Automated visual surface inspection relies on image processing techniques for detection of surface defects on products. Automated quality assurance in industrial setups relies on such techniques to cope with ever-increasing needs for production speed and adaptability to new scenarios. Traditional approaches to automated quality assurance quickly become a bottleneck due to their rigidity to specific scenarios. To solve this more robust and adaptable techniques are needed, which fuels the interest in artificial neural networks. However, for development of such models large quantities of data are required with extensive coverage of edge cases which can be difficult and/or expensive to acquire in real setups.
Our goal is to use synthetically generated data using modern physically based rendering techniques to develop robust and adaptable models for automated visual surface inspection. As synthetic data can differ from real data due to various external influences on the camera and assumptions in the rendering pipeline, we explore techniques for adaptation between data domains. The models tend to overfit on given data thus performing poorly in real setups. To this end we explore multiple techniques of regularization and visualization to produce more robust and trustable models. We expect these research directions to have a strong influence on the development and application of automated visual inspection systems for multiple industries using visual inspection for quality assurance.