Machine vision and inspection

Machine vision has been on production lines for decades, measuring parts, reading codes and checking that components are present. AI changed one part of it: defects that are easy for a person to see but hard to describe with rules, like scratches, dents, stains, uneven welds or cosmetic flaws on molded parts. Deep learning models trained on labeled images can catch many of these where rule-based tools struggled.

The tools here split into two groups. Cognex, Keyence and MVTec HALCON come from industrial vision and run on smart cameras or industrial PCs at the line, combining classic tools with deep learning. LandingLens, Roboflow and Instrumental are newer software platforms built around training models from images, with cloud tools for labeling and review and options to run the model near the line.

Most inspection projects succeed or fail on the setup before any model is trained. Lighting, lens choice, camera position and part presentation decide whether a defect is visible at all. Collect real good and bad parts early, photograph them under production lighting, and agree with your quality team on what counts as a reject. Labels that drift between inspectors produce models that drift too.

Decide in advance how the line will use the result. A model can reject parts automatically, divert them for human review, or simply log defects for analysis. Each choice needs different accuracy, and a false reject costs scrap or rework while a missed defect can reach a customer. Run the model beside your current inspection for a period and compare both on the same parts before you trust it alone.