A point cloud color denoising method based on L0 norm minimization

Abstract

The present application relates to a kind of point cloud color denoising methods based on L0 norm minimization, belong to computer image graphics processing technical field.This method, first decoupling position and color information, respectively create the point cloud RGB vector of length N using extracted color information.Then see the RGB value of position modification point, if the color value of this point and the gradient difference of its neighborhood point color value is larger, then prove that this point is located in the position of color mutation.Form a color smooth surface with this point and its k nearest neighbors, modify the color value of point by L0 norm minimization, and along the position of color smooth point, to sequentially denoise the color value of other points in point cloud.Repeat the process until convergence.This method optimizes color denoising on three-dimensional point cloud model with position information, effectively solves the color mutation problem on the basis of preserving color and geometric features.At the same time, L0 norm minimization method solves the color denoising problem in three-dimensional point cloud denoising.

Description

TECHNICAL FIELD

[0001] The present application relates to a computer three-dimensional image point cloud color denoising method, and belongs to the technical field of virtual reality and computer image graphics processing. BACKGROUND

[0002] In the field of computer graphics and image technology, point cloud is an important representation of objects or scenes in three-dimensional space, which is composed of a set of points that can represent geometric bodies and some attributes related to the model or scene, such as color, normal, transparency and size, etc. Point cloud data with color information can better reflect the surface shape of objects in the real world, and has been widely used in various fields, such as three-dimensional broadcasting, cultural heritage reconstruction, unmanned aerial vehicle navigation and three-dimensional immersive remote presentation, etc.

[0003] With the development of computer graphics, computer vision technology and optical elements, laser scanning sensors and low-cost RGB-D cameras have appeared, such as Astra, Astra S, Astra Pro, Intel RealSense and Microsoft Kinect. By using RGB-D cameras, point cloud form of three-dimensional model can be easily obtained. However, due to various viewing angles, reflective materials or characteristics of object surface, light intensity and limitations of sensors, the point cloud obtained by using these camera devices often contains noise. In color point cloud, noise has randomness, which is not only reflected in space, but also in color. In color distribution, the color of noise points is different from the color of other points in the neighborhood, while the color of target points is similar to the color of other points in the neighborhood.

[0004] In addition to the geometric structure of objects and scenes, the surface color and details of three-dimensional models also play a crucial role in some virtual and augmented reality applications. Color is a basic attribute of point cloud and is considered as a key feature for point cloud segmentation and three-dimensional model retrieval. Noise existing in point cloud color may cause subsequent tasks such as scene understanding and model retrieval to fail. Therefore, color denoising is needed for graphics and image three-dimensional point cloud, that is, to eliminate noise in RGB value of each 3D position in point cloud.

[0005] In the work of three-dimensional point cloud denoising, existing technologies include methods based on moving least squares (MLS) and methods based on local optimal projection (LOP). Among them, the methods based on moving least squares and local optimal projection use local operators, which are easy to cause the results of object denoising to be excessively smooth, so that the original characteristics of the object cannot be maintained. The method based on sparse optimization assumes that the point normal has sparsity and can represent important features of the model. However, when there is too much noise, inaccurate estimation of the model's normal or local plane will lead to excessive smoothing or excessive sharpening.

[0006] Therefore, for the technical problem of three-dimensional point cloud color denoising, there are mainly the following two difficulties.

[0007] First, a significant difference between point cloud and image and mesh is that image or mesh has clear neighborhood pixels and neighborhood vertices, while point cloud does not have connectivity information, and it is difficult to directly apply color denoising technology in two dimensions to point cloud.

[0008] Second, due to the irregularity of point cloud, it is difficult to directly maintain one-to-one correspondence between the geometry and color value of all points in the point cloud, and it is difficult to directly utilize the neighborhood information within the model to effectively denoise the color containing noise. SUMMARY

[0009] The purpose of the present application is to solve the technical problem that the existing three-dimensional point cloud model color denoising method cannot effectively preserve geometric features, and creatively proposes a point cloud color denoising method based on L0 norm minimization.

