Self-supervised extracted contrast network for facial expression recognition

Self-supervised Contrastive learning has recently demonstrated significant performance in Facial Expression Recognition (FER). However, existing methods fail to address inherent challenges such as similar and blurred facial expressions, and low-quality semantic pairs from image cropping. To tackle these challenges, this paper proposes a Lightweight Self-supervised Distilled Contrastive network for FER (LSDC-FER), which consists of a center suppressed cropping module and a multi-scale feature joint distillation module. The former improves the quality of face expression semantic pairs by face image localization and center suppressed cropping. The latter learns fine-grained features using knowledge distillation and explores multi-scale contextual information to improve feature learning. Experimental results show that LSDC-FER achieves 75.65% and 61.16% recognition accuracy on RAF-DB and FER-2013 public facial expression datasets, respectively.