通道注意力机制的局部遮挡人脸表情识别  

LOCAL OCCLUSION FACIAL EXPRESSION RECOGNITION BASED ON CHANNEL ATTENTION MECHANISM

作　　者：莫文彬 伊力哈木·亚尔买买提[1] Mo Wenbin;Yilihamu Yarmamati(College of Electrical Engineering,Xinjiang University,Urumqi 830000,Xinjiang,China)

机构地区：[1]新疆大学电气工程学院,新疆乌鲁木齐830000

出　　处：《计算机应用与软件》2024年第2期145-151,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61866037,61462082)。

摘　　要：针对人脸存在局部遮挡影响人脸表情识别率的问题,提出一种具有可判别性残差网络的局部遮挡人脸表情识别方法。在残差网络的每个残差块内部以及最后一个残差块的后面嵌入通道注意力机制得到新残差网络模型,通过该模型获取具有通道依赖的特征图特征,在残差网络的全连接层引入具有可判别性的island loss函数,该函数与softmax loss函数相结合进行特征的分类。用不同算法对遮挡处理后的表情图像进行识别。结果显示,具有判别性的残差网络在遮挡处理后的Jaffe和CK+数据集上得到的最高识别率分别为97.6%和95.4%,该方法能够在一定程度上有效提高局部遮挡人脸表情的识别率。The part of the face is blocked by various obstructions,resulting in a decrease in the recognition rate of facial expressions.Aiming at this problem,this paper proposes a local occlusion facial expression recognition method with discriminative residual network.The channel attention mechanism was inserted in each residual block of the residual network and after the last residual block to obtain a new residual network model that did not included a fully connected layer,and a channel-dependent feature map was obtained through the model.The island loss function with discriminable characteristics was introduced in the fully connected layer of the residual network,which was combined with the loss function of the softmax to classify the output features.Different algorithms were used to recognize the facial expression after occlusion processing.The results show that the highest recognition rate of the discriminative residual network on the Jaffe and CK+data sets after occlusion processing is 97.6%and 95.4%,respectively.This method can effectively improve the recognition of partially occluded facial expressions to a certain extent.

关 键 词：残差网络 注意力机制 表情识别 特征提取 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]