[0010] The innovation of the present application is that: first, decouple the position and color information, and create a point cloud RGB vector with a length of N using the extracted color information. Then check the RGB value of the position modified point, and if the color value of the point is greatly different from the gradient difference of the color value of the neighborhood point, it is proved that the point is located at the position of color mutation. Then, form a color-smooth surface with the point and its k nearest neighbors, modify the color value of the point by L0 norm minimization, and sequentially denoise the color value of other points in the point cloud along the position of the color-smooth point. Repeat the process until convergence.

[0011] The present application adopts the following technical solutions.

[0012] A point cloud color denoising method based on L0 norm minimization, specifically comprising the following steps:

[0013] Step 1: In a three-dimensional point cloud representation, input a three-dimensional color point cloud P with color noise 0 .

[0014] Wherein, P 0 is the number of vertices, and the position of the point set is P 0 ={P i ,i=0,1,…,N-1}.

[0015] Unlike traditional monochrome point cloud which only contains spatial information, color point cloud contains not only the spatial coordinates (X i ,Y i ,Z i ) of each point P i , but also the color value of each point P iColor three-channel RGB values ​​(R i G i B i ), denoted as P i =(X i ,Y i Z i ,R i G i B i All points P i Constructing a three-dimensional color point cloud P 0 .

[0016] First, extract the initial 3D color point cloud P. 0 RGB values ​​are used as vector L C , denoted as: For L C Each point P in i Color three-channel RGB values ​​(R i G i B i Calculations are performed separately to obtain the RGB value vector of the point cloud after color denoising.

[0017] For the known RGB value L at each point C Since vector W C The results contained therein are all unknown denoising results; therefore, W... C The initial RGB values ​​of all points are denoted as (0,0,0), and the vector L C The RGB value of a point is recorded as the actual value.

[0018] Step 2: Process the input 3D color point cloud P 0 Construct the k-nearest neighbor (kNN) graph.

[0019] Specifically, step 2 includes the following steps:

[0020] Step 2.1: For each 3D color point cloud P 0 ={P i Establish a kdTree for i = 0, 1, ..., N-1.

[0021] A kdTree, also known as a kd-tree, is a high-dimensional spatial index structure used to partition k-dimensional data spaces. Essentially, it's a constrained binary search tree. Binary tree search compares the value of the query node with the split dimension of the splitting node. If the value of the query node is less than the split dimension, it enters the left subtree branch; if the value is equal to the split dimension, it enters the right subtree branch until a leaf node is reached. The search path of the binary tree finds the nearest neighbor, which is the leaf node in the same subspace as the query node.

[0022] Step 2.2: Using binary search, the nearest approximate point along the search path can be found, which is the leaf node. However, the found leaf node is not necessarily the nearest in the global path. The nearest is closer to the query point, so it should be located within a circle centered at the query point and passing through the leaf node. To find the true nearest neighbor, a "backtracking" operation is required: search backward along the search path for a data point closer to the query point. If the nearest neighbor is not found in the space of that node, it is necessary to jump to the space of other child nodes and add them to the search path.

[0023] Step 2.3: Repeat the above process until the search path is empty.

[0024] Step 3: Process the initial 3D color point cloud P 0 The vector of each point in Each point P i Color three-channel RGB values ​​(R i G i B i The color denoising energy function of the 3D point cloud model is called respectively, and the final denoised RGB vector is obtained by minimizing the energy function. As shown in Equation 1:

[0025]

