Automatic Crack Segmentation in Electroluminescence Images of Solar Modules and Maximum Inactive Area Prediction

Cracks on solar cells can cause degradation of photovoltaic (PV) modules, and electroluminescence (EL) images are a common technique for identifying cracks. However, to process a large number of such images it is necessary to develop automated routines for analysis. This article introduces a fast semantic segmentation method (~0.18s/cell) to segment crack lines, cross cracks, busbars, and dark areas on EL images of PV modules. We trained a UN et neural network model on a training set of 1,272 images, and we evaluated its performance on a validation set of 206 images and a testing set of 359 images. We report the performance on the testing set with an average F1 score of 0.875 and an IoU score of 0.782. We introduce our algorithm of predicting the worst-case degradation area with cracks detected. We also demonstrate our automatic preprocessing tool of cropping individual cell images from EL images of PV modules (~0. 72s/module). Our methods are published as open-source software and might be used to segment other kinds of defects or similar types of images by transfer learning in the future.