Deep Learning-Based Detection for Transmission Towers Using UAV Images

Abstract

With two network frameworks, Inception V2 and ResNet50, as feature extraction networks, Faster region-based convolutional neural network (R-CNN) and single-shot multibox detector (SSD) models were used to detect transmission tower targets based on VOC format datasets of forest areas. The results showed that the mean average precision (mAP) of transmission tower detection achieved using the Faster R-CNN model reached 90% and 94% based on the Inception V2 framework and the ResNet50 framework, respectively, while the mAP of the SSD model reached 84% and 92% with the Inception V2 framework and the ResNet50 framework, respectively. The Faster R-CNN had a longer detection time than that of the SSD, but its detection accuracy was higher than that of the SSD. Compared with that of ResNet50, the detection accuracy of Inception V2 was lower, while the detection time was shorter. Deep learning technology can effectively and rapidly detect the positions of transmission tower targets.