Answers are grounded in this paper and cite their sources.
In industrial quality inspection, interactive segmentation provides an efficient semiautomatic annotation paradigm, alleviating the burden of pixelwise labeling. However, existing interactive methods often struggle to integrate prompts into deep feature reasoning, leading to insufficient defect structural understanding. To address this limitation, we propose surface-defect segment anything model (SD-SAM), an interactive framework that establishes a prompt-aware feature recomposition paradigm. Specifically, we design an interactive aggregation encoder (IAE) to explicitly couple sparse prompts with dense visual features. IAE first orchestrates hierarchical interactions between prompt embeddings and visual features, enabling prompt-guided feature updating that propagates interaction cues into visual representations. Furthermore, IAE performs adaptive correlation regulation to align feature responses with prompt intent, amplifying defect-relevant activations while mitigating ambiguities from homogeneous backgrounds. Subsequently, structure-aware dynamic decoder performs dynamic feature recalibration and global relational reasoning to resolve semantic inconsistencies across feature layers. Combined with high-level semantic priors to guide fine-grained spatial refinement, it reinforces the spatial continuity of defect regions. Our SD-SAM achieves state-of-the-art performance on three datasets, and its effectiveness has been further validated on an aero-engine inspection platform, demonstrating strong practical applicability.
No takes yet. Share an insight, caveat, or question.
Meng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: