Machine-vision inspection is the use of cameras, lenses, lighting and image-processing software to judge a manufactured part automatically, without a human looking at it. The camera captures an image of the target, a processor compares measurable features in that image against a defined standard, and the system outputs a pass/fail result or a measurement that a machine, a line controller or a robot can act on. It is inspection reduced to a repeatable decision about pixels: where an edge is, how much light a surface reflects, whether a pattern matches, whether a feature is present.
This explanation covers what the concept includes and excludes, how the inspection loop actually runs on a production line, how system architectures differ, which inspection tasks it performs well, and which limits — resolution, lighting, calibration, false calls and commissioning — decide whether a project succeeds. The aim is conceptual clarity before equipment selection: two systems with the same camera specification can behave very differently once the lighting and the decision logic change.
What the Term Covers — and What It Leaves Out
Machine vision sits at the intersection of three traditions: optical measurement, industrial automation and image processing. A useful working scope has four elements that must all be present:
- An imaging sensor — area-scan or line-scan cameras, 2D or 3D, delivering a digital image rather than a single photodetector signal.
- Controlled illumination and optics — the light and lens that turn a physical feature into contrast the sensor can record.
- A decision algorithm — rule-based tools (thresholding, edge finding, blob analysis, pattern matching) or trained models.
- An industrial output — an OK/NG judgement, a measurement value, coordinates for a robot, or a code read, delivered to a PLC, line controller or database.
What it leaves out matters just as much. A single photoelectric sensor that detects whether a box is present is not machine vision; there is no image. A laboratory microscope with a camera attached is imaging but not automated inspection unless the judgement is made by the system. And a human inspector watching a monitor fed by a camera is remote visual inspection, not machine vision, because the decision remains human. The distinguishing feature is the automated decision, not the camera.
Adjacent terms are often used loosely. "Computer vision" is the broader research field, including applications with no factory context. "Automated optical inspection" is the same idea applied to a specific product family such as printed circuit boards. "Vision-guided robotics" uses the same imaging chain but outputs coordinates instead of a verdict. The concepts overlap; the deliverable differs.
The Inspection Loop, Step by Step
Every inline system, from a self-contained smart camera to a multi-camera PC-based station, runs the same loop. Understanding the loop is what makes it possible to diagnose a system that "worked in the lab" but fails in production.
1. Trigger and part presentation
A sensor — often a photoelectric or proximity switch, or an encoder position from the conveyor — signals that a part is in the field of view. The trigger defines the moment of capture, so its repeatability is part of the measurement. If the part is still moving when the strobe fires, motion blur sets the practical resolution limit; if the trigger jitters, the part drifts in the frame and every downstream measurement inherits that error.
2. Illumination and capture
Light illuminates the target, the lens focuses the reflected (or transmitted) light onto the sensor, and the camera digitises the result. This step is where most inspection problems are actually born. Lighting geometry — backlight, dome, coaxial, low-angle — decides which physical property becomes visible: silhouette, surface texture, specular highlight or topographical relief. For a given part, some defects are only detectable under one of these geometries and invisible under the others.
3. Pre-processing
Before any judgment, the image is conditioned: filtering to reduce noise, contrast enhancement, colour-channel extraction, geometric correction, or subtraction of a reference image. These operations are application-specific and exist to make the feature of interest separable from everything else in the frame.
4. Feature extraction and judgement
The algorithm then locates and measures. Typical families include edge and caliper tools for dimensions, blob analysis for contamination and presence checks, pattern matching for alignment and orientation, optical character recognition and code reading for traceability, and classifiers for surface appearance. Each produces either a measurement with a tolerance or a score compared against a threshold.
5. Output, logging and feedback
The result leaves the vision system as a discrete signal, a fieldbus message or a database record. Modern systems also archive the image alongside the result, which is what makes later tuning possible: a re-tested image with adjusted parameters can show whether a change would have fixed a marginal call or simply moved the failure elsewhere.
How the Architectures Differ
Three arrangements cover most installations. The choice is a trade-off between integration effort, flexibility and lifecycle support rather than a question of which is "better".
| Architecture | What it is | Strengths | Practical constraints |
|---|---|---|---|
| Smart camera / all-in-one | Sensor, processing and I/O in one housing, configured through vendor software | Compact, fast to deploy, no PC in the enclosure, single vendor for support | Limited compute headroom; tool set fixed by the vendor; harder to customise algorithms |
| PC-based | Industrial cameras on standard interfaces feeding an industrial PC running vision software | High compute for deep learning and multi-camera stations; wide choice of libraries and cameras | More integration work; PC lifecycle, OS patching and cooling must be managed on the line |
| Controller-based | Camera heads connected to a dedicated vision controller with proprietary image processing | Deterministic timing, stable behaviour across units, mature tool libraries | Vendor lock-in; configuration knowledge stays with trained staff |
Camera interface standards matter for PC-based designs. GigE Vision and USB3 Vision run over commodity cabling and reach further without repeaters; Camera Link remains common where deterministic high bandwidth is required; CoaXPress serves high-resolution, high-frame-rate applications. GenICam provides a common description layer so that software can talk to cameras from different vendors.
