基于反馈的迭代采样高噪声点云去噪框架

doi:10.11996/JG.j.2095-302X.2025030614

图学学报 ›› 2025, Vol. 46 ›› Issue (3): 614-624.DOI: 10.11996/JG.j.2095-302X.2025030614

• 计算机图形学与虚拟现实 • 上一篇下一篇

基于反馈的迭代采样高噪声点云去噪框架

王昶畅(), 江坤, 姜凯, 张鹏, 苏智勇()

南京理工大学自动化学院，江苏南京 210094

收稿日期:2024-08-19 接受日期:2024-11-25 出版日期:2025-06-30 发布日期:2025-06-13
通讯作者:苏智勇(1981-)，男，副教授，博士。主要研究方向为计算机视觉。E-mail：suzhiyong@njust.edu.cn
第一作者:王昶畅(2000-)，男，硕士研究生。主要研究方向为计算机视觉。E-mail：1812025077@qq.com
基金资助:
国家重点研发计划项目(2022QY0102)

Feedback-based iterative sampling denoising framework for point clouds with high-level noise

WANG Changchang(), JIANG Kun, JIANG Kai, ZHANG Peng, SU Zhiyong()

School of Automation, Nanjing University of Science & Technology, Nanjing Jiangsu 210094, China

Received:2024-08-19 Accepted:2024-11-25 Published:2025-06-30 Online:2025-06-13
Contact: SU Zhiyong (1981-), associate professor, Ph.D. His main research interest covers computer vision. E-mail：suzhiyong@njust.edu.cn
First author：WANG Changchang (2000-), master student. His main research interest covers computer vision. E-mail：1812025077@qq.com
Supported by:
National Key Research and Development Program of China(2022QY0102)

摘要/Abstract

摘要：

在三维点云采集过程中，由于测量异常、边缘散射和被测物体材质等因素的影响，点云数据容易受到噪声的干扰。然而，目前的深度点云去噪算法在高噪声条件下表现较差，且易导致锐利特征的平滑。针对这一问题，提出了一种基于反馈的迭代采样高噪声点云去噪框架，旨在提升现有监督去噪算法在高噪声条件下的表现。首先，用现有的监督去噪网络对噪声点云进行初步去噪，得到预去噪点云；然后，将原始噪声点云和预去噪点云一起输入采样融合模块，得到包含几何细节和边缘特征的融合点云；再次，反馈感知细化网络在预去噪点云的反馈引导下对融合点云进行去噪，得到本轮迭代的去噪结果；最后，通过将当前去噪结果作为下一轮的反馈信息以及采样融合模块的输入，进行多次迭代，逐步去噪，得到最终的去噪结果。实验结果表明，在高噪声条件下，该框架提升了现有监督式点云去噪算法的性能，表现出了较好的去噪效果和特征保持能力。

关键词: 点云去噪, 点云采样, 高噪声, 反馈机制, 迭代

Abstract:

During the 3D point cloud collection process, point cloud data is easily interfered by noise due to factors such as measurement anomalies, edge scattering, and the material properties of the measured object. However, the current depth point cloud denoising algorithms perform poorly under high-level noise conditions and can easily lead to smoothing of sharp features. To address this problem, a feedback-based sampling denoising framework for point clouds with high-level noise was proposed, with the aim of enhancing the performance of existing supervised denoising algorithms under high-level noise conditions. First, the noisy point cloud was denoised using the existing supervised noise network to obtain a pre-denoised point cloud. Second, the original noisy point cloud and the pre-denoised point cloud were jointly input into the sampling module to obtain a fusion point cloud containing geometric details and edge features. Third, the feedback-aware refinement network denoised the fused point cloud under the guidance of feedback from the pre-denoised point cloud to obtain the denoising result for this round of iterations. Finally, by using the denoised result from the current iteration as the feedback for the next round and as input to the sampling fusion module, the process was iterated progressively until the final denoising result was obtained. Experimental results demonstrated that this framework enhanced the performance of existing supervised point cloud noise denoising algorithms under high-level noise conditions, exhibiting excellent denoising effects and feature retention capabilities.

Key words: point cloud denoising, point cloud sampling, high-level noise, feedback mechanism, iteration

中图分类号:

TP391

王昶畅, 江坤, 姜凯, 张鹏, 苏智勇. 基于反馈的迭代采样高噪声点云去噪框架[J]. 图学学报, 2025, 46(3): 614-624.

