Sei sulla pagina 1di 10

American Journal of Scientific Research ISSN 1450-223X Issue 36(2011), pp. 67-76 EuroJournals Publishing, Inc. 2011 http://www.eurojournals.com/ajsr.

htm

Face Detection with the Help of Gabor Wavelets Characteristics and Neural Network Classifier
M. Mohammad Abadi Gonabad Branch, Islamic Azad University, Gonabad, Iran M. Rezapour Birjand Branch, Islamic Azad University, Birjand, Iran R. Jahani Birjand Branch, Islamic Azad University, Birjand, Iran Hadi Chahkandi Nejad Birjand Branch, Islamic Azad University, Birjand, Iran Abstract In this paper a method for detecting desirable number of faces in fixed photo with gray background is proposed. First by using correlation of a window with a desirable face with photo, areas of candidate of face presence is estimated imprecisely and then using step algorithm and referring these areas and around them to the section of extraction of Gabor wavelets characteristics and neural network classifier, resultant areas lead to precise detection of face locations in photo. Finally, the result of estimation of efficiency of proposed method was examined by different tests.

Keywords: Face detection, Gabor wavelets, neural network

1. Introduction
Face is the first and most important factor in human interactions and has been analyzed in different areas for a long time. Face includes important information about a person such as his health, mental condition and and each person interacts with others according to these information. Human has special skill in analyzing and interpreting faces, and so face analyzing has an important role in manmachine relationship and different research areas has been opened in this way. Identity determination, gender determination and face identification are some of these areas [1]. It is obvious that the first step of face processing is detection of face location in the given photo [2]. In spite of that, in many of the researches in face identification, it's supposed that the location and dimensions of face is clear in photo and because of that a deep attention has not been paid to face identification area and conducted researcher are related to recent years [3,4,5]. Face detection includes determining existence or lack of all human faces in photo and also their locations and limits. It's obvious that different hypothesis about face detection problem lead to it's simplicity or complexity. For example it's possible that the aim is just detection of full faces or there are some presuppositions regarding size and angle of face, wearing or not wearing hat, glasses or using artificial tools,

Face Detection with the Help of Gabor Wavelets Characteristics and Neural Network Classifier

68

environmental light and noise in photo. It's clear that providing a method that detects faces in every condition is not a simple task. In addition to mentioned presuppositions, face detection methods can be divided to fixed or moving detection in color photos or in gray backgrounds, Methods which apply to moving photos, use the feature of human dynamism in relation to photo background and simplify the problem of detection in moving areas. In methods to color photos human body's color spectrum can be detected in the photo and detection problem can be simplified to these areas [16]. The most difficult type of face detection is related to fixed photo with gray surfaces which input data of photo is the intensity of the light reflexes Into the camera and in the first sight it's impossible to discriminate probable location of face in the photo background easily. In this article, the aim is providing a method for full face detection in fixed photos with gray surfaces, determined size, maximum variation of 30 degrees and resistant to light variations in optimum time.

2. Literature Review
Face detection methods are divided to two parts from processing perspective. In the first part there are the methods which are applied to gray surface of photos. For example, in [6] using a neural network, photo areas are classified to face and non-face. In study [7], full faces are detected by classifier. These methods have quantity polynomial SVM resistance to environmental light variations. Also in reference [8] a method for face detection independent of relation is provided which uses two neural network, one for detection and another for face angle determination. In this study, all possible blocks are referred from photo to these two neural networks which not only this has a high time cost but also using two separate neural networks cause a high volume of calculations and increases detection time proportionally. The second part includes methods which extract determined characteristics of photo blocks and detects face by help of classification of these characteristics. In reference [9] BIZIN classifier is used. To do so, probability of occurrence of different small blocks in face calculated and used. In reference [10], with the help of extracting local characteristics such as eyes, mouth corners and the structure of nose from photo blocks and comparing it to extracted characteristics from a typical face, detection process is done. In the proposed method, Gabor wavelet change characteristics are used. Using these characteristics highly increases method resistance to variations in environmental light. More ever using correlation qualities in a preprocess class and deleting no face areas in the photo leads to notable decrease in calculation time.

