Holistic foreground–background guided feature learning for few-shot strip steel defect segmentation

Strip steel surface defect (S 3 D) segmentation has become an indispensable technique for surface inspection in strip steel production lines. However, the imbalance between the foreground and background, as well as the complex defect appearances, have become the main challenges in S 3 D segmentation. Moreover, existing methods that adopt background suppression often fail to effectively address these challenges. Motivated by this, we propose a novel foreground–background guided multiscale feature aggregation network (FB-MFAN). It can effectively integrate multiscale features from both the foreground and background information and handle complex defect classes. Our network incorporates two novel modules: Foreground–Background Decoupling Module (FBDM) and Dual Attention Module (DAM). The FBDM initializes a learnable global background prototype to decouple foreground and background features, enabling effective utilization of background cues to guide foreground feature learning and suppress background interference. From the perspective of foreground–background modeling, the DAM introduces a non-local operation based on cosine similarity to capture the long-range dependencies between defect and normal regions, thereby strengthening contextual interactions and achieving precise defect localization and segmentation. Extensive experiments on the FSSD-12 and Surface Defects-4i datasets demonstrate the superiority of our network.