WANG Changchang, JIANG Kun, JIANG Kai, ZHANG Peng, SU Zhiyong. Feedback-based iterative sampling denoising framework for point clouds with high-level noise[J]. Journal of Graphics, 2025, 46(3): 614-624.

图/表 18

图1 基于反馈的迭代采样去噪算法框架

Fig. 1 Feedback-based iterative sampling denoising algorithm framework

图2 反馈感知细化模块结构图

Fig. 2 Feedback perception refinement module structure diagram

图3 FPS流程示意图

Fig. 3 Schematic diagram of the FPS

图4 含有不同类型和等级噪声点云示意图((a) 1.00%高斯噪声点云；(b) 2.50%高斯噪声点云；(c)高斯均匀混合噪声点云；(d)高斯脉冲混合噪声点云；(e)干净点云)

Fig. 4 Schematic diagram of point clouds containing different types and levels of noise ((a) 1.00% gaussian noise point cloud; (b) 2.50% gaussian noise point cloud; (c) Gaussian average mixed noise point cloud; (d) Gaussian pulse mixed noise point cloud; (e) Clean point cloud)

表1 本文方法和现有监督方法在带有1.00%高斯噪声的测试模型上的PSNR对比结果

Table 1 PSNR comparison results between this method and existing supervised methods on test models with 1.00% Gaussian noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	40.19	46.37	46.64	49.42	50.73	50.86
Boxunion	40.46	46.29	46.13	48.59	49.10	49.07
Cube	40.07	46.40	46.47	48.07	48.48	48.54
Fandisk	41.28	47.33	47.26	49.98	50.44	50.84
Horse	38.89	43.35	44.06	45.62	45.68	47.05
Joint	38.52	43.42	44.40	45.78	46.16	46.04
Mask	45.26	50.22	52.05	51.99	51.88	52.76
Screwdriver	36.19	40.57	40.67	43.93	44.96	46.10

表2 本文方法和现有监督方法在带有1.00%高斯噪声的测试模型上的CD( 10 ? 5)对比结果

Table 2 CD comparison results between this method and existing supervised methods on test models with 1.00% Gaussian noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	11.81	6.02	5.92	4.32	3.87	3.83
Boxunion	11.80	6.30	6.52	4.83	4.53	4.55
Cube	12.88	6.42	6.38	5.17	4.98	4.91
Fandisk	11.55	5.94	6.07	4.35	4.02	3.98
Horse	11.69	6.34	6.15	5.04	4.61	4.39
Joint	12.93	7.18	6.65	5.56	5.20	5.23
Mask	13.77	8.23	6.75	6.75	6.31	5.96
Screwdriver	10.91	6.32	6.44	6.44	3.79	3.56

表3 本文方法和现有监督方法在带有2.50%高斯噪声的测试模型上的PSNR对比结果

Table 3 PSNR comparison results between this method and existing supervised methods on test models with 2.50% Gaussian noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	34.46	36.51	36.24	45.26	47.17	47.69
Boxunion	34.65	36.92	35.95	43.07	42.57	43.47
Cube	34.52	37.61	36.53	44.87	44.85	45.97
Fandisk	35.33	37.57	37.07	45.12	45.08	45.13
Horse	32.70	35.75	35.12	41.64	42.39	42.48
Joint	33.67	35.23	34.77	40.20	41.38	41.82
Mask	38.88	46.11	44.15	49.63	50.87	49.98
Screwdriver	30.74	33.04	31.81	38.54	39.70	40.20

表4 本文方法和现有监督方法在带有2.50%高斯噪声的测试模型上的CD( 10 ? 5)对比结果

Table 4 CD comparison results between this method and existing supervised methods on test models with 2.50% Gaussian noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	23.99	17.24	18.51	6.66	5.49	5.11
Boxunion	23.36	16.66	19.57	8.94	8.73	7.65
Cube	25.65	17.00	19.93	7.78	7.39	6.46
Fandisk	22.97	16.76	18.44	7.80	7.01	6.81
Horse	22.93	14.42	16.63	7.70	7.12	7.00
Joint	23.15	17.75	19.96	10.04	8.72	8.14
Mask	24.44	12.40	15.59	8.56	7.37	7.71
Screwdriver	22.16	14.81	18.02	8.08	6.87	6.57

