Hyperspectral Face Recognition Using Block based Convolution Neural Network and AdaBoost Band Selection

Abstract

Hyperspectral imaging, adding discrimination information along spectral dimension, offers a new chance for robust face recognition. To improve the effectiveness of facial feature represented by hyperspectral face data, we proposed a block-based hyperspectral face recognition method using bands selection and convolution neural network (CNN). Firstly, a small convolution neural network is trained to capture discriminative visual information for different blocks in face images. Secondly, an improved AdaBoost algorithm (AdaBoost.MS) is introduced to choose different optimal bands for different blocks. Then, each block label can be determined by the ensemble learning classification. Finally, the recognition result can be gotten by the majority voting principle. The experiment results based on PolyU-HSFD database show that block-level based bands selection can capture the more discriminative spectral features than the method based on image level. The proposed method outperforms the existing state-of-the-art methods.