Boosted of Haar-like Features and Local Binary Pattern Based Face Detection Toan Thanh Do, Khiem Ngoc Doan

Thai Hoang Le, Bac Hoai Le

Department of Computer Science Ho Chi Minh University of Science Ho Chi Minh City, Vietnam [email protected]

Department of Computer Science Ho Chi Minh University of Science Ho Chi Minh City, Vietnam {lhthai, lhbac}@fit.hcmuns.edu.vn

Abstract— Effective and real time face detection has been made possible by using the method of rectangle Haar-like features with AdaBoost learning and cascade of the strong classifiers since Viola and Jones’ work [1]. After that, Rainer Lienhart had improved Viola and Jones’ work by extending set of Haar-like features [3]. However, it still has drawbacks; the detection results often have high false positives. In [2], A. Hadid et al. have used Local Binary Pattern (LBP) method for face description and they applied effectively in face detection problem. However, it is slow. Therefore, it is difficult to apply in real time applications. In this work, we proposed an approach to combine a boosted of Haarlike Features and LBP to achieve a good trade-off between two extreme. The system, which is built from proposed model, is conducted on MIT + CMU test set [7]. Experimental results show that our method performs favorably compared to state of the art methods. Keywords– Face recognition; AdaBoost; Haar-like feature; Local Binary Pattern.

I.

INTRODUCTION

Face detection has a wide range of applications such as automatic face recognition, human - machine interaction, surveillance, etc. Various techniques have been proposed in the literature for this purpose. In general, according to Yang [11], there are four groups of face detecting methods. Knowledge based methods: these methods are based on sets of rules which have been built from experts on standard face structures. These rules are based on relationships between face features. The methods are mostly used to locate positions of faces. Typical researchers are G.Yang (1994) and Kotropoulos (1997) [5]. Invariant feature based methods: these methods focus on finding invariant features which always exist in every condition: facial pose, lighting and expression. Then these features are used only to locate positions of faces. Works which belong in these approaches are K.C.Yow and R.Cipolla (1997), T.K.Leung (1995) [5]. Template matching based methods: In the approaches, to describe faces or individual face features, face templates would be stored. Detecting faces is based on the correlation between input images and the stored templates. These methods are used both to locate and detect faces. Some typical researchers are Craw (1992) and A.Lanitis (1995) [5].

Machine learning based methods: in contrast to template matching based methods, models of the methods will learn from training sets of image. After that these models will be used to detect faces. These approaches are used only to detect faces. There are some machine learning models based on these methods such as Eigenface: M.Turk and A.Pentland (1991) [8]. Probability Distribution Based Model: K.Sung and T.Poggio (1998) [5]. Artificial Neural Network (ANN): H.Rowley (1998) [7]. SVM: E.Osuna (1997) [5]. Bayes Classification: H. Schneiderman and T. Kanade (1998) [4]. Hidden Markov Model: A.Rajagopalan [5]. Reinforcement Learning Model: AdaBoost by Viola and Jones; FloatBoost by Stan Z.Li and Zhen Qiu Zhang (2004) [6]. In same 2004, A. Hadid et al [2] had used LBP method for face description and they applied it successful in face detection problem. In this paper, we only focus on machine learning methods because they eliminate subjective thinking factors from human experience. Moreover, they only depend on training data to make final decisions. Thus, if training data is well organized and adequate, then these systems will achieve high performance without human factors. One of the most popular and efficient learning machine based approaches for detecting faces is AdaBoost approach [1]. Viola and Jones designed a fast, robust face detection system where AdaBoost learning is used to build nonlinear classifiers. AdaBoost is used to solve the following three fundamental problems: (1) learning effective features from a large feature set; (2) constructing weak classifiers, each of which is based on one of the selected features; and (3) boosting the weak classifiers to construct a strong classifier. Viola and Jones make use of several techniques for effective computation of a large number of such features under varying scale and location which is important for real time performance. Moreover, the cascade of strong classifiers which form cascade tree will make the computation even more efficient. Their system is the first real time frontal-view face detector. However, their system still has some drawbacks. Since the detection results depend on weak classifiers, the detection results often have many false positives. To decrease the rate of false positives, it is compelled to increase the number of strong classifiers and Haar-like features in cascade tree, but this will cause a significant increase in the performance time, and detection rate can be decreased. Thus to deal with the issue we should combine AdaBoost with other machine learning techniques to achieve