[0026] Among them, the first item |W c -L c | 2 The first term, |D(C)|0, is a data fidelity term to ensure that the color values ​​of the output point cloud do not differ too much from those of the input point cloud. The second term, |D(C)|0, is a regularization smoothing term that enhances edge contrast by limiting the number of non-zero gradients, while simultaneously implementing smoothing globally. λ is a preset positive scalar used to measure the balance between the two terms in the energy function. |·|0 is the L0 norm, i.e., the number of all non-zero elements in the vector. C represents the RGB values ​​of the input point cloud. D(C) represents the sum of the color gradient differences between a point in the point cloud and its k neighboring points. Given D(C)... i*k+j =C i -Cd(i,j) , C i represents the color value of the i-th point in the point cloud model, C d(i,j) is the position of the j-th point in the k nearest neighbor set of point i. Therefore, D(C) is a vector, the length of which is k times the number of point clouds in the input point cloud P 0 , that is, k*N. Every k vectors represent the gradient difference in k directions of a certain point.

[0027] By minimizing the global energy function, W C has the highest similarity with L C .

[0028] Further, W C is obtained by the following method:

[0029] Step 3.1: First, introduce the auxiliary matrix , . The size of the auxiliary matrix is (k*N)×N, which is obtained by the following process:

[0030] For the input three-dimensional point cloud, construct a kNN graph, for each point P 0 ={P i , i=0,1,…,N-1} in the three-dimensional color point cloud, put the point into the matrix and mark the position of the point in the matrix as 1. According to the constructed kNN graph, find the nearest neighbor of each point, mark the position of the nearest neighbor of the point in the matrix as -1, and mark the remaining positions in the matrix as 0, which are stored as a row in the matrix. Each point is stored in K rows, and a global matrix with K*N rows and N columns is obtained. As shown in equation 2:

[0031]

[0032] Step 3.2: To solve the optimization problem of W C , introduce an auxiliary variable θ. Corresponding to D(C), the length of the auxiliary variable θ is k*N, so that equation 2 is easy to solve, and therefore the color denoising energy function is converted to the following equation:

[0033]

[0034] where the initial value of θ is set as θ i =0, with a length of K*N.

[0035] Step 3.3: Set the initial smoothing weight parameters β0, β max , and the growth rate k rate .

[0036] where the parameters β0, β max are weights that directly control the difference between D(C) and θ, and krate The growth rate of β0is k rate times larger than β0in each iteration. Set the initial value of k rate as k rate = 2β0.

[0037] Step 3.4: Convert the solution of the minimization function into two sub-problems, i.e., solving θ,W C respectively. Specifically as follows:

[0038] Step 3.4.1: Set the initial value of β0as 2, and the initial value of k rate is 2, and β0is enlarged by 2 times in each iteration.

[0039] Step 3.4.2: Fix θ, and solve the W C sub-problem:

[0040] According to equation 4, by derivation of W C , set the derived formula equal to 0, solve the minimization energy function, and get the most expected W C :

[0041]

[0042] Wherein, W e C represents the updated color value vector of W C in the e-th iteration, β e represents the enlarged value of β0in the e-th iteration, θ e represents the updated value of θ in the e-th iteration

[0043] Step 3.4.3: Fix W C , and solve the θ sub-problem:

[0044] Since the color denoising energy function of the three-dimensional point cloud can be decomposed in space, each element can be calculated separately, so the energy function is decomposed as:

[0045]

[0046] According to equation 6, by derivation of θ, set the derived formula equal to 0, solve the minimization energy function, and get the most expected θ as:

[0047]

[0048] Step 3.4.4: When the value of β0reaches a predefined threshold β max , the iteration stops.

[0049] Through the above steps, the new RGB value WC De-noising.

[0050] Advantages

[0051] Compared with the prior art, the method has the following advantages:

[0052] 1. According to the characteristics of the point cloud model, a fast and simple point cloud color de-noising framework is constructed;

[0053] 2. Compared with the existing point cloud de-noising method, the three-dimensional point cloud model kNN graph is constructed, and the color de-noising on the three-dimensional point cloud model is optimized based on the position information, so that the color mutation problem on the basis of preserving color and geometric characteristics is solved; secondly, the L0 norm minimization method is used to solve the color de-noising problem in the three-dimensional point cloud de-noising, and the problem of excessive smoothing or excessive sharpening caused by inaccurate estimation of the normal or local plane of the model when there is too much noise is solved.

