基于残差坐标注意力的人脸表情识别

Facial expression recognition based on residual coordinate attention

针对以往的通道注意力忽略了面部图像中蕴含的坐标信息特征的问题,提出了一种融合坐标信息的人脸表情识别模型。该模型以残差网络(residual network,Resnet)为基础,在网络中嵌入坐标注意力机制,通过在通道注意力中捕获坐标信息辅助生成注意力权重,使得注意力机制不仅考虑不同通道之间的特征,也考虑图像坐标信息和形状特征,进而提高人脸表情识别的准确度。结果表明,该模型在FER2013和CK+表情识别数据集上的准确率分别为74.20%和94.55%,效果优于现有诸多主流方法,在人脸表情识别任务上获得了较优的性能。

A face expression recognition model incorporating coordinate information is proposed to address the problem related to the ignorance of the coordinate information features embedded in facial images in the previous channel attentions.The model is based on a residual network(Resnet)and embeds a coordinate attention mechanism in the network.By capturing coordinate information in the channel attention,the attention mechanism takes into account not only the features between different channels but also the image coordinate information and shape features,thus improves the accuracy of face expression recognition.The results show that the model achieves an accuracy of 74.20%and 94.55%on the FER2013 and CK+expression recognition datasets,respectively,outperforming many existing mainstream methods,which indicates superior performance on the face expression recognition task.