Micro-expression recognition based on direct learning of graph structure

Micro-expressions are brief and subtle facial movements that can reveal an individual's true emotions. In this paper, we introduce the idea of direct learning of graph structures for feature analysis in micro-expression recognition and propose an Optical Flow-based VisionGNN Network (OFVIG-Net). Specifically, The optical flow map leverages the OnsetFrame and ApexFrame of the micro-expression to extract feature information. VisionGNN treats the optical flow map as a graph, where the image is divided into patches treated as nodes, and a graph is constructed by linking neighboring nodes. Feature analysis is then performed based on this graph structure. Together, feature extraction and feature analysis constitute the micro-expression recognition task. We conducted a series of rigorous experiments, and the results show that OFVIG-Net achieves an accuracy of 66.32% on the composite micro-expression databases (SMIC, SAMM, CASMEII). Additionally, it attains the current state-of-the-art (SOTA) accuracy of 88.75% on the CASME3 micro-expression database.