Face Feature Extraction using Bayesian Network School of Computer Sciences Universiti Sains Malaysia 11800 Minden Penang, Malaysia
Rosalina Abdul Salam
Zurinahni Zainol
School of Computer Sciences Universiti Sains Malaysia 11800 Minden Penang, Malaysia
School of Computer Sciences Universiti Sains Malaysia 11800 Minden Penang, Malaysia
Tel: 604-6532486
Tel: 604-6532486
Tel: 604-6532486
[email protected]
[email protected]
[email protected]
Zulkifli Dol
ABSTRACT
2. FACE FEATURE EXTRACTION
Face recognition is highly dependent on two stages that are image preprocessing and classification. Methods for feature extraction and classification have been investigated. Through the investigations a method that uses Bayesian Network for feature extraction and Backpropagation algorithm for classification has been proposed. A prototype of the system was implemented and experiments were carried out. Different set of parameters were used for each experiment. Parameters involved were the learning rate, momentum rate and the number of training cycle. Results were satisfactory. The most outstanding performance shows that 78% successful recognition has been achieved with the feature extraction process and 70% without the feature extraction process.
Selected facial features can be best defined as a unique representation of face that provides only valuable information for face recognition. Facial feature extraction can be grouped into two major categories that are template matching and geometrical local-feature based schemes [4]. First category consists of the Principal Component Analysis (PCA), Independent Component Analysis (ICA) and Linear Discriminate Analysis (LDA) [4]. In Geometrical local-feature based schemes, it consists of knowledge based methods feature invariant approach, template matching and appearance-based [4]. Another method that is Bayesian Network is also another method that is used for feature extraction. From the studies, it shows that Bayesian Network will be able to extract face features and should be able to improve the accuracy and the reliability of the system [3]
Keywords Face recognition, Backpropagation.
feature
extraction,
Bayesian
Network,
2.1 Bayesian Network Approach Bayesian network is a graphical representation that considers probabilistic relationship for variables of interest [5]. Bayesian framework works together with PCA as dimensionality reduction will provide better results especially for classification tasks [5]. There are several advantages of Bayesian Network. One of the advantages is the ability of the model presented by Bayesian network to encode dependencies in each variable. This will avoid the lost of information. The second advantage is the Bayesian network can be used to learn causal relationships among each of the variable. Here it can produce a better understanding of the variables environment and also to study the probability of any changes in the variables. Third, the Bayesian model itself can provide both causal and probabilistic meaning. Besides that Bayesian networks will provide a better result especially when work together with Bayesian statistical methods and it promises the higher quality data. The data analysis gathered by the Bayesian is often represented by ANN, rule base as well as decision trees [6].
1. INTRODUCTION Face recognition is one of biometric methods that have been widely used because of its reliability and accuracy in the process of recognizing and verifying a person identity. The need is becoming important since people are getting aware of security and privacy. There are two important stages in face recognition that are highly dependent on each other. These are the image preprocessing task and the recognition task. These two stages provide a mean of preparing and evaluating data. In the first stage that is image preprocessing one important task has been addressed that is feature extraction. This task plays an important role as to mine the important part of the image, provide such accuracy and reliable information of the images as it will be used in later process that is image recognition [1]. The second stage is the Artificial Neural Network (ANN) and it is the applicable technique for recognition [2]. This paper is based on paper [3] that was the proposed framework for face recognition system. Implementation of the proposed work together with experimental results is presented in this paper.
3. FACE RECOGNITION In the second stage, extracted features from the face images will become the input data. Here only important information of images that is feature of faces will be considered as input. There are several techniques for recognition that has been used successfully and can be found in [7] and [8]. Results show that SVM is very much dependent on what is called as Optimal Separating Hyperplane (OSH) which is use to discriminate between classes [7]. SVM is suitable for general purpose pattern recognition [8].
