HiDeGS: High-frequency detail-enhanced Gaussian splatting for UAV multi-view reconstruction

3D Gaussian Splatting has become a popular emerging technique in 3D reconstruction due to its efficient rendering performance. However, in UAV multi-view reconstruction tasks, it still faces core challenges, such as high-frequency detail loss and insufficient multi-view geometric consistency, which lead to blurred fine structures, geometric distortion, artifacts, and other issues that severely restrict the geometric accuracy of 3D models. To address these bottlenecks, this paper proposes HiDeGS (High-Frequency Detail-Enhanced Gaussian Splatting) for UAV multi-view reconstruction, which improves the geometric accuracy of 3D reconstruction through multi-dimensional joint optimization. First, a multi-scale frequency-spatial joint regularization is constructed to compensate for the loss of high-frequency information, while ensuring the local integrity of geometric structures via spatial gradient enhancement and edge-preserving constraints, overcoming the insufficient detail characterization of traditional pixel-level loss. Second, a high-frequency-aware scale regularization and adaptive loss weighting mechanism are designed to restrict the excessive expansion of Gaussian primitives in high-frequency regions and prioritize the optimization of tiny geometric structures. Finally, an enhanced depth geometric consistency constraint fused with multi-dimensional uncertainty is proposed, which quantifies uncertainty through depth gradient mutations and normal consistency differences, dynamically adjusts the strength of multi-view constraints, and improves the robustness of constraints in low-texture and edge regions. This paper presents the Wuhan University UAV Dataset (WHU_Dataset), which encompasses 8 typical scenes. Experiments show that on the ISPRS benchmark dataset and the self-built WHU_Dataset, compared with the baseline 3DGS, HiDeGS improves the Peak Signal-to-Noise Ratio (PSNR) by 1.34 dB and 1.95 dB, and the Structural Similarity Index (SSIM) by 0.085 and 0.1, respectively. The method can accurately recover fine geometric structures such as building contours and road markings, effectively suppress artifacts and geometric drift, and significantly improve the geometric accuracy of 3D reconstruction in UAV scenarios. The source code is available at: https://github.com/SongJiang-WHU/HiDeGS .