Industrial cameras

An industrial camera is a camera built to produce measurable, repeatable images for automated inspection and control, rather than pleasing images for a person to look at. In practice that means a global or precisely controlled shutter, an external trigger so it captures at an exact moment, a fixed mechanical mount so calibration holds, a standard data interface, and a housing rated for a factory. It also means no automatic enhancement: an industrial camera should return the same pixel values for the same scene every time, because any hidden auto-adjustment invalidates a measurement.

Consumer and security cameras optimise for how the picture looks, using auto exposure, auto white balance, noise reduction, and compression. Every one of those changes the pixels in ways that vary between frames, which is exactly what a measurement cannot tolerate. They also lack a hardware trigger, so you cannot guarantee the image was taken when the part was in position, and they usually lack a rigid mount, so the calibration drifts. A security camera can tell you something happened. An industrial camera can tell you a gap measured 2.03 mm.

A streaming camera sends images somewhere else to be processed, typically to an industrial PC over GigE Vision or CoaXPress. A smart camera does the processing on board and outputs a decision, not an image. Streaming makes sense when you need very high resolution or want many cameras feeding one powerful machine. A smart camera makes sense when you want one device to mount, one thing to configure, and no separate PC to maintain, which is usually the faster path to a working station.

Pick on cable length and data rate, not on preference. GigE Vision runs over ordinary Ethernet, tolerates long cable runs, and uses network hardware you already have, which makes it the common default. USB3 Vision is simpler but short range, so it suits benchtop and in-machine work. CoaXPress and Camera Link exist for cameras producing more data than a network can carry, usually high-speed line scan. If the camera makes its own decision on board, the interface only carries a result, so it stops being a constraint at all.

Megapixels is the wrong starting number. Work from the defect: decide the smallest thing you must detect, then require at least three to five pixels across it, and closer to five if it is subtle or low contrast. Combine that with how much of the part has to be in frame and you get a required pixel count, which then tells you the sensor. Buying resolution without doing that arithmetic is how projects end up with a very expensive camera that still cannot see the defect, because the field of view was too wide.

Usually not, and monochrome is often the better choice. A colour sensor puts a filter over the pixels, so it trades resolution and light sensitivity for colour information. If the defect is a geometric or textural difference, which most defects are, monochrome gives you more usable detail from the same sensor and performs better in low light. Choose colour when colour is genuinely the signal: verifying a correct coloured part, sorting by material, or catching a stain that differs in hue but not in brightness.