Catching defects on the line with vision AI: what works with few defect images

A classifier needs examples of every defect type. On a good line you may see a handful of scratches, porosity or misaligned clips a week, each looking different. By the time you have enough labelled examples, the process has changed.

The MVTec AD dataset, with 15 object and texture categories of industrial images, became the reference benchmark for this approach ^[1]. PatchCore stores features from defect-free images and flags patches that are far from anything in that memory; it reported image-level AUROC of up to 99.6% on MVTec AD ^[2]. Methods like this also produce a heat map showing where the anomaly is, which is what an inspector needs.

• Control the image: fixed lighting, fixtures and camera position remove most false alarms before any model is trained.
• Respect takt time: inference must finish within the station's cycle, so model size and hardware are chosen from the takt, not the benchmark.
• Plan for variants and changeovers: each part number needs its own reference set or a model that handles variants.
• Keep a person on the borderline cases and use their decisions as labels to improve the model.

IATF 16949 requires automotive suppliers to manage nonconformities and corrective actions ^[3]. An inspection system becomes valuable when each finding is linked to the station, shift, lot and process parameters, so quality engineers can find the cause and update the FMEA. Field problems that escape still surface as recalls, which NHTSA publishes for anyone to search ^[4].

Found this useful?

Share the key finding on X.