68
Face Recognition using Local Graph Structure and Support Vector Machine (LGS-SVM)
Aaisha S. R. Al-Shibli, and Eimad Abusham
Abstract--Face recognition has been one of a distinguish research in the field of image processing. Nowadays, face recognition is used in variety of applications for the purpose of providing a high level security. Throughout many years, many methods have been developed and used with different classifiers and techniques in order to classify the images. In this paper, Local Graph Structure is proposed to improve the appearance of images based on image textures, then extracted features based on histogram equalization are fed to SVM for recognition and classification. The proposed method is tested and evaluated on standard benchmark ORL database. The experiments conducted revealed the efficiency of the proposed methods in this paper. Also, proved the effectiveness of the proposed method in reducing the number of input features, while increasing the accuracy of the classification.
Keywords--LGS, SVM, face recognition, pattern recognition and histogram.
I. INTRODUCTION
There have been successful and distinctive applications in all areas that recently, tightening the attention of the majority in the past and present years, that is image processing in general and face recognition in specific. Face recognition is the remarkable part of the ability of perception systems and routine tasks for people, building such systems is still the area of research [1]. The established work of face recognition might be traced back to 1950s in the area of psychology [2] and to the literature of engineering in the 1960s [3], whereas, the real study of face recognition machine established in the 1970s [4].
According to the study that done by Zhao, Chellappa, Phillips and Rosenfeld [1], there is a growth interest in the area of face recognition due to various reasons. The first one is the verity of commercial needs and less mandatory applications, where organizations, companies and people need a secure machine to protect their data. In addition, the availability of new and advanced technologies encourages researchers to engage in area of face recognitions and its applications. Recognizing faces had has some problems and they continue to attract the attention researchers. In the meanwhile, there are wide areas that can be applied for face recognition. These areas are growing fast according to the demands and needs for a secure system that can provide privacy and protection to information, which necessarily needs to be hidden from others. In addition, certain issues are still being debated by psychophysicists and neuroscientists, i.e. whether or not face recognition in the human visual system (HVS) is a dedicated process [5] [6] [7]
[8] and whether it is done holistically or by local feature analysis. Various approaches in face recognition have been
proposed in the literature; these can be classified into three categories, namely feature-based, holistic (global), and hybrid methods. While feature-based approaches compared the salient facial features or components detected from the face, holistic approaches make use of the information derived from the whole face pattern. By combining both local and global features, the hybrid methods attempt to produce a more complete representation of facial images. In this project, the local method is used which is local graph structure (LGS) with the support vector machine (SVM) classifier.
II. LITERATURE SURVEY
In the past, global image processing was used heavily in many studies, which is the first type of image processing. In global approach the system used to identify face image using global representations based on the entire image rather than local feature of the face, but this approach began to decline in recent years. Local approaches process the image first before the recognition or identification of the image. There are several global concepts, which have been used for different algorithms using global image processing [9]. In 2000, LDA/PCA algorithm was proposed for face recognition by Jie Yang, Hnaya and William Kunz, that worked by maximizing the criterion of LDA and PCA [10]. In September 2002, Gian Luca and Fabio Roli [11] described various methodologies in order to enhance the robustness of face recognition systems. One of these methods was fusion of both Principle Component Analysis and Linear Discriminant Analysis (LDA), in his method two types of approaches were identified, which are appearance based and structured-based approach. Appearance- based is able to view all features vector on the images, and structural approaches is used in representing faces [11]. In the year of 2008, the performance of Principle Component Analysis (PCA) was evaluated by different researchers in order to identify the differences of face images of individuals [12].
Second type of image processing is feature based approach (local). According to the results of many experiments indicated that there has been significant improvement in the performance of feature approaches due to the use of accurate alignment of the face [13]. However, the time required to finish the computation is long, because in order to recognize the face, all the images inside the database need to be searched and comparisons made in every image for each feature, which causes a dense computation [13]. In 2005, Matti started to study and analyze images using local binary patterns (LBP) taking in to consideration, the problems of texture analysis [14]. The methods of local binary patterns have been used in variety of applications analyzing biomedical and face images, motion and
69 scene [15]. Matti and his group studied tasks of machine vision adapting methodologies of LBP on many aspects and face recognition[14]. In 2006, Zhao and Pietikainen produced another method in dynamic texture recognition using LBP in specific volume. The method is using three planes, each one of these planes is orthonormal to the other two planes. The LBP code extracted from the three planes joined their features into one histogram that combined appearance and motion [16].
