An Improved YOLOv8 Network Architecture for Enhancing Detection Accuracy of Bird Nests with Small Size and Intricate Background on Power Transmission Lines

Abstract

The harm caused by birds building nests on power transmission lines is enormous, greatly increasing the possibility of short circuits and damages to the transmission lines. Bird nest detection on transmission lines presents significant challenges due to the small size of nests and complex backgrounds. To address these challenges, this paper proposes an improved YOLOv8 model specifically designed for bird nest detection. The model introduces several key innovations: the inclusion of an exponential moving average mechanism in the C2f module to enhance feature extraction and improve detection performance; the development of a dynamic weighted feature fusion network to more effectively integrate contextual features, significantly reducing both parameter count and computational load, thus achieving network lightweighting; and the introduction of weight-CIOU as a bounding box loss function, which better measures the similarity between targets, accelerates convergence, and enhances detection accuracy. Experimental results demonstrate that the improved YOLOv8 model outperforms the original YOLOv8 by achieving a mAP of 96.3%. Visualization of detection results further confirms the model’s robustness in various challenging scenarios, highlighting its potential for real-world deployment in power line monitoring systems.