Machine vision vs computer vision: what buyers need to know

An image-analysis demonstration can recognize a missing component yet still fail as a production inspection. Reflections may hide the feature, motion may smear its edges, or the reject signal may arrive after the part has passed. That gap explains much of the confusion between machine vision and computer vision. For buyers specifying an inspection station, the useful question is not which label wins. It is which parts of the imaging, analysis and control chain the proposed solution actually includes.

The difference is emphasis, not a fixed boundary

Computer vision concerns extracting useful information from images or image sequences. That includes locating objects, measuring geometry, recognizing characters, estimating motion and interpreting scenes. Its methods range from thresholding and geometric fitting to trained neural networks.

Machine vision commonly describes image-based inspection, measurement and guidance within an engineered automation system. The emphasis includes creating a repeatable image, making a decision within a deadline, and connecting that decision to production equipment.

These meanings overlap. An industrial inspection can use computer-vision algorithms within a machine-vision system. A computer-vision application can also include carefully engineered cameras, illumination and real-time interfaces. Neither term guarantees a particular architecture.

Treat this table as a description of common emphasis, not a rule separating two disciplines. Ask suppliers to define their deliverables rather than relying on the terminology.

An algorithm is one part of the inspection chain

Consider a conveyor carrying molded parts. The task is to detect a missing insert and send the correct part to a reject mechanism. The analysis might be a simple intensity measurement inside a located region, or a trained detector if appearances vary substantially.

Either method needs a usable image. Illumination must reveal the insert consistently across normal surface finishes. The lens must cover the inspection area with adequate detail. Acquisition must associate each image with the correct part. The control interface must preserve that identity until the reject action occurs.

A practical chain is:

1. Detect or identify the arriving part.
2. Trigger illumination and image acquisition with known timing.
3. Locate the part and evaluate the inspection feature.
4. Produce a result containing part identity, status and validity.
5. Track that result to the required action position.
6. Confirm the action or report an exception.

Software compatibility belongs in this chain, not at the end of purchasing. The guide to camera acquisition software compatibility and integration options explains the acquisition layer that connects camera hardware to analysis software.

An ordinary inspection result is not a safety-rated protective function. Any personnel-protection requirement needs a separately designed and validated safety system.

Work backward from the smallest feature

Suppose the inspection must reveal a missing-insert indication that is approximately 0.30 mm wide. Assume a horizontal field of view of 120 mm, including placement variation around the part.

The field of view must include the whole inspection region, not just the nominal part position. For example, a centered 100 mm part envelope within a 120 mm field leaves 10 mm on each side. That allowance must accommodate real positioning variation and any required locating features.

Start with a provisional sampling target of 4 pixels across the smallest indication. This is a planning assumption, not a guarantee of detection. Low contrast, blurred edges or measurement uncertainty may require more samples.

Use the camera resolution planning guide to distinguish sampling requirements from measurement accuracy. Although the example here concerns defect visibility, the same field-width arithmetic provides a useful starting point.

Now suppose a candidate acquisition configuration produces an image 4000 pixels wide across that same field. The arithmetic becomes:

That gives more sampling margin than the provisional target, but it does not prove the inspection will work. The lens must transfer the required detail, focus must remain acceptable across part-height variation, and illumination must create sufficient contrast.

Increasing resolution can also increase transfer, memory and processing demands. A larger image is useful only if its additional information improves the decision. Compare candidates through the area scan cameras category and the camera finder after establishing the required field and feature scale.

Motion blur and deadlines change the specification

A stationary image trial does not establish suitability for a moving conveyor. Continue the example with an object sampling of 0.030 mm/pixel and conveyor speed of 600 mm/s.

Assume the initial blur allowance is 0.5 pixel in the direction of travel. This allowance is application-specific; validate it against the feature and algorithm.

This estimate addresses motion during exposure. It does not account for rolling-shutter geometry, vibration, focus error or illumination variation. A short exposure also needs enough light to produce a useful signal without excessive gain.

The control deadline is a separate calculation. If the reject position is 180 mm downstream and speed remains constant:

An illustrative budget might allocate 8 ms to acquisition, 12 ms to transfer, 18 ms to analysis, 4 ms to communication and 30 ms to actuator response. Their sum is 72 ms, leaving 228 ms before the assumed arrival time.

Those values are hypothetical design inputs, not camera or controller specifications. Replace them with measured upper bounds under realistic load. Include timing variation, output scheduling and actuator requirements. If conveyor speed varies, use position tracking rather than relying on a fixed delay.

An average processing time is not enough. Test long processing events, queued images, communication interruptions and missing results. The article on stale conveyor vision data and scene timing explains why a plausible result can still belong to the wrong production moment.