[0054] The method has wide application prospects in the fields of augmented reality, robot navigation and city reconstruction. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a point cloud model example of the Greyc data set in the method;

[0056] Figure 2 is a flow chart of the method;

[0057] Figure 3 is a matrix diagram of the method; DETAILED DESCRIPTION

[0058] The method will be further described in detail below in combination with the drawings and examples.

[0059] This embodiment illustrates the detailed implementation process of applying the method in the road modeling scene.

[0060] ​Urban roads, as spatial links connecting different functional areas of a city and the main carriers of urban spatial information flow, require 3D models that are an indispensable and important component of digital cities. A typical 3D road model consists of a pavement model and its associated structures. 3D laser scanning technology can quickly acquire detailed, high-precision point cloud data of the road surface, and use this as a basis to construct a more refined road model. However, for road modeling, detection in complex environments is a significant challenge. First, for landscapes with complex terrain or abundant vegetation, point clouds are often affected by noise due to inherent limitations of the acquisition equipment or matching ambiguities in image reconstruction. For large objects or large-scale scenes, complete measurements cannot be performed in one go; multiple scans are necessary. Therefore, the scan results are often multiple point cloud data points with different coordinate systems and containing noise, which cannot fully meet the requirements for the realism and real-time performance of digital models. Noise in the point cloud significantly affects downstream tasks such as rendering, reconstruction, and analysis because the underlying structure is distorted. Therefore, point cloud denoising is crucial for related 3D vision applications. However, due to the irregularity and disorder of point clouds, point cloud denoising is challenging.

[0061] This embodiment uses the Greyc dataset, a 3D color model dataset created within the Greyc laboratory. It uses the NextEngine 3D color laser scanner to acquire real objects with different colors, geometries, and textures.

[0062] like Figure 1 As shown, all internal models are positioned along the positive Z-axis. The Greyc dataset contains 425 3D models across 15 classes. Using this as a baseline, the Point Cloud Library (PCL) tool is used to uniformly sample 1024 points on the triangular mesh surface and normalize them to a unit sphere to obtain a point cloud model with color noise.

[0063] like Figure 2 As shown, a point cloud color denoising method based on L0 norm minimization includes the following steps:

[0064] Step 1: In the 3D point cloud representation, input a 3D colored point cloud P with color noise. 0 P 0 The number of vertices is N, and the set of their positions is P. 0 ={P i ,i=0,1,…,N-1}. Unlike traditional monochrome point clouds that only contain spatial information, colored point clouds contain information for each point P. i Spatial coordinates (X i ,Y i Z i It also includes each point P. iColor three-channel RGB values ​​(R i G i B i ), denoted as P i =(X i ,Y i Z i ,R i G i B i ), all scan points P i Constructing a three-dimensional color point cloud P 0 .

[0065] Extracting the initial 3D color point cloud P 0 The RGB values ​​are denoted as vectors as follows: For L C Each point P in i Color three-channel RGB values ​​(R i G i B i The RGB value vectors of the point cloud after color denoising are calculated separately:

[0066] Wherein, for each point, the known RGB value L C Since vector W C It contains unknown denoising results, therefore, W C The initial RGB values ​​of all points are denoted as (0,0,0), and the vector L C The RGB value of a point is recorded as the actual value.