Copyright © 2006 by the Association for Computing Machinery, Inc. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Permissions Dept, ACM Inc., fax +1 (212) 869-0481 or e-mail
[email protected]. GRAPHITE 2006, Kuala Lumpur, Malaysia, November 29 – December 02, 2006. © 2006 ACM 1-59593-564-9/06/0011 $5.00
261
There is also nearest neighbor method which compares distances among class of interest in order to classify patterns [4]. Nearest neighbor approach is also suitable for general purpose pattern recognition since it is dependent to class differences [4.] However, this research focused on neural network since it is more reliable, easy to maintain, develop and minimum maintenance is required [2]. It also has the capability to learn [9].
4.3 Image Transformation In this stage the image will be transformed into binaries. This step is done using the formula [10].
bin – ψk1(f(x)) =
4. METHODOLOGY
1
if f(x) ! k1
0
Otherwise
There are a number of stages involves in this face recognition system. Firstly is the preparation of the face images, image acquisition, preprocessing, and image transformation. Secondly is the feature extraction phase and finally is the recognition phase. All faces are gathered on the same condition in controlled of illumination, resolution, facial view and facial expression. All domain faces used the same condition as to provide standard quality of face images.
Figure 3. Image after transformation.
4.1 Face Image Acquisition In this project a digital camera with 4 megapixels was used. Domain candidates were asked to present only normal facial expression (without any glasses, beard or mustache) and frontal view is the only area considered. The domain candidates will then be applied under the same illumination control and their faces were snapped and stored in folders. The resolution used was 320 x 640. Some samples of the candidates are shown in figure1.
It will only provide valuable information of images and reduced unnecessary information that will be used in later processing. By treating only 0 and 1 bits, it will gives less computationally processing especially in feature extraction procedure that will be discussed in the next sub topic. The sample image from this procedure is shown in figure 3.
4.4 Feature Extraction In this stage the Bayesian network was used. Bayesian network is a Directed Acyclic Graph (ADG) that has structure of network consists of nodes and dependencies probabilistic. Few of models have been studied based on [11]. The most suitable model for this face recognition system that is the Bayesian Network Augmented Naïve-Bayes (BAN) was selected. This is shown in figure 4. C is the class node and X1, X2, X3 and X4 are the features or point of interest.
Figure 1. Sample Images
4.2 Image Preprocessing Three important steps of image preprocessing were performed to all set of images that are illumination equalization, gray scale manipulation technique and edge detection. All the steps were done automatically by using ACDsee 6.0. The final results of the processing will be forwarded to the next step that is image normalization. This can be seen in Figure 2. Each of the images will be applied with a mask that will locate only face area and new face image will have the size of 64 x 64.
C X1 X2
X3
X4
Figure 4. Bayesian Network Augmented Naïve-Bayes (BAN) model. The advantages of the model are: i) it provides independent feature, ii) it promises an acceptable result over large amount of data and iii) it is easy to represents. The image will be scanned in order to detect the point of interest in the search space using the mask. The point of interest will be treated as the network of interrelated nodes. From here the BAN model will be applied. The algorithm is shown in figure 5.
Figure 2. Image preprocessing steps
262
4.5 Face recognition system
5.1 Test Results
There are two stages for this final stage of face recognition system. The first stage is the training process and the second stage is the recognition process. The training algorithm is shown in figure 6 and the recognition algorithm is shown in figure 7.
The performance of the proposed project will be measured in three conditions that are based on the learning rate, the momentum rate and the number of epoch. Each of the conditions is tested with feature extraction process and without feature extraction process.
!
!
!
Phase I (Drafting): i. Computes mutual information of each pair of nodes as a measure of closeness ii. Creates a draft based on this information
1. Begin procedure 2. fire image representation procedure 3. fire point of interest detection
Phase II (Thickening): i. Adds arcs when the pairs of nodes cannot be d-separated ii. The results of this phase is an independence map (I-map) of the underlying dependency model
procedure 4. fire image feature extraction 5. for (final_weight = 1 to 9) 5.1 initialize (final_weight) 5.2 fire sigmoid function 5.3 find final_output
Phase III (Thinning): i. Each arc of the I-map is examined using conditional independence tests and will be removed if the two nodes of the arc can be d-separated ii. The result is the minimal I-map
end for (final_weight) 6. calculate temporary = final output – target value 7.
if (-0.001
