Advancing Micro-Expression Recognition: a Task-Specific Framework Integrating Frequency Analysis and Structural Embedding

Micro-expression recognition (MER) remains a challenging task due to the subtlety and low intensity of facial motion patterns. Existing deep models often struggle with spectral bias and structural mismatches when applied to MER, leading to notable performance degradation. To tackle these issues, we propose a task-specific framework named Adaptive Attenuation Compensated Polarity Network (A2CPN) that introduces two key innovations. First, we conduct a frequency analysis paradigm to reveal the degradation of high-frequency motion cues in deep networks. To mitigate this, we propose a layer-wise compensation method to preserve decision-relevant cues. Specifically, frequency and spatial-domain features are projected into a shared wavelet feature space in a band-wise manner, while the hierarchical compensation mechanism establishes a mapping between frequency specific attenuation levels and network layer indices to adaptively enhance ME motion cues. Second, we embed facial structural priors via a polar coordinate system. It constructs correlations among radial cells, aligning the model's representations with radial motion-distribution patterns. Extensive experiments on four benchmark datasets, CASME II, SAMM, SMIC, and MEGC2019, demonstrate state-of-the-art (SOTA) performance, validating the effectiveness of our frequency-domain and radial analysis paradigms, offering new insights for MER research.