Vision sensors

A vision sensor is a self-contained device that takes an image, judges it against a configured or trained reference, and outputs a simple result, typically pass or fail on a discrete output. It sits between a photoelectric sensor, which only knows whether something interrupted a beam, and a full vision system, which returns measurements and rich data. The defining trait is that everything is in one housing and the setup is meant to be done by the person who runs the line, not a vision engineer.

A photoelectric sensor detects that something is there. A vision sensor detects what is there. A photoelectric sensor works from a single beam, so it can confirm a part broke the beam but not whether the part is the right one, the right way round, or undamaged. A vision sensor looks at an image, so it can distinguish variants, check orientation, count features, and spot a defect. When a photoelectric sensor keeps passing the wrong part, that is the signal you have outgrown it.

Mostly depth of output and how much you can shape it. A vision sensor gives a verdict from a configuration you set up in a guided interface. A vision system gives measurements, coordinates, classifications, and images you can log, and it lets you build a more involved sequence of steps. In practice the line between them has blurred, because modern AI vision sensors run trained models on board and handle work that used to need a PC-based system. The useful question is not which label applies but whether the device outputs what your line actually needs to consume.

Presence and absence, correct variant, orientation, simple counting, label and cap checks, and straightforward surface defects on a consistently presented part. The common thread is a yes-or-no answer about one feature or a small set of them, on a part that arrives in roughly the same place each time. These are also the jobs where the payback is quickest, because the setup is short and the failure being prevented is usually obvious and expensive.

Four situations. When you need a real measurement in millimetres against a tolerance rather than a verdict. When the smallest defect is small enough relative to the part that one field of view cannot give it enough pixels. When the surface is specular or highly variable and needs controlled or multi-directional lighting to make the defect visible at all. And when you need to inspect features that face different directions, which needs more than one view no matter how capable the single device is.

A traditional one struggles, because it compares against a fixed reference and natural variation in colour, texture, or position reads as a difference. This is the single most common reason a vision sensor gets installed, works during commissioning, and then produces false rejects in production. An AI vision sensor is a different proposition, because it learns the acceptable range from example images instead of matching a template, which is what makes variable surfaces and cosmetic judgments workable on a device this simple.