How to choose an industrial camera

Work from the application down, not the spec sheet down

Most poor camera choices start at the wrong end: an engineer browses sensors, megapixel counts and frame rates before defining what the system actually has to see. The reliable method is the reverse. You begin with the inspection task, derive the physical requirements it imposes, and only then translate those into a camera, lens, interface and lighting that satisfy them together. A camera that wins on paper can still fail on the line if its resolution, shutter or housing does not match the real scene.

This guide is the pillar for that process. Each step below corresponds to a concrete decision, and each links to a deeper guide where the detail lives. Treat it as a checklist you walk top to bottom: define the scene, size the resolution, choose the shutter, set the speed and interface, decide area versus line scan, then close the loop with optics and lighting. Get the early steps right and the later ones largely fall out of them.

Step 1 - Define the field of view and the smallest detail

Two numbers anchor everything: the field of view (FOV) you must cover, and the size of the smallest feature or defect you must reliably detect. The FOV is the physical area imaged in one frame, set by the part size plus margin for part-to-part position variation and any handling jitter. The smallest detail is the dimension of the finest crack, mark, edge or character your process cannot afford to miss.

These two numbers are in tension. A large FOV spreads pixels over more area, so each pixel covers more of the part and small defects shrink toward the noise floor. Before touching a datasheet, write both numbers down in millimetres, in the worst-case orientation. If a single camera cannot cover the FOV at the detail you need, that is a signal to consider multiple cameras, a moving stage with line scan, or a tighter mechanical fixture, rather than chasing an ever-higher megapixel sensor.

Step 2 - Turn that into a resolution requirement

Resolution follows directly from the previous step. As a working rule, you want the smallest feature you must detect to span roughly 3 to 5 pixels, not one, because a single-pixel feature is indistinguishable from noise and gives no margin for sub-pixel position, blur or contrast variation. Divide the FOV by the smallest detail, then multiply by that pixels-per-feature factor to get the required pixel count along each axis.

For example, a 100 mm FOV that must resolve a 0.1 mm defect needs about 1,000 feature units across, and at roughly 4 px per feature that is on the order of a 4,000-pixel-wide sensor. Round up to a standard sensor format and add headroom for tolerance stack-up and imperfect optics. Resist the temptation to over-specify: more pixels mean more data to move, more demanding optics, and more cost, all of which ripple into later steps. When detail requirements are extreme, look specifically at high-resolution sensors and the optics that can actually deliver that detail to them.

• FOV / smallest-detail = number of feature units across the image.
• Multiply by ~3-5 px per feature to get required pixels per axis.
• Round up to a standard sensor format and add margin for tolerances and optical blur.
• Confirm your lens can resolve that pixel pitch; a sensor out-resolving its lens wastes pixels.

Step 3 - Match the shutter to the motion

If anything in the scene moves while the sensor is exposing, the shutter type matters. A global shutter exposes every pixel at the same instant, freezing motion cleanly, which is what you want for parts on a conveyor, indexing fixtures or any continuous transport. A rolling shutter exposes rows sequentially, which is cheaper and can offer higher resolution for the price, but it skews and smears moving objects. As a default, choose global shutter for anything in motion and reserve rolling shutter for genuinely static or very slow inspections under steady lighting.

Shutter choice also interacts with exposure and lighting. Freezing fast motion means short exposures, which demand more light, often delivered as a bright, strobed source synchronised to a trigger. So the shutter decision is not isolated: it sets constraints you will revisit when you specify lighting and the camera's triggering and exposure control.

Step 4 - Set the frame rate, then choose the interface

Throughput is dictated by your line: how many parts per minute, how many views each, and therefore how many frames per second the camera must deliver without dropping any. Resolution multiplied by frame rate (and bit depth) gives the raw data rate the link must sustain, and that number is what actually selects the interface. Comfortable margin matters here; running an interface near its ceiling invites dropped frames and intermittent faults that are painful to debug on a production floor.

