【作者】 王兴涛；

【导师】 范晓鹏；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2022， 博士

【摘要】 点云是应用最广泛的3D数据,一直以来备受关注。然而,由于扫描设备的精度限制和重构技术的不完善,所获取的点云往往受到噪声的污染。因此,点云去噪被普遍置于点云获取、处理系统的第一步,其重要意义与价值可见一斑。点云是物体表面的采样点的集合,能够表达物体表面的几何特征。点云噪声表现为点云中的点偏离物体的理想表面。点云去噪,指将偏离的点移动到理想表面上,同时保持去噪后的点分布均匀,从而获得一个能够更好地表达物体表面几何特征的点云。传统的点云去噪方法依赖于人为定义的先验信息。由于人为定义的先验无法有效、完整地表述复杂点云的结构特性,传统去噪方法在复杂点云上表现不佳。近些年,基于深度学习的点云去噪方法实现了明显优于传统去噪方法的性能。然而,基于深度学习的方法仍有进一步提升的空间,具体表现在以下几个方面。第一,现有的点云去噪网络的内部模块,通常没有针对去噪任务的优化设计,从而导致神经网络的利用方式比较低效;第二,点云数据的多样性增加了去噪任务的挑战性,然而现有的工作还没有针对点云的多样性问题提出有效地解决方案;第三,在参数优化期间,去噪网络从带噪点云提取的特征也受到噪声的污染,特征噪声会干扰梯度回传的效率,网络容易收敛到次优的参数。本文针对上述问题,围绕基于深度学习的点云去噪关键技术,从网络结构、损失函数、训练范式等方面进行了深入的探索,提出了有效的解决方案。具体研究内容包括以下三个部分:第一,提出了一种结合传统滤波技术和深度学习技术的点云滤波网络,PointFilterNet。传统滤波技术受限于人为定义的系数,而深度学习在自动学习参数方面展现出卓越的能力。由此出发,本文将滤波技术与深度学习相结合,利用神经网络学习滤波器的系数,提出了PointFilterNet。与之前的滤波方法相比,PointFilterNet利用神经网络学习到的滤波器系数,能够有效应对不同特性的点云,实现了显著的性能提升;与之前的深度学习方法相比,PointFilterNet以滤波器的形式利用神经网络,内部模块经过精心设计,各司其职,提高了神经网络的利用效率。第二,提出了用于点云去噪任务的参数定制网络,Meta-PCD,能够有效地应对点云数据多样性的问题。在进行点云去噪时,现有方法通常使用训练好的网络参数处理所有的点云。然而,点云数据繁杂多样,使用固定的网络参数并非应对点云多样性的合理方案。在本文中,Meta-PCD利用元学习的方法,将每一个点的去噪视为不同的任务,动态地为每个点定制去噪网络的参数,有效地解决了点云多样性的问题。与现有的去噪网络相比,Meta-PCD有效地解决了点云多样性的问题,取得了明显的性能增益。与PointFilterNet相比,前者利用深度学习生成滤波器的系数,Meta-PCD则生成更复杂的神经网络的参数。第三,提出了特征清洁网络(FeatureCleanNet),并引入了师生训练范式,进行特征域去噪,从而改善点云去噪网络的参数优化过程。在点云去噪网络的参数优化过程中,网络从带噪点云中提取的特征也会受到噪声的污染,特征噪声会干扰网络参数的优化,从而影响网络的去噪性能。针对这个问题,本文提出了特征清洁网络,FeatureCleanNet,并引入了一种师生范式来训练网络。与现有的点云去噪网络相比,FeatureCleanNet的参数优化过程中回传的梯度更加稳定可靠;师生训练范式则通过网络内部的引导帮助去噪网络学习无噪特征。以上三个研究内容联系紧密。其中PointFilterNet和Meta-PCD着重于点云去噪网络结构的研究,Meta-PCD是PointFilterNet的进一步优化,两者之间是递进关系。FeatureCleanNet则是针对网络训练过程的研究,可以与PointFilterNet、Meta-PCD结合使用,与前两者为并列关系。更多还原

