A YOLOv11-BASED APPROACH FOR THERMAL FAULT DETECTION IN PHOTOVOLTAIC PANELS USING ZERO-DCE AND SWIN TRANSFORMER

Early detection of thermal anomalies in photovoltaic (PV) panels is crucial for minimizing energy losses and improving maintenance planning. However, thermal images acquired under real-world field conditions often suffer from low contrast, non-uniform illumination, and noise, which significantly hinder the detection of small-scale faults. In this study, the effects of two image enhancement approaches Zero-Reference Deep Curve Estimation (Zero-DCE) and Swin Transformer for Image Restoration (SwinIR) on a YOLOv11-based object detection model are comparatively analyzed. The Zero-DCE method optimizes illumination and contrast without requiring reference images, whereas the SwinIR model focuses on noise suppression and preservation of structural details. To quantitatively evaluate the impact of both enhancement strategies on YOLOv11 performance, a series of experiments were conducted using identical training and testing protocols. The YOLOv11m model trained on raw thermal images achieved mean mAP@0.5 and mAP@0.5:0.95 scores of 0.849 and 0.748, respectively. When Zero-DCE-enhanced inputs were used, the mAP@0.5 increased to 0.943, while the mean recall reached 0.934, indicating an improvement in anomaly detection sensitivity. In contrast, the SwinIR-enhanced model achieved mean mAP@0.5 and mAP@0.5:0.95 scores of 0.843 and 0.755, respectively, demonstrating a limited improvement at higher IoU thresholds. The experimental results clearly indicate that different image enhancement strategies have a significant influence on fault detection performance. Moreover, the selection of an appropriate enhancement method plays a critical role in reliably identifying small and low-contrast anomalies in thermal PV imagery.