融入注意力的残差网络表情识别方法  

RESIDUAL NETWORK EXPRESSION RECOGNITION METHOD WITH ATTENTION

作　　者：史志博 谭志 Shi Zhibo;Tan Zhi(School of Electrical and Information Engineering,Beijing Jianzhu University,Beijing 100044,China)

机构地区：[1]北京建筑大学电气与信息工程学院,北京100044

出　　处：《计算机应用与软件》2023年第9期222-228,共7页Computer Applications and Software

基　　金：2017年度教育部人文社会科学研究专项任务(工程科技人才培养研究)项目(17JDGC015);2019年北京市社会科学基金项目(19JYB003)。

摘　　要：针对现有的卷积神经网络模型算法对人脸表情特征表达能力不足、识别精度不高、模型参数量大的问题,提出一种融入注意力的残差网络人脸表情识别方法。该方法在特征提取部分利用卷积块注意力模块(Convolutional Block Attention Module, CBAM)增强对判别性特征的表示;通过残差结构的卷积层提取表情特征;利用Softmax进行表情分类。实验结果表明,与原ResNet模型方法相比,以较小参数量的增加取得更好的识别效果,在FER2013、JAFFE和CK+数据集上的识别率分别提升了2.68百分点、6.40百分点和6.06百分点,与其他相关方法的对比也证明了其有效性。In order to solve the problem that the existing convolution neural network model algorithm has insufficient ability to express facial expression features,the recognition accuracy is not high,and the model parameters are large,a residual network face expression recognition method with attention is proposed.In the feature extraction part,the convolutional block attention module(CBAM)was used to enhance the representation of discriminative features.The expression features were extracted by convolution layer of residual structure.Softmax was used for expression classification.The experimental results show that compared with the original ResNet model method,it achieves better recognition effect with a small amount of parameters.The recognition rates on FER2013,JAFFE and CK+data sets are improved by 2.68,6.40 and 6.06 percentage points respectively.The comparison with other related methods also proves its effectiveness.

关 键 词：卷积神经网络 人脸表情识别 注意力模块 残差结构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]