[0067] Step 2: For the input 3D point cloud model P with color noise 0 Construct a kNN (k-nearest-neighbor) graph. The specific steps are as follows:

[0068] Step 2.1: For each point P 0 ={P i For each element i = 0, 1, ..., N-1, a kdTree (also called a kd-tree) is constructed. This is a high-dimensional spatial index structure used to partition k-dimensional data spaces; essentially, it is a constrained binary search tree. Binary tree search involves comparing the value of the query node with the split dimension of the splitting node. If the value of the query node is less than the split dimension, the search proceeds to the left subtree branch; if the value is equal to the split dimension, the search proceeds to the right subtree branch until a leaf node is reached. The search path of the binary tree finds the nearest neighbor, which is the leaf node in the same subspace as the query node.

[0069] Step 2.2 The nearest neighbor in the search path can be found quickly by binary search along the search path. The found leaf node is not necessarily the nearest neighbor, which is closer to the query point and should be in the circle domain with the query point as the center and the leaf node as the radius. In order to find the real nearest neighbor, a 'backtracking' operation is needed: the operation searches in the reverse direction along the search path to find whether there is a data point closer to the query point. If there is a possibility, it is necessary to jump to other child node spaces to search (add other child nodes to the search path).

[0070] Step 2.3 Repeat the above process until the search path is empty.

[0071] Step 3: For each point in the initial point cloud P 0 , the color vector of each point P i in the point cloud model is denoted as the color three-channel RGB value (R i , G i , B i ), and the three-dimensional point cloud model color denoising energy function is called respectively to obtain the final denoising completed RGB vector .

[0072]

[0073] The formula is,

[0074] λ is a preset positive scalar, which is used to measure the balance of the two terms in the energy function; |·|0 is the L0 norm, that is, the number of all non-zero elements in the vector; D(C) represents the sum of the gradient difference between the color of a point in the point cloud and its k nearest neighbors. Therefore, D(C) i*k+j = C i -C d(i,j) , C i represents the color value of the i-th point in the point cloud model, and C d(i,j) is the position of the j-th point in the k nearest neighbor set of point i. Therefore, D(C) is a vector, and its length is k times the number of points in the input point cloud P 0 , that is, k*N. Every k vectors represent the gradient difference in k directions of a point.

[0075] By minimizing the global energy function, W C is most similar to L C .

[0076] Specifically, W CBy the following process:

[0077] Step 3. First, introduce the auxiliary matrix Make the size of the auxiliary matrix (k*N)X N is obtained by the following process: construct a KNN graph for the input three-dimensional point cloud, and then put each point P 0 ={P i ,i=0,1,…,N-1} in the matrix, and mark the position of the point in the matrix as 1. According to the constructed KNN graph, find the nearest neighbor of each point, and mark the position of the nearest neighbor of the point in the matrix as -1, and mark the remaining positions in the matrix as 0, and store each point in K rows as a row, and obtain a global matrix with K*N rows and N columns. The auxiliary matrix The schematic diagram is shown in Figure 3 , and the formula is converted to

[0078]

[0079] Step 3.2: In order to solve the optimization problem of W C , introduce an auxiliary variable θ corresponding to D(C), and the length of the auxiliary variable θ is k*N, so that the optimization formula is easy to solve, and therefore the color denoising energy function can be converted to the following formula:

[0080]

[0081] where the initial value of θ is set to θ i =0, and the length is K*N;

[0082] Step 3.3: Set the initial smoothing weight parameters β0, β max and the growth rate k rate ;

[0083] where the parameters β0and β max are weights that directly control the difference between D(C) and θ; k rate is the growth rate of β0, and β0is expanded by k rate times each time;

[0084] The initial value of k rate is set to k rate =2β0.

