AI visual inspection

DefectScope

Spots scratches and dents. No defect photos needed.

No defect labelsPixel heatmapRuns in your browser

Upload a photo of the selected product type: the model only knows that line's normal look. 🔒 Your photo never leaves your device; the AI runs in this page.

Input
Where it looks wrong
Loading model…
Anomaly score / pass-fail threshold

How good is it?

Public MVTec AD test photos the model never saw. One fixed pass/fail threshold per product, set on held-out good photos only.
ProductDefects caughtGood parts wrongly rejected

How it works

  1. Photograph good parts. No defect photos, no labelling.
  2. A pretrained vision network memorises what every small patch of a good part looks like.
  3. Any patch that doesn't match normal lights up. Too far from normal means reject.

For engineers: a from-scratch PatchCore (CVPR 2022) re-implementation, ResNet-18 features, a greedy-coreset memory bank, exported to ONNX. Code and full benchmark.