Data Department builds computer vision models and software for quality inspection, counting, safety and process monitoring in manufacturing and logistics. They run on the cameras you already have or plan to install. We cover model development, the data labelling plan, integration with your MES (manufacturing execution system) and acceptance testing. Cameras, lighting and installation stay with your hardware supplier.
What problems does computer vision solve?
Vision projects should start from a decision on the line, not from a camera. The decisions we work on:
- Pass or reject a part, before it reaches the next station or the customer.
- Which defect type it is, so it goes to the right rework step.
- How many items, boxes or pallets passed a point, without manual counting.
- Whether a label, batch code or expiry date is present and readable, using OCR (optical character recognition).
- Whether a person is in a restricted zone or missing PPE (personal protective equipment).
- Whether fill levels, positions or assembly steps are correct.
Manual visual inspection gets tired, varies between shifts and is hard to audit. A vision system applies the same rule to every item and keeps an image record of each decision.
What a project looks like
Discovery (2 weeks). We review sample images or video from your line and build a defect catalogue with your quality team: what each defect looks like, how serious it is, and where people disagree. We check whether the current camera position and lighting can see the defects at all. If they cannot, we tell you what to ask your hardware supplier for. You get a go/no-go decision and a labelling plan.
Data labelling. Your quality experts define each defect. Labelling can be done by your team, by us, or by a labelling team we have worked with, but your experts review a sample, because the model can only be as consistent as its labels. This step is often the longest in a vision project, and we plan for it honestly.
Pilot (6–10 weeks). We train models and test them on images the model has never seen. Then the system runs in shadow mode on the line: it makes decisions, but people still decide, and we compare the two.
Acceptance testing. Before go-live, we agree two numbers with you on a test set collected on your own line: the share of real defects the system misses, and the share of good parts it rejects. These pull against each other, so the right balance is a business decision. You sign off on measured results, not on a demo.
Production and handover. The model runs on an edge device near the line or on a server, connected to your PLC (programmable logic controller), MES or quality database. We add monitoring and a retraining process for when products, materials or lighting change.
Which vision approach fits which task?
| Task | Typical approach | What you need to provide |
|---|---|---|
| Good or defective | Image classification | Labelled examples of good parts and each defect type |
| Where is the defect? | Object detection or segmentation | Images with defect areas marked |
| Rare or unknown defects | Anomaly detection trained mainly on good parts | Many images of good parts, a few defects for testing |
| Read text and codes | OCR | Images of labels under real line conditions |
| Counting and tracking | Detection plus tracking on video | Video from the counting point |
Proof and experience
We do not yet publish a computer vision case study on this site, and we will not invent one. In a first conversation we can walk you through how we would measure success on your line, and the acceptance test protects you either way: if the system does not meet the agreed numbers on your own parts, it does not go live.
Vision often connects to other work. Defect data can feed production optimisation, and document-heavy quality processes such as certificates are better handled with generative models that read documents. See how this fits a plant in manufacturing.
When is computer vision NOT the right tool?
- The camera cannot see the defect. If lighting, angle or resolution hide the problem, fix the optics first. No model can find what is not in the image.
- A simple sensor does the job. Weight, height or presence checks are often cheaper with a load cell or laser sensor.
- Volume is low. On a slow line with few parts, a trained inspector may be the better choice.
- People disagree on the defect. If your own inspectors label the same part differently, agree the definition before training a model.
- You need a turnkey hardware supplier. We do not supply or install cameras.
Frequently asked questions
How many images do we need?
For supervised models, typically hundreds of examples per defect type. For rare defects, anomaly detection can work with mostly good-part images. Discovery gives you a concrete number for your case.
Do you supply cameras and lighting?
No. We specify what the model needs and work with your hardware supplier, or with the cameras already installed.
Can it run without the cloud?
Yes. Models can run on an edge device at the line or on servers in your plant, with no image leaving your site.
What drives the cost of a vision project?
The number of defect types, how much labelling is needed, the line speed and response time, how many stations and cameras are involved, and how deep the integration with PLC and MES goes.
What if people appear in the images?
Then KVKK (Türkiye's Personal Data Protection Law) or GDPR (General Data Protection Regulation) applies. We design to avoid identifying individuals where possible, agree retention periods, and process images on your premises.