[0085] Step 3.4: Convert the solution of the minimum function into two sub-problems, i.e. θ, W C Two sub-problems are solved

[0086] Step 3.4.1: Set the initial value of β0to 2, and k rateThe initial value of β0 is 2, and β0 is expanded by 2 times in each iteration

[0087] Step 3.4.2: fixing θ, solving W C Sub-problem:

[0088] According to the formula, W is solved by derivation of W C , and the derived formula is equal to 0, and the minimum energy function is solved to obtain the most expected W C :

[0089]

[0090] Wherein, W e C represents the updated color value vector of W C at the e-th iteration, β e represents the value of β0 expanded at the e-th iteration, θ e represents the updated value of θ at the e-th iteration

[0091] Step 3.4.3: fixing W C , solving θ sub-problem:

[0092] Since the color denoising energy function of the three-dimensional point cloud can be decomposed in space, each element can be calculated separately, so the energy function can be decomposed as

[0093]

[0094] According to the formula, θ is solved by derivation of θ, and the derived formula is equal to 0, and the minimum energy function is solved to obtain the most expected θ as:

[0095]

[0096] Step 3.4.4: when the value of β0 reaches the predefined threshold β max , the iteration stops.

[0097] Through the above steps, the new RGB value W C of each point in the point cloud is denoised.

[0098] The experimental environment is Visual 2013 operation platform, and the three-dimensional point cloud denoising method designed by the application mainly depends on the following libraries: PCL1.8.1 (Point Cloud Library), Intel MKL (Intel MathKernel Library), and QT.

[0099] The experiment selects Greyc data set, which is a 3D color model data set established in GREYC laboratory, and uses NextEngine 3D color laser scanner to obtain real objects with different colors, geometrical shapes and textures. The data set contains 15 categories of 425 three-dimensional models in total. Based on this, 1024 points are uniformly sampled on the triangular mesh surface using Point Cloud Library (PCL) tools, and they are normalized into a unit sphere to obtain a point cloud model with color noise.

[0100] In the point cloud color denoising training based on L0 norm minimization, an initial smoothing weight parameter β0=2, β max =20 and a growth rate k rate =2 can be set.

[0101] The above describes the preferred embodiments of the present application, and the present application should not be limited to the content disclosed in the embodiments and the drawings. Any equivalent or modification made without departing from the disclosed spirit of the present application falls within the scope of the present application.

Claims

1. A point cloud color denoising method based on L0 norm minimization, characterized in that, Firstly, decouple the position and color information, and create a point cloud RGB vector with length N using the extracted color information; then view the RGB value of the position modification point, if the color value of the point is greatly different from the gradient difference of the color value of its neighborhood points, it proves that the point is located at the position of color mutation; then, form a color-smooth surface with the point and its k nearest neighbors, modify the color value of the point by L0 norm minimization, and sequentially denoise the color value of other points in the point cloud along the position of the color-smooth point; repeat the process until convergence; The method comprises the following steps: Step 1: In a three-dimensional point cloud representation, input a three-dimensional color point cloud P with color noise 0 ; wherein P 0 The number of vertices is N, and the position of the points is a point set P 0 = {P i , i = 0, 1, …, N-1}. A colored point cloud contains each point P i Spatial coordinates (X i ,Y i Z i It also includes each point P. i Color three-channel RGB values ​​(R i G i B i ), denoted as P i =(X i ,Y i Z i ,R i G i B i ); all points P i Constructing a three-dimensional color point cloud P 0 ; First, an initial three-dimensional color point cloud P is extracted 0 The RGB value in the middle as a vector L C , denoted as: The color three-channel RGB value (R C , G i , B i ) of each point P i in L i is calculated respectively, and the point cloud RGB value vector after color denoising is obtained For each point in the known RGB value L C The W C point initial RGB value is all recorded as (0, 0, 0), and the vector L C The point RGB value is recorded as the true value; Step 2: Compute the input 3D color point cloud P 0 Construct k-Nearest Neighbors, kNN for short. Step 2.1: Build a kdTree for each three-dimensional color point cloud P 0 = {P i , i = 0, 1,..., N - 1} Step 2.2: By binary search, the nearest approximate point in the search path, that is, the leaf node, can be found along the search path, which is located in the circular domain with the query point as the center and passing through the leaf node; Reverse search along the search path whether there is a data point closer to the query point, if the node has no nearest neighbor node in the space, it is necessary to jump to other child node spaces to search, and the other child nodes are added to the search path; Step 2.3: Repeat the above process until the search path is empty; Step 3: Color tri-channel RGB values (R, G, B) of each point P 0 Vector of each point in the middle Each point P i Color tri-channel RGB values (R i , G i , B i ) of each point P As shown in equation 1: wherein the first term |W c -L c | 2 is a data fidelity term to ensure that the output point cloud color values do not differ too much from the input point cloud color values, the second term |D(C)|0is a regularization smoothing term to enhance the edge contrast by limiting the number of non-zero gradients while achieving smoothing in a global manner; λ is a preset positive scalar to measure the balance of the two terms in the energy function; |·|0is the L0norm, i.e., the number of all non-zero elements in the vector; C represents the RGB value of the input point cloud; D(C) represents the sum of the gradient difference of the color between a certain point and its k neighboring points in the point cloud; D(C) i*k+j = C i -C d(i,j) , C i represents the color value of the i-th point in the point cloud model, C d(i,j) is the position of the j-th point in the k nearest neighbor set of point i; D(C) as a vector, the length is k times the number of point clouds in the input point cloud P 0 , i.e., k*N; every k vectors represent the gradient difference in k directions of a certain point; W is determined by minimizing a global energy function, E C Similarity to L C is highest; By the above steps, the new RGB value W of each point in the point cloud is completed C Denoising.

