Face Recognition Based On The Fuzzy Fusion Of Marginal Fisher Analysis And Singular Value Decomposition

Face recognition is widely used in many fields that based on biological characteristic, which attracts more and more attention and research. This paper provides an overview of research background and significance of face recognition. Depending on whether the whole face is used as input, face recognition algorithms can be divided into three categories:holistic methods, local feature-based methods, and hybrid methods. The performance of the face recognition system has reached a satisfactory result under controlled conditions. However, the face recognition rate is decreased greatly when the conditions are not controlled. We introduce the two main contribution of this paper to solve the problems that exist in face recognition.Manifold learning algorithm based on spectral graph theory is proposed by Yan. Then Yang Jian who proposed a two-dimensional marginal Fisher analysis. Since the recognition of 2D Marginal Fisher Analysis is sensitive to the size of neighbors, a novel algorithm is presented which based on the Pearson correlation coefficient. It not only preserves the structural characteristics of the image, but without adjusting parameters. Using training samples can get the best parameters corresponding to the highest recognition rate.Since a sole feature extraction method can’t meet the requirements of further improving recognition rate for variations in expression and pose of facial images. To solve these issues, a novel algorithm is proposed which integrated with 2D Nonparametric Marginal Fisher Analysis (2DNMFA) and SVD. The algorithm extracts algebraic features and identifiable structural characteristics by SVD and 2DNMFA method respectively, then fuse membership degree by using fuzzy decision theory based on the advantages of two characteristics.Experimental results on CIS face database、Texas face database and UMIST face database demonstrate that this algorithm which based on the fuzzy fusion of 2D Nonparametric Marginal Fisher Analysis and SVD improves the recognition rate and is more robust than 2DMFA or SVD.