Many scholars are committed to using deep learning methods to study facial expression recognition (FER). In recent years, FER has gradually been confined to psychology research in the early days to now involves knowledge of many disciplines such as physiology, psychology, cognition and medicine. With the extreme achievement of computer version techniques, various convolutional neural network structures were developed for real-time and accurate FER. There are two main problems in the existing convolutional neural network for handling FER problems: insufficient training data caused over-fitting and expression-unrelated intra-class differences. In this paper, we propose a two-pathway attention network to solve these two problems better. We suppress the intra-class differences efficiently by extracting facial regions based on facial muscle movements driven by facial expressions. We prevent deep networks from insufficient training data by extensively extracting global structures and local facial regions as the training dataset to feed a two-pathway ensemble model. Further more, we weight the whole feature maps from the global image and local regions by introducing an attention mechanism module to reweighs each part according to its contribution to FER. We adopt real-time facial region extraction and multi-layer feature data compression to ensure the real-time performance of the algorithm and reduce the amount of parameters in ensemble model. Experiments on public datasets suggest that our method certifies its effectiveness, reaches human-level performance, and outperforms current state-of-the-art methods with 92.8% on CK+ and 87.0% on FERPLUS.

许多学者致力于使用深度学习方法来研究面部表情识别（FER）。近年来，FER从早期逐渐局限于心理学研究，到现在涉及生理学、心理学、认知和医学等多学科知识。随着计算机版本技术的极端成就，各种卷积神经网络结构被开发出来用于实时和准确的 FER。现有的卷积神经网络在处理 FER 问题时主要存在两个问题：训练数据不足导致过拟合和与表达无关的类内差异。在本文中，我们提出了一个双路径注意力网络来更好地解决这两个问题。我们通过基于面部表情驱动的面部肌肉运动提取面部区域来有效地抑制类内差异。我们通过广泛提取全局结构和局部面部区域作为训练数据集来提供双通路集成模型，从而防止深度网络的训练数据不足。此外，我们通过引入注意力机制模块对全局图像和局部区域的整个特征图进行加权，根据每个部分对 FER 的贡献重新加权。我们采用实时面部区域提取和多层特征数据压缩来保证算法的实时性并减少集成模型中的参数量。公共数据集的实验表明，我们的方法证明了其有效性，达到了人类水平的表现，