Research, Face Recognition Method Based On Canonical Correlation Analysis

As an important application of image recognition, face recognition has been a hot topic in pattern recognition field. Feature extraction is one of the most basic questions in pattern recognition research,for face recognition tasks, extracting the effective facial features is a crucial step. Recently, with the development of feature fusion technology, the method Canonical Correlation Analysis, which will integrate the two feature vectors into a eigenvector with more identifying information, is developing rapidly. This paper discusses Canonical Correlation Analysis and its enhancement algorithms in the application to face recognition. And the main research work and contributions of it are as following:1. From the favoring pattern classification point of view, this paper proposed an improved correlation criterion function named Generalized Discriminant Canonical Correlation Criterion, which has the characteristics both from Generalized Canonical Correlation Analysis (GCCA) and Discriminant Canonical Correlation Analysis(DCCA). Meanwile, this paper provided the theory and algorithms of Generalized Canonical Correlation Analysis (GCCA) and its application to face recognition in detail. Experimental results on the AR, ORL and Yale face database validate the proposed algorithms. The experimental results also show that the method GCCA is better than the other three method CCA, DCCA and GDCCA ,being the best method of the four methods.2. Based on having analysed the theory and algorithms of the two-dimensional traditional Canonical Correlation Analysis (2D-CCA), From the retaining useful information point of view, this paper proposed an improved two-dimensional Canonical Correlation Analysis method. The impoved method uses every train sample in one class and the mean of the class as two feature matrixs of Two-dimensional Canonical Correlation Analysis,and then extracts canonical projection vectors. Experimental results on the AR and ORL face database validate the proposed algorithms. Experimental results show that recognition rate of the improved method improve greatly compared with the traditional one.3. Through analyzing and improving Discriminant Canonical Correlation Analysis(DCCA) based on vector,this paper proposed a new method directly based on image matrix, called Two-dimensional Discriminant Canonical Correlation Analysis (2D-DCCA). Based on the same idea, According to Generalized Discriminant Canonical Correlation Analysis (GDCCA), a novel method of ombined feature extraction called two dimensional Generalized Discriminant Canonical Correlation Analysis (2D-GDCCA) is proposed in this paper. Compared with those methods based on vectors, the methods based on matrix avoid the Small Sample Size problem (SSS), besides, much computational time would be saved. Finally, extensive experiments performed on ORL and AR face database verify the effectiveness of the proposed method.