UCF-EL-GDI

UCF-EL-GDI is a collection of synthetic photovoltaic electroluminescence (EL) images generated from the public UCF-EL-Defect dataset using Generative Defect Isolation (GDI).

GDI uses the original segmentation annotations to remove selected defects through neural inpainting, producing isolated single-defect samples as well as samples with no defect. For this dataset, only images containing at least one defect with an annotated area greater than 10% of the total image area were selected. This threshold is user-configurable in the generation code.

Full methodology, generation instructions, and implementation details are available in the project repository.

Links

• Paper: A Generative Approach for Improving Multi-Label Defect Classification in Photovoltaic Modules
• Code: mueez-overflow/generative-defect-isolation
• Original dataset: ucf-photovoltaics/UCF-EL-Defect

Example

The example below illustrates the GDI process, showing the original defect annotations alongside the generated single-defect and no-defect samples.

In this example, the source image contains four annotated defect classes. Of these, only Contact_NearSolderPad and Crack_Resistive satisfy the configured area threshold, resulting in isolated samples for these two defect classes along with a No_Defect sample.

Citation

If you use GDI, the generated synthetic data, or any part of this work, please cite:

@article{mueez2026gdi,
  title   = {A generative approach for improving multi-label defect classification in photovoltaic modules},
  journal = {Solar Energy},
  volume  = {317},
  pages   = {114943},
  year    = {2026},
  issn    = {0038-092X},
  doi     = {10.1016/j.solener.2026.114943},
  author  = {Abdul Mueez and Yogesh S. Rawat and Shruti Vyas}
}