2. The point cloud color denoising method based on L0 norm minimization of claim 1, wherein, In step 3, W C was obtained by the following method: Step 3.1: First, introduce the auxiliary matrix The auxiliary matrix of size (k*N) x N is obtained by the following process: A kNN graph is constructed for the input three-dimensional point cloud, and for each point P in the three-dimensional color point cloud 0 = {P i , i = 0, 1, …, N-1}, the point is put into the matrix, and the position of the point in the matrix is marked as 1; the nearest neighbor of each point is found according to the constructed kNN graph, the position of the nearest neighbor of the point in the matrix is marked as -1, and the remaining positions in the matrix are marked as 0, which are stored in the matrix as a row, and each point is stored in K rows, to obtain a global matrix with K*N rows and N columns; as shown in formula 2: Step 3.2: To solve the optimization problem of W C Introduce an auxiliary variable θ; corresponding to D(C), the auxiliary variable θ length is k*N, so that formula 2 is easy to solve, therefore, the color denoising energy function is converted to the following formula: where the initial value of θ is set as θ i = 0, length K*N; Step 3.3: Set initial smoothing weight parameters β0, β max and growth rate k rate ; wherein the parameters β0, β max is a weight directly controlling the degree of difference between D(C) and θ, k rate is the growth rate of β0, and β0is enlarged by k rate times each time the iteration is performed; the initial value of k rate is set to: k rate = 2β0; Step 3.4: The solution to the minimization function is converted into two sub-problems, i.e., the solution to θ, W C The two sub-problems are solved separately.

3. The point cloud color denoising method based on L0 norm minimization of claim 2, wherein, Step 3.4 comprises the following steps: Step 3.4.1: Set the initial value of β0 as 2, k rate the initial value of β0 as 2, and each iteration of β0 is enlarged by 2 times in turn; Step 3.4.2: Fix θ, solve for W C Sub-problem: According to the derivation of formula 4 by taking the derivative of W C , setting the derived formula equal to 0, and solving the minimum energy function, the most expected W C is obtained: where W e C denotes the updated value of W C updated color value vector, β e denotes the updated value of β0 e denotes the updated value of θ Step 3.4.3: Fixing W C Solve the θ subproblem: The energy function is decomposed as: According to formula 6, the derivative of theta is obtained, and the derivative formula is equal to 0, and the minimum energy function is solved to obtain the most expected theta as: Step 3.4.4: The iteration stops when the value of β0reaches a predefined threshold β max .

Images