A vision system either has a clean, learnable image of the defect or it does not, and that distinction decides whether the project pays back before a single frame is captured. Most proposals for automated inspection start with an accuracy figure. The question that actually matters comes before it: is the defect the client wants caught one a camera, at a fixed station, under fixed lighting, can reliably see. On lines where the answer is yes, the case is strong: a published machine vision system built to catch missing teeth, surface irregularities and dimensional deviation on gears in mass production reports detection accuracy above 98 percent, with precision and recall both above 96 percent. Those are defects with a fixed geometry, inspected from a fixed viewpoint, at a volume that amortises the camera rig in months.
Exhibit · Where it stalls, where it doesn't
The signature decides the return
of ML vision-inspection implementations stay stuck at prototype or pilot scale
detection accuracy on a published gear-inspection system
A clean-signature, high-volume defect — the case it's built for.
The dimensions that actually decide it, ahead of any accuracy claim a vendor quotes, are the defect signature, the product variability, the run volume, and the cost of a miss.
Exhibit · What decides it
Four conditions, ahead of any accuracy claim
| Condition | Points to vision inspection | Points away from it |
|---|---|---|
| Defect signature | Fixed geometry: missing feature, crack, dimensional deviation | Organoleptic: off-odour, texture, or a defect defined by feel |
| Product variability | A manufactured part with a consistent 'normal' to compare against | A natural or agricultural product where normal itself varies |
| Run volume | High and continuous, amortising the camera rig | Low-volume or high-mix, changing over before the model learns |
| Cost of a miss | Bounded and recoverable — a reject, a rework | Severe enough that a probabilistic system needs a human check anyway |
Photograph the actual defect population first, not a handful of staged examples, and check whether they look alike to a human before assuming a model will find them alike. Price the false-reject rate, not only the miss rate, since a system tuned to catch everything will also flag good product and that cost is not always in the original business case. Confirm the mounting and lighting are fixed on the real line, not the test bench, because a camera that worked in a trial under even light frequently fails once it is bolted next to a motor that vibrates or a window that lets in daylight. And ask what happens on a defect the system has never seen, a hard reject to a person, or a guess with a confidence score attached to it, the second one is the failure mode that costs the most later.
A recent review of machine-learning-powered vision inspection across automotive, aerospace, assembly and general manufacturing found that 77 percent of implementations remain stuck at prototype or pilot scale rather than reaching full production. That figure is consistent with what scoping usually finds: most stalled projects were pointed at a defect closer to the organoleptic end of the spectrum, off-odour, texture, a defect defined by feel rather than appearance, or at a run too low-volume to justify the fixture. The commercial question is not whether a model can be trained to see the defect. On enough labelled examples, most defects can be learned to some accuracy. The question is whether that defect has a signature stable enough, at a volume high enough, that automating it beats the line change needed to mount a camera and the cost of the false rejects it will produce. That is an engineering and economics question, and it is answerable before any code is written, which is also why it belongs in the site walk, not in the vendor's slide deck.
Reference
This piece is adapted for Praxora Lab from the original. Originally published at orionfive.ai ›
