The camera is rarely the hardest part
Industrial vision has existed for decades, but modern computer vision models have changed what is practical. A neural network can classify subtle surface defects, locate components in cluttered scenes or identify patterns that are hard to encode as fixed image rules.
The temptation is to focus on the model. In production, image acquisition is usually the more important engineering problem.
If the lighting changes between shifts, if a part moves a few millimetres, if oil appears on the surface, or if the camera exposure drifts, the model sees a different world from the one used during training.
Define the inspection task precisely
Quality teams use words such as “defect” and “bad part,” but an AI project needs a more precise definition. Is the goal classification, object detection, segmentation or measurement?
A scratch may require segmentation if the defect boundary matters. Missing components may be handled as object detection. A simple binary classification can be enough when the camera view and product family are tightly controlled.
Choosing the task correctly reduces both model complexity and data requirements.
Data collection is where production knowledge becomes visible
A useful dataset is not simply thousands of random images. It should capture normal production variation and the defect variation that engineers care about.
Collect images across shifts, operators, material lots, machine conditions and reasonable lighting variation. Capture borderline cases. Keep examples of parts that were difficult for human inspectors. Those cases are often more valuable than obvious failures.
Most importantly, preserve traceability. Every image should be linked to a part, batch, product variant and production context when possible. Otherwise, it becomes difficult to investigate false decisions later.
The inspection station is a complete system
A computer vision cell typically contains lighting, camera, optics, triggering, image processing, model inference, PLC communication, reject handling and result logging.
The PLC usually needs a deterministic answer such as pass, fail, or retry. The AI model can produce probabilities or several candidate classes, but the final decision logic should be explicit. Timeouts and communication faults need defined behaviour. A lost network packet must not leave the reject mechanism in an ambiguous state.
False rejects can be as expensive as false accepts
Quality projects naturally focus on missed defects. Manufacturing also pays heavily for false rejects. A system that rejects too many good parts can reduce throughput and create unnecessary manual inspections.
The operating threshold should therefore be tuned against the real cost of errors. In some applications, the model should favour sensitivity. In others, it may be better to route uncertain parts to a human review station.
Edge inference is often the practical choice
Quality inspection frequently sits inside a production cycle where seconds or even milliseconds matter. Sending raw images to the cloud introduces network dependency and potentially undesirable data movement.
An industrial PC or edge GPU can run inference locally while sending only structured results upstream. This architecture also makes the inspection cell more independent of corporate network conditions.
Model maintenance does not end at commissioning
Products change. Suppliers change surface appearance. Cameras are replaced. Lighting ages. A model that worked for one product family can quietly degrade after a process change.
Production vision should therefore include performance monitoring and a process for reviewing difficult cases. Model versions should be tied to product variants and deployment dates, and the team should know how to roll back to a previous model.
What a professional deployment looks like
The best systems make the evidence visible. When an inspection fails, the operator should be able to see the relevant image, defect region and decision reason. Engineers should have access to trends and a clear path to flag an incorrect result for review.
Computer vision becomes a serious production capability when the camera, process, PLC, quality workflow and data pipeline are treated as one engineered system. The AI model is important, but it is only one component in that chain.
Lighting, optics and the physical scene
It is useful to think of a vision station as an optical measurement system rather than a camera attached to software. Lighting should make the feature of interest stable. Diffuse lighting can reduce reflections on shiny parts; directional light can make a surface discontinuity more visible. Lens selection, working distance and depth of field determine what the model actually sees.
For a production engineer, this matters because a model cannot compensate for every optical problem. If a scratch is visible only at one angle, a software change is unlikely to create a robust solution. The correct fix may be mechanical or optical.
Integration with the machine sequence
The inspection needs to be synchronised with the machine. The PLC may trigger an image after a sensor detects the part, wait for an inspection result and then route the part accordingly. The vision application should expose explicit states such as ready, busy, pass, fail, uncertain and fault.
A timeout should have a defined result. If the camera does not answer, the system should not silently treat the part as good. Depending on the risk, the correct response may be a controlled stop or a manual review path.
From model score to production decision
Most models produce a probability or confidence score. Production logic needs a policy. Some applications can accept a binary decision. Others need three classes: pass, fail and review. That third state is often valuable because it prevents the model from being forced to make a confident decision in conditions it has not learned.
This is a broader lesson for industrial AI: uncertainty can be represented as an operational state. It does not have to be hidden.
Commissioning checklist for a vision cell
- Verify the camera trigger and part-presence handshake under normal and delayed conditions.
- Test image capture with empty scenes, mispositioned parts and missing parts.
- Test PLC behaviour when the inference service is unavailable.
- Record model version and inspection result with the production identifier.
- Review false-reject cases with quality and manufacturing engineering.
Plan for product changeovers
A production line may inspect several variants that differ only slightly. The inspection logic should know which model or parameter set belongs to the active product, and the switch should be explicit rather than inferred from an ambiguous image.