Most visual inspection systems on production lines today were built for one part, one camera, and one set of conditions. They work well on the day they are commissioned. Then the line changes: a new part revision, a different supplier, a lighting fixture replaced during maintenance. The system that was tuned for yesterday starts flagging good parts, or worse, passing bad ones.
This is not a failure of any single algorithm. It is a consequence of how these systems are built. When inspection logic, camera setup, and decision rules are fused into one custom project, every change becomes another project.
Where rigid systems break
- New part revisions. A small change in geometry or finish can invalidate thresholds that were hand-tuned for the previous revision.
- Changing conditions. Lighting, fixturing, and camera position drift over time. A system with no way to see or adjust each step hides why its results changed.
- No path for feedback. When an operator knows a flagged part is fine, that knowledge usually ends at the station. The system never learns from it.
- Opaque decisions. A pass or fail with no supporting evidence is hard to trust, hard to audit, and hard to improve.
An accurate model is only the beginning. Inspection becomes useful when hardware, software, data, and the people on the line work together as one system that can adapt.
Inspection as a workflow, not a project
Entroptics Vision treats inspection as a composable workflow. Image acquisition, alignment, scoring, classification, and the final decision are separate steps connected in a visual flow. Each step can be examined on its own, so when a result looks wrong, you can see which step produced it.
Because the steps are separate, adapting to a change is a matter of adjusting the step that changed. A new camera view means updating the acquisition and alignment steps. A new defect type means updating classification. The rest of the flow stays as it was, and the new version is compared against the previous one before it reaches the line.
Inference where the parts are
Inspection results are needed at line speed, at the station. Vision runs inference on an on-site edge appliance, managed by site, line, and station, with visibility into each device's health and assigned model version.
Production cannot stop because a network link does. If a station temporarily loses its connection to the platform, an enrolled edge device keeps running local inference with its last assigned pack. Results are buffered and replayed when connectivity returns. Model changes and fleet coordination wait for the reconnection.
Start small, then extend
The most reliable way to evaluate any inspection approach is on a single, well-understood task. Pick one part and one inspection. Gather representative images, examples of acceptable parts and real defects, and your acceptance criteria. Evaluate on held-out examples the model has never seen.
Once that first inspection holds up on the line, the same workflow building blocks carry over to the next part and the next station. That is the difference between a system that is configured and one that has to be rebuilt.