What is machine vision?
Machine vision is the industrial application of image analysis: cameras, lenses, lighting, and software working together to inspect, measure, guide, or identify parts as they move through a production process. The term comes from factory automation. A machine vision station typically sits over a conveyor, triggers a camera when a part arrives, analyzes the image in milliseconds, and sends a pass or fail signal to the programmable logic controller (PLC) that runs the line.
Classic machine vision relies on deterministic algorithms: edge detection, blob analysis, pattern matching, barcode decoding, and gauging against a tolerance. These methods are fast, explainable, and reliable when part position and lighting are controlled. Deep learning is added when the defects are irregular, such as scratches, stains, or cosmetic flaws that no rule describes well.
Machine vision compared with computer vision
Computer vision is the broader field of extracting meaning from any image or video. Machine vision is the engineering discipline of applying it on a production line, where camera, optics, lighting, and timing are designed as one system. The two overlap, and most new machine vision systems include a deep learning model.
| Aspect | Machine vision | Computer vision |
|---|---|---|
| Setting | Controlled: fixed camera, engineered lighting | Uncontrolled: phones, security cameras, scans |
| Methods | Rules and gauging, plus deep learning for complex defects | Mostly deep learning |
| Output | Pass or fail signal or measurement to a PLC in milliseconds | Labels, boxes, text, or descriptions to software |
| Integration | Industrial protocols, line speed, and reject hardware | Software interfaces, databases, and business applications |
Machine vision in practice
A fastener manufacturer runs a sorting line at several hundred parts per minute. A line-scan camera images each bolt as it passes, and software measures thread pitch and head diameter against the drawing tolerance. Out-of-tolerance parts are blown off the belt by an air jet. When the plant added a deep learning model to catch plating defects, the hardest part was collecting enough examples of rare defects; the quality team saved every rejected part for a month and photographed them. Reject counts by shift and defect type now flow into the plant’s data platform for root cause analysis.
What to watch for
- Lighting is the most common cause of failure; a system that works on the bench can fail when afternoon sunlight reaches the line.
- Confirm the cycle time. A model that needs 200 milliseconds cannot keep up with a line that presents a part every 100 milliseconds.
- Plan for product changeovers, since each new part number may need its own recipe or additional training images.
- Ask the integrator who owns the images and the trained model, and whether the model can be exported if the vendor relationship ends.