Human-in-Loop Anomaly Detection Boosts Factory AI Accuracy.

Deploying anomaly detectors in environments with scarce training data, such as new production lines, presents a significant challenge. A novel training-free, human-in-the-loop framework has been developed to address this, enabling domain experts to enhance a PatchCore anomaly detector without requiring retraining, gradients, or access to original training data. The core of this approach involves direct editing of the detector's memory bank. When a false positive occurs, the expert can correct it by inserting the reviewed image's normal patches into the memory bank. A self-calibrating novelty gate ensures that only patches sufficiently distinct from existing "normal" samples are admitted. Starting with a memory bank built from as few as ten "golden" samples, operator corrections closed a median 66% of the performance gap to a fully trained, uncorrected bank, and even surpassed it in some categories. This method significantly improved 12 out of 15 MVTec AD categories without harming any. For already-trained banks, the gains were concentrated where the bank initially undersampled normal appearances. The evaluation used a robust held-out protocol to prevent inflated metrics from memorization, and results showed that gains are attributable to deployment-time label production at 43% of the cost of exhaustive review.