Research on nighttime building image enhancement and high-precision point cloud fusion method for 3D reconstruction

Abstract Under low-light nighttime conditions. Addressing the challenge of poor-quality nighttime building images and the difficulty in achieving high-precision fusion modeling with point cloud data under low-light conditions, this paper introduces a comprehensive solution for nighttime image enhancement and point cloud fusion modeling. We developed an unsupervised learning framework that incorporates building structural priors. Based on RetinexDIP theory, the framework integrates a structural edge enhancement module, effectively restoring details in building facades and outlines. Experimental results demonstrate that our method outperforms baseline approaches on both peak signal-to-noise ratio (12.02 dB, a 0.35 dB gain) and Information Entropy (a 0.08 increase) metrics, while also effectively preserving the integrity of building edge structures. For point cloud registration, we propose a line-feature-constrained registration (LFR) method. Utilizing a planar-and-elevation hierarchical registration strategy, LFR achieved a registration accuracy (RMSE) of 0.04–0.06 m across three test areas, showing significant improvements in accuracy and robustness over traditional methods. Fusion modeling results indicate a maximum increase of 27.6% in the number of image tie points, achieving 100% reconstruction completeness, a surface smoothness improvement to 0.85%–0.94%, and a 35%–48% reduction in overall geometric error. Multi-dimensional validation confirms that our method surpasses comparative approaches in visual quality, geometric accuracy, and model completeness, providing a reliable technical pathway for 3D reconstruction of nighttime buildings.