Corresponding Author: Housam Khalifa Bahsier, Faculty of Computer Information Science & Technology, Multimedia
University, Melaka-75450, Malaysia
E-mail: [email protected]
632
Real Time Face tracker based on Local Graph Structure Threshold (LGS-TH)
1Housam Khalifa Bashier, 2Eimad Eldin Abusham, 1Mohd Fikri Azli Abdullah, 1Liew Tze Hui, 1Ibrahim Yusof and 1Lau Siong Hoe
1Faculty of Information Science and Technology, Multimedia University Jalan Ayer Keroh Lama,
75450 Bukit Beruang, Melaka, Malaysia.
2Faculty of Computing and Information Technology, Sohar University Sohar, Sultanate of Oman
Abstract: Face detection plays an important role in many applications such as face recognition, face image database, and video surveillance. In this paper a novel real time face tracking algorithm has been introduced with accurate results. The method is based on the idea of local graph structure-LGS. LGS-TH features are based on LGS where features are derived from a general definition of texture in a local graph neighborhood. The idea of LGS-TH comes from a dominating set for a graph of the image and then the threshold of the face image is calculated and applied with LGS-TH.
Key words: local graph structure, image processing, pattern recognition, local binary pattern, face detection.
INTRODUCTION
Detection of faces considered one of the natural visual tasks for human where human beings can detect face effortlessly. Face detection is essential step towards many sophisticated biometrics recognition, multimedia and computer vision applications. The challenging difficulty of face detection as it needs to account for all possible appearance variation caused by change in occlusions, illumination, facial features, etc. In addition, it has to detect faces that appear at different scale, pose and localize an unknown number of faces from given a still or video image. Regardless of all these difficulties, remarkable advancement has been made in the last decade and many systems have revealed inspiring real-time performance. The current advances of these algorithms have also made considerable contributions in detecting other objects such as cars and humans movement. The problems of face detection require knowing about the information impeded in the face itself, and the way this information is exploited differs and can be categorized into two main categories the first is features-based and image based. Features-based attempt to extract features of the image and match it against the information of the face features. image-based or view-based approach works directly with the image as a two dimensional array of intensity values and attempts to classify the input windows into two classes, a face or none face class, this category implicitly extract and learns the features from the training set without explicitly identifying the features and their relation to each other, this categorization also applies for general object classification and (M. Yang, 2004). In general, the classification of single image detection methods fall into four categories; some methods obviously have common characteristics are discussed.
1. Knowledge-based methods. These methods aim to find the general knowledge from the typical face image; it takes in the consideration the relationship of the general facial feature.
2. Feature invariant approaches. Feature invariant approaches consider the local structure of the face which is the potential feature this feature does not affect with illumination, pose and angle.
3. Template matching methods. The template matching method is used for face detection and localization (Chang-Woo Park and Mignon Park , 2006), in general template matching is very simple since it’s a two dimension cross-correlation for a given test face image with the template face to estimate the degree of similarity. Template matching algorithm is also used in face recognition as described in (Zhiwen Wang and Shaozi Li, 2011).
4. Appearance-based methods. The appearance –based methods are almost similar to template matching, the only difference is that the appearance –based method is used the training images to extract the features and then these features are learned by any machine learning such as Adaboost-learning; hence these learned features are then used for the detection.
Face Detection Approaches:
633
1996),skin color (Mixture of Gaussian (J. Yang and A. Waibel, 1996)( S. McKenna et al., 1998)) and multiple
features (integrated of skin color size and shape (R. Kjeldsen and J. Kender, 1996)).Template matching is one of the popular approaches in face detection, in this approach there are two ways the first one is predefined face template as it uses shape template (Craw et al., 1992) secondly there’s Deformable templates as there was work
which is Active shape Model (ASM)( A.Lanitis et al., 1995).There’s also appearance-based method where there
were a lot of work been done for instance in Eigneface there was Eigenvector decomposition and clustering (M. Turk and A. Pentland, 1991), distribution-based Gaussian distribution and multilayer perception(K. Sung and T. Poggio, 2002) , machine learning approaches are also been used in Ensemble of neural networks and arbitration schemes(H. Rowley et al., 2002),SVM with polynomial kernel(E. Osuna et al, 2002), joint statistics of local
appearance and position (H. Schneiderman and T. Kanade, 1998), Higher order with Hidden Markov Model (HMM)( A.Rajagopalan et al., 2002) and Kullback relative information(M. Lew , 2002)( A. Colmenarez and T.
