基于代价敏感卷积神经网络的人脸年龄识别方法  

Facial age recognition based on cost sensitive convolutional neural network

作　　者：徐东明 任娅琼 马宣 Xu Dongming;Ren Yaqiong;Ma Xuan(School of Communication&Information Engineering,Xi’an University of Posts&Telecommunications,Xi’an 710061,China)

机构地区：[1]西安邮电大学通信与信息工程学院,西安710061

出　　处：《计算机应用研究》2020年第11期3516-3520,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61571361,61102095);陕西省国际合作交流与面上项目(2017KW-013,2019JM-604)。

摘　　要：当年龄识别被看做分类问题时,基于卷积神经网络(CNN)的方法通常直接采用一般图像分类的CNN进行年龄识别,常常忽略了进行人脸年龄识别时需要考虑的误分类代价问题。基于上述观察,提出一种基于代价敏感卷积神经网络(CS-CNN)的人脸年龄估计方法。具体来讲,基于期望类最大原则(desired class maximum principle,DCMP)提出了一种能够使CNN学习到鲁棒人脸特征的代价敏感交叉熵损失函数(CS-CE),最后通过理论与实验的方法进行验证。相较之前的人脸年龄识别方法,该算法提升的效果是显著的。When regarding age recognition as a classification problem,the method based on convolutional neural network(CNN)used CNN for general image classification to identify the age,and consequently,which ignored the misclassification cost problem that needed to be considered when performing facial age recognition.Based on the above considerations,this paper proposed a facial age estimation method based on cost sensitive convolutional neural network(CS-CNN).Specifically,based on the desired class maximum principle(DCMP),this paper proposed a cost sensitive cross entropy loss function(CS-CE)that enables CNN to learn robust facial representations.Finally,it was verified by theoretical and experimental analysis.Compared with the previous face age recognition method,the effect of the promotion is remarkable.

关 键 词：卷积神经网路 人脸年龄识别 误分类代价 代价敏感性 期望类最大原则 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]