Decoupled Candidate Detection and Multi-Evidence Verification for Tiny Surface Defect Inspection

Tiny surface defects are difficult to inspect because of their small scale, weak contrast, diverse morphology, and similarity to normal textures. To address the conflict between early candidate retention and false-positive suppression, this study proposes a coarse-to-fine inspection framework that separates candidate detection from final verification. You Only Look Once (YOLO) 26s first generates defect candidates, which are locally reconstructed and analyzed by a region-of-interest (ROI)-guided Multi-Task SAM (ROI-MT-SAM) for joint segmentation, semantic prediction, and mask-quality estimation. Multi-Evidence Verification (MEV) then integrates complementary confidence and spatial cues to refine the final decision, while residual-transfer augmentation improves defect diversity under limited-data conditions. Experiments show that ROI-MT-SAM improves mean Intersection over Union (mIoU) from 0.3918 to 0.8869 compared with whole-image SAM, with particularly substantial gains for small defects. MEV further reduces false positives from 245 to 46 and achieves a precision of 0.9565 and an F1-score of 0.9241. These results demonstrate the potential of combining candidate-centered representation with multi-evidence verification for tiny-defect inspection of automotive interior surfaces under the evaluated conditions.