On Improving Linear Discriminant Analysis for Face Recognition with Small Sample Size Problem

• Main Authors: Hsien-Jen Lin, 林咸仁
• Other Authors: Jen-Tzung Chien
• Format: Others
• Language: zh-TW
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/26zux8

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 90 === In the literature of face recognition, LDA (Linear Discriminant Analysis) is a popular linear transformation technique to extract the discriminant feature vector. However, this technique may occur the singularity problem on calculating the inverse of the within-class scatter matrix when the training sample size is small. To overcome the problem, an improved linear discriminant analysis is proposed in this thesis. The proposed method aims to transform the original feature vector to a new feature space with the same degree of scattering and without the singularity problem. Then, We use the LDA to extract the new feature vector. In the experiments on five face databases, the results show that the recognition rates of the proposed method is significantly better than other LDA-based techniques when training data are sufficient. In case of very limited training data, the proposed method achieves desirable recognition performance with moderate training cost.