基于神经隐式场的模糊多视图三维重建  

Blurry multi-view 3D reconstruction based on neural implicit field

作　　者：徐紫慧 王逸群 Xu Zihui;Wang Yiqun(College of Computer Science,Chongqing University,Chongqing 400000,China)

机构地区：[1]重庆大学计算机学院,重庆400000

出　　处：《计算机应用研究》2025年第2期606-611,共6页Application Research of Computers

摘　　要：基于神经隐式表面的重建方法因其能高保真地重建场景而受到广泛关注。然而,这些研究主要集中在理想输入的重建上,对于模糊输入重建效果并不理想。为了解决以上问题,提出了Deblur-NeuS,一种基于神经隐式场的模糊多视图三维重建方法。通过引入模糊核预测模块和隐式位移场来模拟模糊过程以重建模糊视图和表面,并为模糊表面增加模糊点云监督,优化隐式表面的学习。在测试阶段移除模糊核模块与隐式位移场,即可直接提取更清晰的几何表面。在模糊数据集上的实验结果显示,重建的表面质量以及图像渲染的质量都得到了显著提升。该方法增强了网络对模糊输入的鲁棒性,能从运动模糊图像中恢复几何表面细节。Neural implicit surface-based reconstruction methods are widely valued for their high-fidelity scene reconstruction capabilities.However,most of these studies have focused on reconstructing ideal input,yielding less effective results for blurry input.To address this issue,this paper proposed Deblur-NeuS,a method for blurry multi-view 3D reconstruction based on neural implicit fields.The method simulated the blurring process to reconstruct blurry views and surfaces by introducing a blur kernel prediction module and an implicit displacement field.Additionally,it added blurry point cloud supervision for the blurry surface,which optimized the learning of the implicit surfaces.During the testing phase,the blur kernel prediction module and the implicit displacement field were removed,which enabled the direct extraction of clear geometric surfaces.Experimental results on blurred datasets demonstrate significant improvements in the quality of reconstructed surfaces and the quality of image rendering.The method enhances the network’s robustness to blurred inputs and enables the recovery of geometric surface details from motion-blurred images.

关 键 词：神经隐式表面 模糊多视图 三维重建 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]