Deep learning finds the defects you can't write rules for

When you have plenty of labeled rejects, supervised classification or segmentation models locate and categorize each defect type. When rejects are rare — often the case on a well-run line — unsupervised anomaly detection learns what normal looks like and flags anything that deviates, needing only good parts to train. Models are exported through ONNX and optimized with TensorRT to run on an NVIDIA Jetson or industrial GPU PC at the line, inside a hard inference-time budget. In practice most systems are hybrids: deterministic rules handle presence and dimension checks while the network handles appearance.

Before a model touches production it is scored on a held-out image set, with the confusion matrix reviewed alongside your quality team so decision thresholds reflect the real cost of an escape versus a false reject. In production, borderline images route to a human review queue, and that queue becomes the retraining set — so the system improves as your product and process drift instead of quietly degrading.

Feasibility for a learned inspection is a data exercise. We capture a first image set on a bench rig with the lighting and optics we intend to deploy, because a model trained on phone photos or a different camera tells you nothing about production. Then we sit with your quality people and label a sample together, which almost always exposes disagreements about what counts as a reject. Resolving those in writing — with example images attached to each defect class — is the most valuable thing that happens in the whole project. A model can only be as consistent as the labels it learned from.

Deep learning does not excuse poor imaging. Consistent lighting, a fixed working distance, and a stable trigger position shrink the variation a model has to absorb, which means fewer training images and a smaller, faster network. On the line, inference runs inside the same PLC handshake as any other inspection: trigger, capture, infer, and answer within the cycle budget, with a watchdog that flags a missed decision as a reject rather than a pass. Every deployed model carries a version, the dataset hash it was trained on, and its validation scores, so when someone asks why a part was rejected last Tuesday, the answer is on record.