Advancing Building Footprint Extraction with Multi-Stage Regularization Techniques

Abstract. This paper introduces an automated building extraction method combining CNN segmentation with multi-stage regularization. We address urban mapping challenges including boundary inaccuracies and topological errors through: (1) neighborhood matrix processing for local refinement, (2) spectral graph optimization (SBO) for global consistency, and (3) curvature-adaptive contour refinement (ACR) to preserve geometric features. The pipeline converts initial segmentations into precise polygons through hierarchical processing. Experiments show performance matching state-of-the-art methods like PolyWorld, with superior handling of complex geometries. Key innovations include integrated local-global artifact removal and topology-preserving regularization. The curvature-adaptive approach maintains critical architectural features while eliminating noise. Particularly effective for high-resolution imagery, our solution improves geometric fidelity in urban mapping applications. The framework demonstrates robust performance for 3D city modeling and GIS tasks, overcoming common segmentation limitations. Results confirm accurate building outline extraction from satellite/aerial data, advancing automated urban feature mapping.