Abnormal Object Detection with an Improved YOLOv8 in the Transmission Lines

The abnormal objects hanging on the power transmission lines, such as nests, kites and balloons, can cause discharge faults which would endanger the safety of the power grid, especially under severe weather conditions. This paper proposes a power transmission line foreign object recognition algorithm based on the DnCNN network and YOLOv8 network. First, a denoising convolutional neural network (DnCNN) is applied for the image preprocessing to enhance the clarity of the images. Subsequently, the efficient multi-scale attention (EMA) network is applied in the backbone of YOLOv8 model to improve the channel description capability along with better pixel-level attention on the feature maps. At last, an image dataset is developed using the current inspection images and public datasets, and various scenes (such as foggy weather and heavy rainfall) are simulated via the blurring techniques with noises to diversify the dataset. Experiments have been performed to demonstrate that the proposed model can achieve high accuracy with an average precision (mAP) of 82.9%. Experiments have also been conducted in adverse weather conditions with comparison of the original YOLOv8 model to confirm the higher detection accuracy of the proposed object detection algorithm.