Foreign Object Detection Method in Transmission Lines Based on Improved YOLOv8

To address the challenge of limited accuracy in drone-based foreign object detection on power transmission lines, this study proposes an optimized YOLOv8 method. To tackle the issue of non-overlapping between prediction and ground truth boxes in certain cases, the study adopts the WIoU loss function in place of the CIoU loss function. In order to maintain model performance while reducing computational complexity and minimizing information loss, the EffectiveSE attention module is introduced. By integrating the GFPN architecture, the dimensionality of the feature space is enhanced, effectively improving the feature extraction capability in the Neck section and achieving a lightweight Neck network. An ablation experiment for small object detection was conducted using a dataset of 6672 images of foreign objects on power transmission lines, collected in real-world scenarios, which includes four categories: bird nests, kites, balloons, and plastic bags, split into training and validation sets at a ratio of 8:2. The experimental results reveal that compared to the original YOLOv8 algorithm, the improvements in accuracy for detecting bird nests, kites, balloons, and plastic bags are 3.9%, 5.8%, 4.7%, and 2.5%, respectively, with an overall accuracy increase of 4.2%.