Machine Vision Lighting 2026 — Wavelength, Geometry & Trigger-Sync Selection Guide

Camera specifications get all the attention in a vision project: resolution, frame rate, sensor size, interface. Yet the image the camera receives is manufactured upstream, by the illumination. Contrast between a defect and its background is what a detection algorithm actually consumes, and illumination is the only part of the optical chain that can create that contrast where the material itself does not offer it. A sensor upgrade rarely rescues a badly lit scene; a lighting change frequently does.

This guide covers the illumination half of an industrial inspection station: wavelength selection against the material, the geometries and what each reveals, the strobe and trigger timing that separates a repeatable line from an intermittent one, and how all of it changes the controller on the bill of materials.

Why Lighting Decides the Inspection

Detectability is a contrast problem, not a brightness problem. An algorithm needs a measurable grey-level difference between the feature and its surroundings — a signal-to-noise ratio of at least 3:1. Illumination can only create that difference if defect and background differ in how they absorb, reflect, scatter or transmit light; if a grey contaminant and a grey substrate behave identically at every wavelength and angle, no amount of light will separate them. Two consequences follow. Illumination is decided before the camera, because it determines whether contrast exists at all. And ambient light is a defect in the design: factory lighting at 50 or 60 Hz bands the image frame to frame, and skylights change it across a shift.

Wavelength — Contrast Comes From Absorption, Not Brightness

A monochromatic source plus a monochrome camera almost always beats white light plus a colour camera for feature detection. A single wavelength can be chosen to sit where material contrast is greatest, and no demosaicing smears the edge. Silicon sensitivity peaks near 600 nm, so red and green return the most signal per watt, while blue and ultraviolet pack more energy per photon and resolve finer detail at the cost of more optical power.

Two rules follow. Check the camera's quantum efficiency at the chosen wavelength before buying the light: 470 nm blue looks dimmer to a silicon sensor than 630 nm red at the same electrical input, so a blue solution sizes its driver larger. And a bandpass filter matched to the strobe wavelength is usually cheaper than more light, because it removes factory and daylight ambient without touching the geometry.

Geometry — Choosing the Angle That Reveals the Defect

Once the wavelength is fixed, the angle decides which surfaces send light back to the lens. Light reflects at the angle it arrives: a smooth surface lit on-axis returns a specular flash that hides everything under it, while the same surface lit at a grazing angle looks dark except where a defect interrupts the reflection. Choose the geometry so the good surface behaves one way and the defect another.

Line-scan systems need their own class of illuminant rather than a recycled area-scan bar: a line-scan camera at 20 kHz reads a new line every 50 microseconds, so the light must be bright, tightly collimated and uniform along a metre of web, and usually overdriven.

Strobe Control and the Trigger Budget

Continuous illumination is the exception in production vision. Strobing buys two things at once. It freezes motion: at 1 m/s a 100 microsecond pulse limits blur to 0.1 mm, whereas a 4 ms exposure smears the part across 4 mm. And it defeats ambient light — at a 100 microsecond pulse inside a 10 ms frame the duty cycle is 1%, so the strobe delivers roughly a hundred times more light per unit time than the steady overhead lighting the scene also receives.

The other half is overdrive. LEDs tolerate a multiple of rated continuous current when the pulse is short and the duty cycle low, which is why a well-specified head can deliver two to five times its nameplate intensity in a flash. The limit is junction temperature and average power rather than peak current, so 4 A at 5% duty is gentler on the emitter than 1 A continuous. Put the overdrive factor, maximum pulse width and required duty cycle into the specification as numbers.

Two wiring details decide whether that budget survives a real machine. Voltage drop comes first: an overdriven bar drawing several amperes through 10 metres of cable can lose enough volts to dim visibly, which is why remote driver heads sit next to the light and only a control signal travels the long run. Thermal derating comes second: LED output falls as junction temperature rises, so an overdrive factor proven at 25 °C has less headroom at the 45–60 °C ambient of a sealed enclosure in summer.

Sizing the Vision Controller

The controller sequences per-channel lighting, accepts the trigger, and runs the inference or classical algorithm that consumes the image. Channel count, not camera count, is the figure most often underestimated: a multi-angle station inspecting one part may fire a bright-field bar, a dark-field bar and a backlight in three separate exposures per cycle, each needing its own current, pulse width and delay.

Budget power in two numbers, because they are not the same number. Peak current sizes the wiring, connectors and driver: four channels each driving a 5 A bar at 24 V draw 480 W for the duration of the pulse. Average current sizes the supply: the same four channels at 5% duty average roughly 24 W plus the controller's own draw. A 500 W supply for a 30 W average load wastes money and cabinet heat; a 40 W supply chosen on average current alone fails in the field. PoE is not the answer for overdriven strobe bars — 802.3af delivers 15.4 W and 802.3bt Type 4 tops out near 90–100 W at the PSE — so keep lighting on a dedicated 24 V or 48 V rail and reserve PoE for cameras and access points.

Hardening and Seven Selection Rules

In a washdown zone the light is as exposed as the camera. Treat IP65 as the floor and IP67 or IP69K as the requirement in food, beverage and pharmaceutical areas, with 316L stainless housings that survive caustic cleaning agents; a polycarbonate window that clouds after a year of washdown is a maintenance cost, not a saving. Where a fan cannot be used, the emitter's heat must reach the housing by conduction.

• Ask what the defect does to light before asking what the camera does. If defect and background behave identically, change the wavelength or the angle, not the sensor.
• Pick the wavelength from the material's spectral behaviour, then confirm camera quantum efficiency there and size the driver for the extra power blue or UV demands.
• Match geometry to surface: low angle for texture and cracks, coaxial for flat specular surfaces, dome for curved shiny parts, backlight for dimensions.
• Write the strobe numbers into the specification — overdrive factor, pulse width, duty cycle, channel count — rather than accepting a claim of high intensity.
• Trigger from the encoder or PLC in hardware, and treat any software-issued trigger on a moving line as a defect waiting to happen.
• Fit a bandpass filter matched to the strobe wavelength; it is the cheapest ambient rejection available and it sharpens the monochrome image at the same time.
• Size the supply on average current and the wiring on peak current, and keep overdriven lighting off the PoE budget entirely.

QSCompute supplies the compute and camera side of that specification: fanless and wide-temperature industrial PCs, ARM gateways and Jetson-based edge systems with the isolated I/O, hardware trigger inputs and time-synchronised networking a strobed multi-angle station needs. Send us the part, the defect and the cycle time, and we will map the three together.