Enhancing Facial Expression Recognition through Multi-branch Self-distillation with Collaborative Teacher

Efficient facial expression recognition (FER) in real-time applications demands innovative approaches to balance accuracy and computational cost.Multi-branch architecture-based self-knowledge distillation has emerged as a promising solution, enabling earlystop inference in resource-constrained environments.However, existing vanilla methods primarily force shallow branches to mimic deep branch features, neglecting the potential impact of inappropriate learning gaps between branches.To tackle this challenge, we propose a novel Multi-branch Self-distillation with Collaborative Teacher mechanism (MSCT).MSCT transforms the deep branch into an adaptive collaborative teacher, generating tailored attention weights for each student branch, facilitating more nuanced knowledge transfer.Extensive experiments on FER2013 and FER-Plus datasets demonstrate that integrating MSCT with existing multi-branch self-distillation approaches preserving their complexity while yielding substantial improvements, achieving up to 11.06% relative improvement on the FER2013 private test set and 5.74% on FER-Plus.