The product gets measured. The handling doesn't.
On a roll-to-roll line, one part of the process is instrumented and the rest is not. Metrology measures the material as it is produced. Everything before and after that point is handled by people and equipment, and a good deal of what happens there leaves no record at all. That is not a gap in diligence. Those steps were never designed to be measured. They are simply where material gets damaged after it was already made correctly.
The clip below is a method, not a match.
It shows Vaidio running a custom-trained model on a moving line. The material is not lithium film and the industry is not batteries. What transfers is how the model was built and what it was taught to look for.
Recovered material moving on a conveyor, inspected by a model trained on that producer's own images of what is acceptable and what is not. Every item the model identifies is bounded and labeled as it passes. The labels on screen are the class names from that customer's model, not generic ones.
15 seconds. The original runs two minutes and shows the same thing.
The producer defines what counts as acceptable.
A producer supplies image sets showing the material in acceptable condition, and in the specific conditions they want caught. Those images are packaged as datasets and sent for training. The trained model comes back and runs against the camera feed.
Once it is running, it can label and bound objects in new images on its own. That shortens the work of assembling the next round of training data, so the model keeps improving on the producer's own material rather than staying fixed at the accuracy it shipped with.
The important part is who sets the standard. It is not a vendor's general-purpose model deciding what good looks like. It is the producer's own standard, taught from the producer's own images.
Why this clip is not a claim about film.
What a camera can resolve is set by its optics, its sensor, and how far it sits from the subject. A bale two feet across, viewed from several meters away in ordinary light, is well within reach of a standard camera. A defect on a film measured in microns is not. That needs magnifying optics, controlled illumination, and usually a specialized sensor.
Those two requirements pull against each other. A wide field of view and extreme magnification are opposites. It is the difference between a security camera and a microscope. Both are cameras, and neither does the other's job.
So what carries across from the clip is not the camera. It is the platform, the training workflow, and the way results get routed.
Which is why the interesting question isn't about the film.
Metrology on the film is a solved discipline, and it is not somewhere an outside vision platform belongs. The handling stages are different in kind. Their failures are intermittent rather than continuous, so a splice happens once per roll and a forklift catches a roll once a quarter. They happen at ordinary scale, inches and feet rather than microns. They involve people, and in-line metrology does not watch people.
And the useful answer is usually a sequence rather than a number. Not how thick, but what happened, in what order, and when. That is what video is for.
Every one of those steps happens more often as output grows.
What a trained model can check at that scale.
None of the following requires resolving microns. These are all visible at ordinary camera scale, which means a model can be taught to flag them and to log when they occurred:
- Wrinkles, creases, and fold-overs in the web as it runs
- Edge damage and surface scratching
- Splice events: that one happened, when, and where it lands in the finished roll
- Winding faults such as telescoping, starring, edge weave, or an uneven roll face
- Roll condition after handling: impact damage, crushed cores, wrapping torn or missing
- Whether a finished roll is where it should be, and how long it has been there
- Changeover: when it started, when it finished, and whether the sequence was followed
Whether any of it matters is a question we can't answer from outside.
These are the ones worth putting to your own team. If the answers are all comfortable, this isn't for you:
- How do you know where a splice sits inside a finished roll?
- How does scrap from winding and handling compare with scrap from deposition?
- How much of total loss happens after the material was already made correctly?
- How much does changeover time vary between crews and between shifts?
- If a finished roll sat somewhere it shouldn't for two days, how would anyone find out?
- When material comes back from a customer, what record exists of what happened to that specific roll?
You may already have cameras on all of it.
Many roll-to-roll operations do. What they usually find is that the cameras record and nothing interprets. Someone reviews the footage after a problem has already been discovered, which makes it a forensic tool rather than an operational one.
If that is the situation, the question is not more cameras. It is whether anything is watching the ones already installed. Vaidio runs on existing cameras and existing recorders, so the starting point is usually a feed you already have rather than hardware you don't.