Segmentation and Error Detection of PV Modules

An important technique for the quality assurance of photovoltaic (PV) panels during the manufacturing process is the analysis of electroluminescence (EL) images. Various defects can be detected such as scratches, cracks, and short circuit cells. In our paper we introduce a new EL image dataset of monocrystalline panels and compare different loss functions for the detection of defects with state-of-the-art YOLOv5 deep neural network. For the segmentation of cells and to be able to make a geometrical description of panels we use an image-gradient based optimization approach. These information lead to a detailed structural and qualitative characterization of PV modules.