Farhad Navabifar(1), Mehran Emadi (2)Rubiyah Yusof(3),Marzuki Khalid(4)

(1),(2) Faculty of Computer Engineering,Islamic Azad University,Mobarake branch,Esfahan,Iran (1),(2),(3),(4) Center of Artificial Intelligence and Robotics

University Technology Malaysia,International Campus,Jalan semarak,54100,Kuala Lumpur,Malaysia

(f_navabifar@yahoo.com)

Abstract— Detecting faces in images is the first step of any face application such as face recognition ,face localization and face expression. The performance of face detection systems directly affect on correct operating of mentioned applications. Because faces are non-rigid and have high variation in scale ,color ,pose and lighting condition ,designing an automatic system to overcome all mentioned problems is difficult. Machine learning has been shown that is one of the most successful tools to build high performance face detection systems. Due to huge number of studies in this area ,studying all works in this field is tedious. The purpose of this short review paper is to present ,categorize and evaluate some new face detection techniques using four conventional learning machine. The performance and the other evaluation parameters of these methods compare with each other in order to introduce significant techniques and also to state advantages and disadvantages of related works.

Keywords—Face detection, Machine Learning, Feature Extraction

1 Introduction

Automatic human face detection is one of the interesting categories in biometric knowledge which has attracted much attention recently. The goal of this task is to identify faces in a still image or video regardless of variations such as pose, illumination, occlusion and expression. Many applications have adopted face detection. For instance human computer interaction(HCI),computer surveillance ,biometrics , web search engines and digital video indexing are some of the applications which use face detection . The main challenges using face detection are pose ,illumination ,occlusion and facial expression .Another problem is high cost computation time which have faced face detection applications with a lot of problems in real

time purposes. Several different approaches have been presented by the researchers including Template-Base methods ,Feature-Based methods, Appearance-Based methods and Components-Based methods. Template-Based methods use the correlation values between standard pattern and input image to detect face.[1] , [ 2],[ 3] used this approach for locating human faces. Although Template-Based techniques are simple to implement but variation in pose ,scale and shape of images decreased detection rate. The goal of Feature-Based methods is to extract invariant features of images. The relationship between components of faces such as eye ,nose, hair and mouth based on a statistical model determined existing a face in an image.

The major drawback of this approach is if the images influenced by occlusion or noise, detecting faces will be very difficult. The two other approaches , Appearance-Based and Component-Based are more robust against changes in illumination, pose and partial occlusion and have been more successful compared with previous approaches. Both of the mentioned approaches applies learning machine to train examples .Two interesting survey works have been presented in [4 ] and [5 ] until 2002.In these surveys a detailed review of techniques to detect faces in a single image is presented. Also benchmarking database and evaluation criteria are discussed.

In this paper several new face detection techniques are categorized based on four conventional machine learning as a classifier to classify faces and non-faces. The methods are compared with each other to evaluate different techniques. This paper is presented in five sections. Section 2 gives a detailed review of face detection concepts and challenges. Four known learning machine algorithm discussed in section 3. Then in section 4 performance and evaluation criteria are presented. This paper concluded in section 5 for several promising directions for face detection.

A short review paper on Face detection using Machine learning

Figure 1.General Face Detection diagram 2 Face Detection

Face detection is one of the visual task which humans can do effortlessly but in computer vision this task is very difficult. Given a single image , detect and localize the number of faces regardless of pose ,illumination and expression . A simple face detection block diagram is shown in fig. 1. In image preprocessing unit, data prepare for next module. Image normalization and illumination adaptation are some of the processes which is done on data in this module. Feature extraction module, is performed to provide effective information that is useful to distinguish faces and non-faces and stable with respect to the geometrical and photometrical variations. Finally classification module is considered for classify face and non-face images based on the extracted features. In term of high accuracy and low false alarm , appearance based methods which use learning machine algorithms as a classifier are very successful for face detection applications.

In next section four types of conventional learning machines as a classifier which are used in face detection will be presented and will be compared with each other. 3 Machine Learning

Machine learning is a set of algorithms which used a set of data to learn a model. Using learner ,interesting properties capture when probability distribution is unknown. The relationship between observed variables illustrated by data . A major focus of machine learning research is to automatically learn to recognize complex patterns and make intelligent decisions based on data; the difficulty lies in the fact that the set of all possible behaviors given all possible inputs is too large to be covered by the set of observed training data.Four conventional machine learning applied in face detection are categorized and presented in this paper.

