A Novel Transmission Line Inspection Method Based on the Shape-Aware Features Refinement Network

Transmission line inspection (TLI) plays a crucial role in ensuring power industrial stability and safety. Recently, unmanned aerial vehicle (UAV) and deep learning-based inspection systems have achieved remarkable results in the TLI tasks. However, the slender shapes and occlusion problems in TLI tasks present significant challenges (such as accurate locate and recognition) for UAV-based inspection systems. In this article, we propose a novel TLI method based on the shape-aware feature refinement network, which can improve the detection effect of the UAV inspection system. Specifically, to address the mentioned problem of slender object detection, we introduce a shape-aware feature refinement module (SaFRM) in a coarse-to-fine manner. And, the SaFRM can enhance the inspection network's capability to perceive and represent object shapes through the shape-aware attention module and focal border alignment module. In addition, to address the challenge of object occlusion, we introduce a semantically guided feature erasure module, which can reduce the network's dependence on the specific local features and increase the network's inference ability from limited information. Extensive experiments on the transmission line dataset show that our model outperforms the existing state-of-the-art object detection models. In addition, we implement our proposed algorithm in a real-time actual UAV-based inspection system, further demonstrating its excellent ability for real-world TLI tasks.