No model recovers a defect the camera never captured. In machine-vision inspection the accuracy is won or lost in the lighting, the optics and the exposure — long before a model runs. A flaw with no contrast under the installed light simply is not in the image, and nothing downstream can put it back.
| Lever | What it decides | When it is the answer |
|---|---|---|
| Illumination geometry | Whether the defect has contrast | Surface, dent, scratch, relief |
| Light diffusion | Glare on shiny surfaces | Specular metal, glossy plastic |
| Strobe and exposure | Motion blur at line speed | Fast webs and conveyors |
| Optics and depth of field | Resolution and focus range | Small features, curved parts |
| Structured light | Three-dimensional shape | Height, warp, fill, undercut |
| Filters and colour reference | Contrast and colour stability | Print, coating, shade matching |
The instinct when an inspection underperforms is to reach for a better model. Usually the model is not the problem. If the defect has no contrast under the installed illumination, it is not present in the pixels the camera captured, and no amount of training recovers information that was never recorded. A scratch that is invisible under flat front light is obvious under low-angle light that rakes across the surface; the difference is the lighting geometry, and it is decided before a single image is labelled.
This is why the honest sequence is imaging first, model second. The lighting geometry is chosen for the defect physics — low-angle for surface relief, dome for specular and curved, backlight for dimension, structured light for three-dimensional form. The optics are sized so the smallest feature that matters is resolved across the whole field. The exposure and strobe are short enough that motion at line speed does not blur away the resolution the optics bought. Only once the defect is reliably visible in the image does the model earn its place, and then it is doing the job it is actually good at: deciding, not compensating for a bad picture.
Reversing that order is the most expensive mistake in industrial vision. A project that selects cameras and trains models on a poorly lit scene spends months chasing accuracy that the imaging made impossible, and often concludes the task is too hard. The same task, re-imaged so the defect has contrast, becomes straightforward. A short imaging trial on real defective parts, before any hardware is committed, is what separates a deployment from a pilot that never ships.
Because a model can only work with the information in the image. If the lighting, optics or exposure fail to give the defect contrast, that defect is not present in the captured pixels, and no model — however sophisticated — can detect what was never recorded. The accuracy of an inspection is largely fixed by the imaging before the model ever runs, which is why imaging is treated as the core engineering rather than a setup step.
It changes whether a defect has contrast. The same scratch is invisible under flat front lighting and obvious under low-angle light that grazes the surface and casts the defect into shadow. Dome light tames glare on shiny curved parts; backlight turns a dimensional check into a clean silhouette; structured light reveals three-dimensional shape. Choosing the geometry for the defect's physics is what makes the flaw visible in the first place.
No. This is the most common and expensive misconception in industrial vision. A model can classify or locate what is present in an image, but it cannot restore information the imaging destroyed — a defect lost to glare or erased by motion blur is simply absent from the data. Projects that try to fix imaging with modelling spend months chasing accuracy that better lighting would have made trivial, and the reliable path is to get the image right first.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.