A benchmarking study of instance segmentation methods for photovoltaic cell defect detection using electroluminescence images

• One of the largest publicly available Electroluminescence (EL) datasets for defect detection in Photovoltaic (PV) cells. • To the best of the authors’ knowledge, this study is the first to adopt instance segmentation techniques to PV EL images, offering a comprehensive performance comparison of four state-of-the-art models: Mask R-CNN, ConvNeXt, ConvNeXt-V2, and Mask2Former. This analysis is benchmarked against all existing studies using the same dataset, UCF EL Defect. Notably, this work introduces ConvNeXt, ConvNeXt-V2, and Mask2Former for the first time in the context of automated, high-precision defect detection in PV cells. • Furthermore, this work delves into the architectural advancements of the aforementioned instance segmentation models, highlighting the innovations that have led to their superior performance. • While instance segmentation was not adopted by prior works, results show that ConvNeXt-V2 outperforms other segmentation models, achieving a segmentation mean Average Precision (mAP) (Intersection over Union (IoU) = 50 %) of 0.602, segmentation mean IoU (mIoU) of 0.650, and bounding box mAP (IoU = 50 %) of 0.629, making it a promising candidate for defect detection in PV cells. The efficiency of Photovoltaic (PV) panels is often reduced by defects, leading to performance and lifespan issues, and increased financial losses. Manual inspection methods are labour-intensive and depend on extensive data analysis. Automated defect detection using Electroluminescence (EL) imaging has been investigated in the literature. However, to the best of the authors’ knowledge, no previous instance segmentation-based analysis was conducted on PV EL images. Moreover, no benchmarking studies have compared instance segmentation models using a standardised dataset for PV defect detection. This study conducts a comprehensive benchmarking analysis of four instance segmentation models, Mask R-CNN, ConvNeXt, ConvNeXt-V2, and Mask2Former against all prior works that utilised the same benchmarking dataset. For the first time, it pioneers the use of ConvNeXt, ConvNeXt-V2, and Mask2Former for PV cell defect detection analysing their architectural innovations and performance. Utilising UCF EL Defect as a benchmarking dataset (one of the largest pixel-level annotated EL datasets), the research categorises defects into cracks, interconnection issues, contact defects, and corrosions. Experiments were conducted using the MMDetection toolbox, providing a modular and reproducible framework for model implementation and evaluation. While instance segmentation was not adopted by prior works, results show that ConvNeXt-V2 outperforms all other models, achieving a segmentation mean Average Precision (mAP) (Intersection over Union (IoU) = 50 %) of 0.602, segmentation mean IoU (mIoU) of 0.650, and bounding box mAP (IoU = 50 %) of 0.629 making it a promising candidate for PV cell defect detection. The code for this work has been made public to ensure transparency and reproducibility of the results.