Abstract: The change of facial expression is very subtle and usually manifested in the change of some local points and regions in the image. The existing facial expression recognition methods are difficult to capture the subtle changes of facial expression, which does not have robust to interference from non-expressive regions. To obtain an efficient feature representation to describe the changes of facial expression, a facial expression recognition method that integrating key point attributes and attention representation is proposed. Firstly, the key points in face image are extracted by the module with channel attention and spatial attention, which realizes the weight distribution of different dimensions and positions, effectively avoids the interference of non-expressive regions, and obtain the feature representation of the local key points in the image. Then, transformer module is introduced to learn the correlation between different key points and guide the network to build a more distinguishing feature representation, so as to achieve accurate recognition. Finally, the experimental results on CK+, JAFFE and FER2013 public datasets show that the recognition accuracy of the proposed algorithm is up to 99.22%, 96.57% and 73.37% respectively.

Key words: facial expression recognition, key point attributes representation, attention mechanism, convolutional neural network, learning feature map

摘要： 人脸的表情变化非常细微，通常表现在图像中某些局部点区域的改变，现有的人脸表情识别方法难以捕捉到表情的细微变化，对非表情区域干扰不具有鲁棒性。为了获得描述人脸表情变化的高效特征表示，提出了一种融合关键点属性与注意力表征的人脸表情识别方法。通过添加通道注意力和空间注意力的神经网络提取人脸图像中的关键点信息，实现不同维度和位置的权重分配，有效避免非表情区域的干扰，捕获图像中局部关键点的特征表征。引入Transformer模块学习不同关键点之间的相关联系，引导网络构建对表情类型更具分辨力的特征表示，从而实现精准识别。通过在CK+、JAFFE、FER2013三种公开数据集上进行实验的结果表明：提出算法的识别准确率分别达到了99.22%、96.57%、73.37%。

关键词: 人脸表情识别, 关键点属性表征, 注意力机制, 卷积神经网络, 学习特征图