Abstract
In the research on intelligent perception, dynamic emotion recognition has been the focus in recent years. Small samples and unbalanced data are the main reasons for the low recognition accuracy of current technologies. Inspired by circular convolution networks, this paper innovatively proposes an adaptive dynamic micro-expression recognition algorithm based on self-supervised learning, namely MADV-Net. Firstly, a basic model is pre-trained with accurate tag data, and then an efficient facial motion encoder is used to embed facial coding unit tags. Finally, a cascaded pyramid structure is constructed by the multi-level adaptive dynamic encoder, and the multi-level head perceptron is used as the input into the classification loss function to calculate facial micro-motion features in the dynamic video stream. In this study, a large number of experiments were carried out on the open-source datasets SMIC, CASME-II, CAS(ME)(2,) and SAMM. Compared with the 13 mainstream SOTA methods, the average recognition accuracy of MADV-Net is 72.87%, 89.94%, 83.32% and 89.53%, respectively. The stable generalization ability of this method is proven, providing a new research paradigm for automatic emotion recognition.