Features extractions means a decrease in the size of resources needed to characterize the image especially with high categories of data. There has been lots of techniques used to extract features for example, in 2008, Sarjay proposed a system for recognizing faces automatically depending on a detector - corner Harris for multi scale- in order to detect important points. The recognition system worked as follows: 1) separating regions of the face in each image to create segmentations of the face, 2) from the separated regions, the important point’s detected, 3) important points selected and features vector locally extracted, 4) the feature vector normalized against photometric and geometric transformation, 5) the system trained using Eigen approach[17]. In 2010, Chandranda, Vijaya, and Subbaaiah introduced techniques based on extracting the features of texture for face recognition.
The method divided the face image into four parts in order to avoid complexity. Moreover, each one of these parts evaluates its features of texture separately, where the features were derived from parameters called co-occurrence. Then, only the average results of each part are considered as final result [18].
A combination of three algorithms introduced by Yi, Furong and Guoqing, where they proposed a method to extract features based on PCA Principles component analysis (principles component analysis), LBP(Local binary pattern) and MSD (Maximum scatter differences). The aims of the method are to overcome the problems that occurred with PCA and LBP. It worked by segmenting the image into a number of images.
Then, extract the features (histogram) using LBP function, that manage the image’s pixels using 3 x 3 thresholds neighbors by center pixel and binary number (using two bits 1, 0) conceder as a result . After that, the PCA used on the histogram in order to reduce the dimensions of features [19]. Recently, Local Graph Structure (LGS) was introduced for recognizing the faces by Eimad Abusham and Housam Bashir. In LGS, the image represented by a graph of points and it works by selecting a threshold as a target pixel from six points neighbors organized in two regions. The target pixel used to compare with other pixel started from left to right region. If the pixel has higher or same value as target, then 1 (binary value) will be placed on the edge that connecting the two nodes, otherwise 0 will be assigned on the connection. ORL database had been used to test LGS [20].
Many researcher has been used SVM classifier for different aspects. An algorithm of SVM invented in 1963 by A.
Chervonenkis and V. Vapnik [21]. In 1992, Vladimir Vapnik, Guyon and Boser had introduced support vector machine [22], but the basis of Support vector machine was developed by Vapnik. In 1995, SVM publication started [21] where, Vladimir and Cortes used SVM to construct a surface of a linear
decision [23] and extended their study to a set of training data that are not separable [24]. Several studies on SVM have been done since 1995. Some of these studies were trying to find a relation between SVM and any sort of algorithm such as Federico that found a relationship between SVM and Sparse approximation, where both of them gave same solution if both use same set of data [25]. Also, Baudat and Anouar used the idea of SVM to proposed Generalized Discriminant Analysis (GDA), which is the nonlinear variables of input space mapped in to appropriate feature space [26]. Many researchers has participated and collaborated in image processing in general.
Most of them are debated that one type of image processing is better than others. If fact, what is really decide, which type is better is the research-oriented of researcher. The recent research-oriented focus on feature based approach (Local) due to its advantages compared with global. Also, the researchers participated in hybrid of both local and global, but bot global alone.
One of the feature based approach is LGS that has significant advantages such as fast, simple and has better performance. LGS used in this thesis together with SVM.
Therefore, Face recognition model is introduced using LGS and SVM. It is a model based approach using three diminution.
LGS [20] is latest algorithm used to enhance the appearance of images locally. It improves the lightness of face image in short time. Also, it is simple technique. Therefore, LGS is used in this thesis in order to gain its advantage on the used images.
The contribution is to construct histogram features from the enhanced images that has processed by LGS. Then the histogram feed to a classifier to classify image. Due to the wide usage [27] and advantages [28] of SVM such as providing accuracy to data, SVM is used to classify images.
III. FACE RECOGNITION FRAMEWORK Set of face images are taken from ORL database containing about 400 images and divided them randomly into two groups, which are training and testing set. Training set includes 320 face images and testing set includes 80 pictures of face images. All the face images are gray color, and variant in the appearance and subjects. First work done on training images, which is preprocessing operations. The preprocessing of images include process the images using Local Graph Structure (LGS) algorithm to improve the appearance. Then, the discriminate features are extracted. All discriminate features use to display a statistical information (histogram) in which the frequency of item’s data has an equal size (1-255).
The histogram is fed to SVM in order to recognize the face. If the face is recognized, then class label (1-40) is assigned.
Otherwise no match will appear. Secondly, same processes is applied on testing set images and following same method’s steps in order to check the ability of recognition using Support Vector Machine (SVM). The nonlinear and multi SVM is used for classification. Since, there exists 400 face images in ORL, therefore, around 40 classes are used for classification. Figure 1. Illustrates face recognition framework.
