Whatthisisaboutinpractice
Inspecting what comes in or goes out
A batch arrives and someone checks whether anything has to come out. That work is slow, it is done differently at the end of the day than at the start, and two people do it slightly differently.
Counting what is too much to count
How many items are in this photo, how many are in this crate. A person counts that accurately once and no longer on the tenth try.
Sorting into categories
Returns, waste, parts: it is not about pass or fail but about which of six bins. That is a different problem from inspecting, and usually an easier one.
Checking whether something is there
Is the label on, is the cap straight, is a part missing. Often the best starting point, because the question is sharp and the fault is easy to see.
Whatyousupply,andwhythatisthehardeststep
Hundreds of images per category. Not hundreds in total: hundreds per category, and that applies just as much to the category you are trying to catch. Twenty photos of a defect is too few to teach a model what a defect is, however good the rest of your material may be.
Someone has to say what is right and what is wrong. That is not something I can do for you: the model learns your judgement, and that judgement lives in your business. If two colleagues disagree about the same photo, that is not a detail but a problem to settle before training. Usually it turns out the standard itself was never written down.
And the conditions have to resemble what will happen later. Images taken on a phone by the window teach a model something about daylight. If it later runs under fluorescent light at a different height, that is a different problem from the one it practised on.
Howyouknowitworks,inplainterms
One number says nothing here, and that is not modesty but the heart of it. Two numbers do say something. First: of a hundred items the model rejects, how many are genuinely faulty? Second: of a hundred genuinely faulty items, how many does the model catch? Those two pull against each other.
Between them sits a dial that costs no retraining. Where you draw the line decides whether you catch more faults or throw away more good product. That is a business decision, not a technical one: what does a missed fault cost you, and what does a wrongly rejected item cost? With those two amounts the line can be calculated. Without them it is a guess, and then the software's default is as good as anything.
Measurement happens on images held back that the model has never seen, and holding back is done per batch, not per photo. Why that matters is below.
TwothingsIgotwrongmyself
The model that scored best was the worst model. Training automatically picked the version with the lowest error, and that turned out to be a model that simply approved everything. If nine out of ten items are fine, “everything is fine” is right ninety percent of the time. Beautiful on paper, catches nothing in practice. Since then I never look at a single number.
The measurement was an illusion at first. Photos from the same batch sat in both the practice set and the test. The model recognised the batch rather than the defect, and scored beautifully on material it effectively already knew. This is the most common mistake in the field and it is invisible from the outside: the figures do not look wrong, they look too good.
Both mistakes cost you nothing if you find them before delivery, and everything if you find them after. That is why this is a trial project with a measurement at the end, not a build with a handover.
Whatitcosts
One pass-or-fail judgement on images from a single fixed setup: same distance, same light, same background.
Up to five categories, or images that differ by source or by season. The model then has to learn what does not matter, and that costs more examples and more rounds.
Many categories, several models side by side, or exceptions you have to catch one by one because they are rare but expensive.
All amounts exclude VAT.
Whennottostartthis
If cameras, lighting and a link to line controls come into it, this is not the right address. Then you are talking about tens of thousands of euros and maintenance on site, and that is work for a firm with engineers. What I do stops at images you supply and a result you get back.
And not on people. Images where you want to recognise individuals in order to identify them is biometric data under the GDPR, and I do not take those projects on.
Looking for something else?
Questionsthatcomewiththis
- I have thousands of photos, but they are not judged. Does that count?
- That is a good start and not yet a dataset. A model learns from the judgement, not from the image; without it you have the material to practise on but not the answer. Judging can usually start as a sample, so after a few hundred you already know whether it is worth it.
- How accurate will it be?
- I do not know before I have seen your images, and anyone who does tell you is selling a guess with the authority of a number. What I can do: give an honest range after the first round on your own material, and then you decide whether to continue.
- Can the model run without an internet connection?
- Yes. The model is a file; it can run on a machine at your site instead of through an integration. That is a choice you make up front, because it determines how the model is packaged.
- What if my product changes?
- Then the model slips, and it does so without an error message. A model that has never seen a new packaging does not know that it does not know. That is what the care plan is for: measuring again each month on fresh images, and retraining when it slips.