S3DNet: Support-Driven Spatial–Semantic Decoupling for Few-Shot Strip-Steel Surface Defect Segmentation

Abstract

Few-shot strip-steel surface defect segmentation requires a model to identify unseen defect categories and delineate their precise boundaries from only a few annotated support images. This remains challenging because strip-steel defects are often small, elongated, low-contrast, and easily confused with repetitive background textures. To address these challenges, we propose the Support-Driven Spatial–Semantic Decoupling Network (S3DNet), which preserves a frozen DINOv3 encoder while explicitly separating support-conditioned semantic adaptation from fine-grained spatial refinement. Specifically, a low-rank prototype adapter transforms the masked support prototype into an episode-conditioned semantic update of the query tokens, enabling the frozen representation to adapt to the defect category specified by the current support set. In parallel, a lightweight spatial adapter extracts multiscale local cues for recovering weak boundaries and thin structures. A spatial–semantic interaction module subsequently uses support–query semantic similarity to selectively gate these spatial cues before progressive decoding, thereby suppressing irrelevant texture responses while retaining defect-relevant details. Experiments on FSSD-12 demonstrate that S3DNet consistently outperforms representative CNN-based few-shot segmentation methods under both 1-shot and 5-shot settings, while the lightweight DINOv3-Small configuration achieves a favorable accuracy–efficiency trade-off. Ablation and qualitative analyses further validate the effectiveness and complementarity of the proposed semantic adaptation, spatial refinement, and interaction mechanisms.