MultiSolSegment: EL images and masks

(Sandia National Laboratories)

This dataset was used to train MultiSolSegment, a multi-channel segmentation model for photovoltaic defect detection. It contains electroluminescent (EL) images of monocrystalline solar modules from four sources:

• Arizona State University
• Case Western Reserve University (DuPont and SunEdison datasets)
• Lawrence Berkeley National Laboratory (part of the pv-vision effort)
  
  All images are 256×256 RGB and normalized using ImageNet mean/std values. Images were augmented via flips along the x, y, and both axes, increasing the dataset size by a factor of 4. The combined dataset contains 2,340 images.

The classes in the data are as follows:

• Dark regions
• Busbars
• Cracks
• Non-cell areas
  
  Labels were annotated in Supervisely and converted to multi-channel binary masks (NumPy .npy format, 4 channels). Each channel represents one class; pixels can belong to multiple classes.