A Power Line Detection Method Combining Semantic Segmentation and Object-Based Analysis

A new method is proposed to first automatically extract power lines (PLs) by semantic segmentation and then process them. In this method, PLs are first extracted from UAV images using a line feature perception model. The Line Feature Perception Convolution Group in the model has strong multi-scale learning ability and can effectively learn information from long and narrow PLs. The repaired strategy of heuristic broken PLs is then implemented for segmentation results, which use segmentation prediction probability, PLs skeleton line direction, and search distance as information to construct a heuristic function to repair broken PLs. We evaluate our models on publicly available datasets. Experiments show that the proposed method has better segmentation metrics compared with the previous work using deep net of ENet, UNet, ResNet34+UNet, NestedUNet, DeepLabv3_plus, and GCN.