基于二度冗余网络的人脸识别  

Face recognition based on double-redundant network

作　　者：张园林 李晓峰[1,2] 黄飞虎[2] 

机构地区：[1]四川大学计算机学院,成都610065 [2]四川大学视觉合成图形图像技术国家重点学科实验室,成都610065

出　　处：《计算机应用》2017年第A02期160-162,共3页journal of Computer Applications

基　　金：国家重大科学仪器设备开发专项(2013YQ490879)

摘　　要：在人脸识别的研究中,使用传统的深度学习方法需要大量的训练数据和很深层次的神经网络,为此,提出一种通过两个冗余的网络分支协同工作的神经网络设计。首先,该网络在输入层对图像数据随机进行多种仿射变换,模拟生成多组数据,用于扩大数据量;然后,通过一系列的卷积操作和两个相互独立的子网络分支,进行人脸特征预提取;最后,将这两个网络分支拼接到一起,再通过全连接生成最终需要的人脸特征向量。在训练阶段,需要注意先对两个独立的网络分支单独训练,使用softmax loss和triplet loss协同监督,当它们具有较好的收敛效果时再同时训练。通过实验,冗余网络相较于非冗余网络,在LFW公开库和在线实时测试中,识别率分别提升了1.46%和5.56%,ROC曲线的效果都得到了可观的提升。In the research of face recognition, traditional deep learning methods need massive data for training and very deep layers of neural network. Therefore, a neural network with two network branches working together was proposed. Firstly,the image data on input layer was processed by affine transformation to produce simulating data for data base enlargement. And then, the face features were aextracted through a series of convolutional operations and two independent double-redundant subnets. At last, these two sub-nets were concatenated to obtain a face feature vector by inner product layer. In the training phase, two redundant nets were trained independently, softmax loss and tripllet loss were used to loccaboratively supvised until they almost convergent. Compared with non-redundant network structure, experimental result on LFW data base and online real-time test show that the recognition rate is improved by 1. 46% and 5. 56% respectively, and ROC curve is better in this network.

关 键 词：人脸识别 深度学习 神经网络 冗余 模式识别 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]