改进的残差网络和混合注意力的人脸表情识别  

Face expression recognition based on improved residual networks and hybrid attention

作　　者：李弟文 潘伟[1] LI Diwen;PAN Wei(School of Computer Science,China West Normal University,Nanchong Sichuan 637009,China)

机构地区：[1]西华师范大学计算机学院,四川南充637009

出　　处：《智能计算机与应用》2023年第5期9-17,共9页Intelligent Computer and Applications

摘　　要：针对传统残差网络对人脸表情特征提取存在的泛化能力差、识别准确度低等问题,本文实现了基于混合注意力机制和残差网络相结合的方法,从而能够更准确快速的实现对人脸表情进行识别。提出使用CABasicBlock代替ResNet34中的basic block作为骨干网络,并加入空间、通道注意力机制来提取人脸的全局面部特征,以及与表情重要关联的局部面部特征,使得网络收敛更快。在网络的残差模块后加入批量归一化和PReLU激活函数,以及在全连接层前加入了Dropout、全局平均池化,能有效预防训练过程发生过拟合,提升其泛化能力;引入Focal Loss函数平衡训练数据样本均衡性的问题,使用余弦退火策略对训练过程中的学习率进行动态衰减,从而减少训练时间。通过实验结果数据表明,在FER-2013测试数据集上的准确率为73.7%,说明此模型拥有更好的效果。最后,基于以上方法再结合OpenCV和PyQt5,构建了一个实时表情识别的可视化检测结果输出。To address the problems of poor generalization ability and low recognition accuracy of traditional residual networks for face expression feature extraction,this paper implements a method based on a combination of hybrid attention mechanism and residual networks,which can achieve more accurate and fast recognition of face expressions.The CABasicBlock is proposed to replace the basic block in ResNet34 as the backbone network,and the spatial and channel attention mechanisms are added to extract the global facial features of faces and the local facial features associated with expressions to make the network converge faster.The addition of batch normalisation and PReLU activation functions after the residual module,as well as Dropout and global average pooling before the fully connected layer,can effectively prevent overfitting in the training process and improve its generalisation capability.This reduces the training time.The experimental results show that this model has better results,with an accuracy of 73.7%on the FER-2013 dataset respectively.Finally,a visualisation of the detection results based on the above method combined with OpenCV and PyQt5 was constructed for real-time expression recognition.

关 键 词：深度残差网络 CABasicBlock 表情识别 混合注意力机制 余弦退火策略 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]