What a Machine Vision Feasibility Study Actually Looks At

Machine vision is genuinely useful technology. Automated quality inspection, defect detection, measurement, guidance — these are real applications that work in real production environments. But whether a specific system will work in your environment, for your product, with your process, is a completely separate question from whether the technology works in general. That gap is where most vision projects run into trouble.

A feasibility study is how you close that gap before committing a budget. It is not a vendor demo, and it is not a pilot project. It is an independent assessment of whether the proposed solution is technically viable, what the real requirements are, and what risks remain if you proceed. Done properly, it takes two to four weeks and gives you a clear recommendation with the reasoning behind it.

Because Kitron evaluates systems rather than building or selling them, a feasibility study from us is structurally different from what a vendor provides. A vendor's job is to show you that their solution works. Our job is to find out whether it will work for you — including the scenarios they might not volunteer.

The thing most buyers underestimate: data

Of all the factors that determine whether a machine vision project succeeds, data is the one that surprises people most consistently. Not because it is obscure, but because the gap between what companies think they have and what is actually useful for training a vision system is almost always larger than expected.

A system logging images around the clock produces records, not training data. The difference between the two is human effort: someone reviewing each sample, deciding what it represents, and encoding that judgment in a label the model can learn from. If your defects are rare — and in most production environments they are — you may have hundreds of thousands of normal samples and only a few dozen examples of the failure mode you actually care about detecting. That is not a dataset ready for training a reliable classifier. A feasibility study surfaces this early, when it is still a planning problem rather than a project failure.

Environment conditions are the second most common source of problems

What works in a vendor's controlled demonstration environment often behaves very differently on an actual factory floor. Lighting varies throughout the day and across seasons. Surfaces accumulate dust and oils. Products have natural variation that a demo sample set does not capture. Vibrations, temperature changes, and background movement all affect imaging performance in ways that are difficult to predict without looking at the actual production environment.

A good feasibility assessment includes a detailed audit of the imaging environment: what the lighting situation looks like, what the variation range of the target product is, what the line speed and throughput requirements are, and what integration the system needs to make with existing equipment and software. These factors collectively determine whether a vendor's quoted accuracy figures are realistic for your specific case or just for theirs.

What the output of a feasibility study should look like

At the end of a proper feasibility study, you should have a clear go or no-go recommendation, not a list of considerations. The recommendation should be backed by a specific assessment of your data situation, your environment, the technical approaches available, and an honest view of what risks remain if you proceed. If the answer is go, the study should tell you what a realistic implementation looks like, what it will cost, and what the key success conditions are. If the answer is no-go, it should tell you why and whether there is an alternative approach worth considering.

What it should not be is a document that arrives at "it depends" or hedges every conclusion with caveats. The value of an independent assessment is precisely that someone is willing to give you a straight answer based on your specific situation, not a generic framework you could have found yourself.