3. The Proposed Method


The proposed method includes two parts. The first part includes finding windows in which probability of face presence is more. In part 3-1, how these windows are found is described. Each window is known with its center. From now on, when center, pixel or dot is mentioned, imagine windows in which these dots are (this dot is) located and also block, area and window are synonymous. Output of part 3-1 is some dots. These dots are the center of some windows which it's supposed that face is present in them. These dots are called imprecise center because experience shows that these windows either don't have face or their location in a bit different to real face in the photo. Here the second part (part3-2) gets into operation which is the algorithm of space around imprecise center. In this algorithm, to explore space in search of precise centers of faces, a tools is needed which tests windows for presence or absence of face. This tools acts as a function which receives window as an input and the more is the returned amount from it, indicates that the probability of presence of a face in the center of the window is more. This tool is accomplished through extraction of Gabor wavelets characteristics

69

M. Mohammad Abadi, M. Rezapour, R. Jahani and Hadi Chahkandi Nejad

from window (part 3-3) and their reference to a neural network which is trained with the help of windows which have or have not faces (part 3-4). Below these parts are explained in detail. 3.1. Imprecise Determination of Face Location The aim of this part is imprecise determination of windows in which probability of face presence is more. A big photo includes thousands of windows. Because if the window is known by it's center which is a dot or pixel, the number of pixels of window photo can be attained without considering margins approximately. The problem is the great number of window. With the volume of operation which is done on each window, processing all windows in optimum time is not possible. As photos are fixed and no color, skin color information moving area discrimination can't be used. The idea is using correlation coefficients. Correlation coefficient of windows involving a desirable face with faces present in the photo produces greater amount compared to other parts of photo such as background. Correlation equals moving window on the photo and it's internal multiplication. Since this result can be obtained from multiplying two photos in the frequency domain, this part has a high speed. Figure (1) general scheme of this method. The photo is multiplied by windows from photos of desirable faces in the frequency domain which produces greater coefficients in the areas suspected of faces presence. By extracting these local maximums, primary dot batches are got as imprecise centers of windows which are probable of face presence.
Figure 1: General scheme of determination of centers of windows with faces

3.2. Exploring Surrounding of Imprecise Centers If we consider a vertical rectangular inscribed on a face such that in the uppermost line of the pixels of this rectangular starting pixels of hair are located and in the last line lower pixels of chin are located. In general for a typical face, the center of this rectangle is placed on the nose. However experiences indicate that imprecise center obtained in the previous part which are the center of rectangular windows, either are not coincide on the real center of face in the photo or entirely does not include face. In this circumstances an algorithm is needed that necessarily surrounding space of this imprecise centers for finding the best answer for face presence. The following algorithm provides such a possibility: First of all, imprecise centers are marked for referring to network. Until a center for testing is not found: 1. Refer marked centers to neural network. 2. If neural network returned great value than threshold for every center, mark side pixels if they are not marked. Finally just these necessary areas are examined and aforementioned algorithm makes a big saving on face detection time. Figure (2) illustrates examined centers by neural network.

Face Detection with the Help of Gabor Wavelets Characteristics and Neural Network Classifier
Figure 2: Centers of examined windows by neural network

