Vision systems and optical inspection: speed, limits and false calls

What optical inspection does well

A vision system excels at checks that are repetitive, fast and two dimensional. Presence and absence, orientation, hole count, pitch, profile against a template, print position and surface marks are all natural fits. It does not tire, it records an image for every part and it can inspect at the rate of the press or the assembly line.

It is weaker where the question needs depth, touch or judgement. A burr height, an undercut, a thread depth, a crack that runs into a bore or a surface that looks different depending on the angle are difficult. That does not rule vision out, but it changes the design: the system may need a second camera, a structured light projection or a contact check alongside the image.

Resolution, field of view and the pixel budget

The first calculation is the pixel budget. Divide the field of view by the sensor width in pixels to get the millimetres per pixel. A defect of 0.2 mm needs several pixels across to be detected reliably, not one. A common starting point is at least five pixels across the smallest defect, and ten where the edge position matters.

Field of view and resolution trade against each other. A wide view covers more parts but gives fewer pixels per feature. A narrow view gives detail but needs precise part placement. Telecentric lenses reduce perspective error and are useful for measuring hole positions on a flat part. Standard lenses are cheaper and adequate for presence checks where the part is held in a repeatable position.

Lighting is the real design decision

Most vision projects succeed or fail on illumination. A backlight turns the part into a silhouette and gives a stable edge for dimensional and hole checks. A low angle light rakes across the surface and picks up scratches, dents and burrs that a front light hides. A dome light gives soft, even illumination for colour and print checks on a curved part. Coaxial light reflects back along the lens axis and reveals flat, specular surfaces.

The part material decides which of these works. A shiny plated surface will flare under direct light and disappear under a dome. A black moulded part absorbs light and needs more exposure or a different wavelength. A transparent part may need a backlight with a diffuser and a filter. Test the lighting on real parts, including the worst surface finish in the specification, before the camera is fixed.

False calls and escapes

A false call is a good part rejected by the system. An escape is a bad part accepted. The two move in opposite directions. Tighten the thresholds and false calls rise, which pushes operators to switch the system off or wave parts through. Loosen them and escapes rise, which is worse but invisible until a customer complains.

The tuning target is not zero false calls. It is a stable balance that the line can live with, validated against a challenge set that includes good parts at the edge of the process and defective parts at the edge of the defect definition. Keep the images from that set and re-run them after any change to the light, the lens or the software.

Part presentation and the fixture

A vision system sees what the fixture presents. If the part can rotate, the measurement rotates with it. If the part can sit on a burr, the height changes and the focus drifts. The fixture should locate the part on the datum the drawing uses, hold it flat, and repeat the position within the field of view.

For a stamped or formed part, the fixture often needs a nest with a vacuum or a light clamp. Avoid anything that marks the surface the camera inspects. Where the part arrives on a conveyor, a trigger sensor and a reject mechanism have to be timed to the line speed, and the reject confirmation has to be recorded. A system that detects a defect but cannot remove the part is a monitoring tool, not an inspection station.

Where vision inspection fails

Vision struggles with depth and with anything hidden. A chamfer, a countersink angle, a thread depth or an internal bore cannot be judged from a single top view. A crack that opens under load may be invisible when the part is relaxed. A soft material can smear rather than tear, and the mark left behind may look acceptable in one orientation and not another.

It also struggles with defects that are defined by feel rather than by a number. Light scratches, colour variation and slight dents need a boundary sample set that both the supplier and the customer have signed. Without that set, the system is being asked to enforce an opinion.

Validating the system

Treat the vision system as a gauge and study it like one. For a pass or fail decision, run a study with several operators and a set of parts that includes borderline good and borderline bad pieces. Compare each operator's decision with the reference and with each other. Agreement between operators on a borderline part is often poor, and that is a finding about the defect definition, not only about the system.

Then run the system unattended on a challenge set at a fixed interval, and track the false call rate and the escape rate. A sudden change in either is a signal: a lamp aged, a lens moved, a fixture wore or the part supplier changed a tool. Record the software version and the recipe with the results, because a change in either can alter every decision.

• Challenge set with good, borderline and defective parts, images retained.
• Operator comparison on the borderline parts before release.
• False call rate and escape rate tracked as production measures.
• Lamp hours, focus and fixture wear on a planned check.
• Software version and recipe recorded against the lot.