VISYNEXTalk to an engineer
Guide

Anomaly detection vs supervised defect detection, from one photograph

Anomaly detection learns what a good part looks like and flags what departs from it, usually without naming the defect; supervised defect detection names each defect and puts a box on it, and needs labelled examples of each class. Two common ways to teach a camera to find defects differ in what they need from you, and in what they tell you back. This page sets them side by side, then says where synthetic defect images made from one photograph of a good part sit.

What Visynex isVisynex makes synthetic defect images for training machine vision inspection models. Each image is made from one photograph of a good part, and every defect image ships only if a separate detection pass finds the defect ordered.

The two approaches, in general terms

These are general descriptions, not measurements of ours. Results vary a great deal by part, camera and method.

  • Anomaly detection is usually trained on images of good parts only. It learns what normal looks like and flags what departs from it, often as a score or a heat map. It usually does not name the defect.
  • Supervised defect detection (for example a YOLO detector) is trained on labelled examples of each defect class. It returns a named class and a box for each defect it finds. It usually needs examples of every class you want it to name.

When each one tends to fit

Again general guidance, and a starting point rather than a rule:

  • Anomaly detection tends to fit when defects are rare or unpredictable, when you mainly need pass or fail, and when you have many images of good parts but few of bad ones.
  • Supervised detection tends to fit when the line has to say which defect it is and where, for sorting, rework or root-cause counts, and when you can get labelled examples of each class.

The usual obstacle to supervised detection is the second half: collecting and labelling enough real defects.

Where one-photo synthetic data sits

The Visynex Synthesis Engine is aimed at that obstacle. It makes labelled images for supervised detection without needing photographs of real defects.

  • The input is one photograph of a good part. No defect sample goes into it.
  • The output is a supervised dataset: images with a box per defect, in YOLO and COCO format, plus defect-free images with empty labels.
  • The defect classes come from the manufacturing process: the one you tell us, or, if you’re not sure, the one inferred from that photograph, rather than a generic list.
  • Each defect label means the defect you ordered was detected in that image by a detection pass on the finished image. It does not mean the image was searched for other changes.

So, in the terms above: one-photo synthetic data does not replace anomaly detection’s “good parts only” input. It uses a good-part photograph to produce the labelled defect examples that supervised detection would otherwise need from the line.

Reference image of a cast and machined bearing housing, no defect
The reference image
Casting Porosity/Voids on a cast and machined bearing housing, synthetic defect image
Casting Porosity/Voids, synthetic, from one reference image
For the engineer

Try it on your own part

Sign in with your work email, upload one photograph of a good part, and get up to 10 images of it back, labelled in YOLO and COCO. Each takes about a minute. Free, once per email address.

Try it free
For whoever signs off

Talk to an engineer

Thirty minutes with someone who built the pipeline. No upload, no account.

Book a call