70

In the figure (2), green and red parts are parts which have returned amounts higher and lower than threshold (0.4) by network respectively. Violet parts are as red parts with this difference that obtained this amount in the zero phase and their surrounding space is not examined. Black parts have obtained amounts lower than low end of training (-0.9) and these parts have exempted of examination a specified area from them. Yellow parts which are seen are these parts which are exempted of examination by neural network. In each of examined parts (green area), just one center is selected as the center of window that includes face and this center has the greatest amount which is returned from neural network and is greater than threshold amount too. 3.3. Extracting Characteristics from Windows When speaking about referring window to neural network in the previous part, this operation is not done directly but there is a layer for extracting characteristics of window and this is the characteristics vector which goes to neural network for presence or absence of face. These are extracted characteristics which make face detection resistance against variation of environment light. In the layer of characteristic extraction Gabor wavelet is used. Gabor wavelet alteration like other different wavelets alterations is used in areas of photo processing and machine vision because of its unique features. These wavelets have provided a ground for understanding photo frequency and their analysis in the area of place and their greatest advantage is slow variations in frequency area. If L(z) be input photo with values of light intensity of gray surface, when photo wavelet alteration is calculated according to relation (1) from convolution of one of wavelet family members with input photo. O ,v ( z ) = I ( z ) * ,v (1) Wavelet to , the result of O alteration in which * indicates operator of convolution. Equation of magnification applied in alteration of Gabor direction in face indicates a plate simple wave with specified frequency and direction which is trapped under a Gaussian function. This equation can be defined in various forms according to the type of coordinates system wether Cartesian or polar and the following form is the most general form its illustration in various articles [13,14,17]. 2 K ,v 2 z 2 ik ,v z 2 K ,v 2 ,v ( z ) = 2 exp e e 2 (2) In this relation, k indicates the length and direction of wave and is calculate by following relations: k k ,v = kv ei , kv = ma ,v = v f 8 (3)

71

M. Mohammad Abadi, M. Rezapour, R. Jahani and Hadi Chahkandi Nejad As it can be seen from relation (3), is multiplied in
and has produced number phase of k so 8

it will have a integer value from zero to seven. Greater values produce waves with repetitive directions. can have values between zero and four and 40 wavelets with different directions and sizes are obtained from above relation. For analytical understanding of this wavelet, we show one of the members of this better family with dimensions of 128*128 and following parameters in figure (3). v = 4 , = 7, f = 2 , kma = , = 2 .
Figure 3: Illustration of one of the member Gabor wavelet family in the form of: a. 2D b. 3D

As it can be from figure (3), Gabor wavelet is a plate wave with decreasing range. figure (4.a) illustrates all the wavelets of this family and figure (4.b) illustrates a face photo under these wavelets. By paying attention to photos obtained from alteration result it can be found that in figure (4.b) and in the right column only eyes and nose location is extracted. After introducing Gabor wavelets here the method of characteristic extraction from human face photo is discussed. Figure (5) illustrates general structure of phases of characteristic extraction from a face photo. Windows of photo which is supposed to be decided about its being a face is balanced considering their gray surface histogram. In the next step-Fourier transform- in Fourier transforms, Gabor wavelets are multiplied (convolution in time area) then reversed Fourier transform from to obtained photos are placed by each other and in fact produce matrix of specified window characteristic. Here if input window is a rectangular with dimensions of 27*18, then matrix of characteristics will be 144*135.
Figure 4: a. All wavelets of Gabor transformation family. b. human face transformation under wavelets of figure (a)

Face Detection with the Help of Gabor Wavelets Characteristics and Neural Network Classifier

72

It's clear that this number of process characteristics in each classificatory is large and causes slow down in detection process. Here algorithms of space reduction. So by considering process time, here reduction in dimensions of characteristics matrix is done by averaging blocks and instead of 3*3 block. By doing so, characteristics matrix is reduced to 48*45. This operation is done because the maximum and minimum value obtained from output of wavelet transformation is not a determined and clear number and in designing neural network, there is a need limit arrays between two specified value. In this phase, vector of characteristics of examined photo window is ready to refer to neural network.
Figure 5: Phases of extraction of Gabor wavelet characteristics for a face photo

3.4. Neural Network Classifier 3.4.1. Network Structure To classify obtained characteristics vector to two class of face and nonface, different methods can be used. One of these method is using neural network which has developed in recent years and various structures of it are proposed. To select and design the structure of neural networks, various factors should be considered. Three main factors are calculation volume, responding time and generalization capability. Generalization capability means that network is trained with limited number and little training data which is face and nonface vectors here and it's expected that network can provide desirable responses against vectors which have observed them [11]. To improve network response, network structure, training data and extracted characteristics play on important role. As it was observed, after extracting characteristic from windows refered to network, data will be transformed into the form of a vector with 2160 elements. Applied neural network is a two-layer perceptron neural network with 100 neurons in the first layer and one neuron in the second or output layer. Transfer function of each neuron is hyperbolic tangent which its output varies between number of -1 and +1. It has been attempted that the number of layers and neurons be in minimum level so that calculation volume and then response time reduce to its minimum level. Figure (6) illustrates network structure used in this research.

