Abstract: In facial emotion recognition,noisy data caused by poor image quality often degrades recognition accuracy,while limited sample sizes hinder the ability of conventional deep learning models to distinguish noisy from clean facial features.To address these challenges,this paper proposes a novel framework,CAFSC,which integrates an adaptive channel attention strategy with a local-global collaborative mechanism to enhance recognition performance.A noise-robust data augmentation strategy is first introduced,combining Gaussian blur,perspective transformation,and color perturbation with image stitching,flipping,and rotation.This not only preserves subtle facial expression cues but also improves image clarity,dataset diversity,and model robustness.It further designs a Channel Attention Module with Adaptive Channel Reordering(CAM-ACR) that reorders channel features,followed by grouped convolution and concatenation,to capture multi-dimensional local semantics.A local-global feature enhancement mechanism is then employed,where local features guide global feature extraction to strengthen the representation of complex emotional patterns and contextual information.Finally,an improved cross-attention fusion module achieves bidirectional interaction and collaborative enhancement between global and local features.Experimental results show that CAFSC achieves accuracies of 91.21% on RAF-DB,98.31% on CK+,75.54% on FER2013,and 86.74% on FER2013PLUS,demonstrating superior lear-ning efficiency and convergence stability compared to existing methods.

Key words: Facial expression recognition, Local features, Global features, Attention mechanism, Anti-jamming