Human-in-the-Loop Anomaly Detection Bridges Benchmark-to-Deployment Gap

While automated anomaly detection models often show strong performance on academic benchmarks, their effectiveness in real-world industrial settings, especially with challenging data like reflective surfaces and subtle defects, is often inconsistent. A study evaluating 19 models on a manufacturing dataset found that performance was highly sensitive to preprocessing and varied significantly across conditions, with no single model proving universally robust.Motivated by these findings, the researchers developed and deployed a unified human-in-the-loop framework for manufactured-part inspection. This system replaces a manual visual inspection workflow by integrating AI-assisted defect detection with human annotation and validation.The framework supports heatmap-guided defect review, allows inspectors to refine candidate regions using tools like SAM, and maintains a review history for consistency and onboarding. This practical approach highlights the crucial gap between theoretical benchmark performance and the realities of industrial deployment, offering a robust solution for real-world anomaly detection.