Fast cell detection and distortion correction for outdoor electroluminescence images

This paper proposes a fast and robust method for electroluminescence image preprocessing, where lens and perspective distortions are corrected, and individual cells in the module are detected. Our approach works with low-resolution (640 x 512 pixels) images, uses an image-to-image translation neural network, and leverages the geometric properties of a photovoltaic module. The fast computational speed of the neural network allows us to complete image analysis in under 0.5 seconds, which is ten times faster than currently published methods. In addition, the geometry-based postprocessing makes our approach robust to small misdetections in the neural network output.