Lighting Design: Why It Affects Detection More Than the Algorithm

On a production vision project the ceiling on detection accuracy is set at the imaging stage. With the right light the judgement reduces to a simple threshold comparison; with the wrong light even a strong model is only finding patterns in noise. That is why the same algorithm can perform very differently at two stations that look alike — the difference is not in the algorithm but in the lighting and in how much the ambient light is controlled. Lighting deserves to be assessed as its own subject at the design stage rather than patched once the station refuses to stabilise, which is also how the imaging requirements for the Visual AI SOP System are framed.

The goal is stable contrast, not brightness

• Contrast has to be large enough. The grey-level difference between target and background should leave headroom on both sides of the threshold, so that a small lighting drift cannot flip the result.
• Contrast has to be stable. Day shift and night shift, the start and the end of a month, before and after a lens clean — the imaging should land in the same range every time.
• Contrast has to be explainable. The site needs to understand why the light is arranged this way, or the setup will be disturbed without anyone noticing when equipment moves or the station is rearranged.
• A phone photo settles more than it looks like it does. At the design stage, take a phone photo from the intended camera position and look at whether the target edge is clean, whether there are specular highlights, and whether anything in the background interferes. If the phone cannot resolve it, an industrial light may not either — which usually means the detection target itself needs rethinking.

The four basic lighting geometries

In practice the most common combination is a ring or bar light in bright field to judge presence, colour and position, with a backlight or side dark field added where a contour has to be judged — whether a connector is fully seated, for example. Where a rule set includes both position and contour, the two lights are often strobed alternately in sync with the exposure. That is a mechanical design decision, and software cannot make up for it. Camera position limits what lighting is possible, so the two are settled together; see camera mounting and field of view.

Ambient light is the largest variable

• The ambient light is moving while everyone blames the lamp for being too weak. Daylight through a side window changes direction and intensity through the day, ceiling lights run fully or partially at shift change, and a newly fitted light at the next station may shine straight into the lens. The resulting brightness swing is often larger than the difference the system is trying to detect.
• Shroud the lens. A hood or a simple baffle keeps the side light out. It costs very little and the effect is usually immediate.
• Make the dedicated light dominate. Once the active light supplies most of the illumination, the ambient swing is a small share of the total.
• Strobe rather than run continuously, and lock exposure and gain. Strobing only at exposure cuts heat and power and narrows the window in which ambient light matters, and it pairs with the short exposure needed for motion — the trade-off is set out in shutter type and motion blur. Leaving auto-exposure on lets ambient changes turn quietly into image changes and takes the meaning out of every threshold.

Where lighting cannot save the application

• Mirror-finish metal. On a curved metal part the specular highlight moves with the viewing angle, so the same component images very differently as its pose changes. Polarising filters and coaxial light are worth trying, but whether they work depends on surface roughness and cannot be assumed.
• Transparent parts. A transparent material neither reflects nor blocks, and its edge depends on refraction, so contrast is weak and background-dependent. These targets usually need a purpose-built backlight and diffuser, which is markedly more complex than an ordinary station.
• Deep bores. What can be seen inside a bore is limited by the ratio of diameter to depth, and ordinary lights do not reach it; a borescope or fibre-optic arrangement is required.
• Randomly textured surfaces. Castings, textiles and foamed materials have random texture, so the grey-level distributions of target and background overlap and thresholding does not apply.
• The pragmatic alternative is to change the target. Move the judgement from “surface condition” to “did the action happen” — whether a hand reached the pick position, whether the part was placed on the fixture — rather than judging surface defects. Action judgement deals with coarse position and sequence and demands far less of the imaging. This is one reason vision SOP is easier to deploy than cosmetic inspection: the two are not judging the same layer of the problem. Reducing the alarms that do occur is covered in how false alarms are reduced.

Do not overlook long-term lighting maintenance

• Output declines with age. Once the light dims, a threshold that was set with headroom drifts towards the boundary. This typically surfaces six to twelve months after go-live, and machining, moulding and dispensing stations show it first, as the practices described for auto parts illustrate.
• Lenses collect dust and oil mist. Cleaning frequency and an owner need to be written into the daily check rather than left to whoever notices.
• Station changes break the lighting. After equipment moves, a fixture change or a line modification, the light angle is often disturbed without anyone realising. A baseline image captured at acceptance should be archived so the setup can be restored against it.