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

Abstract

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