Micro-expression Recognition Based on Improved MobileViT

Because of its short duration and small movement range, micro-expressions are highly similar and dense in details. Considering that the local receptive field of convolution neural network may lose detailed information, we try to use CA-MobileViT. which is a combination of convolution neural network and Transformer, as the feature extraction network. Self-attention mechanism is used to extract global features, and CA (Coordinate attention) attention mechanism is introduced to enhance the extraction of key location features and strengthen the learning ability of the network, thus improving the accuracy of micro-expression recognition. In this paper, the initial frame and apexframe of micro-expression sequence are used to calculate three optical flow features: horizontal optical flow, vertical optical flow and optical flow strain. The three optical flows are fused as inputs to train the improved network, and then micro-expression recognition is carried out by Argmax classifier. Experiments on CASME and CASME II data sets have achieved good results, which verify the effectiveness of CA-MobileViT network.