3.1 Support Vector Machine(SVM) Support Vector Machine was introduced by Vapnik [6]

as a learning machine .Osuna et al. were presented support vector machine for the first time for face detection [7]. SVM

classifies patterns based on several supervised learning techniques. In other word SVM classifies training examples and belongs to related categories and builds a model to recognize whether a new example belongs to one or other categories. One distinctive advantage this type of classifier is that most of the techniques for classifying examples such as Bayesian, neural network and RBF minimize the training error such as multiple layer perceptrons (MLPs) can produce low error rate on training data, there is no guarantee that this will translate into good performance on test data such as MLPs which minimize the mean squared error over the training data (empirical risk minimization), while SVMs use an additional principal called structural risk minimization .The purpose of structural risk minimization is to give an upper bound on the expected generalization error. As can be seen in fig. 2 suppose two set of examples belong to face and non-face. Line H3 doesn't separate the examples, line H2 can separate examples but with small margin and H3 separate with maximum margin while SVM minimizes an upper bound on expected generalization error. Osuna et al. used two test sets including 10 million patterns of 19˟ 19 pixels for face detection and showed their system can apply about 30 times faster than system was introduced by Sung and Poggio [8].

Also SVM has been applied for detecting faces in wavelet domain in [9],[10],[11].

Figure 2. Classification patterns based on support vectors

Data Acquisition

Image preprocessing

Feature extraction

Classification Face

Non Face

Figure.3 Five states is trained with the sequence of observation vectors[42]

Yungmin Li et al. presented combinational method

using Eigenface and SVM in order to increase speed of detection and pose estimation in multi-view images [12].

C.shavers et al. presented an SVM-based approach for face detection .They used ORL database consist of 200 images. A 20×20 window of pixels(400 dimension vectors) is extracted from each image to create a sample.The hyperplane decision is based on a degree-one polynomial.The decision function is constructed equidistance between support vector archetypes to give an optimal hyperplane decision function. Using one degree polynomial as a kernel function ,the SVM algorithm maps images into higher dimension transform space to construct decision function.During SVM training, input vectors from the training set are input to the multiplicative update algorithm .Only the samples which are support vectors become the archetypes for facial images. Experimental results showed that ninety five percent of the test images were correctly identified [13].

Bernd Heisele et al. proposed a component-based approach for robust face detection against pose and illumination. Two level SVM classifier was considered for classification. The former was responsible for finding face components and the latter generated geometrical configuration of face. Textured 3-D head model[14] used to generate training data. A large number of faces in different poses and various illumination automatically generated using 3-D head models. Experimental results on CMU PIE showed significant improvement comparing global face detection systems across pose and illumination[15].An appearance-Based face detection based on spectral histogram and SVM used by C.A.Waring and X. Liu. For representing spectral histogram, local features of an image are captured using filters such as Gabor filter ,Laplacian of Gaussian filter or Gradient filter based on local structure. A good representation should discriminate face class from non face class.After computing spectral histogram ,it feds to trained SVM for classification using adaptive thresholding . The proposed algorithm used the two standard data set which is applied by Rowly et al. and sung et al.[8] ,[16].They claimed that their results are comparable with

previous works in this area with the same database[17].

3.2 Hidden Markov Model(HMM)

HMM is a statistical model which can be considered as a learning machine to classify examples from unobserved states. The goal of HMM as a learning machine is to find, given an output sequence the best set of state transition and output probabilities. Based on a set of output sequences , HMM estimates the maximum likelihood of the parameters. Using Baum-Welch algorithm or Baldi-chauvin algorithm can be applied for estimating maximum likelihood[21]. The Baum-Welch algorithm is a special case of the Expectation-Maximization algorithm and an example of a forward-backward algorithm. After training HMM ,output probability of an observation determines the class to which is belongs. As an example from fig. 3 it is clearly seen that for training HMM, samples should be converted to observation vector. Observation sequence is produced by vertical scanning with p pixel of overlap. So HMM is trained with five states with sequence of observation.

