Face Recognition System Based On Fuzzy RBF Neural Network

Face recognition is a typical pattern analysis, understanding and classification problem, which is closely related to many disciplines such as Pattern Recognition, Computer Vision, Human-Computer Interaction, Statistics Study, and Cognitive Psychology etc. so it is one of the hot and hard problems in Pattern Recognition. At the same time, AFR technology, as one of the key technologies in Biometrics, is believed to have a great deal of potential and wide applications in digital personal identification, intelligent surveillance, message security, bank finance, etc. However, face recognition techniques are also full of challenges under the non-ideal conditions. Face recognition is definitely affected by illumination and poses in face images as it is based on the photics face images. So there are a lot of key issues to resolve to develop a robust and practical AFR system.This article briefly analyzes the face recognition research at home and abroad in the history and current situation, and describes main methods for image preprocessing, then studies several popular human face feature extraction algorithms and their improvements, lastly, explores the design and implementation of human face image recognition system based on fuzzy RBF neural network. The main works of this paper are as follows:1. The DDCT (Divided Discrete Cosine Transform) transformation of human face images. Firstly, we segment the human face into several blocks, then perform the DCT(Discrete Cosine Transform) transform of each block. Finally, we select the DCT coefficients of low and high frequency components and construct characteristic matrix for each face blocks.2. Compress and reduce the dimensions of DCT features matrix using the TCSVD (Singular Value Decomposition Threshold Compression) method. Since the segmented image blocks contain a lot of redundant information caused by shielding, illuminating and expression changes, so it is necessary to compress and reduce the image singular value dimensionality, and then combine these characteristics of singular value to construct a final combination of characteristics of human face identification.3. Design a face image classifier based on fuzzy RBF neural network. As RBF network is equivalent to fuzzy reasoning in function, we can unified the two different systems, so that the network parameters and operations have a clear meaning. We make the number of cluster centers correspond with and the number of fuzzy rules in RBF network to construct the network environment parameters to make the network have the capability of fuzzy reasoning and classification.4. Fuzzy RBF neural network training. BP learning algorithm is widely applied in fuzzy RBF neural network, which has the low rate of convergence, therefore, an improved learning algorithm named Levenberg-Marquart(L-M) for fuzzy RBF neural network was proposed and it is effectively in the number of learning and the rate of accuracy.5. Application in Face Recognition. Testing the trained fuzzy RBF neural network classifier on ORL face database, Experimental results show the effectiveness and feasibility of this method.