U-Net–Based Semantic Segmentation of Defects in Photovoltaic Panels

This article presents a study on the semantic segmentation of defects in crystalline-silicon photovoltaic cells using U-Net–based models trained on electroluminescence (EL) images. The dataset combines laboratory-acquired images with a publicly available benchmark, both manually annotated to identify cracks, dark zones, and collector-bar discontinuities. Eight model variants were trained with controlled variations in input resolution, encoder depth, and regularization strategies. Performance was assessed using per-class precision, recall, and F1-score, complemented by visual inspection through heatmaps and overlays and by expert validation. Segmentation was robust for defects with well-defined morphology, such as dark zones and busbars; however, cracks remained more difficult to detect due to their sparse pixel representation and irregular geometry. Alternative architectures (U-Net++ and MAU-Net) were also evaluated but did not yield meaningful improvements over the optimized U-Net configuration. Overall, the results support the use of this approach for automated inspection under controlled conditions, while highlighting the need for future adaptation to more diverse operational scenarios.