Visual quality inspection with computer vision: inspect every part, document every decision

Optical inspection tasks in manufacturing are still widely handled by eye or by sampling. We build camera-based inspection that assesses every part, documents each decision traceably and relieves operators of monotonous visual checks.

The situation

Visual inspection does not scale: where every part ought to be checked, time only allows samples, and two inspectors judge the same borderline case differently. Complaints are hard to substantiate later because the inspection decision was never documented.

Camera-based inspection solves this – but only if it copes with series production reality: rare defects, changing lighting, hard cycle times and line IT that has to be integrated.

Typical data basis

Camera images and image streams from the line, complemented by 3D and depth data depending on the task. Labelled defect examples from the past – naturally few for rare defect classes. OK/NOK decisions from the previous inspection as reference.

Crucial is a sufficiently large, cleanly captured set of good parts covering the permissible variant space – it is the basis for anomaly detection. Where synthetic training images make sense, the part's CAD data is added.

Why such initiatives fail

Little data for the defects that matter. Rare defects in particular provide hardly any training examples. The model has to detect reliably even on heavily imbalanced data.

Series production changes the image. Lighting, part variants and contamination shift during operation. Without continuous monitoring, even a strong pilot loses hit rate over the months.

Quality within the line cycle. The inspection has to deliver its decision within the cycle time – added inspection quality must not become the bottleneck.

AI meets production IT. Only the integration into PLC and MES turns the model into a productive solution. That takes close collaboration with the line builder.

Approach

Two routes per defect pattern. For defect classes with sufficient examples we use object detectors, transformer-based (RT-DETR) or from the YOLO family depending on the task. For rare and previously unknown defect patterns the solution works with anomaly detection: a vision foundation model pretrained with self-supervision, such as DINOv3, provides the image features, and assessment is based on the distance to the distribution of good parts. This route needs no examples of every individual defect type and localises the anomalous spot at the same time. Segment Anything speeds up labelling; where real images are missing, synthetic images rendered from CAD supplement the training set.

Measure instead of classify. Gap dimensions and comparable measuring tasks we solve with a measurement method: laser light-section or triangulation, edge extraction from the profile section and conversion back to millimetres via a documented calibration. Measurement system capability is demonstrated according to the customer's procedure.

Stable over time. Monitoring is part of the delivery: normalisation of brightness and colour per capture, continuous comparison of the image feature distribution against the training basis, alerts on deviation, plus a fixed inspection image set ("golden set") that runs daily. New part variants and disputed cases first run in shadow mode. Retraining is a defined, versioned process with sign-off by quality assurance – no silent background updates.

Edge deployment in cycle time. The model runs on industrial hardware, optimised and quantised for the cycle time via ONNX Runtime or TensorRT. Inspection decisions, confidence and image crop go to the line IT and into the archive.

What you get

Every part instead of a sample: spot checks become complete, documented inspection. The same standard for every part: inspection decisions are consistent and reproducible – independent of shift, operator or form on the day.

Every decision stays traceable: image and justification are archived automatically, creating a solid basis for later complaints. People handle the exceptions: staff concentrate on suspect parts and their rework.

The two decisive metrics we define together and measure in the pilot against the current inspection: the false reject rate – good parts the system reports as NOK – and the slippage – defective parts it lets through. The operating point between the two is a quality assurance decision, not a technical constant, and is fixed as the acceptance criterion.