Machine Vision & Camera Systems

Where it pays off

Machine vision takes over the work the human eye does well but cannot sustain: looking at every part, applying the same criterion every time, and leaving a record of having looked. An operator sees the same scratch less clearly at the end of a shift; a camera either never sees it or always sees it — and both of those are manageable conditions.

The station types we build most often:

• Presence/absence and assembly verification. Is the gasket fitted, did the clip seat, does the cable leave the right connector. A cheap check that blocks an expensive failure.
• Metrology. Dimensions, hole position, angle, diameter — turning a sampled caliper check into 100% inspection.
• Surface defect detection. Scratches, dents, burrs, porosity, paint faults: defects that are hard to specify but easy to show.
• OCR/OCV and barcodes. Verifying that the lot code and expiry date were printed and are legible; DataMatrix for part traceability.
• Robot guidance. Telling the robot where the part is, so localisation comes from a camera instead of a rigid fixture.

Lighting comes first

The fate of a machine vision project is decided on day one, when the illumination is chosen. No algorithm rescues bad lighting — if the contrast is not in the image, software cannot invent it. So we try to solve the geometry with light:

• Backlight produces a silhouette. It is the most stable approach for outer dimensions, hole checks and presence, because part colour and surface finish stay out of the result.
• Dome lighting diffuses reflections on glossy and curved surfaces; the default for print and label reading.
• Dark field brings the light in from the side: scratches and burrs on a flat surface glow while the background stays dark.
• Bright field is the general-purpose answer on flat, matte surfaces.
• Infrared (IR) is used to reject ambient light variation and to see through certain paint or ink layers.
• Polarising filters, crossed between the light source and the lens, remove almost all specular glare.

Shielding the scene from ambient light is part of the design too: a station without an enclosure will decide differently in the afternoon, when the sun hits the window next to it.

Optics, resolution and triggering

Lens selection is calculated together with working distance and field of view. Measurement tasks use telecentric lenses because they remove perspective error; with a standard lens a part looks larger as it moves closer to the camera and the measurement drifts with it.

Then comes the resolution budget: how many pixels the smallest feature of interest will occupy. The practical rule is that a defect needs to land on at least 3–5 pixels to be detected reliably. A 5 MP camera across a 100 mm field gives roughly 0.04 mm per pixel, which puts a 0.05 mm scratch right at the edge of what is detectable — that case needs magnification or a second camera. This arithmetic happens before the purchase order, not after the hardware arrives.

Triggering belongs to the control side. The image has to be captured while the part is in position: a photocell or PLC signal for indexed stations, encoder synchronisation on moving lines, encoder-clocked line-by-line acquisition for line scan cameras. The decision travels back with the same discipline: whichever part was rejected, its position in the shift register is what the diverter is told. Get that link wrong and the system finds the right defect and ejects the wrong part.

What we watch for

We do not start a project without a reference sample set. Choosing illumination without having seen the defect is signing off on a result you have not seen.

Thresholds are set by measurement. False reject and escape counts are tallied on the sample set, and the decision is made against that table. “Turn the sensitivity up a bit” is not an engineering decision.

Maintenance is designed in from day one. Lenses collect dust, illumination ages, fixtures drift. A daily or per-shift verification with a calibration sample catches a station that is quietly degrading.

Rejects are logged. If the image and the reason behind a rejected part are not stored, there is no way to discuss the complaint that arrives the next morning: “this machine is throwing away good parts.”

How this connects to the other solution areas

An inspection station does not stand on its own: the control layer governs when the part arrives, what happens with the decision and how the rejection is recorded. That is why we usually build machine vision work together with the industrial automation side. Where the camera guides a robot, calibration and coordinate transformation belong to robotic systems, and the trend of defect rates over time only becomes meaningful with data acquisition and analytics — the decision on a single part is quality control, but the distribution of those decisions is process knowledge.