GigE Vision moves around 1 Gbit/s over standard Ethernet, runs up to roughly 100 m, and can carry power and data on one cable via PoE, making it the pragmatic default for moderate-speed, distributed systems. USB3 Vision delivers higher bandwidth, on the order of 350-400 MB/s, but over short cables of about 3-5 m, suiting a camera close to its PC. When you need multiple gigabytes per second for high-resolution, high-speed or line-scan imaging, CoaXPress (CXP-6, CXP-12 and beyond) is the answer, at the cost of a frame grabber in the host. Brands such as Basler, iRAYPLE and Omron/Sentech (STC) offer cameras across these interfaces, so the link, not the badge, should drive the shortlist.

• GigE Vision: ~1 Gbit/s, up to ~100 m, PoE option; default for moderate speed and longer runs.
• USB3 Vision: ~350-400 MB/s, short cables (~3-5 m); simple single-camera PC setups.
• CoaXPress: multiple GB/s (CXP-6/CXP-12+), needs a frame grabber; high-speed and line-scan work.
• Always size the interface above your peak data rate, not at it.

Step 5 - Decide between area scan and line scan

Most inspections use an area-scan (matrix) camera, which captures a full 2D frame at once and suits discrete parts that can be presented and triggered. Line-scan cameras instead capture one line at a time and build the image as the part moves past, which excels for continuous webs, cylinders and very wide objects where a single area sensor cannot deliver the resolution across the whole width.

The decision usually emerges from Steps 1 and 2. If your FOV is wide and your detail requirement is fine, the pixel count for a single area-scan frame may become impractical, whereas a line-scan system can deliver enormous effective resolution along a moving web. Line scan brings its own demands: precise, encoder-locked motion, careful uniform lighting and typically a high-bandwidth interface such as CoaXPress. If the choice is not obvious, it is worth resolving deliberately rather than defaulting to area scan out of habit.

Step 6 - Close the loop with the lens and the lighting

A camera is only as good as the image reaching its sensor, and that image is formed by the lens and revealed by the lighting. The lens must cover the sensor format, hit the required working distance and magnification for your FOV, and resolve detail finer than the pixel pitch you chose in Step 2; otherwise extra megapixels are wasted. For dimensional gauging where perspective error must be eliminated, telecentric lenses are often the correct, if costlier, choice over standard C-mount optics.

Lighting frequently decides whether a defect is even visible. The geometry, ring, dome, backlight, coaxial or low-angle dark-field, is selected to maximise contrast for the specific feature you must detect, and it must be bright and stable enough for the short exposures your shutter and speed demand. Lighting is a specialist domain in its own right, with brands such as DCM Sistemes covering the common geometries. Plan the lens and lighting alongside the camera, not after it, because changing them later usually forces a re-spin of the whole vision station.

Step 7 - Lock down the package, then put it together

With the imaging chain defined, finalise the physical and environmental fit. Confirm the lens mount (C-mount is common, but high-resolution sensors may need larger formats), the camera's mechanical footprint and mounting holes, cable strain relief, trigger and I/O lines for synchronising with PLCs and lighting, and any ingress protection or temperature rating the cell demands. A camera that performs perfectly on the bench can still fail next to a wash-down station, a hot process or strong vibration if the housing, connectors and standards support (GigE Vision, USB3 Vision, GenICam) do not fit the environment and your software stack.

Walk the steps in order and the camera almost selects itself: the FOV and smallest detail fix the resolution, the motion fixes the shutter, the throughput fixes the speed and interface, the part geometry fixes area versus line scan, and the lens and lighting finish the image. Document each number and its worst case, because that record is what lets you defend the choice and reproduce it for the next station.

When the inspection is genuinely hard, variable surfaces, subtle cosmetic defects, or decisions that resist rule-based logic, the right answer may be a deep-learning approach such as Rosepetal layered on top of correctly specified hardware, or full system integration by a specialist like AIS Vision. If you are unsure whether a single camera can meet your detail and speed targets, send us the part size, smallest defect and cycle time and we will size the imaging chain with you.