Generation of Synthetic Data for Deep Learning in Manufacturing Quality Control Systems

The automotive industry has improved its quality control systems for detecting manufacturing defects through the implementation of deep learning. However, collecting and preparing datasets to train AI models is considered a challenge, given the high diversity of product types and the wide range of defects that can occur. This paper demonstrates the possibility of using synthetic data to train inspection models, and how it reduces the time and the cost involved in data collection, preparation, and cleaning. We describe how we create a virtual environment to generate synthetic images and how we prepare the 3D model of products to render auto-labelled training images with high realism and variety. We demonstrate in two industry use cases that a deep learning model trained with synthetic data can match the accuracy of one trained with real world data.