You are inspecting a moving cavity of shiny metal, hundreds a minute, for a fleck you must not miss. Beverage-can inner-wall inspection is defined by three physical facts — it is a recessed cavity, the surface is specular, and the line is fast — and every one of them fights the camera before the model is even considered.
Inspecting the inside of a beverage can for contamination, coating gaps or foreign matter is deceptively hard. The target is a recessed cavity, the aluminium is specular so it throws glare instead of showing detail, and the line runs at hundreds of cans a minute, so each can offers a few milliseconds of view.
Get the optics and lighting wrong and there is nothing for a model to work with: glare hides a defect, a shadow invents one, and motion blur consumes the resolution. The imaging problem is the inspection problem; the model is downstream of it.
So the system is engineered from the light inward: illumination geometry that reads a curved specular cavity without glare, capture fast enough to freeze the can, and inference sized to answer within the few milliseconds the line allows before the reject.
Stated before any of them had an answer.
The defect is on the inner wall of a recess; getting light in and an image out without glare or shadow is a genuine optical design task.
Polished aluminium reflects the source rather than revealing the surface; the wrong geometry produces a bright image with no usable information.
Hundreds of cans a minute means a few milliseconds per can for capture, inference and decision — the whole loop has to fit that window.
If exposure and strobe are not short enough, the resolution the optics bought is lost to blur, and small defects vanish.
A camera pointed at a shiny cavity sees glare. The engineering is the lighting that turns a specular recess into an image a model can read — inside a few milliseconds.
The scope of the work, rather than a capability list.
Rarely the subsystem that sounds difficult.
This is the crux; a curved, shiny cavity is one of the harder machine-vision imaging problems, and no model recovers a badly lit image.
Capture, transfer, inference and decision all have to fit inside a few milliseconds, which constrains camera, resolution and model simultaneously.
At line speed even a short exposure can blur; strobe timing and exposure are budgeted against the resolution the defect needs.
A verdict is useless unless it reaches the can it belongs to; part tracking on a fast line survives stops and restarts.
A clean, glare-free image of the inner wall that a model can actually work from.
Capture-to-reject inside the cycle the line runs at, not a slowed-down demo.
Small contamination visible because the resolution survived exposure and strobe design.
Every rejected can logged and accounted for, fired against the right can.
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture, confidential deliverables and commercial terms are not disclosed. Where a detail would identify a customer it is omitted rather than approximated. More is available under a non-disclosure agreement, within the limits the customer has agreed.
Every item links to its own page.
Three physical facts make it hard: the target is a recessed cavity, so light and imaging have to reach a curved inner wall; the aluminium is specular, so it reflects the light source as glare instead of showing surface detail; and the line runs at hundreds of cans a minute, leaving milliseconds per can. Each of these fights the camera, and the imaging has to be solved before a model can help at all.
No amount of model quality recovers an image that has no information in it. If glare hides the defect or motion blur erases it, the model has nothing to learn from. The accuracy is won in the lighting, optics and exposure that turn a specular cavity into a readable image; the model is downstream of that and only as good as its input.
By budgeting the cycle first. The few milliseconds each can allows constrain the camera resolution, the strobe and exposure, and the model size together, and inference runs on edge hardware beside the machine so there is no cloud round-trip. Part tracking ensures the verdict reaches the correct can at the reject point even through line stops.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.