Industrial integrated circuit (IC) chip defect inspection is hindered by diverse topologies, micro-scale structures, and the insufficient precision of traditional vision paradigms. While the Segment Anything Model (SAM) offers strong zero-shot capabilities, its heavy reliance on manual prompts and lack of domain adaptability limit its viability in automated production lines. This paper proposes IC-SAM, a highly automated framework tailored for electronic manufacturing quality control. IC-SAM synergistically integrates three core modules: Process Prior Knowledge (PPK), which embeds semiconductor domain constraints to suppress background noise; Self-Driven Semantic Prompting, which leverages CLIP to align visual features with process descriptions for autonomous target localization; and Global Feature Fusion (GFF), which optimizes boundary localization through multi-scale interaction. Extensive experiments demonstrate that IC-SAM outperforms baseline models by approximately 15% in both mIoU and mBIoU across SIC, CIC, and IGBT datasets. The framework achieves 9.6 FPS under the tested 1024 × 1024 input setting while introducing only 2.3 M learnable parameters, indicating a parameter-efficient adaptation strategy with quantified computational cost for precision IC inspection.
Deng et al. (Fri,) studied this question.