Detecting defects on photovoltaic panels using electroluminescence images can significantly enhance the production quality of these panels. Nonetheless, in the process of defect detection, there often arise instances of missed detections and false alarms due to the close resemblance between embedded defect features and the intricate background information. To tackle this challenge, we propose an Adaptive Complementary Fusion (ACF) module designed to intelligently integrate spatial and channel information. This module is seamlessly integrated into YOLOv5 for detecting defects on photovoltaic panels, aiming primarily to enhance model detection performance, achieve model lightweighting, and accelerate detection speed. In order to validate the efficacy of the proposed module, we conducted experiments using a dataset comprising 4500 electroluminescence images of photovoltaic panels. Compared to the cutting-edge detection capability of YOLOv8, our YOLO-ACF method exhibits enhancements of 5.2, 0.8, and 2.3 percentage points in R, mAP50, and mAP50-95, respectively. In contrast to the lightest and fastest YOLOv5, YOLO-ACF achieves reductions of 12.9%, 12.4%, and 4.2% in parameters, weight, and time, respectively, while simultaneously boosting FPS by 5%. Through qualitative and quantitative comparisons with various alternative methods, we demonstrate that our YOLO-ACF strikes a good balance between detection performance, model complexity, and detection speed for defect detection on photovoltaic panels. Moreover, it demonstrates remarkable versatility across a spectrum of defect types.
Read full abstract