Research on precision industrial defect detection assisted by vision foundation models

Abstract

Industrial defect inspection needs to handle micro-scale objects, weak contrast, irregular boundaries and frequent domain shifts. Conventional closed-set detectors are also dependent on category-specific annotations. VFM-PDI is a new high precision inspection method based on a vision foundation model. VFM-PDI is composed of a frozen visual encoder and a defect semantic prompt bank, a normal-prototype adapter, cross-scale patch alignment, boundary-aware mask refinement, and uncertainty-calibrated decision logic. Separately extract transferable visual knowledge and lightweight industrial adaptation. Semantic prompts limit the meaning of defects, normal prototypes suppress acceptable texture variation, and scale-aware fusion retains both hairline and structural defects. Tests on surface, casting, ceramic and printed-circuit-board defects produced an image-level AUROC of 99.2%, a pixel-level AUROC of 98.6%, anAUPRO of 96.8%, a small-defect F1 of 88.9% and an average latency of 28.6ms per image. Based on the above results, foundation-model semantics can be used effectively with industrial priors, boundary constraints and calibrated confidence.