73

M. Mohammad Abadi, M. Rezapour, R. Jahani and Hadi Chahkandi Nejad


Figure 6: Applied network structure

3.4.2. Neural Network Training Neural network are divided to two class of monitored and unmonitored considering training. Applied network in this research is monitored type. A number of photos including face no face and also their desirable responses are displayed for the network. Then network by aid of algorithm after error dispersion, regulates it's weights according to training data. So face photos which have quality of 27*18 like windows dimensions, are sent to characteristics extraction part one by one, and characteristics extraction, is transformed to a vector with 2160 elements. Then a big matrix which every column of it is characteristic of one of training photos is used for network training accompanied by the desirable response of each column which is 0.9 for face vectors and -0.9 for nonface vectors. Face photos are providable easily. The main problem is selection (not providing) nonface photos. It's clear that these photos can be selected optionally but the point is the dependency of network efficiency to selected photos. So it's possible to add windows which wrongly detect face in the network testing process to the set of training nonface photos and repeat training process. In this research 70 face photos and 60 nonface photos are used in training phase. Also for every face photo, its mirror photo and with the angle of 5,10,15 degrees in positive and negative directions and photos with one pixel shift in every four directions are placed in training set for reducing network sensitivity. For no face photos also, their mirror and their 180 degrees transformation is placed in training data.

4. Test and Practical Results


The provided method is simulated in MATLAB environment and is ran on a computer with Corei5 processors and 4 Gigas memory. The results are compiled in table (1) and proposed system has attained to 95% right answer. Considered network with 1000 repetition, is the error limit of 0.0001 and training is sequential (sequentiality of training is not much desirable and it's better to use random method for network training). False positive error includes those locations of photo which has been recognized as face wrongly. It's obvious that these errors are due to neural network inability to classify correctly and for its resolving, network configuration or training process can be corrected. False negative error is equal to 5% and includes locations in the photo in which there is face, however has not been recognized as face. This error can be attributed to neural network or from the beginning they have not been placed in imprecise point. It became clear in examinations that the most part of errors are in estimating imprecise points. In other words, some of the faces which are in the faces, are not recognized as the center of a probable face in primary estimations and in this case presence or absence of face is not examined in them. A number of tested photo in table (1) are depicted in figures (7) and (8). Figure (8.b) displays very good resistance of this method against environmental light changes.

Face Detection with the Help of Gabor Wavelets Characteristics and Neural Network Classifier
Table 1: Results of experiments in 11 pictures that including 148 various faces
Shape 8 a) 7 b) 7 a) 8 b) 8 c) ------------------------Size of used picture 254600 50100 65150 96130 150150 5575 300400 217300 6060 147260 150150 Positive False 12 0 0 0 0 0 1 3 0 1 1 Detection 56 from 57 2 from 3 7 from 7 15 from 15 6 from 7 2 from 2 24 from 24 10 from 11 1 from 1 6 from 9 12 from 12

74

5. Conclusion
In this study, a method for face detection in photos is introduced. To do so, Gabor wavelet characteristics and neural network classifier were used. It's possible to attain a higher percentage of detection by improving primary points estimation method. Another important point is the dependence of network classification capability to the composition of training photo set. For example, when the number of training nonface photos are high, the number of output neurons reduces and results in that even for face photos, output be less than threshold value. In this regard, the composition of training samples or generally the method of network training can be researched. Also characteristics selection can help network by improving training process and reducing input vector size, because all to characteristics obtained from characteristics extraction class are utilized in this method. In some cases that pre-process is done windows with linear functions, a very high speed can be attained by using convolution of weights of input layer of neural network with photos.
Figure 7: Samples of tested photos and the results of detection

