MFPN-ASPP: MobileNet feature pyramid with ASPP for photovoltaic panel defect segmentation

Abstract Photovoltaic (PV) fault detection through thermal infrared imaging is a crucial step for efficient operation and maintenance of large-scale solar plants. Although current deep learning methodologies demonstrate potential, their efficacy is limited by inadequate feature extraction on various scales. This research presents an enhanced semantic segmentation model that combines a feature pyramid network (FPN) with a lightweight MobileNetV2 backbone and an atrous spatial pyramid pooling (ASPP) module. The model is trained, evaluated, and tested on the Photovoltaic Thermal Images dataset. The suggested method achieved competitive performance, with a mean Dice coefficient of 0.8440 and an average IoU of 0.7564, surpassing U-Net, LinkNet, FPN, and Mask-RCNN. The results highlight the efficacy of lightweight encoder–decoder frameworks that incorporate atrous spatial pyramid pooling, enabling accurate and computationally efficient fault detection in photovoltaic installations.