Scoping a machine vision system before you call an integrator

A machine vision project can fail at the scope stage before programming begins. Camera, lens and lighting choices depend on three things: the part, the smallest defect you must catch, and where the camera can physically sit. This guide is for engineers and buyers who are about to request quotes or order components. Work through the numbers here and you will be better prepared to source parts yourself or hand an integrator a specification that makes quotes comparable and reduces disputes about scope.

Write the specification before the shopping list

Start with one page that states the physical facts, not the solution:

• The part: dimensions, material, surface finish (matte, shiny, reflective), and how it arrives (conveyor, fixture, robot gripper).
• The defect or feature list, each with the smallest size in millimetres that must cause a reject.
• The required accuracy: detection only, or measurement with a tolerance in millimetres.
• Line speed or cycle time: parts per second, and whether the part moves during exposure.
• The mechanical envelope: maximum working distance, mounting angle, space for lights.
• The interfaces: PLC protocol, robot, MES, and the pass/fail signal the line expects.

If you cannot fill in the smallest defect size, stop and measure real parts first. Every downstream resolution calculation divides by that value. Vague scopes produce quotes that differ wildly, because each integrator assumes a different difficulty. A clear scope also tells you whether the project is a one-camera check you can own or a multi-camera cell that needs a vision controller and professional programming. For a view of how the work splits between your team and an outside firm, this article on integrator work scope and split-team models is a useful reference when you define contracts and responsibilities.

Field of view and working distance come first

The field of view is the part size plus margin for positioning tolerance. If the part is 40 mm wide and its centre can shift 10 mm either side of nominal, the minimum field of view is 60 mm: 40 + (2 × 10) = 60 mm. Allow additional clearance where needed and verify the result with real parts.

Here m is magnification, WD is working distance and f is the estimated focal length. The focal length expression is a thin-lens estimate; actual coverage depends on lens construction and how working distance is referenced. Use the calculator below for an initial selection, then verify coverage at the intended mounting position.

Worked example

You inspect a 40 mm aluminium fitting for a missing O-ring groove. The groove is 1.2 mm wide. The camera must sit 200 mm away because a gripper passes underneath.

1. Field of view: a 40 mm part plus 10 mm positioning margin on each side gives 60 mm.
2. Assume an active sensor width of 5.1 mm, confirmed from the candidate camera's data sheet. Then m = 5.1 / 60 = 0.085.
3. f ≈ 200 × 0.085 / 1.085 = about 16 mm.

A nominal 16 mm lens is therefore a starting candidate, not a guarantee of coverage. If the calculation lands between standard focal lengths, the shorter one generally provides a wider field of view but fewer pixels across the defect. Moving the camera back also widens coverage. When perspective and magnification changes threaten your measurement error budget, the telecentric lens for gauging article explains when a different optical approach is needed.

From defect size to camera resolution

Resolution starts with a division. Take the field of view, divide by the smallest defect, and multiply by the pixels you want on that defect. This estimates sampling requirements; it does not prove that the defect will be detectable.

Continue the example: 60 mm field of view, 1.2 mm smallest feature, and 4 pixels per feature gives (60 / 1.2) × 4 = 200 pixels across the inspected direction. Many industrial cameras exceed that sampling requirement, but contrast, focus and lighting still need validation.

The opposite case is common too: a 0.05 mm scratch in a 60 mm field at 4 pixels per defect needs 4,800 pixels across that direction. Check the actual horizontal or vertical pixel count rather than selecting by total megapixels alone. The camera resolution calculator below runs this arithmetic for your inspection dimensions.

When the numbers point to a compact one-task inspection, a vision sensor keeps the project small.

A smart camera with an M12 lens mount is another candidate for a stable, single-view check.

Compare the catalog specifications before selecting the inspection hardware.

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

As a selection rule, start with monochrome sensors for defect detection and measurement, and choose colour when colour itself carries the information. Prefer a global shutter when the part moves. A rolling shutter can suit stationary parts, but motion during sensor readout can distort the image. Our industrial camera specifications buyer's guide goes deeper on reading data sheets, and the smart cameras category page filters by resolution, shutter and mount.

