Parallel Feature Extraction through Preserving Global and Discriminative Property for Kernel-Based Image Classification

Kernel-based feature extraction is widely used in image classification, and different kernel methods extract the features based different criterion. KPCA maximizes the determinant of the total scatter matrix of the transformed sample, while KDA seeks the direction of discrimination. KPCA preserves the global property, and KDA utilizes class information to enhance its discriminative ability so as to perform better than KPCA in classifications. To apply the global property and discriminant ability of features, we propose a novel parallel feature fusion method based maximum margin criterion, namely discriminant parallel feature fusion. The advantage of algorithm lies in: 1) A constrained optimization problem based on maximum margin criterion is created to solve the optimal fusion coefficients to be most discriminant in the fused feature space. 2) An unique solution of optimization problem is transformed to an eigenvalue problem, which causes the proposed fusion strategy to perform consistently. Besides of the detailed theory derivation, many experimental evaluations also are presented in this paper.