Face Recognition Based On Independent Component Analysis

The technology of face recognition is one of the most important subjects in the area of computer pattern recognition, which have been widely used in the fields of national security, business etc. For this reason in recent years, more and more attentions have been paid for this research. However, face recognition is also a very diffcult problem because of the particularity of the face image. Though much progress in this research is made, there are still lots of work to do when practical applications. In this paper, some improved face recognition algorithms are proposed based on Independent Component Analysis (ICA).The work includes:1. Based on the characteristic of non-sensitive to illumination of residual image which is retrieved from reconstructing image by ICA, a method of performing Non-negative Matrix Factorization (NMF) in residual image is realized in this paper. The residual images are computed by subtracting the reconstructed images from the original face images, and the reconstructed images are obtained by performing ICA on original images. Then, NMF is applied to the residual images for extracting non-negative subspace and the corresponding coefficient matrices. The two coefficient matrices are combined for face recognition based on the nearest neighbor classifier (NNC). The recognition rate is improved by superior over the lower-dimension by PCA, the independent characteristic of ICA, the non-sensitivity of residual image and the non-negativity of NMF. Finally the simulation experiments illustrate the validity of the method by using the standard ORL database.2. Because Possibilistic C-Varieties (PCV) can apply PCA and data partition which satisfies the premise of linear mixture for ICA, simultaneously, a new method for extracting local independent components is proposed, namely Possiblistic FastICA. Firstly, PCV is applied to the original image data as the preprocessing of FastICA considering the memberships to clusters derived by PCV algorithm. Secondly, the egien-subspace of each cluster is taken as a classifier, and primary recognition is carried on based on NNC. Finally, the recognition results of each class are fused by fuzzy integral to get the final recognition result.Finally, the recognition results of each class are fused by fuzzy integral to get the final recognition result. The computer simulation illustrates the effectivity of this method.