A Comparative Study of U-Net Segmentation and YOLO Detection
Open Access DepositedAutomated Defect Detection in Photovoltaic Modules via Electroluminescence Imaging
YOLOv8x offers efficient and reliable detection suitable for high-throughput screening, whereas UNet16 provides essential pixel-level detail for in-depth spatial analysis despite higher computational costs. These findings provide empirical guidance for selecting appropriate models for automated EL inspection based on the application priorities.
Automated defect detection in Electroluminescence (EL) images is crucial for photovoltaic (PV) module reliability, yet comparative studies between leading deep learning approaches are scarce. This thesis directly compares semantic segmentation using UNet16 and object detection using YOLOv8x for identifying 'crack', 'cross', and 'dark' defects in a public EL dataset (2159 images). Four models (optimal and unified configurations for each architecture) were trained using standardized preprocessing and evaluated based on task-specific metrics (mIoU, mAP_{50}) and computational efficiency.Results show both architectures are effective, with Optimal YOLO achieving mAP_{50}=0.855 and Optimal UNet achieving mIoU=0.785. Task-specific optimization yielded superior results compared to unified settings. YOLOv8x demonstrated significant advantages in computational efficiency, exhibiting fewer parameters (68.2M vs. 143.6M) and faster inference (e.g., 11.2 ms vs. 42.1 ms @ 640 × 640) than UNet16. The study concludes that a clear trade-off exists
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Coulibaly_gwu_0075M_17338.pdf | 2025-07-20 | Open Access |
|