Learning a Reliable Graph Model for Micro-Expression Recognition
作者
Jiateng Liu,Hengcan Shi,Xinlan Xu,Kerong Li,Xiaolin Xu,Yaonan Wang
标识
DOI:10.1109/aihcir67580.2025.11405182
摘要
Exploring the intrinsic relationships in the irregular spatial appearances and temporal dynamics of micro-expressions (MEs) has been an effective way to improve the performance of micro-expression recognition (MER). However, most of the existing MER methods work in a deterministic manner, which is prone to being influenced by the limited available ME data and tends to be trapped in an over-fitting representation. To fully explore the reliable latent information of MEs, we propose a novel reliable graph learning model (RGLM) that models spatial-temporal dependencies probabilistically, providing a more robust and dependable representation. For this purpose, we construct an adaptive spatial-temporal graph convolution network to model the underlying relationships between spatial appearances and temporal dynamics of MEs, employing two probabilistic masks following a multivariate Bernoulli distribution to characterize the discrete relations among MEs further. Moreover, we introduce an uncertainty-aware feature fusion module to aggregate the refined spatial and temporal features into a more reliable and robust representation. Finally, a unique alternative training strategy is used to optimize both the network and the undifferentiable mask structures simultaneously. Extensive experiments across several widely used benchmarks show that our RGLM achieves better performance than state-of-the-art MER methods.