GAN-based data augmentation for crack detection

The application of Machine Learning for manufacturing has become a reality with the ongoing digitalization of factories. Generative Models are a promising but not yet exploited enough technology in manufacturing frameworks. In this work, we apply Generative Models for a quality inspection problem in order to generate synthetic data for training. We show how through the use of a Generative Adversarial Network, synthetic images are generated to augment the training data set of a crack detection system. We show the impact of such data generation in the detection ratios of our system and how for problems with inhomogeneous defect typologies distribution, Generative Models can provide a solution for artificial defect generation to enhance detection capabilities based on neural networks