Infrared Image Segmentation Method Based on DeepLabV3+ for Identifying Key Components of Power Transmission Line

Abstract

—To improve the work efficiency of on-site inspection personnel in diagnosing faults of power transmission lines, in this paper we propose an infrared image segmentation method based on DeepLabV3+ for identifying key components of transmission line. We collected 556 infrared images of transmission lines in our own power supply system, and expanded the original data by data augmentation method. Based on the comparison of the DeepLabV3+ model with three different backbone networks, MobileNetV2 with the best performance is selected as the main backbone network. Compared with FCN, U-Net, and SegNet, the test results show that DeepLabV3+ using MobileNetV2 (compared with ResNet50 and Xception) can segment the five types of key components in power transmission lines from infrared images more accurately and faster. The MIoU on the test set is 0.8624, which is better than the performance of FCN, U-Net, and SegNet. This lays a foundation for improving the work efficiency of on-site inspection personnel and improving the continuous power supply capacity, stability, and safe operation level of the power grid.