75

M. Mohammad Abadi, M. Rezapour, R. Jahani and Hadi Chahkandi Nejad


Figure 8: Samples of tested photos and the results of detection

Suggestions
There are many others methods for face detection and due to expansion of its application in today's world, its optimization or inventing new method is very important. In this project, photo convolution method with a pattern is used at the beginning to determine imprecise location of faces. Convolution operation is very sensitive to changes and so is not suitable so it can be said that today this method is not used any more because of this sensitivity. Using hidden Markov model or fuzzy networks

Face Detection with the Help of Gabor Wavelets Characteristics and Neural Network Classifier

76

accompanied by a neural network have a better response. For reducing errors, face roundness model can be used too. In such a way that photo edges can be determined by edge detection method and then detect face in the photo with regard to specific from of face roundedness and location of eyes, nose and mouth.

References
[1] [2] [3] [4] R.Challappa, C.Wilson, S.Sirohey. Human and machinerecognition of faces: A suravy. PROCEEDINGS OF THE IEEE, VOL 83, NO 5, MAY 1995 G.Donato, M.S.Bartlett, J.C.Hager, P.Ekman and T.J.Sejonowski. Classifying facial actions. IEEE Trans. On Pattern Analysis and Machine Intelligence, 21(10), OCT 1999. E.Hjelmas, B.K.Low. Face detection: A survey. Computer Vison and Image Understanding, IEEE, 83(3), 2001. O.Jesorsky,K.J.Kirchberg,R.W.Frischolz. Robust facedetection using the hausdorf distance. Lecture Notes in Computer Science. Procs. Of the Third International Conference on Audioand Video- Based Person Authentication. 2091:90_95:2001. S.Z.Li, L.Zhu, Z.Zhang, A.Blake, H.Zhang, H.Shum. Statistical learning of multi-view face detection. In European Computer Vision Conference, 2002. K.K.Sung. Learning and example selection for object and pattern recognition. PhD Thesis, MIT, AI Lab, Cambridge, MA, 1996. E.Osuna. Support Vector Machines: Traning and Applications. PhD Thesis, MIT, EE/CS Dept, Cambridge, MA,1998. H.A.Rowley, S.Baluja, T.Kande. Rotation invariant neural network-based face detection. Computer Science Technical Report,CMU-CS-97-201,CMU,Pittsburgh, 1997. H.Schneiderman, T.Kande. Probabilistic modeling of local appearance and spatial relationships for object recognition. IEEE Conference on Computer Vision and Pattern Recognition, 45-51, Santa Barbara, 1998. T.K.Leung, M.C.Burl, P.Perona. Finding faces in cluttered scenes using random labeled graph matching. International Conference on Computer Vision, p:637-644, Cambridge, MA, 1995. Neural Network Toolbox for using with MATLAB,Version 7.0, Release 14. K.A.Ishak, S.A.Samad, A.Hussain, B.Y.Majlis. A fast and robust face detection using neural networks. Kebansaan University. Malaysia D.Paramadihanto, H.Wu, M.Yachida. Invariant face recognition by gabor wavelets and neural network matching. (59-63) 0-7803-3280-6/96/$5.00 k~19 96 IEEE J.Zhu, M.Vai, P.Mak. Gabor wavelets Transform and extended nearest feature space classifier for face recognition, Proceedings of the Third International Conference on Image and Graphics (ICIG04) 0-7695-2244-0/04 $20.00 2004 IEEE F.H.C.Tivive, A.Bouzerdom. A new class of convolutional neural networks and their application to face detection. (2157-2162) 0-7803-7898-9/03/$17.00 02003 IEEE L.Mostafa, S.AbdelAzeem. Face detection based on skincolor using neural networks. (57-62) GVIP 05 Conference, 19-21 December 2005, CICC, Cairo, Egypt Y.Pang, L.Zhang, M.Li, A novel gabor-LDA Based face recognition method, Microsoft Resaerch Asia, China, 2002

[5] [6] [7] [8] [9]

[10]

[11] [12] [13] [14]

[15] [16] [17]

Potrebbero piacerti anche