A modified UNet network bridged with multiscale context fusion for photovoltaic panel image segmentation

The identification and segmentation of photovoltaic panel areas are crucial for precise defect detection and accurate component localization in photovoltaic power plants. To address the challenges of variable ground resolution, complex backgrounds, and ambiguous boundaries of photovoltaic panel areas in inspection images, we introduce a novel multi-scale adaptive photovoltaic panel image segmentation network named EB7-MCF-UNet. Our network draws inspiration from the proven UNet architecture, employing EfficientNet-B7 as the encoder and introducing a Multi-Scale Context Fusion (MCF) module as an encoder-decoder feature bridge. The efficient MCF module bolsters the network's ability to handle multiscale variations effectively. To further enhance model adaptability, we introduce data augmentation techniques to diversify the training data, allowing the model to better accommodate various scenes and lighting conditions. Moreover, the feature matching loss with linear weighting is incorporated alongside the cross-entropy loss for EB7-MCF-UNet to minimize the discrepancy between predicted features and labeled maps. Extensive experiments on practical inspection images photographed in various scenes demonstrate the effectiveness of EB7-MCF-UNet for photovoltaic panel image segmentation.