. The LBP code generation is represented in the below figure.

90 60 10 1 1 0

10 40 30 0 0

50 20 80 1 0 1
LBP Code: 11001010

Figure2: LBP Code Generation.

The LBP code is mathematically represented by the below equation[18]

Where ( , ) represents centre pixel coordinates, ic is value of centre pixel and in is value of
neighbourhood pixel.

Sparse Representation Classification (SRC)

SRC has been widely used for image analysis.SRC [17] has been used for classification
problems. The SRC method for face recognition was developed by Wright et.al [20]. It performs
pixelwise classification of a facial images. The SRC uses sparse based representation and learned
redundant dictionary to classify facial images. The SRC means over all the coefficients of base
vectors, there are only a few entries as non-zero. This representation will naturally select the subset
of base vectors which are more concentrated then rejects the less concentrated vectors. This feature
makes the SRC method more robust for classification tasks.

The main issue in object recognition is to work on labelled training patterns from
different object classes and to accurately measure the class to which a query pattern exists. We
construct the provided training patterns from the class as columns of a matrix
. where of gray image.

when sufficient training patterns of the object class are provided then,
, any query pattern , from like class will almost
exist in the linear span of the training patterns related with object :

For some scalars, .

Initially ith test pattern membership is unknown, the n training patterns of all k classes are
concatenated to represent as a new matrix A.

ISSN: 2005-4238 IJAST 1717

Copyright ⓒ 2020 SERSC
 International Journal of Advanced Science and Technology
Vol. 29, No. 3s, (2020), pp. 1715-1721

Then, the linear representation of y can be written as

Here the coefficient vector whose entries are zero

except which are related with the class. The identity of the test pattern y is encoded by the vector
is obtained by solving the equation of linear system

In robust face verification the problem arises as the value of x is underdetermined in the
equation , this difficulty is overcome by using the minimum -norm solution

The sparse representation is computed from a test pattern y which is taken from the same class in
training set. All non-zero objects in the are arranged in the columns of A from a single object
class i, then test sample y can be assigned to the same class.

For class i, the coefficients related to the class are selected from a charectaristic function . For
is a vector which consists of all the non-zero objects in which are related with
class . The coefficients associated with the class determine the test pattern y as .
Now is classified based on the approximations by giving it to object class which reduces the
residual between and :

Algorithm :
Step 1: Assign an input matrix of training samples
A=[A1,A2,……..AK] Rm×n for k classes ,a test samples y Rm,
Step 2: Solve the -norm minimization problem

Step 3: Compute the residuals

for i=1,2,……,k.
Step 4: Output: identify(y)= arg mini ri(y).

Face Train Images

Feature Extraction

Test Image Dictionary Creation

Feature Sparse
Representation Recognition
Extraction
Classification

Figure 3: Block diagram of Proposed Method.

ISSN: 2005-4238 IJAST 1718

Copyright ⓒ 2020 SERSC
 International Journal of Advanced Science and Technology
Vol. 29, No. 3s, (2020), pp. 1715-1721

III. EXPERIMENTAL RESULTS

For experimental results, ORL database was used in this paper. Fig 3 shows the images of
ORL database. Here, different feature extraction algorithms namely LBP, PCA and LDA were used
and Sparse Representation Classification technique undergoes various pre-processing techniques like
image resizing, face detection and finally recognized images are obtained. The percentage of
identification rate (IR) is shown in Table 1. By the table LBP with SRC has good efficiency than
LDA and PCA techniques.

Table 1: Percentage of Identification Rate

Method % of identification rate
LBP+SRC 98
PCA+SRC 92
LDA+SRC 74

For testing the above techniques under different conditions we choose noise
environment.Table 2shows Identification Rate for different percentages of Noiseenvironment.The
table 2 shows the LBP with SRC has good efficiency than LDA and PCA techniquesin noise
environment and reduction of the recognition rate as percentage of noise is increasesincreased
gradually, by seeing the graph it is identified the up to 3% Recognition Rate reduces slowly, when it
crosses the 4% then the recognition is roll-off.

Table 2:Identification Rate for different percentages of Noise (Salt & Pepper)
% of noise 0 1 2 3 4 5 6 7 8 9 10
LBP+SRC 98 97 95 92 82 76 66 54 48 41 38
PCA+SRC 92 90 89 87 75 68 61 50 35 22 18
LDA+SRC 74 72 69 67 55 48 42 32 23 17 10

The reason for roll-off of the IR is as noise introduced in the image there will be change in
the extracted features, it is known that, change in the extracted features will affectthe Identification
rate.

Fig 4 illustrate the roll-off of the IR while noise is introduced in the different algorithms like
LBP+SRC, PCA+SRC & LDA+SRC. The X-axis is taken as the % of noise introduced in algorithms
and on Y-axis Recognition Rate of the different algorithms is taken.

