Get started
Work

Enhanced discrepancy-aware visual inspection with a single global threshold for industrial assembly

Year: 2026
Type: article
Abstract: Automated optical inspection is widely used in industrial assembly, but collecting sufficient real defective samples and pixel-level annotations is difficult because critical faults such as missing components, loose fastening, and positional deviations are rare in production. This study proposes an Enhanced Discrepancy-Aware Framework (EDAF) for real-defect-free anomaly inspection. EDAF uses a frozen pre-trained teacher network and a student network trained to reproduce normal feature responses. Synthetic texture perturbations generated from normal images are used only as auxiliary pseudo-supervision for localization learning, not as substitutes for real assembly defects. Multi-scale teacher–student discrepancy maps are adaptively fused and refined with non-local contextual modeling, and a decoder converts the resulting anomaly representation into pixel-level masks. For image-level OK/NG classification, kernel density estimation likelihood models the normal anomaly-score distribution and enables a unified global threshold across product types. On MVTec AD, EDAF achieved higher pixel-level AUROC than the DAF baseline in 11 of 15 categories. Runtime evaluation on an NVIDIA RTX 4070 Ti SUPER GPU showed a latency of 50 ms per 256 × 256 image, corresponding to approximately 20 FPS. Evaluation on a six-category industrial assembly dataset further supported stable OK/NG separation under a unified threshold. These results indicate that EDAF can support practical industrial inspection requiring both image-level decisions and interpretable defect localization.
DOI: 10.1038/s41598-026-72116-7
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7218459000
FWCI:
Cites:
Cited by:
Related to: