Unsupervised surface defect detection and localization of transparent vacuum blood collection tubes using region-aware reconstruction and expert fusion
Abstract Automated inspection of transparent vacuum blood collection tubes is complicated by specular reflections, heterogeneous endpoint structures and highly elongated regions of interest. An unsupervised region-aware reconstruction framework is developed using native-coordinate patches, three jointly trained reconstruction experts and region-prior-guided soft fusion. A Regional Multiscale Excess Score (RMES) calibrates normal regional residual tails for image-level detection while preserving the original pixel anomaly map. Evaluation on 100 abnormal ROIs and a reserved pool of 145 normal ROIs across three training seeds yielded Image AUROC of 0.9868 ± 0.0027, macro Pixel AP of 0.8358 ± 0.0710 and micro fixed-threshold Pixel F1 of 0.2499 ± 0.0052. Controlled fusion comparisons and same-map scoring experiments further assess the contributions of expert reconstruction and regional anomaly-score calibration.
Authors
- Zhanmin Zhao
- Xuemei Shi
- Bingyu Bai
- Pengxiang Zhao
- Han Wang (ORCID: https://orcid.org/0009-0002-2777-0702)
Institutions
- Chinese Academy of Medical Sciences & Peking Union Medical College (CN)
- Hebei GEO University (CN)
Publication Details
- Journal
- Engineering Research Express
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1088/2631-8695/aec35f
- Primary Topic
- Industrial Vision Systems and Defect Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00