基于无监督网络框架的文物点云模型去噪  

Denoising of Cultural Relics Point Cloud Model Based on Unsupervised Network Framework

作　　者：刘一萍 周明全[1] 寇姣姣 鱼跃华 海琳琦 李康[1] 张海波[1] Liu Yiping;Zhou Mingquan;Kou Jiaojiao;Yu Yuehua;Hai Linqi;Li Kang;Zhang Haibo(School of Information Science and Technology,Northwest University,Xi’an 710127,Shaanxi,China)

机构地区：[1]西北大学信息科学与技术学院,陕西西安710127

出　　处：《激光与光电子学进展》2022年第12期362-371,共10页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61902317,61731015);国家重点研发计划(2019YFC1521102,2019YFC1521103)陕西省重点产业链项目(2019ZDLSF07-02);陕西省自然科学基金青年基金(2019JQ-166);青海省重点研发计划(2020-SF-142)。

摘　　要：数字化技术在文物保护工作中的应用极大促进了文化遗产领域的快速发展。三维激光扫描设备获得的文物点云数据不可避免包含大量噪声,直接影响点云数据的后续处理。为有效去除无序点云中的噪声点,较好恢复点云数据,基于无监督网络提出了一种新型点云去噪算法。首先通过上层网络分类并去除离群点;然后通过引入空间先验项引导待去噪点云中数据点收敛到流形上多模式中最接近真实点云的最优模式,从去除离群噪声点的点云数据中得到干净点云的分布,实现无监督精细点云去噪;最后通过计算去噪后点云间的chamfer distance来进行定量评价。与一些经典算法的对比分析实验结果表明,所提算法在去噪的同时,能有效保持点云模型的几何特征,对文物点云数据的去噪效果良好,去噪后的点云模型极大程度复原了原始干净点云模型,这对文物数字化保护的后续环节至关重要。The cultural heritage field has developed rapidly based on the use of digital technologies for protecting cultural relics. The point cloud data of cultural relics obtained using three-dimensional laser scanning equipment inevitably contain considerable noise, which directly affects the subsequent processing of the point cloud data. To effectively remove noise points from the disordered point cloud and ensure enhanced recover of point cloud data, a new point cloud denoising algorithm based on the unsupervised network was proposed. First, the outliers are classified and removed from the upper network. Then, a spatial prior term was introduced to guide the data points in the noise cloud to converge to the optimal mode closest to the real point cloud in the multimode on the manifold,enabling the distribution of clean point cloud from the point cloud data of outlier noise points;moreover, the unsupervised denoising of the fine point cloud was realized. Finally, the chamfer distance between the denoised point clouds was estimated for quantitative evaluations. Compared with some classic algorithms, the proposed algorithm can effectively maintain the geometric characteristics of the point cloud model during denoising, and shows a good denoising effect on the point cloud data of cultural relics. The denoised point cloud model considerably restores the original clean point cloud model, which is crucial for the follow-up link of the digital protection of cultural relics.

关 键 词：机器视觉 点云去噪 神经网络 无监督 chamfer distance 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]