On-Premise AI Visual Inspection: Edge Inference With No Cloud
On-premise AI visual inspection with no cloud: DeepVision runs on an edge computer beside the line, about 22 ms inference, images and decisions stored on site.
DeepVision inspects parts on a computer inside your plant. Images go from the camera to an edge computer beside the line, the pass/fail decision goes to your PLC, and the images and results are stored on your own hardware. Nothing in the inspection loop depends on a cloud service or an internet connection.
For the wider comparison of the two approaches, read edge AI vs cloud AI for factory visual inspection. This page sets out how a DeepVision deployment is laid out on site and what that means for your plant.
What runs where in a DeepVision deployment
| Component | Where it sits | What it does |
|---|---|---|
| Camera and lighting | At the inspection station on the line | Captures an image of each part as it passes. |
| Edge computer | Beside the line, inside the plant | Runs the DeepVision model on each image and makes the pass/fail decision. |
| PLC | Your existing line controller | Receives the decision and drives the reject mechanism. Siemens, Allen-Bradley and Mitsubishi PLCs are supported. See PLC integration. |
| Images and records | Your hardware and network, on site | Every decision is stored with its image, so you decide how long records are kept and who can see them. |
The exact hardware, mounting and network layout depend on the line, so we agree them with your engineering and IT teams for each site before installation.
Why on-premise matters on a production line
Latency the line can rely on
Model inference takes about 22 ms on the edge computer. Because the image never leaves the plant, there is no round trip to a remote server and no wait that changes with network traffic. The decision reaches the PLC in time to reject the part at the station, not further down the line.
Plants with no or restricted internet
Many plants keep production networks separate from the office network or have no reliable internet at the line. DeepVision does not need either. It can run on an isolated production network, and inspection keeps running if an outside connection drops, because nothing in the loop depends on a remote server.
IT/OT security reviews
Production images show your parts, processes and volumes. With DeepVision they stay in the plant. Indus Vision is certified to ISO/IEC 27001:2022 for information security management, and we can share the certificate and answer your supplier security questionnaire during evaluation. See security and ISO 27001.
Traceability from the image itself
Each pass/fail decision is stored with the image it was made on. When a customer raises a complaint or an auditor asks about a batch, your quality team can look at the actual part rather than a tick in a log. This is especially useful in automotive and pharma, where an image-backed record supports supplier and audit requirements.
Training and model updates on site
DeepVision uses anomaly detection: it learns what a good part looks like from as few as 200 good images, with no defect labelling. See how anomaly detection works with only good images.
Products and processes change. A new variant, a new supplier’s material or a change in lighting can shift what “good” looks like. DeepVision can be retrained with new images from your own production, so the model keeps up with the line. How retraining is scheduled and who carries it out is agreed with your team as part of the deployment plan.
Questions your IT/OT team will ask
| Question your IT/OT team will ask | Answer |
|---|---|
| Do production images leave the plant? | No. Images are inspected on the edge computer and stored on your hardware and network. |
| Does inspection need an internet connection? | No. DeepVision can run on an isolated production network with no internet access. |
| What happens if the internet goes down? | Inspection carries on. Nothing in the inspection loop depends on a remote server. |
| How does it connect to the line? | Pass/fail decisions go to your PLC (Siemens, Allen-Bradley or Mitsubishi), which drives your existing reject mechanism. |
| What hardware will be installed, and where? | Cameras at the station and an edge computer beside the line. We set out the specific equipment and network layout for your site before installation. |
| Who can access the system and its records? | Records sit on your hardware, so access follows your policies. We review access arrangements with your team during planning. |
| Is the supplier’s information security certified? | Yes. Indus Vision is certified to ISO/IEC 27001:2022. We can share the certificate and complete your questionnaire. |
| How is the model updated? | DeepVision is retrained with new images from your production. The process is agreed with your team for each site. |
Frequently asked questions
Can AI visual inspection run without the cloud?
Yes. DeepVision runs model inference on an edge computer beside the line and sends the result to your PLC. No cloud service is involved in inspecting parts.
Is on-premise inspection slower than cloud inspection?
Usually it is faster for line decisions, because there is no network round trip. DeepVision inference takes about 22 ms per image on the edge computer.
Can DeepVision run on an air-gapped or isolated network?
DeepVision can run on a production network with no internet access. If your site has a formal isolation policy, we will review it with your IT/OT team and set out how the deployment fits it.
How do we keep the model accurate as products change?
Retrain it with new images from your production. Because it learns from good parts, you add images of the new good product rather than collecting defect samples. See our accuracy methodology for how results are measured.
Need inspection that stays inside your plant? See DeepVision AI visual inspection, read about security and ISO 27001, or talk to our engineers about the architecture for your line.