A New Paradigm for Deep Micro-Expression Recognition Under Occluded Conditions

Micro-expression recognition is a pivotal task in facial analysis, focusing on detecting and analyzing subtle, fleeting changes in facial expressions. However, in real-world applications, facial regions are often obscured by objects like glasses, leading to feature loss and a marked decline in recognition accuracy. To overcome this challenge, we introduces a dual-branch neural network model that leverages feature restoration(DBFRNet) for recognizing occluded micro-expression. The first branch leverages the Denoising Diffusion Probabilistic Model to progressively reconstruct occluded facial regions using conditional generation and multi-step sampling, ensuring harmony between the restored and visible regions. The second branch extracts temporal dynamic features of the face using optical flow techniques, while a visual Transformer encoder enhances the spatial feature representation. Through a coordinate attention mechanism, the outputs of both branches are effectively integrated in the feature fusion layer, enhancing the overall performance of micro-expression recognition in scenarios with significant occlusions. Experimental results demonstrate that the proposed method achieves superior performance on multiple simulated occlusion micro-expression datasets, significantly improving both accuracy and robustness.