基于点云重构的自监督点云异常检测方法

doi:10.11772/j.issn.1001-9081.2024091347

《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (10): 3302-3310.DOI: 10.11772/j.issn.1001-9081.2024091347

• 多媒体计算与计算机仿真 • 上一篇

基于点云重构的自监督点云异常检测方法

杨建锋¹^,², 陈斌²^,³^,⁴(), 李雨轩¹^,²

^1.中国科学院成都计算机应用研究所，成都 610213
^2.中国科学院大学，北京 100049
^3.哈尔滨工业大学（深圳）国际人工智能研究院，广东深圳 518055
^4.哈尔滨工业大学重庆研究院，重庆 401151

收稿日期:2024-09-23 修回日期:2024-12-21 接受日期:2024-12-23 发布日期:2025-01-14 出版日期:2025-10-10
通讯作者: 陈斌
作者简介:杨建锋（1998—），男，江西南昌人，硕士研究生，主要研究方向：机器视觉、异常检测
陈斌（1970—），男，四川广汉人，研究员，博士，CCF会员，主要研究方向：工业检测、深度学习 Email:chenbin2020@hit.edu.cn
李雨轩（1999—），男，河南商丘人，硕士研究生，主要研究方向：异常检测、深度学习。

Self-supervised point cloud anomaly detection method based on point cloud reconstruction

Jianfeng YANG¹^,², Bin CHEN²^,³^,⁴(), Yuxuan LI¹^,²

^1.Chengdu Institute of Computer Application，Chinese Academy of Sciences，Chengdu Sichuan 610213，China
^2.University of Chinese Academy of Sciences，Beijing 100049，China
^3.International Institute for Artificial Intelligence，Harbin Institute of Technology （Shenzhen），Shenzhen Guangdong 518055，China
^4.Chongqing Research Institute，Harbin Institute of Technology，Chongqing 401151，China

Received:2024-09-23 Revised:2024-12-21 Accepted:2024-12-23 Online:2025-01-14 Published:2025-10-10
Contact: Bin CHEN
About author:YANG Jianfeng， born in 1998， M. S. candidate. His research interests include machine vision， anomaly detection.
CHEN Bin， born in 1970， Ph. D.， research fellow. His research interests include industrial detection， deep learning.
LI Yuxuan， born in 1999， M. S. candidate. His research interests include anomaly detection， deep learning.

摘要/Abstract

摘要：

面对日趋复杂的工业生产环境，三维点云工业异常检测需求与日俱增。尽管基于预训练网络的二维异常检测方法效果显著，但三维点云预训练网络的泛化能力有限，导致这类点云异常检测方法的效果不佳。为提高三维点云异常检测的性能，提出一种基于点云重构的异常检测方法Point-ReAD（Point cloud Reconstruction for Anomaly Detection），它由异常模拟模块、点云重构网络和异常判别模块这3个核心模块构成。在训练阶段，正常点云图经异常模拟模块产生异常点云送入点云重构网络，正常点云作为自监督信号指导重构网络的学习；点云重构网络使用分组注意力模块（GAM），用于融合点云的复杂结构信息，从而有效地捕捉点云中的几何和语义特征。在推理阶段，测试点云进入重构网络后生成重构点云，使用异常判别模块比较重构前后的点云，从而精确定位异常。实验结果表明，Point-ReAD在MVTec 3D-AD数据集上的点云级AUROC（PC-AUROC）和点级AUPRO（Area Under the Per-Region Overlap）分别达到了95.49%和94.66%，相较于次优方法3DR?M（3D Discriminatively trained Reconstruction Anomaly Embedding Model）分别提升了0.89和1.27个百分点。

关键词: 异常检测, 自监督学习, 点云, 计算机视觉, 重构网络

Abstract:

