Micro-expression Recognition Based on MobileVit-SE Block

Micro-expressions are important information that can understand a person's deepest emotions within their heart and can clearly reflect human real emotions and mental states. They have excellent application prospects in various fields such as medicine, criminal interrogation, and public security. The amplitude of facial movements in micro-expressions is small and their duration is short, making recognition more difficult compared to macros expressions. To identify micro-expressions and execute microexpression recognition tasks on portable devices, this paper optimizes the MobileVit model for visual recognition on mobile devices and adds a MobileVit Block based on the channel attention mechanism (SENet), which is hereinafter referred to as the MobileVit-SE Block. The optimized model replaces the MobileVit Block after the second MV2 down-sampling and the fourth MV2 down-sampling in the original model. Experimental results show that the optimized model achieves an accuracy of 0.817 on the fusion dataset of CASMEII, SMIC, and SAMM, which is slightly lower than the accuracy of 0.82 achieved by directly using the MobileVit model for recognition. However, the processing speed is doubled, meeting the requirements of application scenarios requiring fast detection and recognition of micro-expressions.