Cross-Database Micro-Expression Recognition via Domain-Collaborative Disentanglement Network

Abstract

Micro-expression recognition (MER) has made significant advancements in recent years. Most methods focus on individual-database task, which poses significant challenges when attempting to generalize to complex scenarios. To address this limitation, cross-database micro-expression recognition (CDMER) has garnered increasing attention due to its robustness and generalizability. This work presents a domain-collaborative disentanglement network (DCDN) that leverages domain collaboration to facilitate learning domain-invariant and discriminative emotion representations. DCDN employs a dual-branch disentanglement mechanism to explicitly decouple emotion and domain information. On this basis, a domain information filtering module (DIFM) based on graph neural network adopts multi-granularity and feature relation learning, which can capture potential relations to remove the domain component from the emotional features. Furthermore, a disentanglement regularization module (DRM) introduces information-theoretic constraints to enhance the completeness and robustness of the disentanglement. The experimental results demonstrate that DCDN achieves state-of-the-art performance on twelve CDMER tasks, validating the effectiveness of each component and the overall disentanglement strategy.