HM-FER: Hybrid attention mechanism integrate Multiple scales for Facial Expression Recognition

Facial expression recognition (FER) is of great interest and is widely used in the real world. In this paper, we propose a pure convolutional model with a hybrid attention mechanism and a multi-scale receptive field. First, we introduce the Inception modules in the network's middle layer to aggregate visual information at different sizes, make feature extraction from different scales easier, and prevent overfitting. Second, we employ depthwise separable convolution to reduce the model's computational effort and the number of parameters. Finally, we introduce the hybrid attention mechanism CBAM to enable the network to automatically focus on the important features and regions of facial expression. Our method was experimented on the publicly available dataset fer2013 and achieved an accuracy rate of 72.44 %, which is a high level and some of the training strategies in the experiment are informative for this study.