Micro-Expression Recognition based on TV-L1 based Hierarchical Transformer Network

Accurate human emotion and micro-expression interpretation is essential to enhance human-computer interaction, security system improvement, and supporting psychological studies. Micro-expression analysis enables the system to identify subtle non-verbal behavior, thus enabling empathetic and intuitive communication in security, customer service, and mental illness detection. This paper suggests a novel approach of micro-expression recognition by integrating a Hierarchical Transformer Network (HTNet) and the TV-L1 optical flow method. HTNet hierarchical transformer network efficiently extracts and calculates temporal feature information, while TV-L1 algorithm is stable in identifying more complex facial motion patterns needed for micro-expression analysis. The method is also improved by the hyperparameter optimization of the method's approach as well as the application of accuracy, Unweighted F1-score, Unweighted Average Recall, Matthews Correlation Coefficient, and Expected Calibration Error as measures for evaluation. Experiments on dataset-- SAMM, SMIC and CASME II databases demonstrate the high accuracy, real-time computation capacity, and capability of the model to learn in dynamic real-world environments. The research bridges the existing gap between current state-of-the-art AI models and real-world emotion recognition systems and offers profound insights into human emotional intricacies. The study highlights the paradigm-reversal value of AI in sensing and responding to human emotions and lays the foundation for even further development of human-focused solutions across many disciplines.