Learning more distinctive representation by enhanced PCA network

Subspace learning approaches extract features by a simple linear transformation, which can viewed as a shallow network, and they cannot reveal the deep structure embedded in pixels of image. To solve this problem, a deep principal component analysis (PCA) network, namely enhanced PCA Network (EPCANet), is proposed to explore more distinctive representation for face images. EPCANet adds a spatial pooling layer between the first layer and second layer in the PCANet. The spatial pooling layer reveals more spatial and distinctive information by down-sampling or pixel offset for the first layer output and original images. Extensive experimental results in several databases illustrate the efficiency of our proposed methods.