关键词: Gabor小波, 人脸识别, 深度学习, 深度信念网络, 受限玻尔兹曼机

Abstract: In order to reduce the negative effects of factors such as illumination and posture and solve the problem that shallow learning methods can’t extract the abstract features of face images in face recognition, a face recognition algorithm based on the Gabor wavelet and deep learning was proposed. Firstly, the facial Gabor features of different scales and directions were obtained by Gabor wavelet transform, the dimensionality of Gabor features was reduced availably by downsampling and Restricted Boltzmann Machine (RBM). Secondly, the features of dimensionality reduced were taken as the input of the Deep Belief Networks (DBN), and DBN was trained by the Contrastive Divergence algorithm. Finally, DBN was fine-tuned by labeled data. The Softmax classifier was used to classification for the features extracted, which was implemented at the top layer. The recognition rate reaches 98.72%, 96.51% and 96.13% respectively on ORL, UMIST and Yale-B face databases. The experiment results indicate that the proposed method is markedly better than other existing algorithms in recognition performance and achieves good robustness to changes in illumination and posture.

Key words: Gabor wavelet, face recognition, deep learning, DBN, RBM