As industrial production environments become more and more complex， demand for three-dimensional point cloud industrial anomaly detection is increasing. Although the two-dimensional anomaly detection methods based on pre-trained network have significant effects， generalization ability of the three-dimensional point cloud pre-training network is limited， which leads to poor effect of this kind of point cloud anomaly detection methods. To improve performance of the three-dimensional point cloud anomaly detection， Point-ReAD， an anomaly detection method based on point cloud reconstruction， was proposed. The proposed method consists of an anomaly simulation module， a point cloud reconstruction network， and an anomaly discrimination module. In specific， during training phase， anomalous point clouds were created from normal point cloud maps by the anomaly simulation module， with the normal point clouds served as self-supervised signals to guide the learning process of the reconstruction network； in the point cloud reconstruction network， Group Attention Module （GAM） was used to design complex structural information for point cloud integration， thereby capturing geometric and semantic features in point clouds effectively； in the inference phase， the tested point clouds were input to the reconstruction network to generate reconstructed point clouds， and anomalies were located accurately through the anomaly discrimination module by comparing the point clouds before and after reconstruction. Experimental results show that Point-ReAD achieves the PC-AUROC （PointCloud level Area Under the Receiver Operator characteristic Curve） and the point-level AUPRO （Area Under the Per-Region Overlap） of 95.49% and 94.66%， respectively， on MVTec 3D-AD dataset， which are improved by 0.89， 1.27 percentage points， compared to subprior method 3DR?M （3D Discriminatively trained Reconstruction Anomaly Embedding Model）.

Key words: anomaly detection, self-supervised learning, point cloud, computer vision, reconstruction network

中图分类号:

TP391.4

杨建锋, 陈斌, 李雨轩. 基于点云重构的自监督点云异常检测方法[J]. 计算机应用, 2025, 45(10): 3302-3310.

Jianfeng YANG, Bin CHEN, Yuxuan LI. Self-supervised point cloud anomaly detection method based on point cloud reconstruction[J]. Journal of Computer Applications, 2025, 45(10): 3302-3310.

