Face Recognition Study Based On Gabor Wavelets Transform And Subspace Analysis

Because of face recognition huge application in secure authentication systems, video conferencing, human-computer interaction, the public security system, it will increasingly become the study hotspot in current pattern recognition and artificial intelligence field. In the face recognition, the subspace method has been widely used because of its simple and effective features. The method is studied in depth in this thesis. There are several improvements about the subspace method aim at solving the small sample size and great facial variations. The 2D Gabor wavelet is very conducive to the local feature representation due to its outstanding response in the edge, brightness and position and can improve the robustness of the system. Therefore the 2D Gabor wavelet is introduced to the representation of facial features in the thesis.2DPCA compresses the images only in the row direction, so it needs more coefficients to represent the image. This thesis presents a method called enhanced 2DPCA which compresses the image in row and column direction. The method first uses 2DPCA operated in the row direction and then in the column direction. By this method, the dimension of the feature vector is much smaller than that of the 2DPCA, and the method has sufficient theoretical basis. Making use of Gabor wavelet could express the facial local texture feature, a combination of Gabor wavelet and enhanced 2DPCA is advanced. The method combines the advantages of two methods, and the experiment proves the recognition rate and robustness are greatly improved.Aimed at solving the small sample size in the LDA, this thesis introduces the idea of the image segmentation to the face recognition, and puts forward a 2DLDA method based on Gabor segmentation statistics. This method segmentation the image which transformed by Gabor wavelet, and the sub-block of the images are extracted their feature by 2DLDA as new samples. Lastly, the recognition results are obtained by the general minimum distance classifier. The method extracts local features of face images, reducing the dimensions of the feature matrix. The experimental results show that the performance of this method is better than LDA and 2DLDA.