Resizer-Integrated RepGFPN-Based Multi-Scale Feature Fusion Network with Perturbation Training for Micro-Expression Recognition

Automatic micro—expression recognition holds significant applications in multiple domains. However, its development is impeded by the transitory and subtle characteristics of facial muscle movements. Additionally, the high level of expertise demanded for the collection and annotation of micro—expression data leads to datasets that are small in scale and imbalanced. This paper proposes RRMFNet + PT, a micro—expression recognition approach that combines a resizer—enhanced RepGFPN—based multi—scale feature fusion network with perturbation training. Firstly, an IncepResizer module adaptively modifies the sizes of input images. Subsequently, a lightweight backbone network extracts multi—scale features. Next, the RMFNet module facilitates multi—scale feature interaction and fusion, followed by an MLP classifier. Finally, a perturbation training (PT) strategy alleviates the small—sample problem. On the MEGC2019 composite dataset (three—class verification), RRMFNet + PT attains a UF1 of 0.8916 and a UAR of 0.8913, demonstrating robust performance with effective feature extraction under limited data conditions.