Gabor Feature Based Two-pass Classification Method For Face Recognition

Face Recognition is one of the most toughest problems in Pattern Recognition and Image Processing field, but because its wide range of applications in law, customs, security, and so on, so face recognition attracts more and more research groups and companies.Many subjects such as computer vision, neural network, image processing and pattern recognition are included in the problems of Face Recognition. There are three parts of the face recognition system: face detection, feature extraction and classifier design. This paper focuses on feature extraction and classifier design. Feature extraction is to extract useful information for classification from face images; while pattern recognition is to classify the features that extracted.This paper propose a novel approach named Gabor feature based two-pass classifier: For the feature extraction part, Gabor feature are used; for the classifier design part, Two-pass Classification method with Biomimetic Pattern Recognition and Error Correcting SVMs is used. We do a lot of experiments, for feature extraction part, Gabor feature was compared with PCA and Fisherface; for the classifier part, two-pass classifier was compared with the nearest neighbor, biomimetic pattern recognition and error-correcting SVMs. Gabor wavelets, whose kernels are similar to the response of the two-dimensional receptive field profiles of the mammalian simple cortical cell, exhibit the desirable characteristics of spatial localization, orientation selectivity, and spatial frequency selectivity; two-pass classifier combines the HENN method based on the biomimetic pattern recognition theory and the SVM method with error correction ability. The HENN is used for the first classification to get the intermediate result, and the SVM with error correction method is used to solve the intermediate result and all the training samples for the second classification. It has the advantages of both biomimetic pattern recognition method and the SVM method with error correction ability, the first is that it can avoid the high false recognition rate, and the second is that is has the error correction ability, so it outperforms either of the two methods. Experimental results based on the Cambridge ORL database, the Yale database, the AR sunglass sub-database, and the AR scarf sub-database show with the Gabor featue based two-pass classifier, the recognition rate is improved to 99.75%, 99.001%, 97.9%, 97.62%.