
What this application demands
A machine-vision system is built around the difference between the light you control and the light you do not. The illumination is known — an LED bar, a ring light, a laser line or a structured pattern at a defined wavelength — and everything else is noise. The filter's job is to pass the first and reject the second, and it has to do so across the whole field of view rather than only on the optical axis.
Three requirements follow from that. First, the pass band must sit on the source wavelength with enough margin that production spread and thermal shift do not push the peak off it. Second, the blocking must be deep enough across the rest of the detector's sensitivity range, because a shop floor contains daylight, fluorescent flicker and infrared from hot parts. Third, the filter has to accept the cone of angles the lens delivers without the pass band drifting — and this last point is where many off-the-shelf filters quietly fail.
Specifications we recommend
| Component | Typical specification | Why |
|---|---|---|
| Band-pass filter | FWHM 10–40 nm centred on the illumination, blocked to OD4 across the detector range | Isolates the source without clipping it as the light shifts with temperature |
| Hard-coated filter | Dense dielectric stack rather than dyed glass | Stable centre wavelength and no bleaching under continuous illumination |
| Protective window | Borosilicate or fused silica with broadband AR, 60–40 scratch–dig | Protects the lens while adding negligible reflection into the image path |
| Lens cover glass | Thin AR-coated glass, parallelism below 10 arc-seconds | Prevents ghost images from the two surfaces of the cover |
Failure modes we design against
The most common field problem in vision systems is spectral shift with angle: a filter specified at normal incidence is placed in a converging beam, and the effective pass band moves towards the blue. Where the source is a narrow-band LED, the mismatch silently reduces the signal-to-noise ratio the system was designed around. We design the stack over the real angular range rather than at a single nominal angle, which is why we ask for the lens aperture and working distance rather than just the wavelength.
The second is ghosting. Two parallel surfaces each reflect a small fraction of the signal, and the resulting double image is often mistaken for a focus problem. Specifying parallelism and an AR coating on both faces removes it.
Frequently asked questions
Should I use a dyed-glass or a hard-coated filter?
Hard-coated, in almost every machine-vision case. Dyed glass is cheaper but has broader, less stable edges and can fade under continuous high-intensity illumination.
How tight should the pass band be?
Tight enough to reject ambient light but wide enough to tolerate the source's wavelength shift over temperature. For most LED systems 20–40 nm is the practical compromise; laser-line applications can go much narrower.