Automated visual inspection is increasingly important for repairing and recycling metallic components. AI-based defect detection is a natural fit for this task, but its performance depends heavily on the availability of training images. In practice, those datasets are often scarce and lack diversity, which can limit how well models generalise to real-world defects.
To address this, the paper proposes using ProGAN, a type of generative adversarial network, to create synthetic images for training. The title and abstract indicate that the goal is to supplement existing datasets with generated examples, giving defect-detection models more varied visual input without requiring extensive manual data collection.
The work is described in an arXiv preprint, so it has not yet undergone peer review. Still, the direction is notable: synthetic data generation could be a practical route toward more robust automated inspection in industrial settings where real defect images are hard to obtain.