978-1-4244-4568-4/09/$25.00 ©2009 IEEE

1 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.

the same face detecting ratios but with minimum the number of false positives and the running time. In recent years, A. Hadid et al. had used Local Binary Pattern (LBP) for face description and applied it in face detection problem. Their results had verified the discriminative power of LBP, and LBP for face description is very effective to apply in face detection problem. However, its speed is rather slow. In our experiment, to process an image 20 × 20, it only takes 0.00015 seconds but it takes about 40 seconds to process an image which has size of 276 × 343. The window which used to scan has size of 20 × 20. The scale and step size to slide window are 1.2 and 2 pixels, respectively. The artificial neural network (ANN) is used to classify windows that face or non-face. To overcome disadvantages of AdaBoost and LBP method, in this paper, we present an approach of combining AdaBoost and LBP (called AB-LBP) to detect faces with the purpose of minimizing the number of false positives but still achieving the desired faces detecting rate and can apply in real time face detection. The AB-LBP method includes two stages. In the first stage, we use AdaBoost with a suitable number of strong classifiers and Haar – like features to detect face. The goal of this stage is to get high detection rate by quickly rejecting easy background, but accepting a number of false positives relating to difficult background. Thus, the performance speed of the detector is fast and the detecting rate is guaranteed. However, there are still many false positive faces. The results were detected in first stage called candidate regions. In the next stage, we use LBP method for face description and use ANN in classification stage to filter out false positives in the candidates regions of first stage to achieve better results. Since, in second stage, we only focus in candidate regions which resized at size of 20 × 20 instead of the whole input image. Thus, it takes less time. This allows us to apply it in real time face detection. The rest of this paper is organized as follows. Section II will describe the way to apply AdaBoost and LBP for face detection. Section III will present the combination model of two previous methods to efficiently solve the face detecting problem and discuss experimental results of the proposed system. Conclusion and future works is mentioned in section IV. II.

y k ∈Y = {− 1,+1} is a corresponding label (+1

corresponding to face, -1 corresponding to non face). T weak classifiers hj: X → {-1, +1}. We use AdaBoost algorithm to build a strong classifier. Output: The final strong classifier is

⎞ ⎛ T H ( x) = sign⎜⎜ ∑ α j h j ( x) ⎟⎟ ⎠ ⎝ j =1

(1)

Weak classifiers hj(x) are simple Haar-like features. A large set of very simple “weak” classifiers will be used to classify the image region as face or non face. Viola and Jones proposed 4 basic templates of scalar features for face detection. After, Rainer Lienhart extended to a set of 14 templates [3]. In [1], Viola and Jones had proposed a new image representation called the “Integral Image” which allows the Haar-like features are computed very quickly. The Haar-like features is very simple and can be computed very effective. However, to achieve good results, an AdaBoost based system need a huge number of features. For example, for sub window of size 20 × 20, there can be tens of thousands of such features for varying shapes, sizes and locations. Since then, this significantly decreases the performance speed of the detector. Moreover, final classifier correctness depends on the correctness of weak classifiers (Haar-like features). Thus the performance effectiveness is not high and the detection results often have many false positives. b) Building Cascade of boosted Classifier Most good classifiers need much time to have classification results since they must consider the great number of features of patterns. The cascade structure of strong or boosted classifiers has been suggested in order to reduce performance time, false alarm rates for the classifier. Figure 1 shows structure of cascade of classifiers.

ADABOOST AND LBP FOR SOLVING THE PROBLEM

A. AdaBoost 1) Overview of AdaBoost AdaBoost is a Boosting algorithm. It is proved that a combination of weak classifiers will construct a strong classifier. AdaBoost is very efficient since it combines simple statistical learners while reducing significantly not only the training error but also the more vague generalization error.

Figure 1. Face detection cascade of classifiers. Where rejection can happen at any stage

2) Applying AdaBoost for detecting faces In detecting faces, AdaBoost based approaches has two main steps. In the first step, strong classifiers will be constructed from weak classifiers. Since then in the second step, the strong classifiers will be combined sequentially to create a cascade of boosted classifier. a) Building strong classifier Input: training data X, X = {x1, …, xn}. Where x k = x1k , x k2 ,..., x km ∈ X is a feature vector.

The cascade tree has some stages; each stage is a strong classifier. During the detection process, each pattern is analyzed sequentially by each of the stage classifier; they may reject it or let it go through. During the training stage, each stage classifier is trained by false positive patterns of the previous stage. It means that it will learn difficult background patterns. Thus the combination of classifiers in cascade will decrease false alarm rate. With this structure, the classifier can easily recognize background patterns and reject them with first stages. Hence it solves two problems which are complexity and

{

}

2 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.

performance time. In summary, the cascade structure has partly improved the performance time, but the detection results still depends on weak classifiers.

Figure 4 illustrates facial representation by LBP in our experiment. C. Analyses and Evaluations on AdaBoost and LBP

B. Local Binary Pattern 1) The Local Binary Pattern Operator The local binary pattern (LBP) operator is a non-parametric 3×3 kernel which summarizes the local spatial structure of an image. It was first introduced by Ojala et al. [9]. With each pixel, LBP operator considers its pixel neighbors (8 neighbor pixels), based on the grayscales value of central pixel to change its neighbor to 0 or 1. Therefore, every pixel will be represented in a binary string. The decimal value of this binary string will be value of central pixel in description of LBP. Figure 2 shows an example of LBP calculation.

1) Database for Experiment on AdaBoost To train the detector, a set of face and non face training images were used. To create the face training set, we select 11000 face images from 14051 face images of FERET dataset [13]. With each image which selected, we cropped the regions which contained face and scaled to a base resolution of 20 by 20 pixels. Some typical face examples are shown in figure 3.

Figure 3. Example of face images used for training AdaBoost detector

Figure 2. Example of LBP calculation

Later, Ojala et al. [10] extended their original LBP operator to a circular neighborhood of different radius size. LBPP, R notation refers to P equally spaced pixels on a circle of radius R. In [10], they also noticed that most of the texture information was contained in a small subset of LBP patterns. These patterns, called uniform patterns, contain at most two bitwise 0 to 1 or 1 to 0 transitions (circular binary code). 11111111, 00000110 or 10000111 are for instance uniform u2 patterns. This LPB descriptor is denoted LBPP. R . 2) Applying LBP for detecting faces In [2], A. Hadid et al. proposed a face detection system based on a LBP representation of the face. They had proposed a new facial representation which is efficient for low-resolution images. A specific of this representation is the use of overlapping regions and a 4 - neighborhood LBP operator ( LBP4 ,1 ) is computed on these regions. Additionally, they enhance the holistic description of a face by including the global LBP histogram computed over the whole face image. In this representation, the texture of facial regions is encoded by the LBP while the shape of the face is recovered by the concatenation of different local histograms. In our experiment, we considered 20×20 as the standard resolution. The LBP facial representation as follows: We divide a 20x20 face image into 9 overlapping regions of 10×10 pixels (overlapping size = 5 pixels). From each region, we compute a 16-bin histogram using the LBP4 ,1 operator and concatenate the results into a single 144-bin histogram. Additionally, we apply

LBP8u,12 to

the whole 20×20 face image and derive a 59 – bin histogram which is added to the 144 bins previously computed. Thus, we obtain 59+144 = 203 bin histogram as a face representation.

The non-face sub windows used to train the detector come from 5817 images which were manually inspected and found to not contain any faces. There are about 950 million sub windows within these non-face images. Each classifier in the cascade was trained with the 11000 training faces and 5000 non-face sub-windows (also of size 20 by 20 pixels). To train AdaBoost detector, we used open source Haar training (in OpenCv library) which Created by Rainer Lienhart [3]. 2) Experimental results of AdaBoost We used the model of cascade of boosted classifiers which the number of stage classifiers is 20, and 25 stages. We tested system on the MIT + CMU [7] test set. This dataset consists of 130 images with 507 labeled faces. Table 1 presents performance of AdaBoost detector. One a 2.0 GHz Pentium IV processor, RAM 1GB, the face detector can process a 276 by 343 pixels image in about 0.172 seconds. The scale and step size of slide window are 1.2 and 2, respectively. Figure 5 shows some detection results from AdaBoost detector. In figure 5B, AB25 detects 11 faces (3 faces missed). The time to process this image is 0.278 seconds. Whereas, in figure 5A, with AB20, the faces which detected are 12 faces (2 faces missed). The time to process is 0.230 (less 0.287 of AB25). In two cases, the number of false positives is zero. In figure 5D, although it is processed by AB25, it still has 2 false positives; this is because AdaBoost algorithm is still depended on weak classifiers. 3) Database for Experiment on LBP The face training set to train detector contains 11000 face images from 14051 face images of FERET dataset. It is same as face training set which trained AdaBoost (section II.C.1). To collect non face patterns, we used the “bootstrap” strategy in five iterations [12]. First, we run face detector which trained by AdaBoost algorithm (in sections II.C.1) and collected 3000 non face patterns that were wrongly classified as faces and combine them with 3000 patterns which randomly extracted from a set of images which do not contain faces (an example is shown in figure 6).