Huang , 1997)which it uses information theoretical approach. Some other detection methods had indeed been shown in several papers (S.K. Pakazad et al., 2011) (Chang-Woo Park and Mignon Park, 2006).
Appearance-Based Methods:
As it has been stated above, there are a series of problems that we encounter such as pose, angle, illumination and etc so these unpredictable backgrounds give an attention to face detection in facial features.
Hence the need to handle the problem of detecting faces in several images considering the mess background. From this point of view there’s a new research area that need to be discussed which is treating the face detection as a pattern recognition problem due to the fact that face detection need to solve the problem of illumination, pose, angle, detection of several faces in mess backgrounds and etc. Therefore there’s a need to get the pattern such as the local feature or etc. This eradicates the possible of modeling error due to partial or incorrect face knowledge. The basic approach in processing face detection is via training images where there’s a 2D array which used to store two type of classes, the first class is a face class which contains a series of face features that is used to detect whether the detected region is a face or not and the second class is contains the features of non-face which is stored in this 2D array as well, then the correlation coefficient or Euclidean distance is used to calculate the distance of the input image to classify whether it falls in the face class or in the non-face class.
Image based approach is used where there’s a need for Appearance-based method such as Eigenface and etc, this approach is based on template matching ( Z. Liu and Y. Wang, 2000), due to the fact that this algorithm uses a wind scanning method since it requires the detection of face region inside the input image; but there are variation of implanting this algorithm in almost all based systems. In the following sections the image-based approaches have been divided into three classes, linear subspace methods, neural networks, and statistical approaches. In each division a brief description is highlighted and followed by our proposed method.
Linear Subspace Methods:
Human faces images lie in a subspace of the overall image space. To represent the subspace of images, several approaches can be used. Neural approaches such as artificial neural networks (ANN), but there are also quite a lot of techniques more very much related to standard multivariate statistical analysis which can be applied. In this part we explain some of these techniques, including linear discriminant analysis (LDA), principal component analysis (PCA), and factor analysis (FA). In the late 1980s, Sirovich and Kirby (L. Sirovich and M. Kirby, 1987) developed a technique using principal component analysis to efficiently of faces, represented in terms of eigenvectors (covariance matrix of the distribution). Each subject face in the face set can then be approximated by a linear combination of the largest eigenvectors, which commonly referred to as eigenfaces, using appropriate weights. Turk and Pentland (M. Turk and A. Pentland, 1991) later developed this technique for face recognition. Their method based on nature of the weights of eignefaces exploits the distinct nature of the weights of eigenfaces in individual face representation. Since the face reconstruction by its principal components is an approximation, a residual error is defined in the algorithm as a preliminary measure of “faceness.” This residual error which they termed “distance-from-face-space” (DFFS) gives a good indication of face existence through the observation of global minima in the distance map. Represent human faces. By given a set of different face images, the first step in the method is to find the principal components of the distribution. Principle component analysis is used with the directional filter bank as shown here (Mohammad A.U. Khan et al., 2004)
Local Graph Structure-LGS:
This technique is proposed by Eimad,Housam & Wong in their paper (Eimad Eldin et al, 2010) which is a
634 Local Graph Structure Threshold:
The idea of Local Graph Structure-TH (LGS-TH) is based on LGS which it comes from a dominating set for a graph G = (V, E) is a subset D of V such that every vertex not in is joined to at least one member of D by some edge. The domination number γ (G) is the number of vertices in a smallest dominating set for G.
LGS-TH works with the six neighbors of a pixel, by choosing the target pixel C as a threshold, then we start by moving anti clockwise at the left region of the target pixel C,If a neighbor pixel has a higher gray value than the target pixel (or the same gray value) then assign a value equal to threshold level of the face image plus the large or equal vertex modulus 256, else we keep the last value of the gray pixel. After finish on the left region of graph we stop at the target pixel C and then we move in a horizontal way (clockwise) to the right region of the graph and we apply the same process till we get back to the target pixel C(Eimad Eldin and Housam, 2011).
Direction
Value
Fig. 1: Local Graph Structure-TH (a. Direction, b. Value).
To produce the LGS for pixel (xd, yd) a binomial weight 2p is assigned to each sign s(gd – gn). These binomial weights are summed:
(1)
Where
The difference between LGS and the proposed method is that LGS it takes in consideration the binary values in the vertexes that forming the dominating graph while LGS-Threshold it takes in consideration the threshold of the face image and the Gray value of the vertexes that forming the dominating graph.
