DF-NeRf: A geometrically consistent 3D reconstruction method for blurry images with decoupled and spatial fusion

Abstract

Reliable 3D scene understanding is fundamental to intelligent information systems, but motion blur in visual data often disrupts geometric consistency, making the reconstruction results unreliable. This paper proposes a novel blurred image NeRF framework named DF-NeRF, incorporating three core innovations. Firstly, the Multi-Scale Position Feature Fusion (MSPF) structure performs multi-scale feature fusion within the blurred kernel space to precisely estimate ray displacement ranges, thereby preserving multi-view geometric consistency. Secondly, the Multi-layer Independent Weight and Shifting Network (MIWS) is designed to achieve explicit decoupling between geometry and appearance. Within MIWS, the Dual-head Network (DHNet) employs a separate optimisation strategy, independently learning ray offsets (geometric attributes) and rendering weights (appearance attributes), effectively suppressing geometric distortions caused by motion blur. Meanwhile, the Prior-weighted Network (PWNet) employs prior-guided adaptive weighting to fuse different blur kernels, ensuring high-frequency detail geometric continuity. Finally, an edge-aware loss function is introduced to constrain gradient discrepancies in depth discontinuities, explicitly guaranteeing contour structure consistency and detail fidelity. We evaluated on public benchmarks, including 5 synthetic scenes and 10 real scenes, covering both indoor and outdoor environments. Experimental results demonstrate that DF-NeRF outperforms SOTA on PSNR, SSIM, and LPIPS metrics, with the most notable improvement in LPIPS. In synthetic scenes, DF-NeRF achieves an average LPIPS improvement of 14.32% over SOTA. In real-world scenes, DF-NeRF achieves an average LPIPS improvement of 19.79% over SOTA. DF-NeRF significantly enhances 3D reconstruction quality and robustness under blurred images, making it suitable for applications such as autonomous driving and robot vision. Code is available at: https://github.com/luckhui0505/DF-NeRF.