基于卷积网络注意力机制的人脸表情识别  

Face expression recognition based on attention mechanism of convolution network

作　　者：郭昕刚[1] 程超[2] 沈紫琪 GUO Xin-gang;CHENG Chao;SHEN Zi-qi(Medical Image Processing Technology Innovation Laboratory of Jilin Province,Changchun University of Technology,Changchun 130012,China;School of Computer Science and Engineering,Changchun University of Technology,Changchun 130012,China)

机构地区：[1]长春工业大学医学图像处理吉林省校企联合技术创新实验室,长春130012 [2]长春工业大学计算机科学与工程学院,长春130012

出　　处：《吉林大学学报（工学版）》2024年第8期2319-2328,共10页Journal of Jilin University:Engineering and Technology Edition

基　　金：吉林省教育厅基金项目(JKH20210754KJ);长春市科技局重大专项项目(21GD05);吉林省科技厅重点攻关项目(20210201113GX).

摘　　要：针对表情识别时出现参数量大和识别能力弱等问题,提出一种基于卷积网络人脸表情识别方法。引入改进型残差模块,在减少参数量的同时增强对表情区域的关注;利用通道-空间注意力机制对网络提取的表情区域实现不同维度和位置上的权重分配,专注于表情关键点中细微差别特征信息;利用细节模块进一步提取深度特征信息。为得到更高准确度,引入联合损失函数延长类外距离,缩短类内距离以提高表情识别准确度。本文将此网络运用到数据集FER2013、CK+中,实验结果表明:本算法平均识别率分别为63.91%、97.98%,参数量为11.34 M。与VGG网络、残差网络等对比,该模型不仅提高了识别率,还减少了冗余参数量。A convolutional network based facial expression recognition method was proposed to solve the problems of large reference number and weak recognition ability in facial expression recognition.The improved residual module was introduced to reduce the parameters and enhanced the attention to the expression area;The channel-space attention mechanism was used to assign the weights of different dimensions and positions to the expression regions extracted from the network,and the subtle feature information of the key points of expression was focused on;The refinement module was used to further extract the depth feature information.In order to obtain higher accuracy,the joint loss function was introduced to increase the out-of-class distance and reduced the in-class distance to improve the accuracy of expression recognition.The experimental results showed that the average recognition rate was 63.91% and 97.98%respectively,and the parameter was 11.34 M.Compared with VGG network and residual network,the model not only improves the recognition rate but also reduces the redundant parameters.

关 键 词：面部表情识别 残差模块 通道-空间注意力机制 细化模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]