In the present work, we propose a framework for kernel-based 2D feature extraction algorithms tailored to face recognition .
extending 2D-PCA/LDA in the following two aspects:
(1)kernel technique is incorporated to capture the higher order statistical dependencies among the rows of input images. Thus, face recognition results are considerably improved
(2)A block-wise model for face recognition is introduced and proved to be reliable through our experiments.
2D-KPCA所带来的问题：
（1）Computational complexity: unlike direct extension of 2D-PCA/LDA to kernel-induced feature space which is computationally intractable, proposed B2D-KPCA/GDA perform subspace projection inside each block manifold, significantly alleviating computational cost
（2）Locality: as widely known, distribution of face images is both multimodal and highly nonlinear [7,17]. Direct extension of 2D-PCA/LDA to kernel-induced feature space, overcomes the shortcomings of 2D-PCA/LDA in extracting global nonlinear structure of input data, but fails to learn local characteristics of input data.