Applied Mathematical Sciences, Vol. 6, 2012, no. 40, 1991-1996 A Neuro-Fuzzy Approach for Automatic Face Recognition Mohammed Madiafi Modeling and Instrumentation Laboratory Hassan II Mohammedia-Casablanca University Casablanca, Morocco madiafi.med@gmail.com Abdelaziz Bouroumi Modeling and Instrumentation Laboratory Hassan II Mohammedia-Casablanca University Casablanca, Morocco Abstract The purpose of this paper is to present a neuro fuzzy approach to the problem of automatic recognition of human faces. This approach is based on a Kohonen neural network (ANN), which we have trained, in unsupervised way, using a fuzzy competitive learning algorithm previously designed, implemented and tested on real images. Illustrative examples that demonstrate the effectiveness of this approach will be presented in this paper. Keywords: Artificial Neural Networks, Fuzzy Logic, Vector Quantization, Pattern recognition, face recognition 1 Introduction Face recognition plays an important role in various fields including security. It is extremely easy for a human operator, but still out of reach of modern computers [1]. Automation of this task requires the development of intelligent systems capable of simulating the capabilities of the human brain and artificial neural networks (ANN) is one of the most promising for the realization of such systems [2,3]. The ANN may be defined as mathematical and computer models that simulate two essential skills of the human brain: learning by example and generalization [4]. The purpose of this paper is to present an ANN that we
1992 M. Madiafi and A. Bouroumi Figure 1: Kohonen neural network. designed for the automatic recognition of faces. This is a neuro-fuzzy model in the sense that the network learning is performed using a fuzzy algorithm [5-8]. In the next section we give a brief description of the ANN used in this study. Section 3 is devoted to the concept of unsupervised competitive learning that underpins the learning algorithm used for training of the ANN. The algorithm itself will be also briefly described in this section. Numerical results and discussion are the subject of Section 4. Finally, a conclusion and perspectives for further work are ending the paper. 2 Kohonen neural network A Kohonen network is a special case of ANN (figure 1), which can be trained in two different ways: supervised and unsupervised way [8-10]. Its architecture has only two layers: an input layer which serves to receive vectors that represent the objects to recognize and an output layer where each neuron represents a prototype from the c possible classes to which the received object can belong. The number of neurons in the input layer equals the number of components of the vector objects, and the number of neurons in the output layer equals the number of classes in the database. The first rule that was proposed in the literature to train this kind of ANN is the LVQ algorithm based on the concept of competition [8]. In this algorithm, each time a vector object x k is presented to the network, c neurons of the output layer compete to qualify for the adjustment of their weights by exploiting the information provided by x k. The neuron that wins the competition is the one whose weight vector minimizes the distance with x k, and this vector will be adjusted according to the rule: v j,t = v j,t 1 + η t 1 (x i v j,t 1 ) (1)
A neuro-fuzzy approach for automatic face recognition 1993 Figure 2: Kohonen neural network based system. The disadvantage of this rule is that it adjusts only one prototype at a time, resulting in poor use of information provided by the vector object to be analyzed. To overcome this drawback we proposed a fuzzy generalization of this rule. The algorithm resulting from this generalization, called fuzzy competitive learning (FCL) is described in the section below. 3 Unsupervised fuzzy competitive learning Learning or training is the process of adjusting, in an iterative way, the connection weights between neurons to ensure that the network becomes capable of performing a well defined task. Training an ANN in general requires a set of examples that must be large enough to allow the network to generalize, that is to say to recognize examples not used in the learning stage. When there is no prior knowledge on classes in original sample, which means when these examples are not labeled, learning is said unsupervised in the sense that the network must categorize by itself the examples and identify different classes in the training set. The learning rule we propose in this work is competitive because neurons compete to qualify for the adjustment of their weights. But it is also a fuzzy rule because, contrary to LVQ which allows only one winner, it allows more than one neuron to win the competition but with different degrees. The calculation of membership degrees of the object presented to the network for each class is performed using the relationship: x i v j,t 1 2 u ji,t = cr=1 (2) x i v r,t 1 2 Finally, the adjustment of synaptic weights vectors of winners is done by following the rule: v j,t = v j,t 1 + η t 1 E i,t v j,t 1 (3)
