Visual inspection is one of the oldest and most successful applications of computer vision, and it remains one where the engineering effort sits almost entirely outside the model.
Imaging determines the ceiling
If a defect is not visible in the captured image, no model will find it. Before any modelling work, fix the physical setup:
Lighting. The most important variable by a wide margin. Different defects need different illumination — surface scratches show under grazing light, subsurface inclusions need backlighting, colour variation needs consistent spectrum and no ambient contamination. Getting this right routinely turns an impossible problem into an easy one.
Camera and optics. Resolution must resolve the smallest defect you care about across several pixels, not one. Depth of field must cover part variation. Shutter speed must freeze line motion.
Fixturing and repeatability. A part that presents at a consistent position and orientation reduces the problem enormously. Every degree of freedom you remove mechanically is a degree of freedom the model no longer has to learn.
Teams routinely spend months on model architecture when two weeks on lighting and a fixture would have solved it. Photograph 50 known-defective parts and see whether a human can spot the defect in the image. If they cannot, stop and fix the imaging.
The defect rarity problem
Manufacturing quality is a class imbalance problem in the extreme. At a 0.3% scrap rate, 10,000 parts yields 30 defects — spread across a dozen defect types, which is far too few of each to train or validate a classifier.
Three responses, in order of practicality:
- Anomaly detection instead of classification. Train only on good parts and flag deviation. You have unlimited good examples. It detects novel defects you have never seen, which classification cannot. The tradeoff is that it tells you something is wrong, not what.
- Targeted defect collection. Deliberately retain and photograph every defect found, and where safe, manufacture known defects intentionally.
- Synthetic augmentation. Useful for geometric and lighting variation; less convincing for genuinely novel defect morphology. Validate on real defects regardless.
Setting the operating point
Inspection has two error modes with very different costs. Escapes — defective parts passed — reach the customer. False rejects — good parts scrapped — cost yield.
You must decide the ratio explicitly, because the model cannot. In safety-critical work, escapes are unacceptable and you accept substantial false rejects. In high-volume low-margin work, excessive false rejects destroy the business case.
The usual production pattern is a three-way decision: pass, fail, and review, with the uncertain band routed to a human. This preserves throughput while keeping escape rate near zero, and the review queue doubles as your ongoing training data source.
Deployment realities
Inference runs at the line, not in the cloud. Line speed and network reliability demand local compute. This constrains model size and is a design input from day one.
Drift is physical. A camera vibrates loose, a lens accumulates dust, a bulb ages and shifts colour temperature, a supplier changes material finish. Monitor input statistics, not just output accuracy — the image distribution shifting is your earliest warning.
Operators must trust it. A system that cries wolf gets bypassed within a week. Show operators why a part failed, with the region highlighted, and give them a straightforward override that gets logged and fed back.
Frequently asked questions
How accurate is AI visual inspection?
On well-imaged, well-defined defects, detection rates above 99% with low false-reject rates are routine and often exceed human inspectors, who fatigue. On subtle, variable, or poorly-lit defects, performance can be worse than a skilled human. Accuracy is a property of the imaging setup as much as the model.
How many defect images do we need?
For supervised classification, roughly 100+ examples per defect type is a workable starting point, with several hundred preferred. If you cannot reach that — which is common — use anomaly detection trained on good parts instead, and treat defect images as validation data rather than training data.
Can this replace human inspectors?
It replaces sustained-attention inspection, which humans do poorly over a shift. It does not replace judgement on borderline cosmetic calls or root-cause investigation. Most deployments retain inspectors on the review queue and on process improvement, where they are far more valuable.
What about parts with acceptable cosmetic variation?
This is the hardest category, because "acceptable" is a judgement that varies between inspectors and often is not written down. Before automating, you generally have to force the organization to define the standard — and that alone tends to improve consistency, whether or not you ship the model.
Do we need to modify the production line?
Usually some modification, yes — mounting, lighting enclosure, a reject mechanism, and a trigger signal. This is often the longest-lead item in the project and should be scoped early, because it involves maintenance, safety review, and line downtime.