Two probabilistic methods for face detection were presented by Rajagopalan et. al[18].The first method estimated density using higher order statistics while HMM applied to learn face to non-face and non-face to face transitions in an images in second method. An efficient method for extracting the observation vectors for face detection using the Karhunen-Loeve Transform is presented by A.V.Nefian et al. under variation of lighting conditions ,the accuracy of this method is promising.[19]

Pham and Jo. proposed a boosted skin-color model using HMM to detect color images.They composed modified local binary pattern histogram with embedded hidden Markov model to hierarchical classifier to detect whether skin regions are faces or non-faces. Their proposed system tested on different database and world wide web under different variations of lighting ,occlusion, expression and rotation and they obtained more than 94% accuracy in detection rate [20].

3.3 Neural Network

Neural network is a mathematical model which is derived by the structure and/or functional aspects of biological neural networks including interconnected group of artificial neurons .It computes data using a connectionist approach. As can be seen in fig. 4 the input images are applied to input layers and output layer represented the image as a face or non-face respect to the value of nodes 0 or 1. During training task the network learnt from examples to update network weights. This process will continue until error rate close to zero.

Also there are different neural networks which are non-linear statistical data modeling tools. They are usually used to model complex relationships between inputs and outputs or to find patterns in data.Ther are a lot of researches on biometric which use neural networks.The advantages of using neural network is that ,training complex classification pattern is feasible but designing network need high tuning indexes such as number of nodes ,number of layers ,tuning threshold and learning iteration .Rowley et al. in [16],[22]and [23] present a multi layer network and several arbitration techniques which is able to classify face and non-face . Their system were faster than Sung and poggio's system and are preferable in term of detection rate than sung and poggio's system. The limitation of Rowley's system in [16] is that their system are able to detect only frontal faces but in [23] Rowley present a developed system to detect rotated faces and obtained 76.9% detection accuracy with a small number of false positive. Fraud and Bernier applied neural network to detect faces in images. They used a five layers neural network to perform a non linear component analysis .Their proposed system would able to detect frontal faces and faces with 60 degree in plane rotation[37].

Tsai C.C. et al. presented a face detection system using neural network and eigenface. They applied eigenface to find the candidates of the face regions.Then a neural network including 1 hidden layer with 9 nodes examined the candidates to accept or reject examples as a face or non-face. Finally a template base face model is used to confirm face region from output of neural network for that group of non-faces which could not identified by eigenface and neural network . Their training set consist of 100 frontal and profile face blocks providing by ORL,MIT-CMU and world wide web database. They also have selected 50 images randomly from the world wide web to test system.Their results show that in term of ORL training data set they achieve 100% detection rate and 0 false detection rate .Also 93.23% detection rate and 292 false detection obtained from Bio ID data set in testing condition[24].

C. Garcia and M. Delakis presented a novel image–based approach for face detection using conventional neural network inspired by LeCun et al.[25]. Convolution and subsampling modules used to extract simple map features without any preprocessing on input images.Variable face

pattern including ±20 degrees in image plane and ±60 degree out of plane images utilized for training task. The results have suggested that proposed system are very robustness against noise ,illumination and facial expression and also obtained high detection rate with low false alarm[26].

A face detection system in color images proposed by Seow et al.[27].A three layer neural network in RGB space designed to classify skin and non skin patterns. Although collecting skin samples was easy but non skin patterns collecting was quietly hard.Seow et al. applied the three layered neural network to overcome this problem.Their neural network not only extract the skin regions of the face but also interpolated the skin regions in 3D color cube[28].

Figure 4. Three layer neural network. Input layer is considered for input images ,output layer is considered for determining face and non face images and hidden layer adjusts weights of input layer to minimize error for training task based on desired output.

