Micro-expression recognition algorithm based on improved repVGG

This paper addresses the challenges of micro-expression recognition, which involve subtle facial changes, short durations, and high sensitivity to noise. We propose an enhanced RepVGG network that integrates a polarized self-attention mechanism and adversarial training. During the training phase, the multi-branch architecture of RepVGG strengthens its feature representation, while the polarized self-attention module captures fine-grained local muscle movements and global associations along both channel and spatial dimensions. In addition, we adopt the Projected Gradient Descent (PGD) adversarial training strategy, which applies small perturbations to the input data to simulate real-world noise such as lighting or head movement, thus ensuring the model retains strong recognition capabilities under adversarial conditions. Experiments conducted on the CASME II and SAMM datasets demonstrate that the proposed approach significantly outperforms baseline methods in terms of accuracy, precision, recall, and F1 score, with peak accuracy exceeding 95%. Meanwhile, only a modest increase in network parameters and computational overhead is incurred, underscoring its deployment efficiency. Overall, our method achieves a favorable balance between fine-scale micro-expression analysis and noise robustness, offering a promising technical solution for complex emotion recognition scenarios.