FlowMarket
Where to find machine vision suppliers: FlowMarket is a B2B sourcing network powered by an AI agent. You describe the inspection — the part, the defect types, the cycle time, how parts are presented and what a false reject costs you — and the agent matches you with vision suppliers and integrators who solve that kind of problem, then requests quotes and lead times so you can compare them in one place.
Sourcing machine vision
Vision projects fail more often than any other kind of factory automation, and almost never for the reason people expect. The camera is rarely the problem and the software is rarely the problem. Projects fail on lighting and on part presentation — on a shiny surface that reflects differently every time, a part that arrives at a slightly different angle, a defect that is only visible from one direction nobody thought to light from.
That has a direct consequence for how you buy. A supplier who quotes from a written description without seeing your parts is guessing, and their price reflects the guess. A supplier who asks for samples and runs a feasibility study before quoting is doing the only thing that reliably predicts whether the system will work.
Start with a feasibility study
For anything beyond reading a barcode, a feasibility study is the cheapest insurance available. Send real parts — good ones and, more importantly, the defective ones, including the marginal cases you argue about internally. The supplier builds an imaging setup, demonstrates that the defect is visible and repeatable, and reports what accuracy is realistic.
Expect to pay for it, and treat a supplier who offers it free as a supplier planning to recover the cost elsewhere. What you receive is images, a stated detection rate and an honest statement of what cannot be seen — which is worth more than an optimistic quote, because it arrives before the purchase order rather than after.
What to specify
- The defects, with samples — named, categorised and physically available. 'Surface defects' is not a specification; a box of rejected parts is.
- False reject versus false accept — which error costs you more, and what rate is acceptable. No system achieves zero of both, and this trade-off drives the whole design.
- Cycle time and presentation — parts per minute, whether they stop or move, and how tightly their position is controlled. Position variation is the hidden cost driver.
- What happens on a reject — sorted automatically, marked, or flagged to an operator. This is an integration question that arrives late and expensively if left out.
- Environment — ambient light changes, vibration, coolant mist, wash-down. All of these break systems that worked in the demonstration room.
Rule-based or deep learning
Traditional rule-based vision measures things: edges, distances, blob areas, pattern matches. It is fast, deterministic, explainable, and the right answer for measurement, presence checking and code reading. If your inspection can be expressed as a measurement with a tolerance, this is what you want.
Deep learning classifies things that resist measurement — cosmetic defects on varied surfaces, natural products, anything a person recognises but cannot define. It needs training images, including enough examples of each defect, and it produces a probability rather than a number. That makes validation different and, in regulated industries, harder. Many real systems use both: rules for the measurements, learning for the judgement calls.
What a vision system costs
Camera and lens are a small part of it. Lighting, mounting, enclosure, the industrial PC or smart camera, software licences, integration with the PLC and the reject mechanism, and the engineering time to make it robust across shifts and seasons make up the rest. Systems that work in February and fail in July usually failed on ambient light through a skylight nobody accounted for.
Budget for the tuning period after commissioning too. A vision system reaches its final accuracy after it has seen a few weeks of real production, including the parts nobody remembered to include in the samples. Agree what the acceptance criteria are and when they are measured, or that period turns into a dispute.
How the AI agent matches suppliers
The agent reads your requirement rather than your keywords. It searches the FlowMarket network and connected industrial directories together, scores every candidate on how well it answers what you described — inspection type, industry, cycle time, integration scope, geography — and drops anything that does not clear the bar. Each match comes back with a plain-language reason, which distinguishes an integrator who has solved your class of problem from a camera distributor.