Sensor format follows the task. Area-scan sensors suit discrete parts presented one at a time. Line-scan cameras build an image one row at a time and therefore suit continuous webs, sheet material and cylindrical surfaces, where a 2D area camera would either lose resolution or suffer uneven lighting across the field. 3D techniques — laser triangulation, structured light projection, stereo — add height data, which turns a surface defect into a measurable depression or a missing feature into a measurable height difference.
Where Machine Vision Earns Its Place
- Presence and assembly verification — is the fastener, label, gasket or connector there, and in the right orientation?
- Dimensional gauging — distances, diameters, angles and hole positions measured against tolerances in the image plane.
- Surface defect detection — scratches, dents, stains, inclusions and coating defects, where lighting design does most of the work.
- Character and code reading — lot codes, date codes, direct-part marking, for traceability and label verification.
- Positioning and guidance — locating a part for a robot or a pick-and-place head, where output is a coordinate rather than a verdict.
- Colour and material checks — verifying a specified colour or detecting a wrong component in a mixed assembly.
The common thread is that these tasks are repetitive, high-volume and defined in advance. Machine vision is strongest where a human inspector would be bored, rushed or inconsistent — and weakest where the acceptance criterion has never been written down.
Resolution, Accuracy and the Limits That Decide the Project
The single most misused number in machine-vision planning is camera megapixels. What matters is resolution at the part: the field of view divided by the number of pixels across it, usually expressed in millimetres per pixel. A generous sensor aimed at a wide field of view can resolve less detail than a modest sensor on a narrow field. The rule of thumb used throughout the industry is that a reliable edge measurement needs several pixels across the feature of interest — a one-pixel feature cannot be measured, only guessed at.
Accuracy and repeatability are not the same thing. Repeatability describes how tightly repeated measurements of the same feature cluster; accuracy describes how close that cluster sits to the true value. A system can be highly repeatable and consistently wrong if it is poorly calibrated, which is why calibration against a traceable reference and periodic re-verification belong in the commissioning plan rather than in the troubleshooting phase.
Other boundaries that repeatedly determine feasibility:
- Contrast, not magnification, limits detection. A defect that produces no difference in brightness, colour or shadow across the image cannot be found by software, however sophisticated. Rule-based threshold tools are especially sensitive to slow drift in LED output, ambient light leakage and reflections; learned models tolerate moderate variation better but shift their behaviour when the input distribution moves.
- False calls have a cost in both directions. Overkill — rejecting good parts — erodes the business case faster than underkill in many lines, because operators learn to bypass the system. Setting thresholds therefore needs data on the natural spread of acceptable parts, not just examples of defects.
- Determinism versus generalisation. Rule-based logic returns identical outputs for identical pixels, which simplifies validation and auditing. Learned models generalise across cosmetic variation but are harder to explain, so hybrid designs — learned detection plus rule-based confirmation — are common where traceability matters.
- Golden samples age. A model or rule set tuned on parts from one tool, one material batch or one season of ambient conditions may not hold a year later. Maintenance is a recurring task, not a one-time handover.
- Sampling rate versus cycle time. Image transfer, processing and I/O all consume time. A fast camera does not guarantee a fast decision; the slowest element in the loop sets throughput.
Misconceptions Worth Retiring
"A better camera will fix it." Usually the problem is contrast or geometry. Changing optics and lighting resolves more marginal inspections than upgrading the sensor.
"Deep learning removes the need for lighting design." Learned models can absorb more variation, but the training set can only contain what the imaging chain makes visible. Poor illumination means the defect never appears in any training image.
"If the system passes the sample parts, it is ready." A system is ready when it has been challenged with marginal parts, known defects, deliberately wrong assemblies and the natural drift of the process — and when the operator understands what to do when it flags something ambiguous.
"Inspection means the parts are good." Machine vision verifies the features it was configured to verify. Anything outside that specification is simply not checked, and the system's verdict carries exactly the scope of its definition.
Commissioning: Where Most Projects Are Won or Lost
The engineering content of a machine-vision project is largely front-loaded. Before hardware is ordered, the inspection specification should be written as a measurement problem: what feature, what tolerance, what defect size must be caught, what rate of false calls is acceptable, what happens to a rejected part. Building the imaging chain — lens choice for the required field of view and working distance, lighting geometry chosen for the feature rather than for convenience, mounting rigid enough that vibration does not move the camera relative to the part — follows from that specification.
Acceptance should be measured, not observed. Run the system against a statistically meaningful population of good parts, known-bad parts and borderline parts; record repeatability by measuring the same feature repeatedly; verify calibration against a traceable artefact; log results with images so that future drift can be diagnosed rather than argued about. Document the settings and back them up — a vision configuration is a piece of intellectual property that walks out of the door with the engineer who built it unless it is written down.
Machine-vision inspection is, in the end, a discipline of making a decision explicit. The camera does not decide what "good" means; the specification does. When a project fails, it is more often because the definition of good was never pinned down than because the pixels were insufficient.