Research questionHow can industrial visual anomaly detection generalize to small, weakly contrasted defects in imperfect factory images?Production images contain structured patterns, permitted design variation, and imaging imperfections, while defects may be small and weakly contrasted. Benchmarks dominated by optimal imaging conditions can overstate how well anomaly detectors will perform in deployment. Latest papersRecent research connected to this question, newest first.ISP-AD: A Large-Scale Real-World Dataset for Advancing Industrial Anomaly Detection with Synthetic and Real DefectsThe evidence concerns industrial screen-printing inspection using a dataset with synthetic and real defects, including factory-floor examples, structured patterns, and separate unsupervised and supervised splits. Reported experiments cover unsupervised anomaly detection and mixed supervised training that adds weakly labeled real defects to synthesized defects, including later integration of emerging real defective samples.research paper · Sep 3, 2026DPA: Decoupling Product-Agnostic Anomaly Representations for Zero-shot Anomaly GenerationThe source concerns image-based industrial anomaly detection using real anomalies from existing products to support unseen target products. It reports experiments on MVTec-AD, VisA, and a dedicated anomaly-transfer benchmark, including downstream zero-shot and few-shot detection.research paper · Sep 2, 2026