Complementary feature fusion network for few-shot segmentation of steel surface defect

Steel surface defect segmentation plays a critical role in improving production efficiency and ensuring product quality traceability. However, existing segmentation networks, which depend on large amounts of labeled samples for training, still face challenges in segmenting certain sparse defect types. To address this, a multi-source fusion framework is proposed in this work for few-shot segmentation of steel surface defects. This framework proposes a multi-source complementary information extraction module that integrates global-local semantics and cross-region interactions between support and query features, enabling accurate capture of complex structures and variations in images. Meanwhile, a multi-scale spatial-channel attention module is introduced to highlight foreground defect semantics while suppressing irrelevant background noise in support features. Finally, a multi-source information fusion module is proposed to consolidate these complementary features with the query feature and the support prototype for generating a comprehensive defect representation. Additionally, a support decoder is integrated into the framework to generate the auxiliary support mask prediction, while a dual-loss training strategy is employed to bridge the gap between query and support features learning. Comparative experiments against state-of-the-art methods on the FSSD-12 dataset demonstrate that our framework achieves the best segmentation performance, outperforming the second-best model by 2.9% (1-shot) and 3.4% (5-shot) in mean intersection over union (mIoU), with corresponding foreground-background-IoU (FB-IoU) gains of 1.3% and 2.5%. Meanwhile, ablation studies validate the synergistic contributions of our proposed modules, showing that our full model surpasses the baseline by 14.2%/11.9% in mIoU and 12.8%/7.2% in FB-IoU improvement for 1-shot/5-shot settings, respectively.