图5 带有1.00%高斯噪声的测试模型去噪结果对比图

Fig. 5 Comparison chart of denoising results of the test model with 1.00% Gaussian noise ((a) Boxunion; (b) Cube; (c) Fandisk; (d) Tetrahedron)

图6 带有2.50%高斯噪声的测试模型去噪结果对比图

Fig. 6 Comparison chart of denoising results of the test model with 2.50% Gaussian noise ((a) Boxunion; (b) Cube; (c) Fandisk; (d) Tetrahedron)

表5 本文方法和现有监督方法在带有高斯均匀混合噪声的测试模型上的PSNR对比结果

Table 5 PSNR comparison results between this method and existing supervised methods on test models with Gaussian uniform mixed noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	36.52	40.94	40.04	47.22	48.78	49.46
Boxunion	37.23	41.31	39.79	44.81	45.63	46.56
Cube	36.71	41.86	40.10	46.26	46.88	47.70
Fandisk	37.67	41.66	40.74	46.39	47.49	48.06
Horse	32.92	36.53	35.62	41.71	42.94	42.98
Joint	35.33	38.63	37.34	42.86	43.90	44.51
Mask	35.62	40.14	38.48	47.97	49.16	47.88
Screwdriver	32.53	35.59	34.21	39.39	41.87	41.97

表6 本文方法和现有监督方法在带有高斯均匀混合噪声的测试模型上的CD( 10 ? 5)对比结果

Table 6 CD comparison results between this method and existing supervised methods on test models with Gaussian uniform mixed noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	17.92	10.48	12.03	5.39	4.71	4.34
Boxunion	17.67	10.54	12.96	6.97	6.39	5.76
Cube	19.25	10.56	12.98	6.30	5.94	5.36
Fandisk	17.54	10.67	12.29	6.28	5.55	5.10
Horse	22.94	13.45	15.81	7.61	6.68	6.68
Joint	19.05	11.91	14.33	7.53	6.80	6.26
Mask	40.27	22.81	28.54	10.11	8.88	9.93
Screwdriver	18.00	11.00	13.47	7.14	5.49	5.43

图7 带有高斯均匀混合噪声的测试模型去噪结果对比图

Fig. 7 Comparison chart of denoising results of the test model with Gaussian uniform mixed noise ((a) Boxunion;(b) Cube; (c) Fandisk; (d) Tetrahedron)

表7 本文方法和现有监督方法在带有高斯脉冲混合噪声的测试模型上的PSNR对比结果

Table 7 PSNR comparison results between this method and existing supervised methods on test models with Gaussian pulse mixed noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	37.00	41.66	40.75	47.62	49.40	50.02
Boxunion	37.37	42.02	40.48	45.51	46.42	47.23
Cube	37.08	42.60	40.95	46.80	47.13	47.88
Fandisk	38.06	42.37	41.50	47.00	47.86	48.12
Horse	33.70	37.15	36.19	42.21	43.52	43.26
Joint	35.73	39.45	37.98	43.57	44.60	44.98
Mask	36.10	41.16	39.33	48.10	49.36	48.62
Screwdriver	32.73	36.27	34.89	39.93	42.89	42.67

表8 本文方法和现有监督方法在带有高斯脉冲混合噪声的测试模型上的CD( 10 ? 5)对比结果

Table 8 CD comparison results between this method and existing supervised methods on test models with Gaussian pulse mixed noise

模型	噪声点云	DMR	Score	PCN	PF	本文方法
Tetrahedron	17.22	9.75	11.14	5.16	4.43	4.05
Boxunion	16.91	9.72	11.95	6.47	5.89	5.36
Cube	18.43	9.73	11.79	6.02	5.78	5.23
Fandisk	16.75	9.86	11.27	5.82	5.31	4.97
Horse	21.96	12.44	14.76	7.35	6.36	6.44
Joint	18.25	10.92	13.14	7.03	6.34	6.03
Mask	38.31	20.50	26.11	9.97	8.67	8.98
Screwdriver	17.14	10.07	12.46	6.74	4.97	5.01

图8 带有高斯脉冲混合噪声的测试模型去噪结果对比图

Fig. 8 Comparison chart of denoising results of the test model with Gaussian pulse mixed noise ((a) Boxunion;(b) Cube; (c) Fandisk; (d) Tetrahedron)

图9 真实噪声模型去噪结果对比图((a)真实点云场景1；(b)真实点云场景2)