ISSN: 2005-4238 IJAST 1719

Copyright ⓒ 2020 SERSC
 International Journal of Advanced Science and Technology
Vol. 29, No. 3s, (2020), pp. 1715-1721

Fig 4: Graph of recognition rate vs % of noise

IV . CONCLUSION

This paper is novel in identifying the performance to algorithm like LDA+SRC, PCA+SRC
and LBP+SRC. Here performance is calculated by applying salt & pepper noise to the three hybrid
algorithm and finding the recognition rate. It is identified that recognition rate is will be roll-off after
4% in every algorithm, due to the drastic change in the extracted feature by applying the noise.
Performance of the LBP+SRC is high because of its less tendency to the change in the extracted
features due to add of noise, PCA takes the more features when compared to the LBP so variation the
features, so that it had high roll-off when compared to the LBP

Sparse Representation Classification had de-noising technique which reduce the effect of the
noise on the features variation, so it will reduce roll-off of recognition rate. For this reason, SRC is
applied to feature extraction algorithm. SRC had the one more advantage it classifies the large
number of features into the smaller number of classes, so that processing time will be reduced when
compare to other classifiers like decision tree, support vector machine and nearest neighbor.

References
1. Bo Yang & Songcan Chen, “A comparative study on local binary pattern (LBP) based face
recognition: LBP histogram versus LBP image,” Elsevier, Neurocomputing, 3 April 2013, pp.
365-379.
2. Turk, A.; Pentland, A.P. Face recognition using eigenfaces. In Proceedings of the IEEE
Conference onComputer Vision and Pattern Recognition, Maui, HI, USA, 3–6 June 1991; pp.
586–591.
3. Jian, Y.; Zhang, D.; Frangi, A.; Yang, J.-Y. Two-dimensional PCA : A new approach to
appearance-based facerepresentation and recognition. IEEE Trans. Pattern Anal. Mach. Intell.
2004, 26, 131–137.
4. Belhumeur, P.N.; Hespanha, J.P.; Kriegman, D.J. Eigenfaces vs. Fisherfaces: Recognition
using class specificlinear projection. IEEE Trans. Pattern Anal. Mach. Intell. 1997, 19, 711–
720.
5. Vapnik, V.N. The Nature of Statistical Learning Theory; Springer: New York, NY, USA,
1995.
6. Bach, F.; Jordan, M. Kernel independent component analysis. J. Mach. Learn. Res. 2002, 3,
1–48.
7. Wiskott, L.; Fellous, J.M.; Kuiger, N.; von der Malsburg, C. Face recognition by elastic
bunch graph matching.IEEE Trans. PAMI 1997, 19, 775–779.
8. Lades, M.; Vorbruggen, J.C.; Buhmann, J.; Lange, J.; von der Malsburg, C.; Wurtz, R.P.;
Konen, K.Distortion invariant object recognition in the dynamic link architecture. IEEE
Trans. Comput. 1993, 42,300–311.

ISSN: 2005-4238 IJAST 1720

Copyright ⓒ 2020 SERSC
 International Journal of Advanced Science and Technology
Vol. 29, No. 3s, (2020), pp. 1715-1721

9. Ahonen, T.; Hadid, A.; Pietikainen, M. Face recognition with local binary patterns. In
ComputerVision—ECCV 2004; Springer: Berlin/Heidelberg, Germany, 2004; pp. 469–481.
10. Viola, P.; Jones, M.J. Robust real-time face detection. Int. J. Comput. Vis. 2004, 57, 137–
154
11. Lowe, D.G. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis.
2004, 60, 91–110.
12. Cho, H.; Roberts, R.; Jung, B.; Choi, O.; Moon, S. An efficient hybrid face recognition
algorithm using PCAand GABOR wavelets. Int. J. Adv. Robot. Syst. 2014, 80,
doi:10.5772/58473.

13. Nefian, A.V.; HIII, M.H. Face detection and recognition using hidden Markov models. In
Proceedings of theIEEE International Conference on Image Processing, Chicago, IL, USA,
4–7 October 1998.
14. Kar, A.; Bhattacharjee, D.; Nasipuri, M.; Basu, D.K.; Kundu, M. High performance human
face recognitionusing gabor based pseudo hidden Markov model. Int. J. Appl. Evol. Comput.
2013, 4, 81–102.
15. Chihaoui, M.; Bellil, W.; Elkefi, A.; Amar, C.B. Face recognition using HMM-LBP. In
Hybrid Intelligent Systems;Springer: Cham, Switzerland, 2015; pp. 249–258.
16. Chihaoui, M.; Elkefi, A.; Bellil, W.; Ben Amar, C. A Survey of 2D Face Recognition
Techniques. Computers 2016, 5, 21.
17. Sunitha, A. et al. “Recognition of Facial Emotions Based on Sparse Coding.” (2017).
18. P. Ajay Kumar Reddy, G.N Kodanda Ramaiah, M.N Giri Prasad “Efficient Facial
Recognition System using LBP Technique” International Journal of Recent Technology and
Engineering (IJRTE) ISSN: 2277-3878, Volume-8, Issue-1, May 2019.
19. T. Ojala, M. Pietik¨ainen and D. Harwood, “A comparative study of texture measures with
classification based on feature distributions” Pattern Recognition vol. 29, 1996.
20. J. Wright, A. Yang, A. Ganesh, S. Sastry, Y. Ma, "Robust face recognition via sparse
representation", IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 31, no. 2, pp.
210-227, 2009.

ISSN: 2005-4238 IJAST 1721

Copyright ⓒ 2020 SERSC