3.4 AdaBoost One of the interesting works on face detection

presented by Viola and Jones. In contrast to many face detection systems which focused on pixels intensities, they applied a new image representation called the "integral image" to compute selected features very rapidly. Different shape and size of rectangular features namely Haar like features used to capture features from images. Some of the rectangle features is shown in fig 5. A small set of Haar like features selected by AdaBoost algorithm to classify faces and non-faces in training task. Each weak classifier ,classify the limit number of examples but the number of examples which have not classified correctly by the detector ,transferred to next weak classifier. So A linear combination of weak classifiers constructed strong classifier in order to increase accuracy and reducing computation time.Fig.6 shows the structure of cascade classifier. The results showed that the approach of Viola and Jones detect 93.7% of examples with 422 false positive using MIT+CMU

database that is more accurate than Rowley et al. proposed system. On the other hand viola's system was 15 time faster than Rowley et al.'s system [29].

A fast multi view face detection was developed following their previous work of viola and Jones which utilized frontal faces for detection. The first reliable method for non-upright face detection system belongs to Rowley et al[30].They designed two neural network classifiers, the former was considered for pose detection and the latter was only a conventional face detector .A similar two stage face detector including decision tree classifier and face classifier was designed by Viola and Jones for multi-view face detection. The proposed system divide spaces to different classes based on various poses and then training applied for each poses. For non face windows any detector return false. The results of Viola and Jones in comparison with work of Rowley et al. shows similar accuracy and false positive but in term of speed Viola and Jones acts better [31].Li et al. presented a real time multi view face detection inspired by viola and Jones research. They used floatboost algorithm which was useful for steady training of AdaBoost algorithm. Also they applied coarse to fine and simple to complex architecture in order to increase detection rate in non-frontal faces [32].

A component-Based approach was developed by Goldman et al. to detect faces in the presence of partial occlusion. In the first step facial components extracted individually then face localization and topology verification applied in the next step .Left eye, right eye ,nose and mouth were detected by the component detector They used Haar like features and AdaBoost [33]as a classifier due to one of the fastest algorithm for training data. A specific Graph model are considered to detect relationship between face components in term of size and position..

To evaluate proposed system, AR and BioID databases are utilized and TPR(True Positive Rate) and FPR(False Positive Rate) are calculated. Their results are shown that component-Based approach achieved higher robustness especially in presence of partial occlusion and out of plane rotation in comparison with holistic approach face detection developed by Viola and Jones[34].

Figure5. The sum of the pixels captured in white sections subtracts from grey pixels captured sections.(a) and(B) two

rectangular features,(C) three rectangular feature and (D) four rectangular features.

Figure.6 schematic diagram of cascade classifier. The first classifier reject a large number of negative examples. The next classifier try to reject misclassified examples from the previous classifiers.

An statistical learning method presented by Peng and Quiang(2007) to extract features and constructing classifiers for multi-view face detection.They used a recursive nonparametric discriminant analysis (RNDA).The RNDA relaxes Gaussian assumption of Fisher discriminant analysis which handle more general complexity.Combining RNDA features with AdaBoost form a multi-view face detector.The time consumption to train examples is a drawback of this method because of large vector elements for an image.To solve this problem RNDA proposed to search feature in transformed features space which is lower dimensionality than original space.They analyze the multi-view face detection on CMU profile test set containing 208 images with 347 profile faces.The result showed that they obtained 90 % detection rate with over 800 false detection[38].

Ruan and Yin(2009) applied facial features and AdaBoost algorithm for multi-pose face detection.In their method using skin color information , region of interest segmented.After detecting eyes and mouth by geometric features the face candidates are verified by AdaBoost algorithm.Their experimental results show that detection accuracy improved and also their proposed method is robust to facial expression and occlusion[39].

Huang et al.(2005)proposed a new boosting algorithm namely Vector Boosting as a novel tree-structured multi-view face detector.Thier algorithm covers large RIP and ROP images.Using coarse-to-fine strategy ,the entire face space divided to smaller subspace. To train the nodes of tree structure ,Vector Boosting applied.Thier results demonstrate high speed and performance detection[40]. Li et al.(2004) proposed a FloatBoost algorithm to backtrack all the learned features and also learning a

boosted classifier for achieving minimum error rate [41]. Besides these algorithm a novel statistical model used to lead fewer weak classifiers than original AdaBoost which provides lower error rates in both training and testing task. Their multi-view face detection system achieved a speed of 5–8 frames per second with detection rate lower than frontal face detection. Besides the commonly used local features, they showed that global features can increase high accuracy for detection.

