Improved Faster R-CNN Algorithm for Transmission Line Small Target Detection

Abstract

AbstractBird’s nests, as well as suspended foreign objects such as plastic and rags, are serious potential safety hazards on transmission lines. Because the Bird’s nests and suspended foreign objects in the unmanned aerial vehicle (UAV) images often belong to small targets with less pixels, more noise and easy to be disturbed, the detection of these objects puts forward higher requirements for the detection algorithm. In this paper, an deep learning-based algorithm are designed for the detection of these two kinds of small targets. By adding the attention mechanism module to the backbone network in this algorithm, the importance of each part of the feature map extracted from UAV image are refined in two different dimensions, and sufficient context learning are carried out to improve the detection of small targets. Further more, a post-processing algorithm based on Soft-NMS are designed to prevent small targets from being filtered and further improve the detection of small targets. Compared with the benchmark algorithm Faster R-CNN, the proposed algorithm achieves \(4.7\%\) improvement in average precision (AP).KeywordsTransmission line detectionAttention mechanismSoft-NMS