What computer vision means
Computer vision is the field concerned with getting useful information out of images and video. A camera records a grid of numbers, one brightness or colour value per pixel. Software then tries to answer a question about the scene: is this part present, where is that edge, how far is the object, is there a crack?
It is worth being clear about what this is not. The software does not “see” the way a person does. It computes from pixel values, and it answers only the question it was built for. Everything else in this article follows from that.
How an image becomes an answer
Most systems follow a similar pipeline, whatever the application.
- Capture. A camera, lens and light source produce an image. This stage matters more than newcomers expect, because later software cannot recover detail that was never recorded.
- Preprocessing. The image may be cropped, corrected for lens distortion, adjusted for brightness or reduced to greyscale.
- Extraction. The system finds features such as edges, corners, blobs or regions of similar colour.
- Interpretation. Rules or a trained model decide what the features mean: pass or fail, object type, position.
- Output. The result goes to a person, a database or a control system.
Older approaches rely on hand-written rules, for example “measure the distance between two edges and compare it to a limit.” Newer approaches use machine learning, where a model learns patterns from many labelled example images instead of following explicit rules. Neither is always better. Rule-based methods are easier to explain and check; learned models cope better with varied, messy images but need good training data.
Common task types
It helps to separate the kinds of question a vision system can answer, since each has different needs:
- Classification: what is in this image, or is this part good or bad?
- Detection: is there an object, and where is it?
- Segmentation: which pixels belong to the object?
- Measurement: how large is it, or how far apart are these features?
- Tracking: where does the object move over successive frames?
Typical engineering uses
Inspection. Cameras are widely used to check products on a line for missing components, wrong labels, surface marks or misplaced parts. This supports quality control, though the system is usually one layer of checking rather than the only one.
Measurement. With a calibrated camera, software can estimate dimensions without touching the part. The accuracy depends on calibration, lens quality and lighting, and the result carries its own measurement uncertainty that should be stated rather than assumed to be perfect.
Robotics and automation. A robot arm may use a camera to find where a part sits so it can pick it up, or a mobile robot may use images to help with navigation and obstacle avoidance. Here the vision output feeds a control loop, so delay and error in the image result affect motion.
Reading and identification. Barcodes, printed characters and markings can be read automatically to track parts through a process.
Condition monitoring. Images of equipment, structures or products can be reviewed for visible wear or damage, sometimes with a person confirming what the software flags.
Known failure modes
Vision systems often work well in a demonstration and then struggle in a real setting. Common causes include:
- Lighting changes. Shadows, glare, reflections from shiny surfaces and variation between day and night can change pixel values enough to confuse a method tuned for steady light. Controlled lighting is often the cheapest improvement.
- Differences from the training data. A learned model may perform poorly on parts, backgrounds or camera positions that look unlike what it was trained on.
- Drift over time. A camera can shift slightly, a lens can get dirty, or a supplier can change a surface finish. Performance can degrade without any obvious error message.
- Rare or unseen defects. A model can only recognise what its examples taught it. A new kind of fault may pass unnoticed.
- Overconfidence. A system may report a confident result on an input it has never encountered.
- Blur and motion. Fast-moving parts and long exposure times smear detail.
- Limited resolution. A defect smaller than a pixel footprint on the part simply cannot be found.
Because errors go in both directions, it helps to track both false alarms (good items rejected) and misses (bad items passed). Which matters more depends on the application.
Planning for use in practice
A few habits reduce risk. Define clearly what the system must detect and how often an error is tolerable. Test on images collected from the real setting, not only clean samples, and keep a separate set of examples that were never used to tune the system. Build in a way for people to review uncertain cases and to notice performance changes over time. The idea of separate training and test data is covered further in the discussion of machine learning for engineers.
Where a vision system supports a safety-related decision, it should be treated as part of a wider safety design assessed by qualified people, following the rules and standards that apply in your country and industry. Nothing here replaces that judgment. As with any automated tool, the broader effect of artificial intelligence on technical work comes down to understanding what the tool does well, where it fails and who remains responsible for the outcome.