GMR-Net: Road-Extraction Network Based on Fusion of Local and Global Information

Road extraction from high-resolution remote-sensing images has high application values in various fields. However, such work is susceptible to the influence of surrounding environment due diverse slenderness and complex connectivity roads, leading false judgment omission during extraction. To solve this problem, a road-extraction network, global attention multi-path dilated convolution gated refinement Network (GMR-Net), proposed. The GMR-Net facilitated by both local information. A residual module with an mechanism first designed obtain other aggregate information for each location’s features. Then, (MDC) approach used extract road features at different scales, i.e., achieve multi-scale feature Finally, units (GR) are proposed filter out ambiguous gradual details. Multiple methods compared study using Deep-Globe Massachusetts datasets. Experiments on these two datasets demonstrate that method achieves F1-scores 87.38 85.70%, respectively, outperforming approaches segmentation accuracy generalization ability.