Lighting is where projects are won

A defect is visible only if the lighting makes it differ from the good part. Before choosing hardware, classify your task:

• Geometry and presence: test bright-field illumination; low-angle illumination can help when edges or raised features matter.
• Scratches and embossed marks on flat surfaces: test dark field, with light grazing across the surface.
• Print, labels, contrast marks: test diffuse bright field to reduce glare.
• Shiny, curved parts: test a dome or another diffuse source, with polarising filters where appropriate.

Budget as much engineering time for lighting as for the camera choice. Choosing a light after the mechanical design is fixed can leave no room for the angle the defect actually needs. If your inspection is glossy or reflective, read the glass bottle defect detection article: the fix-lighting-first method there applies to other specular parts, and the fabric defect detection guide covers continuous web inspection. Browse the machine vision lighting finder by shape and wavelength once you know the geometry you need.

Sizing the processing: sensor, smart camera or controller

Processing hardware follows from the number of cameras and the algorithm, not the other way round.

• One simple check with deterministic tools: a vision sensor or smart camera, no separate processing computer required.
• Multiple cameras, custom logic, PLC communications and logging: consider a vision controller.
• AI classification or segmentation: check supported tools, deployment format and measured inference time. Some smart cameras can handle these tasks; other applications need a controller, with training planned separately.

For a multi-camera cell, port count and light channels are practical differentiators.

For more demanding processing, compare controller resources against measured application requirements rather than assuming that a larger controller will meet the cycle time.

If your application leans on trained models, plan the deployment format early. The ONNX anomaly detection article shows how detection maps become pass/fail decisions in practice, and the open source machine vision software guide explains what you take on when you build on non-commercial tools. Use the vision system configurator to plan the component selection.

What to source yourself and what to hand over

You can own these decisions with the arithmetic above: field of view, focal length, resolution, shutter type, lighting geometry, and the camera or controller family. Order samples, set the parts on a bench, and confirm the defect is visible with the light you chose. A bench test with a C-mount camera, a lens and a lamp can answer questions that component specifications alone cannot resolve.

An integrator should own the agreed integration deliverables: cycle-time guarantees, mechanical mounting and guarding, the PLC or robot handshake, and acceptance criteria written as measurable pass/fail performance. Machine safety responsibilities and validation must be explicitly assigned to qualified parties. When you request quotes, attach your one-page specification and ask each bidder to state assumed defect contrast and assumed part positioning. Missing assumptions need clarification before you compare proposals. Before any integration meeting, our article on vision system diagnostics to PLC is worth reading, because the alarm and status mapping you agree on early helps keep the cell serviceable later.

Checks before you order

Run through this list with your specification in hand:

1. Field of view includes the worst-case part position, verified with a physical mock-up or at least the stroke dimensions of the conveyor or fixture.
2. Resolution calculation uses the smallest defect, not the average one. Treat 3-5 pixels per defect and 20-50 percent additional resolution as starting estimates, then validate with representative samples.
3. Working distance fits the mechanical envelope with room to focus and to fit the light.
4. Lighting geometry is proven on sample parts, good and bad, on a bench before anything is mounted.
5. Exposure is short enough to limit motion blur. For less than one pixel of travel, exposure must be shorter than object-space pixel size divided by speed. At 100 mm/s and 0.05 mm/pixel, that is under 0.5 ms.
6. Interface agreed: protocol, handshake, and what happens on a timeout, documented before delivery.
7. Spare capacity: consider one unused camera port on the controller and mechanical space for one more light where expansion is likely.

When the usual answer is wrong

Three scope patterns should change your plan:

• The defect has no visual signature. Hidden internal cracks, some coating-thickness requirements or torque errors may need a different sensing method. More image pixels cannot reveal a hidden property; define the observable measurement and consult a specialist before buying cameras.
• The part position is uncontrolled by more than the field of view. Fix the part or add a wider field of view and more resolution; do not plan on software finding a part that can sit outside the image.
• The line runs faster than the inspection allows. Check acquisition, exposure, transfer and processing separately. Pixel count × frame rate gives pixel throughput; multiply by transmitted bytes per pixel to estimate image payload rate. Continuous web inspection may suit line scan, but changing camera type does not automatically remove processing limits. The line scan line rate calculator below helps estimate the required acquisition rate.

Finally, resist scoping by camera megapixels. A well-lit image with adequate sampling is more useful than a higher-resolution image that hides the defect in glare or blur. The lighting, lens and mechanics you specify today determine whether the inspection can meet its acceptance criteria.

Bring your specification and representative good and bad samples to our applications support team before finalising the component list.