AbstractLack of training data is one of the main problems when realizing optical surface inspection systems. In the best case, provision of enough representative training samples is difficult and most of the time expensive. In some cases, it is not possible at all. Here we present an alternative method where the surface defects are simulated. Thereby, we focus on metal surfaces in the microscale where diffraction phenomena start to play a major role. Ray tracing and scalar diffraction approximation methods are applied and compared.
Automatic Optical Surface Inspection of Wind Turbine Rotor Blades using Convolutional Neural Networks