Fig. 9 Comparison chart of real noise model denoising results ((a) Real point cloud scene 1; (b) Real point cloud scene 2)

表9 本文方法在带有2.5%高斯噪声的测试模型上的消融实验

Table 9 Ablation experiment of this method on a test model with 2.5% Gaussian noise

迭代次数	反馈机制	PSNR均值
1		42.93
2		43.02
3		43.52
2	√	44.12
3	√	44.59
4	√	43.67

参考文献 26

[1]	CHEN H, SHEN J. Denoising of point cloud data for computer-aided design, engineering, and manufacturing[J]. Engineering with Computers, 2018, 34(3): 523-541.
[2]	ALEXA M, BEHR J, COHEN-OR D, et al. Computing and rendering point set surfaces[J]. IEEE Transactions on Visualization and Computer Graphics, 2003, 9(1): 3-15.
[3]	LIPMAN Y, COHEN-OR D, LEVIN D, et al. Parameterization-free projection for geometry reconstruction[J]. ACM Transactions on Graphics, 2007, 26(3): 22-es.
[4]	ZHANG D B, LU X Q, QIN H, et al. Pointfilter: point cloud filtering via encoder-decoder modeling[J]. IEEE Transactions on Visualization and Computer Graphics, 2021, 27(3): 2015-2027.
[5]	RAKOTOSAONA M J, LA BARBERA V, GUERRERO P, et al. Pointcleannet: learning to denoise and remove outliers from dense point clouds[J]. Computer Graphics Forum, 2020, 39(1): 185-203.
[6]	LUO S T, HU W. Score-based point cloud denoising[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 4563-4572.
[7]	CASAJUS P H, RITSCHEL T, ROPINSKI T. Total denoising: unsupervised learning of 3D point cloud cleaning[C]// 2019 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2019: 52-60.
[8]	LUO S T, HU W. Differentiable manifold reconstruction for point cloud denoising[C]// The 28th ACM International Conference on Multimedia. New York: ACM, 2020: 1330-1338.
[9]	YOUSAF S, QIN S Y. Closed-loop restoration approach to blurry images based on machine learning and feedback optimization[J]. IEEE Transactions on Image Processing, 2015, 24(12): 5928-5941. DOI PMID
[10]	MA R J, ZHANG B, ZHOU Y C, et al. PID controller-guided attention neural network learning for fast and effective real photographs denoising[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 33(7): 3010-3023.
[11]	MA R J, LI S Y, ZHANG B, et al. Towards fast and robust real image denoising with attentive neural network and PID controller[J]. IEEE Transactions on Multimedia, 2022, 24: 2366-2377.
[12]	MA R J, LI S Y, ZHANG B, et al. Meta PID attention network for flexible and efficient real-world noisy image denoising[J]. IEEE Transactions on Image Processing, 2022, 31: 2053-2066. DOI PMID
[13]	FLEISHMAN S, COHEN-OR D, SILVA C T. Robust moving least-squares fitting with sharp features[J]. ACM Transactions on Graphics (TOG), 2005, 24(3): 544-552.
[14]	HUANG H, LI D, ZHANG H, et al. Consolidation of unorganized point clouds for surface reconstruction[J]. ACM Transactions on Graphics, 2009, 28(5): 1-7.
[15]	PREINER R, MATTAUSCH O, ARIKAN M, et al. Continuous projection for fast L₁ reconstruction[J]. ACM Transactions on Graphics, 2014, 33(4): 47.
[16]	AVRON H, SHARF A, GREIF C, et al. ℓ₁-sparse reconstruction of sharp point set surfaces[J]. ACM Transactions on Graphics (TOG), 2010, 29(5): 135.
[17]	SUN Y J, SCHAEFER S, WANG W P. Denoising point sets via L₀ minimization[J]. Computer Aided Geometric Design, 2015, 35-36: 2-15.
[18]	ZHENG Q, SHARF A, WAN G W, et al. Non-local scan consolidation for 3D urban scenes[J]. ACM Transactions on Graphics (TOG), 2010, 29(4): 94.
[19]	ROVERI R, ÖZTIRELI A C, PANDELE I, et al. PointProNets: consolidation of point clouds with convolutional neural networks[J]. Computer Graphics Forum, 2018, 37(2): 87-99.
[20]	GUERRERO P, KLEIMAN Y, OVSJANIKOV M, et al. PCPNET learning local shape properties from raw point clouds[J]. Computer Graphics Forum, 2018, 37(2): 75-85.
[21]	WANG X T, FAN X P, ZHAO D B. PointFilterNet: a filtering network for point cloud denoising[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023, 33(3): 1276-1290.
[22]	LEHTINEN J, MUNKBERG J, HASSELGREN J, et al. Noise2Noise:learning image restoration without clean data[EB/OL]. [2024-06-19]http://proceedings.mlr.press/v80/lehtinen18a.html.
[23]	ZHAO H S, JIANG L, JIA J Y, et al. Point transformer[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 16259-16268.
[24]	SERNA A, MARCOTEGUI B, GOULETTE F, et al. Paris-rue-Madame database:a 3D mobile laser scanner dataset for benchmarking urban detection, segmentation and classification methods[EB/OL]. [2024-06-19]https://hal.science/hal-00963812.
[25]	ZHANG S K, TAM H, LI Y K, et al. SceneDirector: interactive scene synthesis by simultaneously editing multiple objects in real-time[J]. IEEE Transactions on Visualization and Computer Graphics, 2023, 30(8): 4558-4569.
[26]	ZHANG S K, LIU J H, LI Y K, et al. Automatic generation of commercial scenes[C]// The 31st ACM International Conference on Multimedia. New York: ACM, 2023: 1137-1147.

