Edge-Aware and Deformable Feature Fusion for Steel Surface Defect Segmentation

Steel surface defect segmentation plays a crucial role in industrial quality inspection and intelligent manufacturing. In this paper, we propose an enhanced semantic segmentation framework based on Segformer to improve the precision and edge awareness of defect localization. Specifically, we introduce a novel adaptive fusion feature module (AF-Fuse) to effectively fuse multi-scale features with adaptive attention. To enhance spatial context representation, atrous spatial pyramid pooling (ASPP) is integrated into the encoder. Furthermore, we adopt deformable convolutional layers to handle irregular defect shapes and incorporate a mid-level edge focus module (EFM) branch to reinforce boundary localization. Extensive experiments on NEU-Seg datasets demonstrate that our improved model outperforms better in terms of mIoU and boundary accuracy, while maintaining computational efficiency. The proposed approach provides a promising solution for high-precision industrial defect segmentation.