DPENet: Dual-path extraction network based on CNN and transformer for accurate building and road extraction

Ziyi Chen, Yuhua Luo, Jing Wang, Jonathan Li, Cheng Wang, Dilong Li

Abstract: Highlights•We propose a novel dual-path object extraction network that combines the strengths of CNN and Transformer. The network integrates a spatial detail branch and a global semantic branch, leading to exceptional precision in extracting building and road objects.•To enhance the network's reconstruction capability for building and road objects, we introduce a multi-view fusion-based multi-sampling mechanism in the reconstruction branch. This mechanism ensures that the network captures finer object details, significantly improving its ability to reconstruct these objects.•We employ a deep supervision strategy by designing segmentation heads at different levels in the reconstruction branch. This strategy allows for comprehensive supervision of the reconstruction process for building and road object features at different resolutions.