一种结合Gabor小波与深度学习的人脸识别方法

A Face Recognition Algorithm Based on Gabor Wavelet and Deep Learning

针对人脸识别中识别效果易受光照、姿态等因素影响和浅层学习方法不能有效提取人脸图像抽象特征的问题,提出一种结合Gabor小波与深度学习的人脸识别方法。该方法首先利用Gabor小波变换获取不同尺度和方向的人脸Gabor特征,通过下采样和受限玻尔兹曼机(RBM)对Gabor特征进行有效降维;其次将降维后的特征作为深度信念网络(DBN)的输入,并使用对比散度算法训练DBN;最后利用标签数据对DBN进行有监督微调,网络顶层附加Softmax分类器对提取后的特征进行分类。所提方法在ORL、UMIST和Yale-B人脸库上的识别率分别达到了98. 72%、96. 51%和96. 13%,实验结果表明所提方法不仅识别效果明显优于其他现有方法,而且对光照、姿态变化具有很好的鲁棒性。

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.