多分支精简双线性池化的人脸表情识别  

Multi-branch Compact Bilinear Pooling for Facial Expression Recognition

作　　者：王彬 徐杨 石进 张显国 WANG Bin;XU Yang;SHI Jin;ZHANG Xian-guo(School of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;Guiyang Aluminum Magnesium Design and Research Institute Co.,Ltd.,Guiyang 550009,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵州贵阳550025 [2]贵阳铝镁设计研究院有限公司,贵州贵阳550009

出　　处：《计算机技术与发展》2023年第3期27-33,共7页Computer Technology and Development

基　　金：贵州省科技计划项目(黔科合支撑[2021]一般176)。

摘　　要：针对人脸表情识别研究中特征提取不充分、难以辨别人脸表情细微的类间差异等问题,提出了一种多分支精简双线池性化的人脸表情识别方法。该方法以ResNet-18为基础,在避免大幅度增加计算复杂度的前提下提升ResNet-18的特征提取能力,提出了一个新的多样化分支块(diverse branch block)对ResNet-18进行改进;为使改进后的ResNet-18更方便地聚焦人脸图像中产生表情区域的特征,提出了残差空间注意力;为了减少人脸表情细微的类间差异带来的不利影响,增强人脸表情类间的区别性,设计了多分支精简双线性池化结构。最后用所提的方法分别在公开的人脸表情数据集CK+、RAF-DB进行实验,识别率分别达到了98.46%、82.99%。实验结果表明,该方法的识别率优于DLP-CNN、MA、DeepExp3D等诸多的表情识别方法,具有一定的竞争性。Aiming at the problems of insufficient feature extraction and difficulty in distinguishing subtle inter-class differences in facial expression recognition in the study of facial expression recognition,a multi-branch compact bilinear pooling facial expression recognition method is proposed.Based on ResNet-18,this method improves the feature extraction capability of ResNet-18 without greatly increasing the computational complexity,and proposes a new diverse branch block to improve ResNet-18.To make the improved ResNet-18 more convenient to focus on the features of expression regions in face images,the residual spatial attention is proposed.In order to reduce the adverse effects of subtle inter-class differences in facial expressions and enhance the distinction between facial expression classes,a multi-branch compact bilinear pooling structure is designed.Finally,the proposed method was used to conduct experiments on the public facial expression datasets CK+and RAF-DB,and the recognition rates reached 98.46%and 82.99%,respectively.The experimental results show that the recognition rate of the proposed method is better than that of DLP-CNN,MA,DeepExp3D and many other expression recognition methods,with certain competitiveness.

关 键 词：人脸表情识别 多样化分支块 残差空间注意力 多分支精简双线性池化 ResNet-18 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]