Answers are grounded in this paper and cite their sources.
Building extraction using deep learning is an important research direction in the field of remote sensing. HRNet is able to maintain high-resolution representations throughout the whole process, providing a powerful tool for dense prediction tasks. However, HRNet is not specifically designed for building extraction, leading to poor performance when applied directly. In addition, it suffers from excessive parameter counts and high computational complexity. To address these problems, this paper proposes LBEHRNet, an innovative lightweight network for building extraction. LBEHRNet is achieved by reconstructing the framework of HRNet and developing three novel core functional modules. First, the network framework is reconstructed to boost the overall computational efficiency and preserve more spatial details. Second, a lightweight star convolutional module is proposed by reconstructing the classic star convolution unit and further improving it with asymmetric convolutions and attention mechanisms, which significantly reduces computational overhead during the encoding stage while enhancing building feature representation. Next, an efficient cross-resolution assisted hybrid attention module is constructed to strengthen the collaborative enhancement across multi-resolution features through explicit interaction between spatial and channel attention mechanisms. Finally, a star operation-based multi-resolution fusion prediction head is designed to reinforce the model’s capability for local detail reconstruction of building features and global decoding. Experimental results on three public building extraction datasets, Inria, WHU-S and WHU-A, confirmed the effectiveness of LBEHRNet in terms of both accuracy and efficiency.
No takes yet. Share an insight, caveat, or question.
Shao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: