Abstract
BACKGROUND: The identification of foreign objects on transmission lines is crucial for their normal operation. There are risks and difficulties associated with identifying foreign objects on transmission lines due to their scattered distribution and elevated height. METHODS: The dataset for this paper consists of search material from the web, including bird nests, kites, balloons, and rubbish, which are common foreign objects found on top of transmission lines, totaling 400 instances. To enhance the classical U-Net architecture, the coding component has been substituted with a ResNet50 network serving as the feature extraction module. In the decoding section, a batch normalization (BN) layer was added after each convolutional layer in the decoder to improve the model's efficiency and generalization capacity. Additionally, a combined loss function was implemented, merging Focal loss and Dice loss, to tackle class imbalance issues and improve accuracy. RESULTS: In summary, RB-UNet, a novel semantic segmentation network, has been introduced. The experimental results show a mIoU of 88.43%, highlighting the significant superiority of the RB-UNet approach compared to other semantic segmentation techniques for detecting foreign objects on transmission lines. The findings indicate that the proposed RB-UNet algorithm is proficient in detecting and segmenting foreign objects on transmission lines.