图/表 15

参考文献 28

[1]	LIU J， XIE G， WANG J， et al. Deep industrial image anomaly detection： a survey［J］. Machine Intelligence Research， 2024， 21（1）： 104-135.
[2]	BERGMANN P， JIN X， SATTLEGGER D， et al. The MVTec 3D-AD dataset for unsupervised 3D anomaly detection and localization［C］// Proceedings of the 17th International Joint Conference on Computer Vision， Imaging and Computer Graphics Theory and Applications — Volume 5： VISAPP. Setúbal： SciTePress， 2022： 202-213.
[3]	PANG Y， WANG W， TAY F E H， et al. Masked autoencoders for point cloud self-supervised learning［C］// Proceedings of the 2022 European Conference on Computer Vision， LNCS 13662. Cham： Springer， 2022： 604-621.
[4]	FAN H， SU H， GUIBAS L. A point set generation network for 3D object reconstruction from a single image［C］// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2017： 2463-2471.
[5]	ROTH K， PEMULA L， ZEPEDA J， et al. Towards total recall in industrial anomaly detection［C］// Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2022： 14298-14308.
[6]	HORWITZ E， HOSHEN Y. Back to the feature： classical 3D features are （almost） all you need for 3D anomaly detection［C］// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2023： 2968-2977.
[7]	WANG Y， PENG J， ZHANG J， et al. Multimodal industrial anomaly detection via hybrid fusion［C］// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2023： 8032-8041.
[8]	ZHAO H， JIANG L， JIA J， et al. Point Transformer［C］// Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2021： 16239-16248.
[9]	BERGMANN P， SATTLEGGER D. Anomaly detection in 3D point clouds using deep geometric descriptors［C］// Proceedings of the 2023 IEEE/CVF Winter Conference on Applications of Computer Vision. Piscataway： IEEE， 2023： 2612-2622.
[10]	HECKLER L， KÖNIG R， BERGMANN P. Exploring the importance of pretrained feature extractors for unsupervised anomaly detection and localization［C］// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway： IEEE， 2023： 2917-2926.
[11]	DENG J， DONG W， SOCHER R， et al. ImageNet： a large-scale hierarchical image database［C］// Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2009： 248-255.
[12]	BENGS M， BEHRENDT F， KRÜGER J， et al. Three-dimensional deep learning with spatial erasing for unsupervised anomaly segmentation in brain MRI［J］. International Journal of Computer Assisted Radiology and Surgery， 2021， 16（9）： 1413-1423.
[13]	SIMARRO VIANA J， DE LA ROSA E， VANDE VYVERE T， et al. Unsupervised 3D brain anomaly detection［C］// Proceedings of the 2020 International MICCAI Brainlesion Workshop， LNCS 12658. Cham： Springer， 2021： 133-142.
[14]	LI W， XU X， GU Y， et al. Towards scalable 3D anomaly detection and localization： a benchmark via 3D anomaly synthesis and a self-supervised learning network［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 22207-22216.
[15]	ZAVRTANIK V， KRISTAN M， SKOČAJ D. DRÆM — a discriminatively trained reconstruction embedding for surface anomaly detection［C］// Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2021： 8310-8319.
[16]	刘永江，陈斌. 基于多尺度记忆库的像素级无监督工业异常检测［J］．计算机应用， 2024， 44（11）： 3587-3594.
	LIU Y J， CHEN B. Pixel-level unsupervised industrial anomaly detection based on multi-scale memory bank［J］. Journal of Computer Applications， 2024， 44（11）： 3587-3594.
[17]	YANG M， WU P， FENG H. MemSeg： a semi-supervised method for image surface defect detection using differences and commonalities［J］. Engineering Applications of Artificial Intelligence， 2023， 119： No.105835.
[18]	SONG J W， KONG K， PARK Y I， et al. AnoSeg： anomaly segmentation network using self-supervised learning［EB/OL］. ［2023-07-02］..
[19]	ZHANG X， LI N， LI J， et al. Unsupervised surface anomaly detection with diffusion probabilistic model［C］// Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2023： 6759-6768.
[20]	ZAVRTANIK V， KRISTAN M， SKOČAJ D. Keep DRÆMing： discriminative 3D anomaly detection through anomaly simulation［J］. Pattern Recognition Letters， 2024， 181： 113-119.
[21]	ZAVRTANIK V， KRISTAN M， SKOČAJ D. Cheating depth： enhancing 3D surface anomaly detection via depth simulation［C］// Proceedings of the 2024 IEEE/CVF Winter Conference on Applications of Computer Vision. Piscataway： IEEE， 2024： 2153-2161.
[22]	YANG Y， FENG C， SHEN Y， et al. FoldingNet： point cloud auto-encoder via deep grid deformation［C］// Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2018： 206-215.
[23]	YU X， TANG L， RAO Y， et al. Point-BERT： pre-training 3D point cloud Transformers with masked point modeling［C］// Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2022： 19291-19300.
[24]	VASWANI A， SHAZEER N， PARMAR N， et al. Attention is all you need［C］// Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2017： 6000-6010.
[25]	QI C R， YI L， SU H， et al. PointNet++： deep hierarchical feature learning on point sets in a metric space［C］// Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2017： 5105-5114.
[26]	HONG C Y， CHOU Y Y， LIU T L. Attention discriminant sampling for point clouds［C］// Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision. Piscataway： IEEE， 2023： 14383-14394.
[27]	ZHANG R， GUO Z， FANG R， et al. Point-M2AE： multi-scale masked autoencoders for hierarchical point cloud pre-training［C］// Proceedings of the 36th International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2022： 27061-27074.
[28]	RUDOLPH M， WEHRBEIN T， ROSENHAHN B， et al. Asymmetric student-teacher networks for industrial anomaly detection［C］// Proceedings of the 2023 IEEE/CVF Winter Conference on Applications of Computer Vision. Piscataway： IEEE， 2023： 2591-2601.

方法	bagel	cable_gland	carrot	cookie	dowel	foam	peach	potato	rope	tire	mean
Voxel AE^［2］	69.30	42.50	51.50	79.00	49.40	55.80	53.70	48.40	63.90	58.30	57.18
PatchCore^［5］	62.40	68.30	67.60	83.80	60.80	55.80	56.70	49.60	69.90	61.90	57.18
PC-FPFH^［6］	82.00	53.30	87.70	76.90	71.80	57.40	77.40	89.50	99.00	58.20	75.32
3D-ST^［9］	86.20	48.40	83.20	89.40	84.80	66.30	76.30	68.70	95.80	48.60	74.77
M3DM^［7］	94.10	65.10	96.50	96.90	90.50	76.00	88.00	97.40	92.60	76.50	87.36
AST^［28］	88.10	57.60	96.50	95.70	67.90	79.70	99.00	91.50	95.60	61.10	83.27
3DRÆM^［20］	98.90	73.10	99.60	98.60	98.80	86.60	96.70	96.90	97.70	99.10	94.60
Point-ReAD	98.44	87.07	98.11	97.74	96.90	87.81	96.27	97.14	98.38	96.99	95.49

