Good Product Training: How AI 2.0 Inspection Handles Defects You Have Never Seen

KeyeTech built the core technology of the industrial vision inspection AI 2.0 era, the Good Product Training Model, on two deep learning modules, supervised and unsupervised, combined with practical production experience. It divides into anomaly inspection and anomaly defect inspection.

The problems it targets are the ones that delay real projects: small defect samples, long model construction cycles, low product changeover efficiency and a complicated labelling process. Major defects do not require labelling, the product model can be changed quickly, sample training time is fast, unknown defects can be detected, defect sample collection time is short, and serious defects are not missed.

The anomaly detection model needs only good quality images. From those it performs pixel level detection and whole image classification of known and unknown defects, which achieves fast online verification. The anomaly defect detection model then addresses small defect scenarios by adding annotated data to optimise the detection model further and improve flexibility.

Combining the good product inspection model with existing defect models, classification models, segmentation models and object detection models improves detection performance and comprehensive recognition. Whatever the size of the defect, known or unknown, identification is fast and accurate, the missed detection rate falls, and the bidirectional training model raises overall detection efficiency.