Recently A.W.Mohemmed et al. have been proposed a PSOAdaBoost algorithm ,a combination of an optimization method called Particle Swarm Optimization and AdaBoost algorithm to develop AdaBoost algorithm for increasing accuracy and speed of computation .AdaBoost algorithm uses exhaustive search to find the best features to construct weak classifier .Combination of the small set of weak classifier constructs strong classifier which is used to classify face and non face patterns. Time consumption of exhausting search from AdaBoost algorithm is a serious problem for real time object detection purposes.PSO is an optimization tools developed by Kennedy and Ebehart[35].The best features and threshold decision determine by PSO.Comparing with the other forms of

AdaBoost , PSOAdaBoost shows increment in computation speed and detection rate on a test set [36] .

TABLE I EXPERIMENTAL RESULTS OF FACE DETECTION METHODS USING DIFFERENT MACHINE LEARNING

Approach Type of machine learning Detection Rate False

Detection Test database

Database Description

Component-Based [34] AdaBoost 90.80% 4 Bio ID and AR

143 images under different lighting ,expreesion and occlusion

Appearance-Based [33] AdaBoost 93.90% 167 MIT+CMU

130 gray scale images with a total of 507 frontal faces and 23 images from MIT database

Appearance-Based [17] SVM 96.70% 67 MIT+CMU

Appearance-Based [19] HMM 90% N/A MIT

Two sets of high and low resolution gray scale images with multiple faces in complex background

Appearance-Based [36] PSO+AdaBoost 97.63% 2.25% N/A 2424 positive and 3480 negative image

Image –Based [26] Convolutional neural architecture 93.10% 167 CMU

130 gray scale images with a total of 507 frontal faces

Appearance-Based [16] Neural Network 90.10% 167 CMU

Appearance-Based [32] Floatboost 90.20% 31 CMU

4 Discussion and Evaluation In this paper we have presented a short review paper of

new face detection techniques. Unfortunately due to lack of standard tools to evaluate approaches and different databases used by researchers, exact comparing methods is not feasible. To discuses and to assess presented methods, some of evaluation parameters are considered for comparing techniques with each other.

Table 1 describes the reported performance of some different approach and different machine learning as a classifier for face detection systems .As can be seen from this table, various databases have been used. From table 1 it is clear that the performance of all methods over 90% and this means that face detection for frontal images have been solved largely but in term of partially occlusion or multi-view face detection a few of the works have been successful.

According to table 1 Comparing SVM and neural network shows that SVM obtain better performance in term of detection rate and false detection.To detail this comparison we can say that SVM achieves generalization performance but some kind of neural networks such as multiple layer perceptrons (MLPs) are able to minimize error rate on training examples, there is no guarantee that this will translate into good performance on test data.On the other hand SVMs use structural risk minimization while MLPs minimize the mean square error on training examples or use empirical risk minimization .In other word using structural risk minimization gives an upper bound on the expected generalization error.

To minimize expected classification error ,SVM separate hyperplane while estimating optimal hyperplane implicated to solve linearly constrained quadratic programming problem which is time and memory consumption.On the other hand neural network is a powerful tools to classify complex class conditional density of image patterns but designing network including number of nodes and layers and training rate is complicated.

From table 1 we can see that HMM obtain 90% detection rate on MIT database. Although HMMs as a classifier have been applied in a wide range of applications specially in 1D application such as voice recognition but due to have two major drawback including high sensitivity to image noise and large dimension of the observation vector that causes complexity computation and time consumption, used less in face detection and recognition.

AdaBoost is one of the fastest algorithm which took popularity for object detection systems and has been used in a wide range of image applications recently[31]. Due to utilizing new image representation called "integral image" to compute the values of pixels in a specific area , AdaBoost reduces time of computation and consequently speed detection will be increased.But exhaustive search for searching small set of efficient features and finding optimum threshold is still disadvantage of AdaBoost algorithm.

According to table 1 adding one optimization technique such as PSO can increase performance because PSO select the best features for classification and also decrease feature space which affect on speed detection. Due to using various type of computer to train and test patterns, it is impossible to evaluate methods base on speed and computation time but it seems proposed techniques based on boosting algorithm obtain high speed detection. 5 Conclusion

In this paper we have summarized new methods for face detection systems based on four conventional machine learning algorithm .We have found that although detecting multiple frontal faces in images have been applicable but still detection faces in complex background with arbitrary viewpoint and different expression and occlusion need more attention.Also some optimization method can be applied in face detection algorithm to find best features and reduce time of feature selection. Among existing methods for detecting faces in images, techniques using boosting algorithm are more effective for real time object detection.

