Point cloud denoising method for ring forging based on local density and improved fuzzy C-means
By combining radius filtering, mutual neighborhood degree division and Marshall distance clustering based on local density and improved fuzzy C-means, the problem of denoising noise in the ring forging point cloud data is solved, and the non-uniform distribution of data is achieved The effective denoising and smoothing of the noise improves the denoising efficiency and accuracy.
Patent Information
- Application Number
- CN202210081509.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-24
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-01-24
AI Technical Summary
When processing three-dimensional point cloud data, especially in ring forging point cloud data, there is a problem that noise points are difficult to denoise, especially in non-uniformly distributed data environments, traditional methods such as statistics, clustering, and distance. The density method has limitations, making it difficult to achieve effective noise removal and smoothing.
Using the point cloud denoising method of ring forgings based on local density and improved fuzzy C mean, noise is gradually removed through parameter adaptive radius filtering, mutual neighborhood degree division, local outlier factor calculation and Mahayana distance fuzzy C clustering Points, realize the denoising and smoothing of point cloud data.
Improve the local outlier calculation efficiency and the accuracy of the fuzzy C clustering algorithm, realize reliable denoising of non-uniformly distributed point cloud data, and has a wider range of application, and overcomes the use of radius filtering, outlier factor method and fuzzy alone. The limitations of C clustering method achieve better noise removal effect.
Smart Images

Figure CN114494059B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of three-dimensional point cloud outlier detection and denoising, in particular to a ring forging point cloud denoising method based on local density and improved fuzzy C-means, belonging to a denoising method based on local density and fuzzy C-means clustering. Background Art
[0002] In the forging manufacturing industry, ring forgings are the basic parts for manufacturing major equipment. The casting level is an important indicator of a country's industrial development level. The size detection of ring forgings can verify the forming size of the forgings and verify the rationality of the corresponding process parameters and forming process. Since the ambient temperature is high and the risk is high when rolling ring forgings, laser measurement can effectively improve the measurement efficiency and safety.
[0003] When using 3D laser scanning to obtain point cloud data, due to the complex environment of large forgings during the rolling process, there will be vibration interference during the measurement process, and there will be temperature change interference when the forging is hot, which makes the measured 3D data have a certain proportion of noise. It is inevitable that there will be noise points in the original data obtained by the test, which is not conducive to the subsequent point cloud processing. Therefore, it is necessary to perform denoising preprocessing on the obtained point cloud data.
[0004] The current methods for processing outliers in scattered point clouds are mainly divided into methods based on statistics, clustering, distance, and density. However, these methods have their own limitations. For example, the statistical method requires that all data in the sample space conform to a certain distribution or data model; the clustering method focuses on finding clusters, and has high requirements for the setting of cluster initial parameters when using this method, and also has restrictions on the cluster shape; the distance-based method requires that the data be evenly distributed. When the data is unevenly distributed, the distance-based outlier detection cannot be effective for global data; the density-based method can effectively detect global and local outliers, but for large amounts of data, frequent neighborhood queries have a greater impact on computational efficiency. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a ring forging point cloud denoising method based on local density and improved fuzzy C-means, to achieve denoising and smoothing of the model, to achieve reliable denoising of point cloud models with different density distribution areas, to achieve smoothing of the ring forging point cloud model while maintaining features, and to have a wider range of applications, and to be able to be used for denoising of scattered point clouds with non-uniform distribution, to achieve better denoising effects.
[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0007] The ring forging point cloud denoising method based on local density and improved fuzzy C-means includes the following steps:
[0008] Step 1: Use the parameter-adaptive radius filter denoising method to denoise outliers that deviate from the point cloud model structure and obtain the point cloud data after the initial denoising;
[0009] Step 2: Using the density-based point cloud denoising method, the data set is divided into a core set and an alternative set based on the value of mutual neighborhood degree; different outlier factor calculation methods are used for the data characteristics of the two data subsets, and large-scale outliers are screened out by outlier degree;
[0010] Step 3: Use the elbow method to calculate the number of clustering divisions c for the core set in step 2, propose a method to determine the initial cluster center position based on the size of the mutual neighborhood degree, and then achieve denoising and smoothing through the fuzzy C clustering method based on the Mahalanobis distance.
[0011] A further improvement of the technical solution of the present invention is that the radius filtering denoising method using parameter adaptation in step 1 includes the following steps:
[0012] Get the point cloud data set D of the laser scanner, and calculate the appropriate radius filter parameters according to the overall distribution of the point cloud data; where the number of data points is n; the number of neighboring points of data point p within the specified radius is q; the k points closest to p in the data set are the nearest neighbor points NN k (p);
[0013] Calculate any data point p i The average distance between points in the k nearest neighbors Take the average of the k-neighborhood average distances of each point to get the average distance of each point in the overall point cloud data.
