Two Discriminant Analysis Methods And Their Application To Face Recognition

Compared with other biologics, face recognition has the merits of natural,friendly,non-intrusive and simple to collection. So it has became a hot topic in biologic characters recognition field. Aim at the merits and demerits of traditional method, the dissertation proposes a new approach and an improved algorithm in face recognition. The main points include the following several parts:Firstly, in chapter1 of the dissertation, it narrates the backgroud,researchful signification of face recognition and its development. Also including the presentation of face database,classifier and distance measurement. In chapter2, it introduces the development of Fisher discriminant analysis and correlative theory foundation of locality preserving projects. Based the LPP, then it introduces the method of Laplacianfaces in face recognition.Secondly, to combine the merit of Fisher discriminant analysis and laplacian graph: Linear discriminant analysis can obtain the best projecting direction and laplacian graph can preserve the local structure. In chapter3, a new method of general laplacian discriminant analysis is proposed in face recognition. Research shows that the local structure commonly includes important discriminant information in pattern classification problem. Two standard databases from Yale University and Oliveti research laboratory are selected to evaluate the recognition accuracy of the proposed method with traditional method. From the result of the experimental data, it shows that our method has better recognition effect than the traditional method.At last, aim at the deficiency of traditional discriminant analysis in multicategory case. An improved algorithm of weighted linear discriminant analysis is proposed in chapter4. Because its between-class scatter matrix pays more attention to the classes which between-class distance are far. It overlaps the classes which between-class distance are close with each other. The main idea of the algorithm is that redefines the between-class scatter by adding a weight function according to the between-class distance. It...