Research On Face Recognition Technique Based On Kernel Fisher Discriminant

Face recognition is an important branch of biologic feature identification.Because of its advantages comparing to other biologic features, considerable attentionhas been paid to face recognition. Due to the importance of human being in themultimedia information, the recognition based on man's biometrics information is oneof the important topics in computer vision and pattern recognition in past 20 years.Many approaches to face recognition problem have been devised, from the earlygeometry based methods to statistics based methods. Although linear techniques havebeen fully developed, they are still inadequate to describe the complexity of real faceimages because of the illumination, facial expression and pose variation. Hence it'snecessary to extend the linear techniques to the nonlinear ones. In this dissertation, theresearch to focuses on kernel Fisher discriminant analysis. The emphasis is on theextension. The primary contributions and original ideas included in this dissertation aresummarized below:1. Because it is inadequate for linear subspace analysis methods to describe thecomplex relations of real face images, such as pose, illuminant, expressionvariations. Kernel based Fisher Discriminant Analysis (KFDA) is proposed forface recognition, which combines the nonlinear kernel trick and Fisher LinearDiscriminant Analysis (FDA). According to the algorithm of kernel fisherdiscriminant analysis (KFDA), a concept, named kernel sample set, is introduced.Based on this concept, KFDA is equivalent to performing FDA on kernel sample set.Then the nonlinear algorithm is converted to a linear one. A great deal of researchhas been on linear approaches. And the kernel sample set extends the linear methodsto nonlinear ones that can improve the result. Experimental results show that it cangive higher accurate recognition rate than linear subspace analysismethod and KPCA.2. Two enhanced kernel fisher discriminant models (EKDA-1 and EKDA-2) areproposed which use a strong combination of enhanced FDA and kernel tricks. It takes the over-fitting dilemma into account, and diagonalizes the whthin class andbetween class scatter matrix simultaneously. Experimental results show that theproposed method has an encouraging performance compared with FDA and EFM.