[0014] Traverse each data point, within the filter radius Within the range of , if the number of neighboring points in its radius neighborhood reaches the specified minimum number of neighboring points Q, the point is retained, otherwise the point is recorded as an isolated point and deleted.
[0015] A further improvement of the technical solution of the present invention is that in step 2, the data set is divided into a core set and an alternative set, including the following steps:
[0016] Based on the k-neighborhood data of all objects counted in step 1, the farthest distance between p and o is determined as k-dist(p) among the k-neighboring points o of p, and the points in o that are k-neighboring points with p are taken as the mutual neighboring points of p, which are recorded as the mutual neighboring point set MNS of p, as shown in the following formula:
[0017] MNS k (p) = {o|o∈NN k(p)∧p∈NN k (q)}
[0018] The proportion of mutual neighboring points of object p in the k nearest neighboring points is defined as the mutual neighboring degree When the value of MND is greater than the set threshold μ, the data is marked as the core set D of point cloud data D a , the number is recorded as n a ; Otherwise, mark it as candidate set D b , the number is n b , as shown below:
[0019]
[0020] A further improvement of the technical solution of the present invention is that in step 2, calculating the outlier factors of the core set and the candidate set includes the following steps:
[0021] After dividing the point cloud data D into a core set and an alternative set, different methods are used to denoise the two data sets according to their different data characteristics;
[0022] The local outlier factor (LOF) method is used to calculate the outlier degree of each data point on the core set data. k The higher the value of (p), the more likely p is an outlier, and the calculation method is as follows:
[0023]
[0024] *When calculating the outlier factor for the candidate set data, not only the nearest neighbors of the data point should be considered, but also the anti-nearest neighbor RNN k (p) data, calculate the comprehensive outlier factor value INFLO k (p), as shown in the following formula:
[0025]
[0026] Among them, IS k (p) is the kth distance neighbor NN of object p k (p) and the inverse k-th distance neighborhood RNN k (p), den(p) refers to the local density of object p, expressed as the inverse of the k-distance of object p.
[0027] A further improvement of the technical solution of the present invention is that in step 3, the specific process of using the improved elbow method to determine the optimal number c of clustering divisions and the set V of the c cluster centers includes the following steps:
[0028] Sort the mutual neighborhood degrees of the core points from large to small, select the reference cluster centers in order, and use them as the selection order of the reference points;
[0029] Calculate the sum of the squares of the distances from each sample point to the reference cluster center in the nearest core set to which it belongs; obtain the sum of squares of errors SSE corresponding to each cluster quantity c through iteration, and determine the quantitative relationship between quantity c and SSE; obtain the slope between adjacent points during the iteration process, and then calculate the ratio r of the slope k between two adjacent points. The calculation method is shown in the following formula:
[0030]
[0031] If the slope ratio r i When the value is less than the threshold ε or c = θ in the iterative calculation, the iteration is stopped and r is taken from the calculated result. i The point with the largest value of is the elbow point, and the optimal number of clusters is obtained from it;
[0032] Then, sort the points from large to small by the mutual neighborhood degree MND value, and select the maximum value point as the first cluster center c1 in this order; select the points p with smaller MND one by one in order from large to small, and screen them according to the conditions determined by the cluster center set below
[0033] a) Determine the data point p and the i cluster centers v1 to v1 that have been determined i The Euclidean distance of the coordinates of each point in the cluster. When the distance between point p and other cluster centers is greater than the preset threshold α, condition b is judged. Otherwise, the point is skipped to avoid the initial cluster center selection being too concentrated.
[0034] b) Point p to be determined i The local density is greater than p i The local density of the mutual neighboring points is the cluster center point of the load condition, as shown in the following formula (10):
[0035]
[0036] A further improvement of the technical solution of the present invention is that in step 3, a fuzzy C clustering algorithm based on Mahalanobis distance is used to smooth the small-scale noise, including the following steps:
[0037] Based on the obtained initial number of cluster centers and cluster center coordinates, the FCM algorithm of Mahalanobis distance is constructed to determine the objective function as shown in the following formula (11):
[0038]
[0039] u in the formula ij The value of must meet the corresponding constraints;
[0040] Then the Lagrange multiplier method is used to obtain the membership of each point relative to the cluster center, and the obtained membership is used to smooth the small-scale noise. The filter function is shown as follows:
[0041]
[0042] Where p' is the point after smoothing and denoising; λ is the smoothing coefficient, and its value range is λ∈[0,1]. When the smoothing coefficient is larger, the filtering process is smoother, and when the value is smaller, the model will not change much; n p is the direction of the normal vector at point p; u i is the cluster center v of point p i The membership value between i Take the cluster center with the maximum membership degree in V corresponding to point p.
