Study On Linear Subspace Analysis For Face Recognition

Face recognition has become a hot research field of pattern recognition and imageprocessing because of its broad application prospects. Extracting effective feature is thekey to improve the performance of the face recognition system. Due to its simpleness andhigh recognition rate, subspace analysis method has now become one of the mainstreammethods of facial feature extraction and face recognition. We have mainly studied thelinear subspace-based face recognition algorithms, specific research is as follows:(1) Based on the in-depth research of the PCA and LDA for face recognition, theirimproved methods were described in detail, including2DPCA, modular2DPCA(M2DPCA), Maximum Scatter difference Discriminate analysis (MSD) andTwo-dimentional MSD(2DMSD). Then we analyzed advantages and disadvantages ofeach algorithm.(2) A method of combination of improved modular2DPCA and2DMSD wasrealized. Firstly, the improved modular2DPCA was applied to the original facial imagefor feature extraction. Then2DMSD was applied to the sub-images of these obtainedfeature images in which way the final feature images were obtained. This method can notonly exploit local features of original image and discriminate information but also totallyavoid the problem of singular value decomposition of matrix.(3) In order to make use of the local features, inspired by the idea of matrix block, amethod called Modular2DMSD (M2DMSD) was realized in this paper. The proposedmethod performs better than traditional2DMSD in recognition rate and robustness.(4) In order to make Enhanced Fisher Disriminant Criterion (EFDC) avoid ofimpairing the discriminating information caused by PCA dimension reduction, based onEFDC, we proposed Two-dimensional Intra-class Diversity Preserving (2D-IDP) for facerecognition. In this method, we built a more robust discriminate criterion which canmake the data points of different class as distant as possible and simultaneously preservethe intra-class compactness and variation and thus avoided the over-fitting problem. Atthe same time, we redefined parameter t in the neighbor graph of EFDC which made it change adaptively according to different samples. Thus it avoided the problem ofperformance degradation caused by inappropriate choice of t.