方法	bagel	cable_gland	carrot	cookie	dowel	foam	peach	potato	rope	tire	mean
Voxel AE^［2］	69.30	42.50	51.50	79.00	49.40	55.80	53.70	48.40	63.90	58.30	57.18
PatchCore^［5］	62.40	68.30	67.60	83.80	60.80	55.80	56.70	49.60	69.90	61.90	57.18
PC-FPFH^［6］	82.00	53.30	87.70	76.90	71.80	57.40	77.40	89.50	99.00	58.20	75.32
3D-ST^［9］	86.20	48.40	83.20	89.40	84.80	66.30	76.30	68.70	95.80	48.60	74.77
M3DM^［7］	94.10	65.10	96.50	96.90	90.50	76.00	88.00	97.40	92.60	76.50	87.36
AST^［28］	88.10	57.60	96.50	95.70	67.90	79.70	99.00	91.50	95.60	61.10	83.27
3DRÆM^［20］	98.90	73.10	99.60	98.60	98.80	86.60	96.70	96.90	97.70	99.10	94.60
Point-ReAD	98.44	87.07	98.11	97.74	96.90	87.81	96.27	97.14	98.38	96.99	95.49

类别	bagel	cable_gland	carrot	cookie	dowel	foam	peach	potato	rope	tire	mean
Voxel AE^［2］	26.00	34.10	58.10	35.10	50.20	23.40	35.10	65.80	1.50	18.50	34.78
PatchCore^［5］	70.10	54.40	79.10	83.50	53.10	10.00	80.00	54.90	82.70	18.50	58.63
PC-FPFH^［6］	97.20	84.90	98.10	93.60	96.30	69.30	97.50	98.10	98.00	94.90	92.79
3D-ST^［9］	95.00	48.30	98.60	92.10	90.50	63.20	94.50	98.80	97.60	54.20	83.28
M3DM^［7］	94.30	81.80	97.70	88.20	88.10	74.30	95.80	97.40	95.00	92.90	90.55
3DRÆM^［20］	97.60	87.90	98.30	94.60	94.80	81.20	97.90	98.40	85.30	97.90	93.39
Point-ReAD	97.75	92.02	97.77	94.89	95.11	78.89	96.88	98.52	98.80	95.96	94.66

类别	bagel	cable_gland	carrot	cookie	dowel	foam	peach	potato	rope	tire	mean
Voxel AE^［2］	26.00	34.10	58.10	35.10	50.20	23.40	35.10	65.80	1.50	18.50	34.78
PatchCore^［5］	70.10	54.40	79.10	83.50	53.10	10.00	80.00	54.90	82.70	18.50	58.63
PC-FPFH^［6］	97.20	84.90	98.10	93.60	96.30	69.30	97.50	98.10	98.00	94.90	92.79
3D-ST^［9］	95.00	48.30	98.60	92.10	90.50	63.20	94.50	98.80	97.60	54.20	83.28
M3DM^［7］	94.30	81.80	97.70	88.20	88.10	74.30	95.80	97.40	95.00	92.90	90.55
3DRÆM^［20］	97.60	87.90	98.30	94.60	94.80	81.20	97.90	98.40	85.30	97.90	93.39
Point-ReAD	97.75	92.02	97.77	94.89	95.11	78.89	96.88	98.52	98.80	95.96	94.66

方法	bagel		cable_gland		carrot		cookie		dowel		foam
方法	PC-AUROC	AUPRO	PC-AUROC	AUPRO	PC-AUROC	AUPRO	PC-AUROC	AUPRO	PC-AUROC	AUPRO	PC-AUROC	AUPRO
Mask	94.67	97.67	79.50	90.33	89.02	97.32	96.67	93.52	92.26	91.56	84.96	75.40
Point-ReAD	98.44	97.75	87.07	92.02	98.11	97.77	97.74	94.89	96.90	95.11	87.82	78.89
方法	peach		potato		rope		tire		mean
方法	PC-AUROC	AUPRO	PC-AUROC	AUPRO	PC-AUROC	AUPRO	PC-AUROC	AUPRO	PC-AUROC	AUPRO
Mask	93.85	95.69	97.23	95.67	94.75	97.55	92.37	94.78	91.53	92.95
Point-ReAD	96.27	96.88	97.14	98.52	98.38	98.80	96.99	95.96	95.49	94.66

