Dynamic Self-Distillation Former for Weakly Supervised Semantic Segmentation

To address the issues of pseudo-label noise and overfitting in weakly supervised semantic segmentation (WSSS), this paper proposes a Dynamic Self-Distillation Former (DSDF-WSSS) to improve model stability and segmentation accuracy under a Transformer architecture. DSDF-WSSS adopts a progressive gradient-free teacher parameter evolution (PGTPE) mechanism and a dual-branch pixel-level distillation loss for robustness against noise and better boundary modeling. PGTPE builds a stable teacher model via progressive parameter fusion and delayed distillation to guide student learning. The pixel-level distillation loss offers fine-grained supervision on category distribution and mask structure, improving semantic consistency and structural alignment. On PASCAL VOC 2012 and MS COCO 2014 datasets, DSDF-WSSS achieves mIoU of 82.75% and 56.51%, achieving state-of-the-art performance under low-resource requirements.