Micro Expressions (ME) are involuntary facial expressions that reveal a person’s emotions and psychological state and are among the most significant external indicators of these internal states. Experts in fields such as medicine, education, and psychology have used the Facial Action Coding System (FACS) to analyze the relationship between Action Units (AUs) and emotional expressions. Although this system has significantly improved the accuracy of micro expression recognition, it still lacks precision. The challenge arises from the localized and fleeting nature of micro expressions, which makes them difficult to isolate from broader head movements or blinks, and the small size of ME features results in limited database quantities.
This study makes three primary contributions. First, it establishes the Local Temporal Pattern (LTP). When a micro expression occurs, the grayscale values in the key areas change. The study calculates these changes within a 300-millisecond window and finds that when a micro expression occurs in the Regions of Interest (ROIs), an “S-shape” pattern is formed from the start to the peak (see Figure 1). This finding indicates a higher degree of alignment between LTP and micro expressions, further developing a specific pattern that occurs during micro expressions.
Figure 1 – S Pattern
The second contribution involves differentiating broad head movements and blinks. As described earlier, LTP can be identified locally through ROIs. This study also uses a facial integration system to distinguish between micro expressions and other facial expressions, with the S-pattern further explaining similarities among various micro expressions.
The third contribution focuses on improving recognition functionality through filtering S patterns and data augmentation. Due to the small size of micro expression databases and the lack of precise positional labeling, the positioning performance and model accuracy are significantly restricted. This study uses the Hammerstein model to simulate parameters for each S pattern, generating additional data that can be effectively filtered to produce similar S patterns, thus enhancing the reliability of model training.
Previous limitations in micro expression data have led to low recognition accuracy. This study identifies common patterns occurring when micro expressions are generated and uses these patterns to more effectively differentiate between macro and micro expressions, reducing errors introduced by macro expression data. Additionally, data augmentation is used to expand the micro expression dataset, incorporating more common patterns into the model for filtering and training. The detection results on CASMEI and CASMEII indicate that the proposed LTP outperforms existing recognition methods in F1-score (see Figure 2). Further integration systems and data augmentation could enhance positioning performance. This model is expected to improve future academic and practical applications in developing micro expression models, increasing accuracy in micro expression localization and recognition.
Figure 2 – Detection Results