Visual inspection is one of the most practical industrial uses of AI: cameras and models spot defects consistently, at line speed.
Typical Tasks
- Detecting scratches, cracks, dents and contamination.
- Checking assembly: missing parts, wrong orientation.
- Reading labels, codes and serial numbers.
- Measuring dimensions.
Get the Imaging Right
Lighting, camera position, lens and background matter more than the model. Consistent lighting that makes defects visible is the single biggest factor in success.
Collecting Data
Defects are rare, so collecting examples takes time. Save images of every defect found, capture the full range of normal variation, and involve quality engineers in labelling.
Model Approaches
- Classification: good or defective.
- Object detection or segmentation: where the defect is and how big.
- Anomaly detection: trained mostly on good parts, flagging anything unusual — useful when defect examples are scarce.
Running It Reliably
- Measure false rejects (good parts scrapped) and escapes (defects missed); both cost money.
- Route borderline cases to human inspectors.
- Monitor for drift when lighting, materials or products change.
- Retrain as new defect types appear.
Integration
Connect inspection results to production systems so defects trigger actions and trends reveal process problems upstream.