3 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.

Figure 4. Facial representation: feature histogram is made by a concatenation of a global and a set of local LBP histograms

TABLE I.

PERFORMANCE OF DETECTING ON MIT + CMU TEST SET OF ADABOOST DETECTOR

Method

Number of stages

Number of Haar like features had used

Face detected

False detections

Detection rates

Mount of time to process 130 images (second)

Average time to process an image (second)

AB20

20

1925

467

202

92.11%

23.325

0.179

AB25

25

2913

452

40

89.15%

25.337

0.202

Figure 5. Detection examples from MIT + CMU test. If the large of number of stages are used then the performance time will increase and some face can be missed. Some difficult background patterns are wrong classified (false positive)

4 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.

Figure 6. Some windows size of 20×20 is randomly extracted from an image which does not contain face to make negative training samples

Thus, we have 6000 non face patterns for first training. Then, at each iteration we trained the system, run the face detector, and collected 3000 patterns that were wrongly classified as faces and used them for training. Overall, we obtained 6000 + 15000 non face patterns as negative training examples. To train and test LBP detector, we used an artificial neural network (ANN). ANN which used is three-layer feed forward neural network with back propagation algorithm. The number of input neurons is 203; it is equivalent to the number of bin histogram that extracted by LBP method. The number of output neurons is just 1; it will return true if the scan window contains a human face and false if it does not. The number of hidden neurons will be selected basing on the experiment; it equal to 10 in our experiment. The activation function is Tanh function with the learning rate = 0.3 4) Experimental results of LBP To detect faces in a given target image, a 20×20 sub window scans the image at different scales and locations. The scale and step size to slide window are 1.2 and 2 pixels, respectively. At each iteration, the LBP representation is computed from the sub window and fed to the ANN classifier to determine whether it is a face or not. To merge multiple detections and remove the false ones, we apply similarly as in [2]. We tested system on the MIT + CMU face test set [7]. Table 2 presents performance of LBP detector. 5) Evaluations of AdaBoost and LBP AdaBoost method is one of today’s fastest algorithms. However, the cascade of boosted classifier still depends on weak classifiers. Thus, the detection results often have many false positives. In the number of stages that we evaluated, AdaBoost get highest performance with 92.11% detection rate when using 20 stages. However, the number of false positives is so many with 202. If we increase the number of stages in cascade tree to 25 stages, the number of false positives will go down, but performance time will go up and it is the most important thing that the detecting rate will be decreased. Each search-window must pass through all stages in cascade tree to draw a final decision. Thus if a search-window is a difficult positive pattern, in many cases, it can be rejected by last stages instead of passing through. Mean that if we use many stages in cascade tree, the detection rate can be decreased. However, if we use number of few stages, the detection results will have many false positives. We had tested with 15 stages in cascade tree. The detection rate is 92.50% but the number of false

positives is very big with 1472. (This is same proved in [14]). Thus, if using LBP detector to verify candidate regions of AdaBoost, it will take a lot of time. The LBP method gets high performance with 93.49% detection rate and the number of false positives is 17. This is very small in comparison with 202 of AB20. However, it takes 27635 seconds to process all image and averagely, it takes 212.58 seconds to process an image. This is very slow in comparison with AdaBoost; thus, if only using LBP method in face detection problem, it cannot apply in real time detection. To take full advantage of AdaBoost method about processing speed; LBP method about accuracy and solve drawbacks of these method, we had suggested a hybrid model of AdaBoost and LBP in section III. III.

