What a Machine Vision Feasibility Study Should Prove Before You Buy Hardware

Anyone can produce an impressive image of a clean sample under ideal light. The question a feasibility study answers is different: does the imaging separate acceptable from unacceptable product across the full range of variation the line will actually present, at the speed the line actually runs?

That means the study is designed around the marginal samples, the ugly substrate variant, the chilled product, the worst-case part position — not around the sample the sales engineer would have chosen.

Before any image is captured, write down what must be caught, what must not be rejected, and what is out of scope. Ambiguity here is the leading cause of projects that technically succeed and are commercially rejected.

• Each defect class named, with an example and a rough frequency if known.
• The acceptance boundary for each class: at what size or severity does it become a reject?
• Explicitly out-of-scope conditions, so nobody expects them at acceptance.
• Confirmation that experienced inspectors agree on the marginal samples — settled before, not during, the project.

Everything in the imaging chain is sized from the smallest feature you must judge: a stroke width, a hole diameter, a scratch width, a label edge offset. Measure it on real parts. Datasheet values and drawings are frequently optimistic relative to what production actually produces.

Minimum feature

Measured on real samples, with observed variation, not a nominal specification.

Field of view

Worst-case feature position plus presentation tolerance plus margin.

Pixels per feature

Design assumption stated explicitly (and later verified), because it drives sensor and optics cost.

Depth of field

Checked at both height extremes with the aperture that the illumination budget allows.

Blur budget

A stated fraction of the minimum feature, which sets maximum exposure at line speed.

• Good product spanning suppliers, lots, colors and seasonal variation where relevant.
• Defective product for each class in scope, retained physically and labeled.
• Marginal units, which are the hardest to obtain and the most decisive.
• Environmental variants: condensation, dust, residue, packaging that has been handled.
• For learned-model candidates, enough volume to build training and held-out evaluation sets — and a labeling standard.

The core of the study. Vary illumination geometry, wavelength, polarization, optics and exposure on the real samples, and record what the resulting images show — not just whether an algorithm passed.

• Multiple illumination geometries compared on the same samples: diffuse dome, low angle, coaxial, backlight, and polarized variants.
• Wavelength trials where contrast is color-dependent, including IR or UV where the substrate justifies it.
• Optics candidates checked for resolution at the field edges, not only at center.
• Exposure at line-speed-equivalent motion, either on a motion rig or with a strobe at the target duration.
• The winning configuration documented well enough to be reproduced: geometry, distances, angles, part numbers, settings.

The imaging can be perfect and the project still fail on time. Add up trigger latency, exposure, transfer, processing, verdict communication and reject actuation, and compare the worst case against the available window at maximum line speed.

Illustrative arithmetic. A 600 units/minute line gives a 100 ms unit period; if two views must be acquired and processed and the rejector sits 400 mm downstream, the decision budget is bounded by the encoder distance, not by the unit period. Working this out on paper early is what prevents a mechanical surprise at commissioning.

Acquisition

Trigger latency, exposure, readout, transfer time per view.

Processing

Worst-case, not average — including the slowest defect-class path.

Communication

Verdict to PLC, including fieldbus cycle time.

Actuation

Reject mechanism response and the distance available to act.

Margin

Explicit headroom, plus the specified behavior if the budget is exceeded.

The criteria should be written before the experiments, so the study cannot be retroactively declared a success. A no-go with a clear reason — the required contrast does not exist under any tested geometry, the resolution needed exceeds the available working distance, the reject window cannot contain the decision — is genuinely valuable output.

Often the honest result is conditional: feasible if part presentation is constrained, if the coder is relocated, if a dwell station is added. Those conditions belong in the report, because they are the difference between a system that works and a dispute.

• Representative images of each defect class under the recommended configuration, with the settings recorded.
• The recommended imaging configuration: optics, illumination geometry, sensor class, working distance, trigger method.
• Measured or estimated sensitivity on the sample set, with the sample count stated — never a rounded marketing figure.
• Timing and throughput budget with worst-case numbers.
• Station architecture concept: views, mounting, access, environment, safety and cleaning considerations.
• Integration requirements: trigger source, reject mechanism, recipe source, data interface.
• Explicit assumptions, exclusions and remaining risks.
• A validation plan for acceptance, including what the labeled panel must contain.

A feasibility study should be willing to conclude that a smart camera or vision sensor is sufficient. Single view, single SKU, stable presentation, generous reject window, no traceability requirement — that is a product purchase, not a system project, and saying so is part of doing the study honestly.

It is a systems project when views must be coordinated, when recipes come from a work order, when records must be retained, when the reject interaction needs a specified fail-safe, or when the imaging itself required non-obvious engineering to exist at all.