Class-aware edge-assisted lightweight semantic segmentation network for power transmission line inspection

The demand for real-time efficient scene comprehension has been increasing rapidly in the drone-based automatic inspection of power transmission lines (PTL). The extensive application of semantic segmentation in urban scenes proves that it can meet the requirements for scene understanding. However, existing methods have difficulty adapting to changes in the scene, which leads to problems of performance degradation and fuzzy contours of segmented objects. To overcome the existing problems, a class-aware edge-assisted lightweight semantic segmentation network is proposed in this paper. Class-aware edge detection is introduced as an auxiliary task, and a two-branch network is designed to locate instances and refine contours. Specifically, hybrid graph learning uses task-specific graph-based structures to reason attention information of region and edge features. Based on the complementary characteristic of region and edge features, cascaded shared decoders adopt specific interaction functions to enhance the ability of region features to locate targets and the ability of edge features to improve contour details. In addition, to verify the effectiveness of the proposed method, we construct two datasets named the transmission tower component recognition dataset (TTCRD) and the transmission line regional classification dataset (TLRCD). Comprehensive experiments on TTCRD and TLRCD prove that the proposed method can accurately refine the contour of objects and overcome the challenges in the two datasets. Comparison experiments and ablation experiments also demonstrate the superior performance of the proposed method and the effectiveness of each component in our architecture.