THE HYBRID MODEL OF ADABOOST AND LOCAL BINARY PATTERN FOR SOLVING THE PROBLEM

A. Proposed model Our hybrid model is named AB – LBP. This is the model of combining AdaBoost and LBP for detecting faces. In this model, AdaBoost has a role to quickly reject non face images; then the candidate regions will be represented by LBP method and classified by ANN to filter out false positive images to achieve better results. The final result is face/non face. Figure 7 illustrate the process of detecting face of AB – LBP. With each candidate region, we consider the windows which got by increasing and decreasing size of candidate regions to 120% and 80% (one time increase or decrease 2%). Thus, corresponding with each candidate regions, we consider (120 – 80)/2 = 20 windows. If one of them is verified as face, the candidate region will be concluded is face. Otherwise, it will be concluded is non face. In [2], Abdenour Hadid et al. had used SVM method to classify the bin histograms which are represented by LBP method. However in our work, we decide to use ANN in classification step because the fact that SVM is slower. Particularly, when trained using dataset mentioned in section II.C.3, SVM method has created 17604 support vectors and it takes 0.031 seconds to scan an image of size 20 × 20. From table1, with AB20, the detection rate is highest with 92.11%, but the number of false positives is 202; the number of candidate regions after using AdaBoost are 467 + 202 = 669, means that, averagely, we have about 669/130 = 5 candidate regions in an image. If we use SVM method in to verify candidate regions, the average time to verify candidate regions in an image is 0.031 x (5 + 20/2) = 0.465 seconds. Clearly, it is time-consumption compared to 0.179 seconds of AdaBoost. Whereas, ANN just takes 0.00015 seconds to scan an image of size 20 × 20, thus, average time to verify the candidate regions in an image is 0.00015 x (5 + 20/2) = 0.00225 seconds; this is an insignificant quantity in comparison with 0.179 seconds of AdaBoost. Thus, AB – LBP model can apply in real time detection. B. Experimental results and Discussion AdaBoost and LBP detector are trained same on section III.C.1 and III.C.3, respectively. These experiments were done on MIT + CMU test set [7]. Table 3 presents performance of AB – LBP detector.

5 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.

TABLE II.

PERFORMANCE OF DETECTING ON MIT + CMU TEST SET OF LBP DETECTOR

Method

Face detected

False detections

Detection rates

474

17

93. 49%

LBP4,1 +

LBP representation u2 8 ,1

LBP

Mount of time to process 130 images (second)

Average time to process an image (second)

27635 ( ≈ 7.67 hours)

212.58

(classified by ANN)

Figure 7. The process of detecting faces of AB – LBP

TABLE III. Name

PERFORMANCE OF DETECTING ON MIT + CMU TEST SET OF AB – LBP DETECTOR

AdaBoost structure Number of strong classifiers

Number of Haar like features

AB20-LBP

20

1925

AB25-LBP

25

2913

LBP structure

LBP4,1 +

LBP8u,12 representation (classified by ANN)

We can see that our approach compares favorably against the approach of AdaBoost in section III.C.2. All AB20-LBP and AB25-LBP get detection rate approximate with AB20 and AB25, respectively but with less than the number of false detection. AB20-LBP gets the detection rate 91.91%; it is approximate with 92.11% of AB20 and higher detection rate of AB25. The number of false positives only is 14. It is very small in comparison with 202 of AB20 and smaller than 40 of AB25. Further more, the processing time to process 130 images is 24.576 seconds. It is approximate with time of AdaBoost detector. In case of AB25-LBP, it gets detection rate 88.95%. This equals to AB25 but the number of false positives is only 3; this is very small in comparison with 40 of AB25. The processing time of our detector for a 276 × 343 image is about 0.174 seconds on a P4 1.8 GHz PC. Table 4 shows performance of our detector with the detector of H. A. Rowley et al [7]. TABLE IV. False detections Detector Ours Rowley et al.

DETECTION RATES FOR VARIOUS NUMBERS OF FALSE POSITIVES ON THE MIT+CMU 3

10

14

31

95

167

89.15%

-

91.91%

-

-

-

-

83.2%

-

86.0%

89.2%

90.1%

Face detected

False detections

Detection rate

Mount of time to scan 130 images (second)

Average time to process an image (second)

