LUMIN: Lifecycle-Utility-aware and Metadata-Informed Network for Visual Anomaly Detection
The LUMIN network improves visual anomaly detection by evolving from UniADet to a more lightweight architecture. It introduces PSP, a novel four-stage sampling pipeline using 18D pixel metadata and visual plugins, which achieves high sampling accuracy with near-random construction costs. This is paired with two engineering optimizations—parallel computation and stratified sampling—to drastically reduce inference memory and evaluation time.