融合注意力机制改进ResNet的人脸表情识别  

Improved Facial Expression Recognition in ResNet by Integrating Attention Mechanism

作　　者：张栋昱 赵磊[1] ZHANG Dong-yu;ZHAO Lei(School of National Cyber Security,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学国家网络安全学院,湖北武汉430072

出　　处：《计算机技术与发展》2023年第5期130-137,共8页Computer Technology and Development

基　　金：国家自然科学基金联合基金项目(U1936122);武汉市应用基础前沿项目(2018010401011295)。

摘　　要：鉴于现有人脸表情识别方法在表情识别过程中存在的诸多痛点,比如对有效特征提取不够、泛化能力不强、识别准确性不高等,提出了一种改进残差网络的人脸表情识别方法。首先,引入卷积注意力机制,对网络中间的特征图进行重构,强调重要特征,抑制一般特征;其次,使用激活函数PReLU替换ResNet中原有的ReLU,在提高模型拟合复杂数据能力的同时,避免出现在负值区域的梯度永远为0,进而导致模型训练时无法执行反向传播的问题;然后,在网络输出层的avgpool与fc之间加入Dropout抑制过拟合,以进一步增加网络模型的鲁棒性与泛化性;最后,在公开数据集CK+上的仿真实验结果表明,该方法的准确识别率达到96.12%。与现有多种经典算法,以及baseline算法即ResNet101相比,改进的网络模型具有更好的识别效果,证明了该方法的有效性与优异性。In view of many pain points in the process of expression recognition of the existing facial expression recognition methods,such as the extraction of effective features is not enough,generalization ability is not strong,recognition accuracy is not high,a facial expression recognition method with improved residual network is proposed.Firstly,the convolutional attention mechanism is introduced to reconstruct the feature map in the middle of the network,emphasizing important features and suppressing general features.Secondly,the activation function PReLU is used to replace the original ReLU in ResNet,which can not only improve the ability of the model to fit complex data,but also avoid the problem that the gradient in the negative area is always zero,which leads to the failure of backpropagation in model training.Then,Dropout is added between avgpool and fc in the output layer of the network to suppress overfitting,so as to further increase the robustness and generalization of the network model.Finally,the simulation experiments on the open dataset CK+show that the accurate recognition rate of the proposed method reaches 96.12%.The improved network model has better recognition effect than the existing classical algorithms and baseline algorithm,namely ResNet101,which proves the effectiveness and excellence of the proposed method.

关 键 词：人脸表情识别 深度学习 残差网络 卷积注意力机制 DROPOUT 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]