466

14

91.91%

24.576

0.189

452

3

89.15%

25.510

0.196

It is clear that our approach gets better performance in comparison with approach of H. A. Rowley et al. both detection rate and the number of false positives. Figure 8 shows the output of our face detector on some test images from the MIT+CMU test set. In figure 8A, AB20 – LBP detector detected 12 faces which same to AB20 (12 faces) and higher AB25 (11 faces) (in figure 5A, 5B). In figure 8B, it detected 6 faces with the number of false positives is zero whereas AB20 and AB25 had 5 and 10 false positives, respectively (figure 5C, 5D). The AB20 detector detected 6 faces in figure 8C and the number of false positives is 1 whereas AB25 detector just detected 5 faces (figure 8D). Notice that this missed face is a difficult positive pattern because it is darker than detected faces. In spite of that, AB20 – LBP detector detected 6 face with the number of false positives is zero. Similarly in figure 8F and 8G, AB25 detector detected 2 in 3 faces whereas AB20 – LBP detected 3 faces. In figure 8H and 8I, all AB25 and AB20-LBP detector detected 3 in 3 faces. However, AB25 has 3 false positives but the number of false positives of AB20 – LBP is zero. It is important to note that in all cases, AB20 – LBP detector takes less time than AB25 detector. These examples summarize the main aspects of our detector. The results show that AB-LBP not only gets approximate detection rate and processing time AdaBoost detector but also minimizes false detections.

6 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.

Figure 8. Output of our face detector on a number of test images from the MIT+CMU test set

7 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.

IV.

CONCLUSION

We have presented an approach for face detection which simultaneously achieving high detection accuracy and minimizing computation time and the number of false detections. The approach used a boosted of Haar-like features and cascade of the strong classifiers with suitable structure in first stage. Then the candidate regions will be represented by LBP method and classified by ANN. The experiments are done on a difficult face detection dataset which has been widely studied. The results show that AB-LBP not only gets approximate detection rate and processing time AdaBoost detector but also minimizes false detections. AB – LBP had solved the drawbacks of AdaBoost [1] [3] and LBP [2] detector. From the achieved experiment results and theoretical analyses presented, we have reached a conclusion that proposed model is efficient and has a practical meaning in the problem of detecting faces.

[3]

[4]

[5] [6]

[7]

[8]

[9]

[10]

ACKNOWLEDGEMENT The author would like to thank to Dr. A. Hadid at University of Oulu, Finland for many helpful discussion on his results [2] and LBP problem. REFERENCES [1]

[2]

P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features”, IEEE Conference on Computer Vision and Pattern Recognition, pp. 511–518, 2001. A. Hadid, M. Pietikäinen, and T. Ahonen, “A discriminative feature space for detecting and recognizing faces,” IEEE Conference on Computer Vision and Pattern Recognition, pp. 797–804, 2004.

[11]

[12]

[13]

[14]

R. Lienhart and J. Maydt, “An extended set of haar-like features for rapid object detection”, IEEE International Conference on Image Processing, 2002. H. Schneiderman and T. Kanade, “A statistical method for 3D object detection applied to faces and cars”, In International Conference on Computer Vision, 2000. S.Z. Li, and A.K. Jain, Handbook of face recognition, Springer Press, 2005. S.Z. Li, and Z. Zhang, “Float Boost learning and statistical face detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1112-1113, 2004. H. A. Rowley, S. Baluja, and T. Kanade, “Neural network-based face detection”, IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 23 – 38, 1998 M. Turk and A. Pentland, “Face recognition using eigen faces”, IEEE Conference on Computer Vision and Pattern Recognition, pp. 586–591, 1991. T. Ojala, M. Pietikäinen, and D. Harwood, “A comparative study of texture measures with classification based on featured distributions”, Pattern Recognition, vol. 29, pp. 51–59, 1996. T. Ojala, M. Pietikäinen, and T. Mäenpää, “Multiresolution gray-scale and rotation invariant texture classification with local binary patterns”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, pp. 971–987, 2002. M.H. Yang, D.J. Kriegman, and N. Ahuja, “Detecting Faces in Images: A Survey,” IEEE Transactions on Pattern analysis and Machine Intelligence, pp. 34-47, 2002. K.-K. Sung and T. Poggio, “Example-based learning for view-based human face detection”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, pp. 39–51, 1998. P. J. Phillips, H. Moon, P. J. Rauss, and S. Rizvi, “The FERET evaluation methodology for face recognition algorithms”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No. 10, pp. 1090 - 1104, October 2000. Thai H. Le, Len T. Bui, “A hybrid approach of AdaBoost and Artificial Neural Network for detecting human faces”, IEEE International Conference on Research, Innovation and Vision for the Future (RIVF), pp. 79-85, 2008.

8 Authorized licensed use limited to: UR Rennes. Downloaded on February 25,2010 at 09:56:20 EST from IEEE Xplore. Restrictions apply.