"Defect detection" names four different products: rule-based automated optical inspection (AOI), classical computer vision, AI classification, and AI classification running on edge hardware. Classical CV and rule-based AOI suit stable, high-volume lines where the part and the defect rarely change; AI classification — especially AI plus edge — suits small-batch, high-mix lines where new parts and new defect types arrive faster than anyone can hand-code a rule.
Most RFQs conflate the four, which is why vendors talk past buyers. A procurement team writes "AI defect detection" and means "catch our defects"; three of the four approaches are not AI at all, and the one that is comes in two very different deployment shapes. Treating them as one product is how a plant buys a rule-based system for a high-mix line, then spends a year wondering why every new variant breaks it.
The four approaches, side by side
| Approach | What it is | Handles new/variable defects | Data/setup burden | Best fit |
|---|---|---|---|---|
| Rule-based AOI | Hand-coded thresholds on fixed features | Poorly — every new defect is a new rule | Low data, high engineering per rule | Stable, high-volume, fixed part |
| Classical CV | Algorithmic feature detection (edges, blobs, templates) | Limited — brittle to lighting and variation | Moderate tuning, no training set | Well-controlled, repeatable scenes |
| AI classification | A trained model learns defect classes from images | Well — retrain on new examples | ~1,000 images per class, labelling | High-mix, changing defects |
| AI + edge | Same model, running on-floor hardware | Well, at line speed and sub-second | As above, plus edge integration | High-mix and high-volume together |
The column that reorganises the decision is "handles new/variable defects". If your part and defects are frozen, the cheaper rule-based or classical routes are often the right answer and AI is over-buying. The moment your defect set moves — new SKUs, new materials, cosmetic calls that drift — the rule-based and classical approaches start failing in a way more tuning cannot fix, because they were never built to learn.
Why small-batch and high-volume pull in opposite directions
A small-batch, high-mix plant changes parts constantly, so its binding constraint is adaptation cost: how fast and cheap it is to teach the system a new part or defect. Rule-based AOI punishes exactly that, because each new variant is an engineering ticket. AI classification inverts it — reaching usable accuracy from on the order of 1,000 images per class, retrainable by the plant's own QA team, so a new part is a short training pass, not a project.
A high-volume plant with a frozen part has the opposite constraint: throughput and determinism. Here classical CV or rule-based AOI can be ideal, because a fixed scene with a fixed defect does not need a model that learns. The trap is the plant that is high-volume today and high-mix tomorrow; that plant needs AI plus edge — learning for the mix, edge hardware for the speed — and buying a rule-based system for its current volume mortgages its next product launch.
How to place yourself on the map
Two questions settle most of it. First: does your defect set change — new parts, new materials, new cosmetic standards — on a timescale shorter than a year? If yes, you are in AI territory; if no, classical or rule-based may be cheaper and perfectly adequate. Second: do you need the verdict at full line speed with a hard cycle time? If yes, you are specifically in AI-plus-edge territory, not cloud AI.
The honest boundary: this map is a buyer's framing, not a claim that one approach wins everywhere — the right answer is the one matched to how much your defects move and how fast you need the verdict. The image counts and speed characteristics cited are platform specs on suited tasks; the way to confirm the fit is a sample of your own parts.
Internal links: the machine vision vendor evaluation checklist and our solutions.
Frequently asked questions
What are the four approaches to defect detection? Rule-based automated optical inspection, classical computer vision, AI classification, and AI classification on edge hardware. Only the last two learn from examples; the first two are hand-coded or algorithmic.
Which approach suits a small-batch, high-mix line? AI classification, usually with edge hardware, because its binding constraint is adaptation cost — a new part is a short retraining pass from about 1,000 images, not a new engineering rule.
When is rule-based or classical CV the better buy? On stable, high-volume lines with a fixed part and fixed defect set, where a learning model is over-buying and a deterministic, well-tuned algorithm is cheaper and adequate.
Send us the parts you inspect and how often your defect set changes, and within two weeks we will return which of the four approaches fits your line and where the crossover sits, before any commitment. Start at https://apac.hypernology.net/contact.