70
Fig. 1. Framework of Face Recognition
In LGS algorithm, the image is represented by a graph, which consists of points and it works by choosing a threshold as a target pixel from neighbors of six points organized into two regions. The idea is to compare between the existing points one after another from left to right until all points are scanned, searching for points of high value. The steps are:
The target pixel used to compare with other pixel started from left to right region
If the pixel has higher or same value as target, then 1 (binary value) will be placed on the edge that connecting the two nodes, otherwise 0 will be assigned on the connection.
The result from the above steps are a clear image with good performance compared with other algorithms.
Extract pixel density values, which is known as histogram, where each graph represents values of image’s pixel at different density values. Then the resulting values were fed into SVM for classification.
ORL database had been used to test the proposed LGS-SVM, where 40 variant subjects exist with 10 images for each. In specific, 2 images used for testing and 8 images used for training from the existing image face database. Figure 2. Illustrates sample of faces.
Fig. 2. Sample of Faces from ORL database
IV. LGS-SVM
Training and testing data set are processed using local graph structure.
The output of the processed images is used to construct the discriminant features (histogram).
The histogram values used as input to SVM.
The output value between 1 and 40 which is the class label.
V. EXPERIMENTS
Stage 1: 320 images has been chosen as a training set with eight face images for each subject and the remaining 80 faces were used as testing set.
Stage 2: LGS algorithm is performed on both samples of images.
Stage 3: Histogram extracted for each image.
A. SVM
Support vector machine is used for the classifications based on histograms of the processed images. The images are represented by numerical value. The declaration of images is as each subject image belongs certain class as they appear in the database. For example, in the training set, images from 1-8 belongs to first subject, so the representation of the images will be (11111111) for the class label and so on.
In the training set, the image files are declared in a groups, where the images belonging to same subject are having same class value. An experiments were conducted for recognition and the results are shown in table 1 below.
71
TABLE 1
Subject Number Range of Images Classes it belongs
1 1-8 1
2 9-16 2
3 17-24 3
4 25-33 4
5 34-41 5
6 42-50 6
7 51-59 7
8 60-68 8
9 69-76 9
10 77-84 10
11 85-93 11
12 94-102 12
13 103-111 13
14 112-120 14
15 121-129 15
16 130-138 16
17 139-146 17
18 147-154 18
19 155-162 19
20
…
163-170
…
20
…
LGS-SVM perform well and recognition rate is shown in table 2. Although, two images were wrongly classified.
Table 3 below illustrates the recognition rate of testing images.
TABLE 2
Recognition Rate % Error Rate%
97.5% 2.5%
Table 3 shows the LGS-SVM algorithm applied on two images from each class and recognition rate of each image is tabulated below in the figure. All the classes were used.
TABLE 3
Class No Test Image Recognition Rate
1 1
2
100%
100%
2 1
2
100%
100%
3 1
2
100%
100%
4 1
2
100%
100%
5 1
2
100%
100%
6 1
2
100%
100%
7 1
2
100%
100%
8 1
2
100%
100%
9 1
2
100%
100%
10 1
2
100%
100%
11 1
2
Error Error
12 1
2
100%
100%
LGS-SVM
VI. CONCLUSION
Face recognition is a technique used to recognize face images through search over certain face image database.
Throughout the years, lots of methods invented and developed.
In this paper, face recognition model based on fusion of Local Graph Structure (LGS) and Support Vector Machine (SVM) is proposed. LGS is a simple and fast method used to recognize and improve the appearance of face image. LGS is applied to process train and test then, the histogram is extracted to process. Finally, the for recognition SVM is used. All the face images that are used in the training set and testing set are from ORL face database.
REFERANCE
[1] W.ZAHO, R.CHELLAPPA, P.J.PHILLIPS and A.ROSELFELD, "Face Recognition: ALiterature Survey," ACM ComputingSurveys, 2003.
[2] B. I. S. and T. R. , The perception of people, Lindzey: Handbook of Social Psychology, 1954.
[3] B. W. W., "The model method in facial recognition," Panoramic research Inc, Palo Alto, 1964.
[4] M. D. KELLY, "Visual identification of people by computer," Stanford AI Project, Stanford, 1970.
[5] I. BIEDERMAN and K. P. , "Neural and psychophysical analysis of object and face recognition," In Face Recognition: From Theory to Applications, pp.
3-25, 1998.
[6] H. D. ELLIS, "Intoduction to aspects of face processing: Ten questions in need to answer," In aspects of face processing, pp. 3-13, 1986.
[7] G. I. B. M. and T. M. , "Can face recognition really be dissociated from object recognition," J. Cogn. Neurosci, pp. 11, 349-370, 1999.
