DVDNet: a dual-view decoding network for road extraction

Abstract Remote sensing image road extraction faces challenges due to complex road shapes and background interference. To address the trade-off between detail preservation and global context understanding, this paper introduces a dual-view decoding network (DVDNet). The network comprises three key modules: (1) Multi-Branch Feature Guidance Module (MFGM), which extracts discriminative features via parallel multi-scale, direction-sensitive, and context-enhanced branches; (2) Heterogeneous Dual-View Collaborative Decoder (HDCD), designed with two specialized paths for local detail restoration and global contextual understanding; (3) Path Importance Self-Evaluation Module (PISM), which adaptively generates fusion weights through lightweight analysis for end-to-end path collaboration. Experiments on the DeepGlobe and RoadTracer datasets demonstrate that DVDNet achieves competitive performance and outperforms the compared methods in terms of F1-score and mIoU. All necessary resources, including code, trained models, configurations, and data processing details, will be provided to ensure full reproducibility of the experimental results, and will be publicly available at https://github.com/Huayang122/DVDNet.