基于改进的局部卷积网络的多姿态人脸识别

Multi-pose face recognize based on Improved locally convolution

针对人脸姿态变化对人脸识别性能的影响以及人脸姿态恢复局部纹理信息不够清晰,提出一种改进的局部卷积神经网络的人脸识别方法。该方法在网络中引入分离层将中间层的信息分离成两个通道,一个通道用于产生中间层的重构函数,另一个通道进行进一步的局部卷积操作;最后将中间层的重构函数作为另一路网络顶层代价函数的一个惩罚项进行网络训练;网络训练完后,将姿态恢复的人脸图片采用LDA进行特征提取,并采用最近邻分类器来做识别。实验结果表明,提出的方法不仅能够很好的进行姿态恢复,且恢复的图片具有较好的纹理信息;同去除姿态的传统方法和较为典型深度学习方法以及3D方法相比,识别性能具有较大的提升。

To reduce the influence of pose variance on face recognition performance and to make the unclear information of local texture in the process of facial pose recovery clearer,an improved local convolution neural network face recognition method is proposed.Separation layer is introduced in the network,which separates the intermediate layer information into two passages,one is used to generate the reconstruction function of the middle layer,the other is for further locally convolution;Finally,considering the reconstruction function of the middle layer as a penalty term of the top layer loss function of another channel in the network,and putting it to the network training.After the network training,LDA is applied to extract the face features,meanwhile,the nearest neighbor classifier is used for the recognition of the pose-recovered face.The experimental results show that the proposed method works well on the pose recovery,also the restored image covers clearer texture information.Compared with the traditional deep learning and 3D method,the proposed method achieves a better recognition performance.