What Is Machine Vision?

Machine vision is the use of cameras, lighting and a computer to inspect, measure, identify or guide objects automatically, usually on a production line or in a robot cell. A machine vision system captures an image, processes it and makes a decision, such as pass or fail, many times a second, and then acts on it: rejecting a part, sorting a product, adjusting a process or telling a robot where to pick.

Machine vision replaces subjective manual checks with the same objective decision on every product. It does not get tired at the end of a shift, it can measure to fractions of a millimeter, and it keeps an image record of every item it inspects. That is why it is standard in electronics, automotive, pharmaceutical, food and packaging production, and increasingly in logistics, agriculture and recycling.

Every system has the same core parts, and each one matters as much as the software.

Traditional systems use rule-based measurement tools: edge finding, blob analysis, pattern matching, gauging and code reading. They are fast, explainable and very precise for measurements and for consistent parts, such as checking that a hole is in the right place or a label is present.

Modern systems add deep learning models that learn from labeled example images. They can judge variable products, such as baked goods, wood, textiles or natural materials, where fixed rules struggle, and find defects that are easy to recognize but hard to describe, such as scratches on a textured surface. Most production systems now combine the two: rules for measurement and AI for classification and anomaly detection. Running AI at line speed needs a GPU or AI accelerator, which is why vision computers increasingly look like industrial GPU computers. Our AI in manufacturing FAQ covers where this fits in the wider factory.

Machine vision AI usually means one of four deep learning tasks. Classification sorts an image or region into classes, such as good, scratched or dented. Object detection finds and counts items, such as missing components on a board. Segmentation marks the exact pixels of a defect, so its size can be checked against a tolerance. Anomaly detection is trained only on good parts and flags anything that differs, which suits lines where defects are rare and varied and there are too few bad samples to train on. Deep learning OCR adds reading printed dates and lot codes on curved or shiny packaging, where rule-based tools struggle.

An AI inspection project also runs differently. Instead of tuning tools by hand, the team collects and labels images, typically hundreds to thousands per class, trains the model and validates it against a held-back test set before it goes on the line. Images the model gets wrong in production are fed back into training, so accuracy improves over time. At run time, inference has to finish within the line’s cycle time, often tens of milliseconds per part, and that sets the model size and the processor the vision computer needs. Our edge AI computers span NPU, Hailo, NVIDIA Jetson and GPU systems for that work.

The camera interface sets how much data the system can move, how far the camera can be from the computer and whether a frame grabber is needed.

For moving parts, choose global shutter sensors, which expose every pixel at once and avoid skewed images. Our frame grabbers cover Ethernet, PoE, USB and GMSL cameras, and where several cameras must capture at the same instant, IEEE 1588 PTP keeps them in step.

The two terms overlap, and many people use them interchangeably. Computer vision is the wider field of making computers understand images; machine vision is its industrial application, built to control a process rather than describe a scene.

The line is blurring. Deep learning models from computer vision now run inside machine vision systems for defect classification, and the same edge computers often host both.

Our food factory case study shows machine vision PCs inspecting every muffin and adjusting the ovens automatically, a task where natural variation makes fixed rules hard to apply. The same approach applies to checking seals and labels in packaging, solder joints in electronics and surface finish on machined parts. Vision computers on the line need fanless cooling, wide temperature ratings and, in food and pharmaceutical plants, sealed housings that survive washdown.

Not every task is 2D. 3D machine vision uses laser triangulation, structured light, time-of-flight or stereo cameras to measure height and volume, which suits checking fill levels, weld beads and part flatness, and guiding robots to pick randomly placed parts from a bin. 3D systems generate far more data per part, so they put more load on the computer and the network than a 2D camera running the same line.

We supply machine vision computers with frame grabbers, GPUs and camera drivers fitted as standard, and validate each camera input and trigger before the system ships.