1994 M. Madiafi and A. Bouroumi Figure 3: Steps of data extraction. with : so that : c E i,t = u ji,t x i v j,t 1 2 (4) j=1 v j,t = v j,t 1 +2cη t 1 (u 2 ij,t )(x i v j,t 1 ) (5) where v j,t 1 and v j,t denote the prototype vector of class j before and at iteration t, and η t the learning rate at iteration t. 4 Numerical results and discussion To illustrate the ability of the model described in the previous section to classify and recognize data images of faces, in this section we present the results of its application to dataset of test commonly used by experts in the field to test and validate their models. These data come from the data images of the project computer vision [11]. We choose from this dataset 200 different images of faces of 20 different individuals where each individual is represented by a sample of 10 images. These images are scanned as bitmap files with a horizontal resolution of 180 pixels and 200 vertical pixels. Before extracting data table from each image, we proceed to a conversion to gray level (Figure 3). Figure 4 gives an idea of the evolution of the quality of learning during the iterations, which means the model s ability to recognize the data used as training examples. Examination of this figure shows that through iterations the quality of learning increases gradually to expect stabilization. This demonstrates the effectiveness of the fuzzy competitive learning technique proposed in this work. Figure 5 shows, for eight different individuals randomly chosen from the database, how prototypes change with the iterations during the learning stage. It is clear from this figure that the quality of prototypes increases with time, reflecting the model s ability to learn well the characteristics of classes and produce prototypes of these classes that can serve as a basis for the rule making or learning to be used in the generalization to recognize pictures of faces not seen in the learning stage, which is needed in practice to optimize the recognition process.
A neuro-fuzzy approach for automatic face recognition 1995 Figure 4: Evolution of the quality of classification over the iterations. Figure 5: Examples of prototypes generated by FCL during the iterations. 5 Conclusion In this paper, we outlined simulation results of a neuro-fuzzy model for automatic recognition of facial images. We opted for an architecture of Kohonen neural network and an unsupervised learning algorithm and we placed in the most sensitive case, but also the most encountered in practice, where no prior information on the images to classify and recognize. Examples of results that have been presented in this paper, using a test dataset used by other researchers, are very encouraging for the satisfactory continuation of this study, for example to study the possibility of optimization of parameters such as learning rate, the initialization protocol, and this in order to increase the robustness of the model and the possibility to use it for other types of applications.
1996 M. Madiafi and A. Bouroumi References [1] T. Pavlidis, 36 years on the pattern recognition front, Pattern Recognition Letters, 24 (2003), 1-7. [2] C. M. Lee, S. S. Yang and C. L. Ho, Modified back-propagation algorithm applied to decision-feedback equalisation, Vision, Image & Signal Processing, 153 (2006), 805-809. [3] S. Furao, T. Ogura and O. Hasegawa, An enhanced self-organizing incremental neural network for online unsupervised learning, Newral Networks, 20 (2007), 893-903. [4] J. C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms, New York: Plenum, (1981). [5] A. Bouroumi, M. Limouri, and A. Essaid, Unsupervised Fuzzy Learning and Cluster Seeking, Intelligent Data Analysis, 4 (2000), 241-253. [6] G.E. Tsekouras, M. Antonios, C. Anagnostopoulos, D. Gavalas and D. Economou, Improved batch fuzzy learning vector quantization for image compression, Information Sciences, 178 (2008), 3895-3907. [7] M. Biehl, A. Ghosha and B. Hammer, Learning vector quantization: The dynamics of winner-takes-all algorithms, Neurocomputing, 69 (2006), 660-670. [8] M.T. Martin-Valdivia, L. M. Uren-Lopez and M. Garcia-Vega, The learning vector quantization algorithm applied to automatic text classification tasks, Newral Networks, 20 (2007), 748-756. [9] A. K. Jain, M. N. Murty, and P. J. Flynn, Data clustering: A review, ACM computing surveys, 31 (1999), 264-323. [10] A. K. Jain and R. C. Dubes, Algorithms for Clustering Data, Prentice Hall, (1988). [11] Computer vision project database, http://cswww.essex.ac.uk/mv/allfaces/index.html. Received: July, 2011