Using high-resolution generative models for PV cells and module extraction: large-scale evaluation

Individual cell defect counts in electroluminescence (EL) images are often used as evaluation criteria for photovoltaic (PV) plant inspections. However, the significant variation in appearance and quality makes it challenging to extract individual PV cells accurately. To address this, we utilize a high-resolution generative neural network, reducing the problem of cell identification to a more straightforward analysis of a binary grid image. The network is trained on a relatively small (90 images), manually labeled dataset, with augmentations such as cell shuffling significantly improving model performance. We evaluated our method on a dataset of 102000 EL images, including low- and high-resolution images and various module technologies. Our findings show that even a limited number of carefully selected training images, combined with effective augmentations, can achieve a model performance exceeding 95%. Notably, the method remains effective for low-resolution half-cut-cell images, where the separation between module parts is just a few pixels.