【Abstract】 As the most widely used 3D data,point clouds have attracted much attention.However,due to the imperfection in scanning and reconstruction processes,the acquired point clouds are often accompanied by noise.Therefore,point cloud denoising is usually placed in the first step of point cloud acquisition and processing systems,which reflects the significance and importance of point cloud denoising.Point clouds are sets of points sampled on the ideal surface of objects,revealing the geometric characteristics of the ideal surface.Point cloud noise refers to the points in the point cloud deviate from the ideal surface.Point cloud denoising is to move the noisy points to the ideal surface while maintaining the distribution uniformity of points,so as to obtain a point cloud that can reveal the ideal surface better.Conventional point cloud denoising methods rely on handcrafted priors.These methods don’t perform well on complex objects as the handcrafted priors can not reveal the structures comprehensively.Recently,deep-learning-based point cloud denoising methods have achieved significantly better performance than conventional methods.However,there is still room for improvement in point cloud denoising methods based on deeplearning,which can be summarized as: First,the modules of existing point cloud denoising networks are usually lack of specific design for the denoising task,resulting in that the adoption of networks is inefficient.Second,the diversity of point cloud data increases the challenge of denoising task,but existing works have not put forward effective solutions to the diversity of point clouds.Third,during parameter optimization,features extracted by denoising networks from noisy point clouds are also contaminated by noise.The feature noise will interfere with the efficiency of backpropagation,making networks converge to the sub-optimal parameters.In view of the above problems,this dissertation conducts an in-depth exploration on the key technologies of deep-learning-based point cloud denoising in terms of network structures,loss functions,and training paradigms.The specific research contents can be summarized as the following three parts:First,this dissertation proposes a point cloud denoising network called PointFilterNet,which combines conventional filtering techniques and deep-learning techniques.Conventional filtering techniques are often limited by manually defined coefficients,while deep-learning has shown excellent ability in automatically learning parameters.Therefore,this dissertation combines conventional filtering techniques and deep-learning techniques and proposes PointFilterNet.Compared with previous filtering techniques,PointFilter Net applies a neural network to learn the filter coefficients,which are better at dealing with point clouds of different characteristics,achieving significant performance gains.Compared with existing deep-learning-based methods,PointFilterNet adopts the neural network in the form of filters and employ carefully-designed modules,improving the utilization efficiency of the neural network.Second,this dissertation proposes a parameter-customized network called Meta-PCD for point cloud denoising,solving the point patch diversity problem effectively.Existing methods usually use well-trained network parameters to process various point clouds while denoising.However,point clouds are complex and diverse.Adopting fixed network parameters is not a reasonable solution to deal with the diversity of point clouds.In this dissertation,Meta-PCD takes the denoising of each point as a unique task and dynamically customizes denoising parameters for each point,solving the point patch diversity problem effectively.Compared with the existing denoising network,Meta-PCD effectively solves the problem of point cloud diversity and achieves significant performance gains.Compared with PointFilterNet,Meta-PCD generates the parameters of a more complex neural network.Third,this dissertation proposes a feature cleaning network called FeatureCleanNet and introduces a novel teacher-student training paradigm to conduct denoising in the feature domain,improving the parameter optimization of the point cloud denoising network.During parameter optimization,features captured by a denoising network from noisy point clouds are usually contaminated by noise.The feature noise will count against the network parameter optimization,thereby affecting the denoising performance.To this end,this dissertation proposes a feature clean network called FeatureCleanNet,and introduces a novel teacher-student paradigm.Compared with existing point cloud denoising networks,FeatureCleanNet makes the back-propagated gradients more stable and reliable.The teacherstudent paradigm helps the denoising network learn noise-free features through internal guidance.The above three research contents are closely related.PointFilterNet and Meta-PCD are both research on the structures of point cloud denoising networks,and Meta-PCD is the improved method based on PointFilterNet.There is a progressive relationship between them.FeatureCleanNet is a research on network training process.It can be used in combination with PointFilterNet or Meta-PCD.It is in parallel with the former two research contents.更多还原