Transmission Tower Re-Identification Algorithm Based on Machine Vision

Abstract

Transmission tower re-identification refers to the recognition of the location and identity of transmission towers, facilitating the rapid localization of transmission towers during power system inspection. Although there are established methods for the defect detection of transmission towers and accessories (such as crossarms and insulators), there is a lack of automated methods for transmission tower identity matching. This paper proposes an identity-matching method for transmission towers that integrates machine vision and deep learning. Initially, the method requires the creation of a template library. Firstly, the YOLOv8 object detection algorithm is employed to extract the transmission tower images, which are then mapped into a d-dimensional feature vector through a matching network. During the training process of the matching network, a strategy for the online generation of triplet samples is introduced. Secondly, a template library is built upon these d-dimensional feature vectors, which forms the basis of transmission tower re-identification. Subsequently, our method re-identifies the input images. Firstly, we propose that the YOLOv5n-conv head detects and crops the transmission towers in images. Secondly, images without transmission towers are skipped; for those with transmission towers, The matching network maps transmission tower instances into feature vectors. Ultimately, transmission tower re-identification is realized by comparing feature vectors with those in the template library using Euclidean distance. Concurrently, it can be combined with GPS information to narrow down the comparison range. Experiments show that the YOLOv5n-conv head model achieved a mean Average Precision at an Intersection Over Union threshold of 0.5 (mAP@0.5) score of 0.974 in transmission tower detection, reducing the detection speed by 2.4 ms compared to the original YOLOv5n. Integrating the online triplet sample generation into the matching network training with Inception-ResNet-v1 (d = 128) as the backbone enhanced the network’s rank-1 performance by 3.86%.