Block-based Kernel Semi-supervised Discriminant Projection

The paper developed a new feature extraction method, termed Block-based Kernel Semi-supervised Discriminant Projection (BKSDP). Its main idea is that each face image is partitioned into smaller sub-blocks which are reshaped into a large column vector. Then all these column vectors are mapped into a higher dimensional feature space through a predefined kernel function. Finally, the semi-supervised discriminant analysis is performed over the feature space by integrating the intrinsic manifold structure of the labeled and unlabeled data. Experiment results conducted on ORL and YALA face databases demonstrate the effectiveness of the proposed method.