Micro-Expression Recognition Based on Multi-task Learning and Resnet18

Micro-Expression recognition (MER) based on deep learning has broad research prospects. Under the background that the current mainstream micro-expression recognition (MER) algorithms mostly use single-task learning models, to achieve an effective breakthrough in model performance, we apply multi-task learning to micro-expression recognition (MER). This paper proposes an attention model based on multi-task learning, which consists of a single ResNet18 network as the shared network, and an attention module is inserted among the intermediate layers of the network to improve the ability of image feature extraction. Specifically, we take the task of facial action unit detection as an auxiliary task, and jointly train the two tasks of micro-expression recognition (MER) and facial action unit detection. We selected two datasets, CASME II and SAMM, and designed microexpression recognition (MER), facial action unit detection, and ablation experiments for performance verification. According to the experimental results, the performance of the multi-task learning model we designed is significantly improved on the two evaluation indicators of UAR (81.97%) and UF1 (81.98%), and it has both adaptability and robustness.