Selective Re-Inference for PCB Defect Detection via Image-Specific Template-Difference Density

摘要

Small and subtle printed circuit board (PCB) defects can be missed by direct full-image detection, whereas dense slicing improves local resolution at the cost of additional detector views and more false positives. We propose a selective re-inference framework that uses image-specific template-difference density to allocate a single additional local detector pass. Given an inspected image and its paired defect-free template, the method computes a template-difference map, aggregates the difference responses over candidate 512×512 windows, selects the highest-density region of interest (ROI), applies the same detector to this crop, and fuses the remapped local detections with the full-image detections. On a locked pair-level split of the DeepPCB dataset, the proposed method improves F1 from 0.9453 to 0.9535 relative to direct full-image detection. Under the same two-view detector-inference budget, it achieves a higher F1 score and fewer false positives than random, grid, and mismatched-difference ROI controls. These results indicate that the gain is not attributable solely to adding a local crop, but to placing that crop using the image-specific paired template difference. The method is best interpreted as a recall-oriented selective re-inference strategy for paired PCB inspection.