基于残差扩展网络的性别识别方法研究  

Research on gender recognition method based on residual expansion network

作　　者：黄丽雯[1] 黄俊木 宋涛[1] 陈兵[1] 王勃 黄铮 HUANG Liwen;HUANG Junmu;SONG Tao;CHEN Bing;WANG Bo;HUANG Zheng(Chongqing Key Laboratory of Modern Photoelectric Detection Technology and Instrument,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学光纤传感与光电检测重庆市重点实验室,重庆400054

出　　处：《激光杂志》2020年第8期77-82,共6页Laser Journal

基　　金：国家自然科学基金青年基金(No.61701056);重庆市科委基础与前沿研究计划项目(No.CSTC2016JCYJA0308)。

摘　　要：针对利用面部静态图像实现的性别识别方法在远距离情况下失效的问题,提出了一种基于步态分析的性别识别方法。首先以步态剪影的高宽比特性确定步态周期并生成步态能量图作为神经网络输入,其次,提出残差扩展模块以构建残差扩展网络,对步态能量图进行深度训练提取步态深度特征,最后用SVM代替神经网络全连接层作为分类器进行性别识别。CASIA-B数据集实验表明该算法具有较高的性别识别率,并对视角、背包协变量具有较强鲁棒性。Aimed at the problem of gender recognition method using facial static images in long-distance situation,a gender recognition method based on gait analysis is proposed.Firstly,the gait cycle is determined by the aspect ratio of gait silhouette and the gait energy image is generated as neural network input.Secondly,the residual expansion module is proposed to construct residual expansion network.The deep training of gait energy image is used to extract gait depth features.Finally,the SVM,instead of neural network full connection layer,is set as a classifier for the gender recognition.The CASIA-B dataset experiments show that the algorithm has a high gender recognition rate and has good robustness to the perspective and backpack covariates.

关 键 词：性别识别 残差扩展模块 残差扩展网络 支持向量机(SVM) 深度特征 

分 类 号：TN91[电子电信—通信与信息系统]