Machine Vision Lighting: Why Every Failed System Has the Same Root Cause

There is one lesson machine vision has taught us year after year, machine after machine, for two decades: if the lighting is not perfect, you will spend the rest of the project - and the rest of the machine's life - chasing it with mechanical patches and software patches. Every failed vision installation we have ever been called in to rescue had the same root cause. It was never the camera. It was never the software. It was the lighting.

The patch cascade always looks the same. The image is marginal, so someone adds a filter to the code. Then a threshold that gets tweaked every shift. Then a shroud gets bolted on after the machine is built, where a shroud never quite fits. Then a fixture to present the part at exactly the angle the weak lighting needs. Then the model gets retrained, again, because winter sun comes through the bay door at a different angle than summer sun. None of these patches fix anything - they compensate. Photons you never captured cannot be recovered downstream. Contrast is made at the part, not in the processor.

Here are three machines that taught us to fix it in photons instead.

The eight-foot field of view: a linelight that gets brighter at the ends

Years ago we built a series of camera gantries for a consumer-robotics company - overhead cameras watching a play field roughly eight feet across. Lighting a field that wide sounds simple until you do the geometry: light from a uniform source falls off hard toward the edges of a wide field, because the ends are farther away and the rays arrive at a shallower angle. Expose for the center and the corners go dark; expose for the corners and the center blows out. Software gain maps just amplify noise where there is no signal.

The fix was a 96-inch linelight from Metaphase, custom built with a compensated intensity profile - the LED density increases toward the ends, so the light gets brighter exactly where the geometry fans the rays out. The camera sees a flat field edge to edge. No gain map, no per-zone thresholds, no seasonal retuning. One custom light erased what would have been a permanent software workaround.

White on white at 15 meters per minute: overdriving the LEDs

A membrane-materials plant needed microscopic defects found in an expanded-PTFE web - a white, translucent, texture-free material - moving at 15 meters per minute under a line-scan camera. White-on-white at speed is the worst case in this business: there is almost no native contrast, and the exposure window per scan line is measured in microseconds.

The answer was brute photonic force, applied precisely. We ran the LED linelights far past their continuous ratings - strobed and heat-sunk so they survive it - because at those exposure times, only an obscene amount of light puts enough electrons in the sensor to see anything at all. Once the raw image had real contrast, the software got simple: the defects are just there, dark against a bright, even background. We tuned the light for weeks so we would not have to tune the algorithm for years.

The payoff on the monitor: membrane panels on the backlit web at full line speed, edges crisp, background even. Nothing in this image was rescued in software.

Deep-drawn cans: iterate the light, not the code

Deep-drawn metal cans are little cylindrical mirrors. Every feature you need to see - wall defects, rim condition, draw lines - hides inside glare that moves every time the can sits a millimeter differently. On a feasibility study for a high-volume can line, we did what we now consider the only honest process: we put real cans on the bench and iterated the light itself, round after round - our own design, revised between test sessions - instead of freezing a light early and asking the software team to cope.

One of the test artifacts from that study is still one of our favorite images: cans standing on a printed radial-line target, shot from directly above. The fan of lines shows in one frame exactly what the lighting geometry does to the can wall, where the contrast lives, and where it dies. Every revision of the light moved more information into the image before a single line of code ran.

The same discipline paid off again when we built an argon-shielded laser-welding cell: the process glow drowned every off-the-shelf light we tried, so we designed our own illuminator for the chamber. The customer later ordered ten spares. Nobody orders spares of a light that almost works.

What imperfect lighting actually costs

Lighting is a rounding error on a vision project - hundreds to a few thousand dollars against a system costing two orders of magnitude more. But it is the one component whose shortfall gets paid for continuously: every false reject, every shift-change threshold tweak, every service call that ends with someone taping cardboard over a skylight. The cheapest engineering hours you will ever buy are the bench hours spent getting the light right before the frame is welded.

The short version of lighting technique

The full technique catalog fills textbooks, but the working set is small. Darkfield rakes light across the surface so scratches, burrs, and engraving glow against black. Dome light wraps shiny and curved parts in cloudy-day illumination so specular surfaces go readable. Backlight turns measurement into silhouette. Cross-polarization strips glare off film and wet product. Wavelength is a contrast knob - red light erases red print, infrared sails through some plastics, and UV fluorescence finds what white light cannot, which is how our high-speed web inspection system finds invisible coating defects. And strobing freezes motion optically - on our onion topping-and-tailing line, ten cameras image product on a moving shaker table only because the flash is shorter than the motion.

Every one of those choices happens on a bench with real parts - including the ugly ones - before anything is designed around it. Anyone who quotes you a lighting technique without your parts in front of them is guessing.

Get the checklist

The lighting-selection matrix we use on the bench - technique versus surface type versus defect class - lives in the Saber knowledge base, along with the pattern libraries behind our other articles. If you have a vision system that only works on cloudy days, tell us through the contact form - we have probably lit that part before.