References [1] V. Govindaraju, “Locating Human Faces in hotographs” Int’l J.Computer Vision, vol. 19, No. 2, pp. 129-146, 1996. [2] V. Govindaraju, D.B. Sher, R.K. Srihari, and S.N. Srihari, “Locating Human Faces in Newspaper hotographs” .Proc. IEEE Conf.Computer Vision and Pattern Recognition, pp. 549 -554, 1989. [3] V. Govindaraju, S.N. Srihari, and D.B. Sher, “A computational Model for Face Location”. Proc. Third IEEE Int’l Conf. Computer Vision, pp. 718-721, 1990. [4]M.H.Yang,D.J.Kriegman,N.Ahuja,"Detectiong Faces in Images:A Survey",IEEE trans. On Pattern Analysis and Machine Intelligence ,Vol.24,No.1,Januarey 2002 [5]E.Hjelmas and B.K.Low,"Face Detection:A Survey ". Computer Vision and Image Understanding 83,pp.236-274 ,2001 [6] V.N. Vapnik. "Estimation of Dependencies Based on Empirical Data".springer-Verlag, 1982 [7]E.Osuna,R.freund,and F.Girosi,"Training Support Vector Machines:An Application to Face Detection" ,proc.IEEEconf.computer vision and pattern recognition ,pp.130-136,1997. [8]K.-K.Sung and T.Poggio,"Example-Based Learning for View Based Human Face Detection". Technical Report AIMemo 1521,Massachosetts Ins.of technology AI Lab ,1994 [9] M. Oren, C. Papageorgiou, P. Sinha, E. Osuna, and T. Poggio,“Pedestrian Detection Using Wavelet Templates” Proc. IEEE Conf.Computer Vision and Pattern Recognition, pp. 193-199, 1997.

[10] C. Papageorgiou, M. Oren, and T. Poggio, “A General Framework for Object Detection,”. Proc. Sixth IEEE Int’l Conf. Computer Vision, pp. 555-562, 1998 [11] C. Papageorgiou and T. Poggio, “A Trainable System for Object Recognition,” Int’l J. Computer Vision, vol. 38, No. 1, pp. 15-33, 2000 [12 ]Y. Li, S.Gong, H. Liddell, "Support vector regression and classification based multi-view face detection and recognition", IEEE International Conference on Automatic Face and Gesture Recognition, pp. 300–305,2000 [13] C.shavers et al.”An SVM-based approach to face detection”.Proceeding of the Thirty-Eighth Southeastern Symposium on Digital Object Identifier , PP. 362 – 366,2006 [14]T.Vetter."synthesis of novel views from a single face", international journal of computer vision ,28(2),pp.103-116,1998 [15]B.Heisele et al.,"Component-Based face detection ",in proc. IEEE Int. Conf. Computer Vision and Pattern Recognition, Vol. 1,pp.657-662,2001 [16]H.Rowley,S.Baluja,and T.Kanade,"Neural Network-Based Face Detection '',proc.IEEE conf. computer vision and pattern recognition,PP.203-208,Jan.1998 [17]C.A. waring and X.Liu,"Face detection using Spectral Histograms and SVMs",IEEE Trans. On Systems,Man and Cybernetics-part B:Cybernetics vol.35No.3,June 2005 [18]A.Rajagopalan,K.Kumar,J.Karlekar,R.Manivasakan,M.Patil,U.Desai,P.Poonacha, and S.Chaudhuri,"finding Faces in Photographs",Proc.sixth IEEE Int'l Conf. Computer Vision ,pp.640-645,1998 [19]A.V.nefian and M.H.HIII,"Face Detection and Recognition Using Hidden Markove Models"proc.IEEE Int'l conf.Image Processing ,vol,pp.141-145,1998 [20]P.-T.Pham. and K.-H. Jo ," Color-based Face Detection using Combination of Modified Local Binary Patterns and embedded Hidden Markov Models",International Joint Conference",PP.5598-5603,2006 [21]L.R.Rabiner and B.H.Jung,"Fundamental of speech Recognition",Printice Hall ,1993 [22]H.Rowley, S .Baluja and T.Kanade ,"Neural Network-Based Face Detection '', IEEE Trans. pattern Analysis and Machine Intelligence,vol 20,No.1,pp.23-38,Jan.1998 [23] H.Rowley,S.Baluja,and T.Kanade,"Rotation Invariant Neural Network-Based face detection,proc.IEEE conf.computer vision and Pattern Recognition,pp.38-44,1998 [24]C.Tsai,et al.,"Face detection using eigenface and neural network", 2006 [25]Y.LeCun,L.Bottou,Y.Bengio,and P.Haffner, "Gradient-Based Learning Applied to document recognition", proc.IEEE ,vol.86,No.11,pp.2278-2324,1998 [26]C.Garcia , M.delakis ,"Convolutional Face Finder:A Neural Architecture for Fast and Robust face Detection",IEEE Trans. On Pattern Analysis and Machine Intelligence",vol.26, No 11, November 2004 [27]M.J.seow,D.Valaparla,V.K.Asari,"Neural Network-Based Skin Color Model for Face Detection",Proceeding of