[8] I. GAUTHIER and L. N. K. , "Is face recognition so unique after all?," J.
Cogn. Neurosci, pp. 17, 125-142, 2000.
[9] M. M. S. AG, "The ML Programming Guide," 2013. [Online]. Available:
http://www.mevislab.de/docs/2.5/MeVisLab/Resources/Documentation/Publi sh/SDK/MLGuide/ch04s04.html. [Accessed 4 12 2014].
[10] J. Y. H. Y. and W. K. , "An Efficient LDA Algorithm for Face Recognition," anthropomatik.kit, p. 6, 2000.
[11] M. G. and R. F. , "Fusion of LDA and PCA for Face Recognition," in 8th Congress of Italian Association for Artificial Intelligence - See more at:
http://pralab.diee.unica.it/en/node/837#sthash.bflpvtBz.dpuf, Siena (Italy), 2002.
[12] C. S. P. V. R. D. ,. P. G. M. R. D. H. and L. P. R. , "A comparative performance evaluation using pcs for face recognition," THE ICFAI University Journal of Science & Technology, vol. 4, p. 4, 2008.
72
[13] C. G. and . X. , Pattern Recognition:Face recognitionbasedonthemulti- scalelocalimagestructures, Singapore: ScienceDirect, 2011.
[14] M. PIETIKAINEN, "Image Analysis with Local Binary Patterns,"
Springer- Verlag Berlin Heidelberg, pp. 115-118, 2005.
[15] T. M¨aenp¨a¨a and &. M. P. , "TEXTURE ANALYSIS WITH LOCAL BINARY PATTERNS," 13 May 2004. [Online]. Available:
http://www.ee.oulu.fi/research/image/texture/. [Accessed 20 October 2014].
[16] G. Z. and M. P. , "Local Binary Battern Descriptors fo Dynamic Texture Recognition," in The 18th International Conference on Pattern Recognition, 2006.
[17] S. A. Pardeshi, "Face Recognition Using Local Invariant Features,"
Journal of Computational Intelligence in Bioinformatics, vol. 1 Number 1, p.
73–81p, 2008.
[18] M. C. K. V. V. and S. K. V. , "A New Method of Face Recognition Based on Texture Feature Extraction on Individual Components of Face,"
International Journal of Signal & Image Processing, vol. 1, no. 2, p. 69, 2010.
[19] Y. L. F. L. and G. Q. , "Research of Face Recognition Methods Based on Bindingh Feature Extraction," Applied Mechanics and Materials Vols, Vols.
568-570, pp. 668-671p, 2014.
[20] H. K. b. and . E. E. A. A. , "Face Recognition Using Local Graph Strucrure (LGS)," Journal of Measurement Science and Instrumentation, pp. 169-175, 2011.
[21] wikipedia, "Support vector machine," wikipedia.org, 1 2 2015. [Online].
Available: http://en.wikipedia.org/wiki/Support_vector_machine. [Accessed 15 2 2015].
[22] B. E. Boser, I. M. G. and V. N. V. , "A Training Algorithm for," in 5th Annual ACM Workshop on COLT, Pittsburgh, 1992.
[23] C. CORTES and V. V. , "Support-Vector Networks," Kluwer Academic Publishers, p. 20, 1995.
[24] C. Corinna and V. V. , "Support-Vector Networks," Machine Learning, pp. 273-297, 25pages, Septamber 1995.
[25] F. Girosi, "An Equivalence Between Sparse Approximation and Support Vector Machine," Neural Computation, vol. 10, p. 26pages, 1998.
[26] G. B. and F. , "Generalized Discriminant Analysis Using a Kernel Approach," Neural Computation, vol. 12, p. 20 pages, 2000.
[27] C. F. C. R. and W. G. , "Research of Web-page Multi-Class Classification System Based on Support Vector Machine," Applied Mechanics and Materials, Vols. 713-715, pp. 2312-2316p, 2015.
[28] V. Jakkula, "Tutorial on Support Vector Machine (SVM)".
Aaisha S. R. Al-Shibli, Master in Computer Science, Lecture in Faculty of Computing and Information Technology in Sohar University,Sohar, Sultanate of Oman, Email [email protected]
Eimad Abusham, Ph.D Information Technology, Multimedia University, Malaysia. M.Sc Computer Science, University Putra Malaysia, Malaysia. B.Sc Computer Science, Sudan University Of Science And Technolgy, currently working as assistant professor in Faculty of Computing and Information Technology at Sohar University, Sohar, Sultanate of Oman, [email protected]
