CBRNet: Corner-Guided Boundary Refinement Network for High-Precision Building Extraction From Remote Sensing Imagery

Accurate building extraction from high-resolution remote sensing imagery remains essential for urban planning, disaster assessment, and geographic information system updates. Existing methodologies face three critical challenges: (1) substantial scale variations across diverse urban contexts, (2) boundary ambiguity induced by shadows and spectral similarities, and (3) progressive loss of fine-grained structural details during hierarchical feature aggregation. We propose CBRNet, a Corner-guided Boundary Refinement Network that introduces structural geometric reasoning into deep building extraction. Our key insight is that building corners serve as anchor points for boundary delineation-by detecting corners first, boundary precision naturally follows. CBRNet features four synergistic innovations: (1) Content-Adaptive Scale Convolution (CASC) that dynamically adjusts receptive fields based on local content; (2) Corner-Guided Feature Enhancement (CGFE) that integrates differentiable Harris corner detection as learnable geometric priors; (3) Local Window Cross-Attention (LWCA) enabling cross-scale interaction with linear complexity; and (4) Dual-Stream Boundary Refinement (DSBR) that decouples semantic and boundary learning with bidirectional feature exchange. Additionally, we employ BiFPN for efficient multi-scale feature aggregation. Extensive experiments demonstrate state-of-the-art results: 91.35% IoU on WHU Aerial, 76.86% on HUST-ABS, and 84.28% on WHU Satellite I at 512 × 512 resolution. Notably, CBRNet generalizes robustly to 1024 × 1024 resolution on Inria, achieving 81.87% IoU without architectural modification. CBRNet achieves 79.8 FPS with only 32.81M parameters, establishing an optimal accuracy-efficiency trade-off for real-world deployment.