A benchmark dataset for defect detection and classification in electroluminescence images of PV modules using semantic segmentation

Abstract

Electroluminescence (EL) images enable defect detection in solar photovoltaic (PV) modules that are otherwise invisible to the naked eye, much the same way an x-ray enables a doctor to detect cracks and fractures in bones. This paper presents a benchmark dataset and results for automatic detection and classification using deep learning models trained on 24 defects and features in EL images of crystalline silicon solar cells. The dataset consists of 593 cell images with ground truth masks corresponding to the pixel-level labels for each feature and defect. Four deep learning models (U-Net_12, U-Net_25, PSPNet, and DeepLabv3+) were trained using equal class weights, inverse class weights, and custom class weights for a total of twelve sets of predictions for each of 50 test images. The model performance was quantified based on the median intersection over union (mIoU) and median recall (mRcl) for a subset of the most common defects (cracks, inactive areas, and gridline defects) and features (ribbon interconnects and cell spacing) in the dataset. The mIoU measured higher for the two features compared to the three defects across all models which correlates with the size of the large features compared to the small defects that each class occupies in the images. The DeepLabv3+ with custom class weights scores the highest in terms of mIoU for the selected defects in this dataset. While the mIoU for cracks is low (25%) even for the DeepLabv3+, the recall is high (86%), and the resulting prediction masks reliably locate the defects in complex images with both large and small objects. Therefore, the model proves useful in the context of detecting cracks and other defects in EL images. The unique contributions from this work include the benchmark dataset with corresponding ground truth masks for multi-class semantic segmentation in EL images of solar PV cells and the performance metrics from four semantic segmentation models trained using three sets of class weights.