Machine vision: lighting, optics and why a camera sees less than you think

Lighting geometry is the application

How light strikes a surface determines what appears in the image. Direct frontal illumination flattens texture and flares on gloss. Low-angle grazing light throws scratches, embossing and edges into sharp relief. Backlighting gives clean silhouettes for dimensional work and tells you nothing about the surface. Diffuse domes suppress specular reflection on curved metal. Structured illumination extracts height where flat imaging cannot. Choosing the wrong geometry produces images where the defect is present but marginal, and the resulting inspection then depends on thresholds that drift. Trial the lighting on real parts, including the ones your process makes at its worst, before anything else is specified.

Resolution, field of view and the trade you cannot escape

Covering a wide area and resolving a fine feature pull in opposite directions. A sensor covering a whole panel devotes very few picture elements to a small mark, and detecting something reliably needs it to span several elements, not one. Options are a longer lens with less coverage, a higher-resolution sensor with more data to move and process, or multiple cameras with the calibration and cost that implies. Depth of field is the other constraint: the aperture that gives you tolerance for part height variation also reduces the light reaching the sensor. These trades are physics and cannot be negotiated with the supplier.

Presentation repeatability decides how hard the software has to work

A part that arrives in the same position, the same orientation and the same plane every time allows a simple, fast, robust inspection. A part that arrives anywhere within a wide window forces location-finding, perspective handling and tolerance for scale variation, all of which are solvable and all of which make the result more fragile. Money spent on a fixture or a nest that constrains the part is generally cheaper than money spent on making the vision cope. This is where an inspection borrowed from a supplier demonstration falls over: their sample sat on a jig, and yours arrives on a moving belt.

The factory attacks the image

Ambient light from a roller shutter, a skylight or an overhead fitting swamps a carefully designed illuminator. Dust, oil mist and coolant film the lens and the light, dimming the scene gradually until the inspection quietly changes behaviour. Vibration blurs. Temperature shifts focus. Someone knocks the camera mount with a pallet truck. Enclose the station where you can, mount rigidly and independently of vibrating structure, put lens and illuminator cleaning on the maintenance routine, and give the system a way to report that its own image quality has degraded rather than letting it silently pass everything or fail everything.

Being honest about what imaging can judge

Cameras are excellent at presence, position, orientation, dimension in a known plane, code reading and colour comparison under fixed illumination. They are much weaker at anything with no visual signature: an internal void, a joint that looks correct but is unbonded, a defect whose appearance overlaps entirely with acceptable cosmetic variation. Learning-based approaches extend the range but need many labelled examples of defects you may not have collected, and they answer with a confidence rather than a measurement. Before committing, sit with the people who currently judge these parts and establish whether their decision is genuinely visual, or whether they are using feel, sound and context as well.

Frequently asked questions

Why does a vision system that worked in the demonstration fail in production?

Because the demonstration controlled everything the factory does not. The samples were typical rather than extreme, the part sat still in a fixed position, ambient light was constant and the optics were clean. Production supplies parts from several tool cavities and suppliers, presented with variation, under changing light, on a lens gathering oil mist. Insist that trials use a spread of real parts including known rejects, run in conditions resembling the intended location, before any purchase commitment.

How many defect examples does a learning-based inspection need?

More than most plants have, and covering the full variety of the defect rather than many pictures of one instance. That is the practical barrier: rare defects are rare, so the examples do not exist, and manufacturing deliberately reduces them further. Where a defect can be described geometrically, conventional measurement is usually more dependable and far easier to justify. Learning-based methods earn their place on surface and cosmetic judgements that resist explicit rules, and they still need a validated review process for borderline results.

Should the camera be on the line or at a separate station?

In-line inspection catches problems immediately and can drive a reject mechanism, but it must keep up with the cycle and tolerate whatever presentation the line provides. A dedicated station gives control over lighting, position and timing, at the cost of extra handling and a delay before the result is known. A common compromise is in-line inspection for straightforward checks and an off-line station for the difficult measurement, with the line result triggering a sample to the station.