Abstract: In order to overcome the noise interference of illumination, pose, color in face recognition, this paper combines the advantages of convolutional neural networks with twin neural network, puts forward an improved CNN network structure. The structure is composed of two convolutional neural network’s twin network and sharing the weights, using the Discriminative Deep Metric Learning（DDML） algorithm in the training of the network. The convolutional structure can effectively remove external noises and automatically extract homologous features by the trick of sharing weights. It evaluates the method on the ORL, YaleB and AR face data sets and shows that it outperforms other approaches and its recognition rate up nearly 5 percent compared to the PCA and CNN algorithm.

Key words: face recognition, convolutional neural network, twin network, Discriminative Deep Metric Learning（DDML）, deep learning

摘要： 为了克服人脸识别中存在光照、姿态、颜色等噪声的干扰，融合了卷积神经网络与孪生神经网络的优点，提出了一种改进的CNN网络结构，该结构由两个卷积神经网络组成，且共享网络权值，在该结构的训练中采用了差异深度度量学习（DDML）算法。卷积结构有效地去除外界噪声干扰，且在非线性降维中权值共享结构能够自动提取相同特征，DDML算法增加了提取特征的有效性。在ORL、YaleB和AR人脸数据库上实验结果表明，与PCA、CNN等算法相比，识别稳定度高，识别率提升近5个百分点。

关键词: 人脸识别, 卷积神经网络, 孪生网络, 差异深度度量学习（DDML）, 深度学习