Instance Segmentation of Transmission Line Images Based on an Improved D-SOLO Network

Under the background of unmanned aerial vehicle inspection of transmission lines, in order to solve the problems of overlapping detection boxes and label adhesion in the mainstream object detection and semantic segmentation methods, in this paper, we propose a transmission line instance segmentation dataset and an optimized instance segmentation network, based on the characteristics of transmission line pictures. It is derived from the decoupled-SOLO (D-SOLO). The network combines the transformer attention module and deformable ConvNets v2 (DCNv2) module to get the Transformer and DCNv2 (TAD) modules according to the residual method. Extensive experiments on our dataset show the superiority of the proposed transformer and DCNv2 decoupled-SOLO (TAD-D-SOLO) network over four representative methods, and thereby prove the effectiveness of the proposed TAD components.