Synthetic data generation can effectively replace scarce real-world defect images for training industrial quality control systems, achieving strong performance on real data while eliminating costly manual annotation.
This paper presents a synthetic data generation framework that automatically creates realistic images of printing defects (creases, streaks, misregistration) with annotations for training object detection models. The framework solves a critical problem in rotogravure printing: the extreme scarcity of real defect images needed to train deep learning models.