the 32nd Workshop on Applied Imagery Pattern recognition,2003 [28]P.Kakumanu,S.Makrogiannis,N.Bourbakis ,"A Survey of Skin-Color Modeling and Detection Methods" ,Pattern Recognition 40 ,pp.1106-1122,2007 [29]P.Viola,M. Jones,"Robust Real Time Object Detection" ,IEEE Workshop on Statistical and Theories of Computer Vision,Vancouver,Canada ,July 2001 [30]H.Rowley,S.Baluja and T. Kanade,"Rotation Invariant Neural Network-based Face Detection".In Proc.IEEE Conf. on Computer Vision and Pattern Recognition,PP. 38-44, 1998 [31]P.Viola,M.Jones"Fast Multi-view Face Detection",July 2003 [32]S.Li,L.Zhu,Z.Zhang and A.Blake,H.Zhang, H.Shum, "Statistical Learning of Multi-View Face Detection" Proc.Seventh European Conf. Computer Vision,pp.67-81,2002 [33]P.Viola and M.Jones,"Rapid Object Detection using a Boosted cascade of Simple Features ",In Proc. IEEE Conf. Computer Vision and Pattern Recognition,2001 [34]L.Goldman et al.,"Component and their Topology for Robust Face Detection in the Presence of Partial Occlusion"IEEE Trans. Information Forensics and Security ,Vol.2,NO. 3,September 2007 [35]J.Kennedy and R.C.Eberhart,"Particle Swarm Optimization",in Neural Networks, ,pp.1942-1948, 1995 [36]A.W.Mohemmed ,M. Zhang and M. Johnston"Particle Swarm Optimization Based AdaBoost for Face Detection,"IEEE 2009 [37] R. Feraud, O.J. Bernier, J.-E. Villet, and M. Collobert, “A Fast and Accurate Face Detector Based on Neural Networks,” IEEE Trans. On Pattern Analysis and Machine Intelligence, vol. 22, no. 1, pp. 42-53,Jan. 2001. [38] P. Wang, Q. Ji, "Multi-view Face and Eye Detection using Discriminant Features ",Computer vision and Image Understanding,pp.99-111,2007 [39]J. Ruan and J.Yin," Multi-pose Face Detection using Facial Features and AdaBoost Algorithm ",Second international workshop on computer science and engineering,pp.31-34,2009 [40] C. Huang,H.Z. Ai,Y.Li,and S.H. Lao,"Vector Boosting for Rotation Invariant Multi-View Face Detection",proc.10th IEEE int'l conf.Computer Vision,2005 [41] S.Z. Li, Z. Zhang, "FloatBoost Learning and Statistical Face Detection", IEEE Transactions on Pattern Analysis and Machine Intelligence 26 (9) , pp. 1112–1123,2004 [42]F.S.Samaria,"Face Recognition using Hidden Markov Models",PhD thesis, University of Cambridge ,1994