Dual Similarity and Progressive Multi-Scale Feature Enhancement for Few-Shot Surface Defect Segmentation

Accurate surface defect segmentation is critical for product quality and production efficiency. However, Few-Shot Semantic Segmentation (FSS), which relies on a small number of defect samples for training, faces challenges when dealing with defects that vary in scale, have unclear boundaries, or resemble the background. This happens because the small number of defect samples used for training cannot fully represent all defect regions, leading to poor generalization. To address these challenges, we propose a method combining dual similarity and progressive multi-scale feature enhancement for few-shot surface defect segmentation. Our approach includes a Dual Similarity Module (DSM), which improves the model’s ability to distinguish defects from the background by computing feature similarities. Additionally, the Multi-Scale Feature Fusion (MSFF) module enhances segmentation accuracy for defects of various sizes through feature up-sampling and fusion. Finally, the Multi-Scale Cross Attention (MSCA) mechanism distinguishes defects with blurred edges from the background by associating multi-scale features. Experiments conducted on the MSD, Magnetic Tile Surface Defect and MVTec-FS datasets demonstrate the effectiveness of the proposed method. Compared to the baseline, it achieves average mIoU improvements of 15.7% (1-shot) and 11.0% (5-shot), and FB-IoU improvements of 4.0% (1-shot) and 2.2% (5-shot) on the MSD dataset. On the Magnetic Tile dataset, the method yields relative mIoU gains of 15.2% (1-shot) and 12.7% (5-shot), along with FB-IoU gains of 14.8% and 11.6%, respectively. On the MVTec-FS dataset, it outperforms the baseline with mIoU improvements of 6.8% (1-shot) and 7.1% (5-shot), and FB-IoU improvements of 4.1% and 2.8%, respectively, validating its robustness across diverse surface defect types.