Good product training and AI model deployment

Traditional inspection projects stall in the same place: you need labelled examples of every defect before the model works, and on a high-yield line those examples arrive slowly. Changeover to a new product restarts the problem, and the labelling process itself is slow and error-prone.

The Good Product Training Model, built on supervised and unsupervised deep learning, inverts this. The anomaly detection model uses only good-quality images to perform pixel-level detection and whole-image classification of both known and unknown defects, so fast online verification is possible without a defect library. Major defects do not require labelling, product changeover is quick, sample training time is short and serious defects are not missed.

For small-defect scenarios, an anomaly defect detection model adds annotated data on top to optimise the result further. Combined with existing defect, classification, segmentation and object detection models, the bidirectional training approach improves comprehensive recognition regardless of defect size and reduces the missed detection rate.