The order matters
After fifteen-odd vision systems shipped, the single piece of advice we give every new engineer is this: pick the sensor first, the lens second, and design the lighting third. The other order doesn't work — once you've committed to a lens or a lighting concept, the sensor decision ends up forced and you spend the rest of the project working around it.Below is the checklist we walk through ourselves on every new project.
Sensor decisions you can't undo later
- Mono vs colour. Almost always mono, unless colour itself is the feature you're inspecting. Mono gives you 2-3x more usable signal, simpler lighting, and cheaper lenses (no Bayer filter loss).
- Pixel size and resolution. A 5MP camera with 3.45 µm pixels and a 5MP camera with 2.4 µm pixels are not interchangeable. The bigger pixel collects more photons, which means cleaner edges, lower exposure, and more headroom on a moving line.
- Global vs rolling shutter. Anything moving faster than ~50 mm/s on a continuous line is a global shutter problem. Rolling shutter cameras are cheap and sharp on static scenes, but they will distort a part travelling under them.
- Frame rate vs interface. GigE is a fine default. USB3 if cable run is short. CoaXPress when you genuinely need >200 fps at full resolution. We've seen too many projects pick a camera with no clear interface plan and have to redesign the cable trays after the fact.
Lens math people skip
The relationship is simple:FOV = sensor_size × working_distance / focal_length. Pick three of those four and the fourth is decided. What we see go wrong most often is engineers picking two (focal length and working distance) and then being surprised the FOV doesn't fit.For measurement work — anything where you're reporting a number in millimetres — a telecentric lens is worth the price tag. The geometric distortion of a regular entocentric lens will eat your error budget before you've even calibrated.
Lighting last (but most ignored)
Lighting is where careers get made. We'd rather use a $200 camera with the right ringlight than a $4000 camera under bad fluorescent. A few patterns we reach for:- Diffuse dome for shiny surfaces (coated metal, glass, polished plastic).
- Coaxial when you need surface defects on flat metal to pop without highlights.
- Backlight for silhouette / dimensional work — the easiest, most stable option when geometry allows it.
- NIR (850 nm) when you need to see through cling film, light dust, or dye.
Worked example
"Detect 30 µm dents on coated steel sheet, 200 mm/s line speed, FOV 80 mm." That spec lands us on:- A 12 MP global-shutter mono camera, 3.45 µm pixels, GigE.
- A 35 mm telecentric lens at ~120 mm working distance.
- A coaxial 24 V LED head running at 50 % duty, strobe-synced to an encoder pulse.
Total optical bill of materials: under $5k. Total time saved by deciding in this order: about three weeks per project, in our experience.
