A machine vision system is a camera-based setup that lets a machine inspect, measure, or guide by "seeing." It has five parts working together, lighting, a lens, an image sensor, a processor, and an output, and the quality of the picture those first three produce sets the ceiling on everything the software can do.
Most machine vision problems are lighting and optics problems wearing a software costume. This guide walks the hardware side: the five components, the lighting geometries that make a defect visible or invisible, how to choose a lens and sensor, and where classical rule-based vision beats AI vision and where it does not. For the AI decision-making side of inspection, see our companion guide on AI quality control.
What are the parts of a machine vision system?
A machine vision system is five functional blocks in a chain. Light hits the part, the lens forms an image of it on the sensor, the sensor turns light into pixels, the processor analyzes those pixels, and the output tells the line what to do. A weak link anywhere caps the whole system, the best algorithm cannot recover a defect the lighting never revealed.
Notice how much of the system is optical, not computational. Two of the five blocks are pure hardware, and a third, the sensor, is the bridge. Teams that struggle with machine vision almost always underinvested in the first three and overinvested in the last two.
Why is lighting the most important choice?
Lighting is the highest-leverage decision in machine vision because the geometry of the light, where it comes from relative to the part and the camera, decides whether a feature shows up at all. The same scratch is invisible under one lighting setup and glaring under another. Getting the light right often turns a hard software problem into a trivial one. Two more variables ride alongside geometry: the wavelength (color) of the light and whether it strobes. Choosing a light color that contrasts with the part, blue against a copper trace, red against a green board, can separate features the eye barely distinguishes, and a strobed light synchronized to the trigger freezes fast motion far more cheaply than a faster shutter. None of that is software; it is all decided at the fixture.
| Lighting | Geometry | Best for |
|---|---|---|
| Backlight | Behind the part, camera opposite | Dimensions, presence, edges, the part reads as a black silhouette |
| Dark-field | Low angle from the sides | Scratches, engraving, surface texture, raised or recessed defects |
| Bright-field | Direct, near the camera axis | General surface inspection on matte parts |
| Dome / diffuse | Wrapped around the part | Shiny or curved surfaces where glare must be killed |
| Coaxial / on-axis | Through a beamsplitter, down the lens axis | Flat, reflective surfaces read evenly |
How do you choose a lens and sensor?
The lens and sensor together set your field of view, your resolution, and how accurately you can measure. The lens forms the image; the sensor samples it into pixels. Three ideas do most of the work.
First, resolution has to fit the smallest feature you must detect. A common working rule is that the smallest defect should span at least a few pixels, one pixel is not enough to detect reliably. Work backward from your smallest feature and field of view to the sensor resolution you need, rather than buying megapixels for their own sake.
Second, telecentric lenses matter for measurement. An ordinary lens shows perspective: parts farther away look smaller, so an object shifting in the field of view appears to change size. A telecentric lens keeps magnification constant across its depth, removing that parallax error. If you are measuring dimensions to a tolerance, a telecentric lens paired with backlighting is often the difference between a gauge and a guess.
Third, match the sensor to the motion. A global-shutter sensor captures every pixel at once, which stops motion cleanly on fast lines; a rolling shutter can smear a moving part. Area-scan sensors image a rectangle at a time; line-scan sensors build the image one row at a time and suit continuous web or cylindrical products. The part's speed and shape drive the choice more than the spec sheet does.
Rule-based or AI vision: which do you need?
Machine vision software comes in two broad families, and matching the family to the task matters as much as the hardware. They are not rivals so much as tools for different jobs.
- Rule-based (classical) vision uses explicit algorithms: find edges, measure distances, match a template, count blobs, read a code. It is deterministic, fast, needs no training data, and is easy to validate, but it demands controlled, repeatable conditions and struggles with parts that vary in acceptable ways.
- AI (deep-learning) vision learns from labeled examples. It handles cosmetic defects, natural variation, and "you know it when you see it" judgments that resist explicit rules, but it needs a labeled dataset, more compute, and careful validation, and it can fail in ways that are harder to explain.
The practical answer is usually "both, for different checks." Measure a machined dimension with a rule-based tool and a telecentric lens; judge a cosmetic surface finish with a trained model. Reaching for AI when a caliper-style rule would do adds cost and fragility; forcing rules onto a naturally variable defect leads to endless tuning. The decision belongs in the design, not as an afterthought.
How do you design a machine vision setup?
A reliable vision application is designed in a fixed order, hardware first. Skip a step and you pay for it in the software:
- Define the exact task and defect. Name the specific thing you must detect or measure, its smallest size, and the pass/fail rule. "Inspect the part" is not a spec.
- Choose the lighting geometry. Pick the geometry that makes your defect obvious, backlight for dimensions, dark-field for scratches, dome for shiny parts. Prove it with sample images before anything else.
- Choose the lens and working distance. Set field of view and depth of field; use a telecentric lens if you are measuring to a tolerance.
- Choose the sensor and resolution. Work back from the smallest feature and the line speed to resolution and shutter type.
- Choose rule-based or AI per check. Match each inspection to the software family that fits it, and gather labeled data early if any check needs AI.
- Validate on real variation. Test against good parts, known defects, and the full range of normal variation, including the edge cases the line will actually throw at it.
- Wire the result into action. A pass/fail that does not reject, sort, or log changes nothing. Connect the output to the line and to your records.
The order is not arbitrary. Each choice constrains the next, and lighting constrains all of them. This is also why computer-vision quality inspection succeeds or fails long before the software is written.
Where does machine vision fit in the bigger picture?
A vision system is a superb sensor, and like any sensor its data is only worth what you do with it. A camera that logs every reject but whose data never leaves the station tells you nothing about why rejects cluster on Tuesdays, or which upstream change caused a spike. The value shows up when inspection results join the rest of the plant's data, machine states, downtime, changeovers, and the operator's notes.
That is where an operational layer earns its place: take the pass/fail and defect data a vision system produces and connect it to everything else, so a quality trend can be traced to its cause instead of sitting in a silo. It is the wedge Harmony takes as a manufacturing operating system read the signals your PLCs and inspection stations already produce, add the paper and tribal knowledge that explain them, and connect what you run with no rip-and-replace, the way plants like CLS made their records live. From there, vision data feeds the same picture as machine learning for maintenance AI quality control and broader smart-factory analytics, one operation, not a rack of disconnected cameras.
Machine vision systems by the numbers
Primary references for the hardware side of machine vision:
- Cameras and image sensors for machine vision are characterized under EMVA 1288 the standard that defines how to measure and state sensitivity, noise, linearity, and defect pixels, so specs can be compared across vendors (European Machine Vision Association, EMVA 1288).
- The standard's Release 4 splits camera characterization into separate documents for linear and general sensors, reflecting how varied modern sensors have become (EMVA 1288 downloads).
- A durable engineering rule of thumb: the smallest feature you must detect should span at least a few pixels on the sensor, which is what ties your required resolution to your field of view and defect size rather than to a megapixel number.
Read those honestly: machine vision is an optics and lighting discipline first. Get the light, lens, and sensor right and the software gets easy; get them wrong and no algorithm will save the application.