Choose hardware for the image and interfaces

Processor type does not define either discipline. CPUs, GPUs, FPGAs and DSPs can support image-processing workloads. The useful selection questions concern latency, throughput, memory movement, supported software, electrical interfaces and maintainability.

Camera selection

The MV-CU020-90GC is a global-shutter, GigE color camera. The LS-CMR-0166 is a rolling-shutter, USB3 camera. They illustrate different acquisition choices, not interchangeable solutions for the worked example.

For moving parts, evaluate shutter behavior alongside exposure time. A global shutter helps avoid motion-related row-to-row geometric skew, but it does not eliminate blur during exposure. A rolling shutter may suit stationary or suitably controlled scenes, but moving-part suitability needs a specific timing assessment.

Choose color only when color information helps the inspection. Confirm native image dimensions, trigger behavior, required frame rate, host interface and software support before selecting a camera. Do not infer those capabilities from the general product category.

Lens selection

The LS-FFL-0567 has a 16 mm focal length and a C-mount. Treat it as a candidate for optical assessment, not a confirmed match to the example geometry.

Lens selection requires sensor dimensions, available working distance, field of view, image-circle coverage and acceptable distortion. Confirm the camera mount separately, and check performance across the actual aperture and focus range.

Use the machine vision lens selection guide to build that specification. Then compare machine vision lenses in the lens finder. A matching mount alone does not establish optical suitability.

Processing and control integration

The MV-VC2140-128G40-1T is a vision controller with built-in light outputs. Its suitability still depends on the intended acquisition and control architecture.

When evaluating vision controllers, confirm camera connectivity, supported operating system, software licensing, image-buffer capacity and control communications. For light outputs, check electrical ratings and timing functions against the proposed illumination hardware.

A built-in connector does not establish end-to-end synchronization. Ask which events are hardware-timed, which depend on software scheduling, and how faults are reported.

Select the simplest analysis that meets the requirement

A missing insert with stable appearance may be detected using part location, a defined inspection region and a contrast or area threshold. This approach can make the decision easy to explain and debug.

A learned model may be more suitable when acceptable appearances vary substantially or defects cannot be described reliably with a few geometric rules. It still needs representative data and controlled acquisition. Learning does not remove the need for suitable lighting and optics.

For either approach, separate development data from evaluation data. Include acceptable variation in finish, position, height and orientation. Include borderline defects, dirty backgrounds, changed illumination and partial occlusions where these are credible operating conditions.

Report false accepts and false rejects separately. For example, rejecting every part could avoid accepting defective parts while making the station useless. An overall accuracy percentage can hide that distinction, especially when defects are rare.

Preserve original images and relevant settings when diagnosing failures. For threshold-based inspections, a live binary image preview can help reveal unstable segmentation, but it should not replace inspection of the source image.

When the usual answer is wrong

The usual purchasing shortcut is to ask for a camera and an algorithm. Sometimes the largest improvement comes from neither.

If reflections hide the insert, changing illumination geometry may produce a clearer decision than increasing model complexity. If the feature is visible only from an oblique direction, camera placement may matter more than additional pixels. If the result reaches the wrong part, tracking and timing need attention before retraining.

An offline image-analysis tool may be sufficient when an operator reviews results and no machine action depends on a deadline. Conversely, an apparently simple threshold inspection can require substantial integration when several parts are simultaneously between the camera and reject mechanism.

There are also cases where a different sensing method is more appropriate. If the required property has no reliable visible signature, image analysis cannot infer it dependably merely because the processor is powerful.

Specify observable evidence and acceptable operating conditions first. Then choose the sensing and processing approach that can demonstrate the required performance.

Checks before ordering and commissioning

Use this checklist to turn terminology into a purchasing specification:

• Inspection requirement: Define the smallest relevant indication, acceptable variation, false-accept limits and false-reject limits.
• Image formation: Record field of view, working distance, height variation, lighting geometry and ambient-light conditions.
• Acquisition: Confirm shutter behavior, exposure range, trigger support, interface compatibility and image identity handling.
• Optical compatibility: Check mount, image circle, focus range and detail across the complete inspection region.
• Timing: Document acquisition-to-result latency, variation, tracking method and the action deadline.
• Software: Confirm supported versions, deployment licensing, startup settings and access to diagnostic images.
• Fault handling: Define responses to missing images, unavailable results, lost communication and restart conditions.
• Acceptance testing: Run representative parts at production speed and record both inspection outcomes and control actions.

Test recovery as well as normal operation. The guide to camera acquisition stalls under load provides a useful framework for separating acquisition faults from downstream processing problems.

Start your component selection with the camera configurator, or send the field of view, feature size, conveyor speed and required machine interface through application support. A clear specification is more useful than choosing between two overlapping labels.