PCBClip: Vision-Language Defect Detection Model for Low-Sample Inspection Systems

This paper addresses critical challenges in implementing vision-language models to industrial visual inspection systems, where domain-specific terminology, limited labeled data, and annotation noise impair cross-modal alignment and detection performance. We propose PCBClip, tailored for printed circuit board (PCB) defect detection with three key innovations: (1) Anchors by Patches (ABP), a transformer-compatible region proposal method that eliminates auxiliary structures while maintaining localization precision, particularly robust to coarse-grained industrial annotations; (2) Semantic Bridging Prompt (SBP) systematically connects domain terminology to open-domain visual knowledge, enabling interpretable weak supervision; and (3) Antithetical Contextual Learning (ACL) treats defect-free samples as negative constraints for robust normal-state learning. Experiments demonstrate superior data efficiency (94.1% AP50 with 10% training data vs. 83.24% baseline), training efficiency, and competitive real-time inference. While SBP requires expert input, results establish PCBClip as a practical model for industrial defect detection with limited data.