LGS-TH Algorithm:
Start moving from the target pixel in anti-clockwise direction to the first neighbour.
If the move from large or equal vertex to small vertex we assign a value = level Gray + Large vertex%256
Loop until get back to the target pixel.
Then move to the right side of the graph and compare the target pixel with its neighbor in the right side.
Assign the target pixel to the next pixel on the right hand side of the target pixel.
Move a clock wise and apply the same as in step 2.
635 Experiments and Results:
Local graph structure thresholds (LGS-TH) have proved to be useful in a variety of image processing and pattern recognition tasks. The basic idea is illustrated in Fig. 1
In the initial work of face processing using LGS-Threshold, can be seen in Fig 3, is an example of new generated images from original images and figure 4 shows an example of the histogram.
Fig. 2: Example of an original images. Fig. 3: Example of new generated images.
Fig. 4: Example of the histogram.
We test our method using the public ORL face database. The database consists of 400 faces from 40 persons. The faces were captured with the subjects in a straight, frontal position against a dark identical background, and with acceptance for some sloping and regular change of up to 20 degrees. Image variations of five individuals in the database are illustrated in Fig 5.
Fig .5: ORL Database.
To assess the performance of the three proposed LGS-TH method on texture classification where face recognition is taken as an example 5 subjects have taken for our experiments, 8 images for training and the remaining 2 images for testing, first both methods LGS-TH are applied to find the histograms for the entire training set, for testing, the histogram of test image is calculated and then find the between the training images histograms and histogram of test image by applying the correlation functions which it compares two histograms. The correlation is calculated based on the below formula.
636
The recognition rates obtained with LGS-TH are shown in Table 1. LGS-TH is robust with respect to variations of illumination. We believe that the main explanation for the better performance of the local graph structure over other texture descriptors, it takes in consideration the relationship between the pixels that form the local graph of the target pixel C, threshold of the face image, the value of the gray vertex and consideration tolerance to monotonic gray-scale changes and the threshold.
Table I: Recognition Rate.
Method LGS-TH
Recognition Rate 88.75%
The similarity results for the proposed method are in shown Table 2. Table 2: Similarity Rate.
Subjects Testing Image Similarity
(with Training) LGS-Th
Subject 1 1 99.98%
2 99.99%
Subject 2 1 99.98%
2 99.99%
Subject 3 1 99.97%
2 99.98%
Subject 4 1 99.99%
2 99.99%
Subject 5 1 99.99%
2 99.99%
As shown in table 1 and 2 the proposed method is quite acceptable for texture classification where we can carry this prove for the next step which is the face detection at the real time since the first step of face detection has been shown where we use the texture classification then on below figures we see the real time face detection images;
Fig. 6: Detected faces. Conclusions:
637
ACKNOWLEDGMENT
The authors thank the Olivetti Research Laboratory (United Kingdom) for using their Olivetti database (formally known ORL database).
REFERENCES
CBCL Face Database #1, MIT Center for Biological and Computation Learning, http://www.ai.mit.edu/projects/cbcl
Chang-Woo Park and Mignon Park, 2006. Fast Template-based Face Detection Algorithm Using Quad-tree Template. Journal of Applied Sciences, 6: 795-799.URL: http://scialert.net/abstract/?doi=jas.2006.795.799.
Colmenarez, A. and T. Huang, 1997. "Face detection with information-based maximum discrimination," pp: 782.
Colmenarez, A. and T. Huang, 2002. "Maximum likelihood face detection," pp: 307-311. Craw, et al., 1992. "Finding face features," pp: 92-96.
Dai, Y. and Y. Nakano, 1996. "Face-texture model based on SGLD and its application in face detection in a color scene," Pattern recognition, 29: 1007-1017.
Eimad Eldin Abdu Abusham and Housam Khalifa Bashier, 2011. “Image-Preprocessing using Local Graph Structure-Threshold (LGS-TH)” International Conference on Information and Communication Technology– Muscat-Oman, 22-23 March 2011.
Eimad Eldin Abdu Abusham, Housam Khalifa Bashier, 2011. “Face Recognition Using Local Graph Structure”. Human-Computer Interaction. Interaction Techniques and Environments Lecture Notes in Computer Science, Volume 6762/2011, 169-175, DOI: 10.1007/978-3-642-21605-3_19.
Eimad Eldin Abdu Ali Abusham and Wong Eng Kiong, 2009. Non-Linear Principal Component Embedding for Face Recognition. Journal of Applied Sciences, 9: 2625-2629 URL: http://scialert.net/abstract/?doi=jas.2009.2625.2629.
Guoqiang Wang and Weijuan Zhang, 2011. Neighborhood Preserving Fisher Discriminant Analysis. Information Technology Journal, 10: 2464-2469,DOI: 10.3923/itj.2011.2464.2469; URL: http://scialert.net/abstract/? doi=itj.2011.2464.2469.
Hjelmås, E. and B. Low, 2001. "Face detection: A survey," Computer vision and image understanding, 83: 236-274.
Kirby, M. and L. Sirovich, 2002. "Application of the Karhunen-Loeve procedure for the characterization of human faces," Pattern Analysis and Machine Intelligence, IEEE Transactions on, 12: 103-108.
Kjeldsen, R. and J. Kender, 1996. "Finding skin in color images," fg, pp: 312.
Krishnamoorthy, R. and R. Bhavani, 2007. Face Recognition System Based on Orthogonal Polynomials. Journal of Applied Sciences, 7: 109-114,DOI: 10.3923/jas.2007.109.114; URL: http://scialert.net/ abstract/?doi=jas.2007.109.114.
Lanitis, A., et al., 1995. "Automatic face identification system using flexible appearance models," Image
and Vision Computing, 13: 393-401.
Lew, M., 2002. "Information theoretic view-based and modular face detection," pp: 198-203.
Lin, et al., 2002. "Face recognition/detection by probabilistic decision-based neural network," Neural
Networks, IEEE Transactions on, 8: 114-132.
Lowe, D., et al., "Combined Bottom-Up And Top-Down Search Using Compound Templates."
McKenna, S., et al., 1998. "Modelling facial colour and identity with gaussian mixtures," Pattern recognition, 31: 1883-1892.
Mohammad, A.U. Khan, Muhammad Khalid Khan, Muhammad Aurangzeb Khan, Muhammad Talal Ibrahim, Muhammad Kamran Ahmed and Jahanzeb Afzal Baig, 2004. Principal Component Analysis of Directional Images for Face Recognition. Information Technology Journal, 3: 290-295.
Osuna, E., et al., 2002. "Training support vector machines: an application to face detection," pp: 130-136.
Pakazad, S.K., F. Hajati and S.D. Farahani, 2011. A Face Detection Framework based on Selected Face Components. Journal of Applied Sciencess 11: 247256,DOI: 10.3923/jas.2011.247.256; URL:http://scialert.net/abstract/?doi=jas.2011.247.256.
Rajagopalan, A., et al., 2002. "Finding faces in photographs," pp: 640-645.
Rowley, H., et al., 2002. "Neural network-based face detection," Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 20: 23-38.
Samal, A. and P. Iyengar, 1992. "Automatic recognition and analysis of human faces and facial expressions: A survey," Pattern recognition, 25: 65-77.
Schneiderman, H. and T. Kanade, 1998. "Probabilistic modeling of local appearance and spatial relationships for object recognition," pp: 45-51.
638
Sung, K. and T. Poggio, 2002. "Example-based learning for view-based human face detection," Pattern Analysis and Machine Intelligence, IEEE Transactions on, 20: 39-51.
Turk, M. and A. Pentland, 1991. "Eigenfaces for recognition," Journal of cognitive neuroscience, 3: 71-86. Turk, M. and A. Pentland, 2002. "Face recognition using eigenfaces," pp: 586-591.
Wenhui Li, Yifeng Lin, Huiying Li, Ying Wang and Wenting Wu, 2011. Multi-channel Gabor Face Recognition Based on Area Selection. Information Technology Journal, 10: 2126-2132, DOI: 10.3923/itj.2011.2126.2132; URL: http://scialert.net/abstract/?doi=itj.2011.2126.2132.
Yang, G. and T. Huang, 1994. "Human face detection in a complex background," Pattern recognition, 27: 53-63.
Yang, J. and A. Waibel, 1996. "A real-time face tracker," pp: 142-147.
Yang, M., 2004. "Recent advances in face detection," IEEE ICPR 2004 Tutorial.
Yow, K. and R. Cipolla, 1997. "Feature-based human face detection," Image and Vision Computing, 15: 713-735.
Zhiwen Wang and Shaozi Li, 2011. Face Recognition using Skin Color Segmentation and Template Matching Algorithms.Information Technology Journal, 10: 2308-2314, DOI:10.3923/itj.2011.2308.2314,URL:http://scialert.net/abstract/?doi=itj.2011.2308.2314.
