A Comparative Study of PCA, LDA and Kernel LDA for Image Classification

Although various discriminant analysis approaches have been used in content-based image retrieval (CBIR) application, there have been relatively few concerns with kernel-based methods. Furthermore, these CBIR applications still applied discriminant analysis to face images as face recognition did. In this paper we concerns images with general semantic concepts. We use our presented symmetrical invariant LBP (SILBP) texture descriptor to extract image visual features. We then explored effectiveness of principal component analysis (PCA), fisher linear discriminant analysis (LDA), and kernel LDA algorithms in providing optimal discrimination features. Following it, we present an LDA based framework to carry out kernel discrimiant analysis in our application. By taking advantage of the efficiency in nonlinear condition of kernel-based methods and the simplicity of LDA, the proposed approach can improve the retrieval precision of CBIR. The experimental results validate the effectiveness of the proposed approach.