基于点云重构的自监督点云异常检测方法

Self-supervised point cloud anomaly detection method based on point cloud reconstruction

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 15

参考文献 28

相关文章 15

编辑推荐

Metrics

注意力模块	无仿射变换		有仿射变换
注意力模块	PC-AUROC	AUPRO	PC-AUROC	AUPRO
标准	89.23	93.84	92.39	94.08
GAM	93.30	94.39	95.49	94.66

策略	重构局部	重构全局	重构整体	PC-AUROC	AUPRO
I	√	√		95.49	94.66
Ⅱ	√			89.31	92.29
Ⅲ		√		79.62	84.12
Ⅳ			√	92.04	93.03

F	PC-AUROC/%
F	I=0.2	I=0.3	I=0.4	I=0.5	I=0.6
0.2	93.05	94.31	91.52	86.30	77.95
0.4	94.88	95.49	94.30	88.07	76.55
0.6	90.43	91.14	86.01	81.33	/
0.8	73.23	75.28	74.50	/	/

方法	bagel	cable_gland	carrot	cookie	dowel	foam	peach	potato	rope	tire	mean
AST_RGB	94.70	92.80	85.10	82.50	98.10	95.10	89.50	61.30	99.20	82.10	88.00
M3DM_RGB	94.40	91.80	89.60	74.90	95.90	76.70	91.90	64.80	93.80	76.70	85.05
3DRÆM_RGB	94.20	90.60	81.60	66.60	78.30	91.90	65.60	78.40	98.60	92.50	83.83
Point-ReAD	98.44	87.07	98.11	97.74	96.90	87.81	96.27	97.14	98.38	96.99	95.49

[1]	李维刚, 邵佳乐, 田志强. 基于双注意力机制和多尺度融合的点云分类与分割网络[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 3003-3010.
[2]	刘超, 余岩化. 融合降噪策略与多视图对比学习的知识感知推荐模型[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 2827-2837.
[3]	谢劲, 褚苏荣, 强彦, 赵涓涓, 张华, 高勇. 用于胸片中硬负样本识别的双支分布一致性对比学习模型[J]. 《计算机应用》唯一官方网站, 2025, 45(7): 2369-2377.
[4]	闫龙博, 毛文涛, 仲志鸿, 范黎林. 面向城市排水管网缺陷诊断的鲁棒无监督多任务异常检测方法[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1833-1840.
[5]	吴宗航, 张东, 李冠宇. 基于联合自监督学习的多模态融合推荐算法[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1858-1868.
[6]	李维刚, 李歆怡, 王永强, 赵云涛. 基于自适应动态图卷积和无参注意力的点云分类分割方法[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1980-1986.
[7]	王文鹏, 秦寅畅, 师文轩. 工业缺陷检测无监督深度学习方法综述[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1658-1670.
[8]	陈子和, 陈斌. 基于多表征融合的无监督点云异常检测[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1677-1685.
[9]	陈凯, 叶海良, 曹飞龙. 基于局部-全局交互与结构Transformer的点云分类算法[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1671-1676.
[10]	杨光局, 罗天健, 王开军, 杨思琪. 多分支多视图的时间序列上下文对比表征学习方法[J]. 《计算机应用》唯一官方网站, 2025, 45(4): 1042-1052.
[11]	朱俊屹, 常雷雷, 徐晓滨, 郝智勇, 于海跃, 姜江. 基于最小先验知识的自监督学习方法[J]. 《计算机应用》唯一官方网站, 2025, 45(4): 1035-1041.
[12]	潘理虎, 彭守信, 张睿, 薛之洋, 毛旭珍. 面向运动前景区域的视频异常检测[J]. 《计算机应用》唯一官方网站, 2025, 45(4): 1300-1309.
[13]	薛振华, 李强, 黄超. 视觉基础模型驱动的像素级图像异常检测方法[J]. 《计算机应用》唯一官方网站, 2025, 45(3): 823-831.
[14]	任志强, 陈学斌. 基于历史模型更新的自适应防御机制FedAud[J]. 《计算机应用》唯一官方网站, 2025, 45(2): 490-496.
[15]	文诗佳, 金世俊. 结合目标检测和特征点关联的动态视觉SLAM算法[J]. 《计算机应用》唯一官方网站, 2025, 45(2): 610-615.