Face Recognition Based On Gabor Wavelets And RBF Network

Humans can detect and identify faces in a scene with little or no effort, building an automated system that accomplishes such objectives is the aim of many researchers in computer. Machine recognition of human face from still and video images has become an active research area in the communities of pattern recognition and computer vision, this interest is motivated by the urgent need of real-time and acute status-discrimination in the society.In this paper, the frequently used paradigms for face feature extraction and categorization are introduced and classified, and a novel approaches based on Gabor wavelets and RBF network is presented. The main content is summarized below:An improved symmetry transform for eye location gives accuracy above 94% on ORL database.Make lots of experiments to examine the importance of frequency and orientation of 2D Gabor wavelets, the conclusion: in multi-pose face recognition, no orientation should be omitted, but can choose certain frequencies.To reduce the image dimension of Gabor feature, wavelets decomposition is more useful than sampling.A novel algorithm to determine the initial hidden structure by clustering algorithm can make the RBF network's architecture approximately optimum according to Moody theory, the experiments proves that our algorithm has high generalization ability without over-training.Combining the gradient paradigm and the linear least square paradigm to adjust the RBF network parameters, and the iterative formula for the gradient paradigm has been proved. Propose an applied approach to compute the learning rates for width and center modification, so the RBF network can work without auxiliary parameter. Only 16 Gabor wavelets are used for image analysis, on the ORL database, five images are randomly chosen for training from the ten images available for each subject, for original images or the preprocessed images the correct face recognition rate is above 99%; If each subject contains one face with rotation in-depth at least, the success rate of recognition reaches 100%. To our knowledge, this is one of the best results on this database up to now.