Abstract:Knowledge distillation enhances student model performance by transferring knowledge from the teacher model. However, for lightweight student models, it is challenging to fully absorb the implicit knowledge contained in the teacher feature maps. To address this, this study proposes a spatial-channel attention refinement-based collaborative distillation (SCAR-KD) method, which extracts key discriminative information from the original feature maps to mitigate the semantic discrepancy between the student and teacher. Specifically, we adopt a multi-scale spatial-channel attention (SCSA) module to refine discriminative attention-enhanced features from the teacher feature maps along both the channel and spatial dimensions, while decoupling the spatial attention map to weight the original feature maps for dynamic distillation. This approach achieves dual knowledge transfer. Experimental results demonstrate that the mAP@0.5:0.95 of YOLOv8n-SCAR-KD model improved from 64.1% to 65.3% on the VOC dataset and from 20.7% to 21.6% on the VisDrone dataset, outperforming current mainstream distillation methods and validating the effectiveness of the proposed method.