Skip to content
Conference

CAS-YOLO: a lightweight detection network for photovoltaic panel defects in electroluminescence images

Oct 2026 · The International Conference on Optical Imaging and Deep Learning · Vol 14494, pp. 144940H - 144940H-9 · 0 citations · 14 references
Engineering

Abstract

Defect detection in photovoltaic (PV) panels is of great significance for ensuring power generation efficiency and the safe operation of power stations. To address the unstable performance of existing detection methods in identifying small defects, handling defect scale variations, and dealing with complex polycrystalline silicon background interference, this paper proposes an improved model based on YOLO11n, named CAS-YOLO. This method makes three key improvements focusing on the feature extraction and fusion processes. First, a Convolutional Haar Discrete Wavelet Transform (CHDWT) module is introduced in the shallow layers of the backbone network to enhance the preservation of high-frequency details of small defects during down-sampling. Second, the ADown module is employed in the deep down-sampling path to smoothly aggregate semantic information and suppress the propagation of grain boundary noise. Third, a C3k2_Star module is integrated into the high-resolution feature fusion branch in the neck network to enhance weak defect responses and fine-grained feature expression capabilities. Experimental results on the PVEL-AD dataset demonstrate that the mAP@0.5 and mAP@0.5:0.95 of CAS-YOLO reach 91.39% and 66.28%, respectively, representing improvements of 1.63 and 3.50 percentage points over the baseline model. Furthermore, the precision achieves 86.70%, the recall reaches 89.20%, and the parameter count is only 2.25 M. The results indicate that the proposed method effectively improves PV defect detection performance under complex backgrounds while maintaining the lightweight nature of the model, exhibiting strong potential for engineering applications.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.