基于反馈的迭代采样高噪声点云去噪框架

Feedback-based iterative sampling denoising framework for point clouds with high-level noise

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 18

参考文献 26

相关文章 15

编辑推荐

Metrics

本文评价

[1]	李赛赛, 孙博文, 李迪嘉, 潘文娟. 认知机制视域下的机器人手术系统界面设计迭代优化[J]. 图学学报, 2025, 46(2): 469-478.
[2]	倪云昊, 黄雷. 基于数据表示不变性的域泛化研究[J]. 图学学报, 2024, 45(4): 705-713.
[3]	王艳会, 卢元杰, 伊文卿, 于涛, 宛旭. 基于模型的虚实结合多机协同鉴定性试验技术[J]. 图学学报, 2024, 45(2): 369-373.
[4]	王浩, 郑德华, 刘存泰, 程宇翔, 胡创. 基于 SHOT 特征描述子的自动提取球形标靶方法研究 [J]. 图学学报, 2022, 43(5): 849-857.
[5]	周晨 , 陈伟 , 刘渊 , . 基于渐进迭代逼近的矢量地图曲线化简方法[J]. 图学学报, 2021, 42(6): 979-986.
[6]	黄茹, 李亚娟, 邓重阳. 基于衍生多边形的混合坐标[J]. 图学学报, 2021, 42(4): 659-663.
[7]	刘荧 , 李垚辰 , 刘跃虎 , 付靖文 , 马飞 . 基于相关熵的多视角彩色点云配准[J]. 图学学报, 2021, 42(2): 256-262.
[8]	于万波，范晴涛. 基于三角函数迭代的视频数据特征提取[J]. 图学学报, 2020, 41(4): 512-519.
[9]	张肇轩 1，王诚斌 1，杨鑫 1，朴星霖 2，王鹏杰 3，尹宝才 1 . 基于模板替换的室内场景建模方法研究[J]. 图学学报, 2020, 41(2): 270-276.
[10]	霍彦妏，蔡占川 . 基于多结点样条磨光函数的几何迭代法[J]. 图学学报, 2019, 40(1): 15-23.
[11]	李健 1，杨静茹 1，何斌 2 . 一种应用于大角度变换点云的配准方法[J]. 图学学报, 2018, 39(6): 1098-1104.
[12]	张志林，苗兰芳 . 基于深度图像的三维场景重建系统[J]. 图学学报, 2018, 39(6): 1123-1129.
[13]	叶颖，张琪，苏智勇 . 基于哈希的二维工程 CAD 图纸检索技术[J]. 图学学报, 2018, 39(5): 963-969.
[14]	赵夫群1,2，周明全 2,3，耿国华2. 基于局部特征的点云配准算法[J]. 图学学报, 2018, 39(3): 389-394.
[15]	李秋宇1，张玉明 2，杨福猛 3，詹曙 1. 一种视频微表情检测的改进光流算法[J]. 图学学报, 2018, 39(3): 448-452.