Micro-expression Recognition Based on Optical Flow Maps and Attention Mechanism

To address the challenge of detecting subtle changes in local regions during micro-expression occurrence, this paper proposes a micro-expression recognition method based on optical flow maps and attention mechanisms. First, the frame difference method is used to identify the apexframe in micro-expression samples that lack apexframe labels. Next, the TV-L1 optical flow method is employed to extract the horizontal and vertical components of the optical flow between the onsetframe and the apexframe, and these optical flow component maps are then used to derive the corresponding optical flow strain pattern maps. Subsequently, the three optical flow component maps are stacked together to form the optical flow feature map of the micro-expression. Finally, an attention module is introduced into a convolutional neural network built on the Inception module, allowing the model to focus more on the subtle changes in the motion areas when the micro-expression occurs during the feature extraction process. Experimental results show that the proposed method achieves an UF1 score of 0.860 and a UAR score of 0.854 on the CASME II dataset, surpassing many existing methods in terms of UF1 and UAR scores on other datasets as well.