PSTL-Net: A Patchwise Self-Texture-Learning Network for Transmission Line Inspection

A stable and safe smart grid ensures our daily and fundamental activity and the inspection of transmission lines (TLs) plays an essential role in this process. However, aerial images suffer from weak texture representation and target occlusion caused by the unmanned aerial vehicle's (UAV) measurement constraints, which pose challenges to existing TL inspection (TLI) methods. This article proposes a novel patchwise self-texture-learning network (PSTL-Net), which can effectively realize the inspection process of TLs by incorporating the self-texture-learning module (STLM) and a patch-aware spatial attention module (PSAM). Specifically, to enhance texture representations, the STLM enhances the deep network's texture detail perception ability of TL targets via channelwise and patchwise texture-enhanced modules. In addition, to reduce the effect of occlusion, the PSAM improves the spatial position attention of areas with high semantic similarity to the TLs target patch. Finally, comparison experiments were conducted on a collected real-world dataset, and the results showed that the proposed PSTL-Net has outperformed the detection accuracy of other advanced methods, including the highest performance on the link groups (71.3%), vibration hammers (82.8%), bird nests (82.1%) categories, and second-highest performance in terms of the missing insulators, suspension clamps, and missing vibration hammers. Furthermore, the ablation and generalized experiments also proved the effectiveness of the proposed STLM and PSAM. In conclusion, the proposed PSTL-Net can effectively complete the TLI tasks based on the aerial images. Combined with the UAVs, it can provide a deep learning-based solution for the automatic TLI of real-world power scenes in the future. The source codes and datasets are available on GitHub.