Improving the Accuracy of Pixel-level Defect Classification of Electroluminescence Images

This work tackles one particular challenge encountered when applying machine learning model to automatically detect defects in electroluminescence images of PV modules. We propose a way to add an synthetic dataset generated by physics-based simulations to solve the class imbalance issue. The results show that the precision can be improved from 0.66 to 0.76, recall can be improved from 0.62 to 0.78 and f1-score can be improved from 0.64 to 0.77 when adding the simulated images to the training dataset.