Research On Point Cloud Denoising Algorithm Based On Non-local Self-similarity And Manifold Learning

3D scanning technology is an emerging high-tech measurement technology that can quickly obtain 3D data of an object’s surface without contact.It has attracted the atten-tion of many scholars and has wide applications in industrial manufacturing,aerospace,medical imaging,archaeology and other fields.Point cloud data is obtained by 3D laser scanners and is characterized by simple structure,high information content,and easy ac-quisition.This makes it suitable for expressing complex and irregular object surfaces.However,due to the accuracy limitation of scanning equipment,differences in reflectiv-ity of the object surface,and interference from the external environment,there are often noises in the actual acquired point cloud data.These noises may seriously affect down-stream tasks such as point cloud registration,feature extraction,3D reconstruction,3D target recognition,etc.Therefore,as a primary and fundamental part of 3D point cloud processing,point cloud denoising aims to weaken the influence of noise.This is of great significance in the application of point clouds.This thesis analyzes the noisy point cloud as the research object and provides insight into the denoising process.The main contents are as follows.(1)The structure of point clouds and the causes of noise generation are classified.Different denoising algorithms are categorized according to point cloud types and denois-ing principles.The mechanisms of several common point cloud denoising algorithms are introduced,and their advantages and disadvantages are analyzed.(2)This thesis proposes a multi-patch collaborative point cloud denoising algorithm based on the theory of non-local self-similarity.The algorithm characterizes the structural similarity between different patches by low rankness and constructs a non-local tensor with self-similarity.The denoising problem can be formulated as a nuclear norm mini-mization problem and solved by a weighted singular value soft thresholding algorithm.To improve the ability of local blocks to characterize the detailed structure,a robust ex-traction algorithm for local structure patches based on the random sampling consistency concept is designed.To avoid local points aggregating and forming filamentary struc-tures during iterative updating and improve the visualization results after denoising,the algorithm introduces a regular term with repulsive effect during coordinate updating to mitigate this phenomenon.The effectiveness of the proposed algorithm is experimentally verified.(3)In response to the proposed algorithm’s limitation that the feature structure is not well maintained,this paper proposes a denoising algorithm based on manifold learning.By introducing the locally linear embedding algorithm,the denoising algorithm ensures that the linear relationship in the local neighborhood of points before and after denois-ing does not change as much as possible,thus enhancing the ability to represent local features.In addition,the paper proposes a kernel matrix fusion strategy of different man-ifold learning algorithms from the perspective of fusion kernel.This strategy realizes the effective fusion of global information and local features in the point cloud denoising algo-rithm,which further improves the denoising performance and enhances the representation capability for the detailed features of the point cloud model.The proposed algorithm’s effectiveness is experimentally verified under different noise environments,demonstrating that the algorithm can maintain the detailed structural features of the point cloud model and has superior denoising performance.