一种改进的残差网络的人脸识别方法  

A Face Recognition Method Based on Improved Residual Net

作　　者：唐风高 伍雪冬[1] TANG Fenggao;WU Xuedong(School of Information and Electronics,Jiangsu University of Science and Technology,Zhenjiang 212000)

机构地区：[1]江苏科技大学电子信息学院,镇江212000

出　　处：《计算机与数字工程》2020年第9期2248-2253,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61671222)资助。

摘　　要：针对卷积神经网络在人脸识别中会随着网络加深而造成训练精度退化和分类效果不理想的问题,论文提出了一种基于改进的深度残差网络人脸识别算法。该方法先利用多任务级联卷积神经网络(MTCNN)算法对人脸图像进行检测对齐,再利用残差网络提取人脸图像特征,然后用Angular Softmax(A-Softmax)来代替普通Softmax层使得神经网络能够学习到更具辨别性的角度特征。在CASIA-WebFace、LFW(Labeled Face in the Wild)和YTF(Youtube Face)人脸库上进行仿真实验,结果表明:与普通的卷积神经网络相比,该算法不仅能够学习更加深的人脸特征,而且可以有效提高人脸的识别率。Aiming at the degeneration of training accuracy and the poor classifying performance on deepening convolution neural network in face recognition process,a face recognition method based on improved Residual Net is proposed.Firstly,the method uses multi-task convolution neural networks(MTCNN)algorithm to detect and align face images,and applies residual network to extract facial image features.Finally,utilizing Angular Softmax(A-Softmax)instead of the normal Softmax enables the neural network to learn more discriminative angular features.The study is carried out on the CASIA-WebFace,LFW(Labeled Face in the Wild)and YTF(Youtube Face).The experimental results suggest that compared with the common convolutional neural network,this algorithm not only can learn deeper facial feature,but also effectively improve the face recognition rate.

关 键 词：卷积神经网络 网络加深 人脸识别 改进的残差网络 角度特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]