[0043] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:
[0044] The present invention improves the point cloud filtering model from two aspects of improving the calculation efficiency of local outliers and improving the accuracy of the fuzzy C clustering algorithm, thereby realizing denoising and smoothing of the model. The present invention realizes reliable denoising of point cloud models with different density distribution areas, realizes smoothing of annular forging point cloud models on the basis of maintaining features, and has a wider scope of application. The present invention is a denoising method for scattered point clouds that can be used for non-uniform distribution.
[0045] The present invention can realize denoising layer by layer in a coarse to fine order according to the deviation degree of noise points. Compared with the effects of radius filtering, outlier factor method and fuzzy C clustering method used alone, the method can overcome the limitations of the use conditions of each method and achieve better denoising effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a basic flow chart of the present invention;
[0047] Figure 2 This is the point cloud model of the ring forging after adding noise;
[0048] Figure 3 This is the denoising result after removing the outliers using the method in step 1 of this article;
[0049] Figure 4 This is the denoising result after removing large-scale outliers using the method in step 2 of this article;
[0050] Figure 5 This is the denoising result after removing small-scale outliers using the method in step 3 of this article. DETAILED DESCRIPTION
[0051] The present invention is further described in detail below in conjunction with embodiments:
[0052] like Figures 1 to 5 As shown, the present invention provides a method for denoising a ring forging point cloud based on local density and improved fuzzy C-mean, comprising the following steps:
[0053] Step 1: Use the parameter-adaptive radius filter denoising method to denoise outliers that deviate from the point cloud model structure, and obtain the point cloud data after the initial denoising, including:
[0054] Get the point cloud data set D of the laser scanner, where the number of data points is n; let the number of neighboring points of data point p within the specified radius be q; the k points closest to p in the data set are the nearest neighbor points NN k (p).
[0055] Calculate any data point p i The average distance between points in the k nearest neighbors Take the average of the k-neighborhood average distances of each point to get the average distance of each point in the overall point cloud data.
[0056] Traverse each data point, within the filter radius Within the range of , if the number of neighboring points in its radius neighborhood reaches the specified minimum number of neighboring points Q, the point is retained, otherwise the point is recorded as an isolated point and deleted.
[0057] The noise points that deviate from the main body of the point cloud model can be removed based on the parameter-adaptive radius filtering method.
[0058] Step 2: Use the density-based point cloud denoising method to divide the data set into a core set and an alternative set based on the value of mutual neighborhood degree. Use different outlier factor calculation methods for the data characteristics of the two data subsets to filter out large-scale outliers through outlier degrees, including:
[0059] By using the outlier classification detection based on local density, the outliers around the point cloud model in the result of step 1 are denoised to obtain the point cloud data with large-scale noise removed.
[0060] Based on the k-neighborhood data of all objects counted in step 1, the farthest distance between p and o is determined as k-dist(p) among the k-neighboring points o of p, and the parts of the points o that are k-neighboring points with p are taken as the mutual neighboring points of p, which are recorded as the mutual neighboring point set MNS of p, as shown in formula (1).
[0061] MNS k (p) = {o|o∈NN k (p)∧p∈NN k (q)} (1)
[0062] The proportion of mutual neighboring points of object p in the k nearest neighboring points is defined as the mutual neighboring degree When the value of MND is greater than the set threshold μ, the data is marked as the core set D of point cloud data D a , the number is recorded as n a ; Otherwise, mark it as candidate set D b , the number is n b , as shown in formula (2).
[0063]
[0064] After dividing the point cloud data D into a core set and an alternative set, different methods are used to denoise the two data sets according to their different data characteristics.
[0065] The local outlier factor (LOF) method is used to calculate the outlier degree of each data point on the core set data. k The higher the value of (p), the greater the degree to which p is an outlier; specifically, the following steps are included:
[0066] For the core set D a For any object p in the data, the reachable distance is calculated as:
[0067] rdist k (p,o)=max{k-dist(o),||po||} (3)
[0068] Calculate the local reachability density lrd of data point p k (p), the expression is as follows:
[0069]
[0070] In the formula (xx), |NN k (p)| is the number of objects in the k nearest neighbors of point p.
[0071] Calculate the local outlier factor LOF k (p), which is the k-distance neighborhood NN of object p k The average of the ratios of the local reachability density of other objects in (p) to the local reachability density of object p. The calculation method is shown in formula (5).
[0072]
[0073] When calculating the outlier factor for the candidate set data, not only the nearest neighbors of the data point should be considered, but also the anti-nearest neighbor RNN k (p) data, calculate the comprehensive outlier factor value INFLO k (p), as shown in formula (6).
[0074]
[0075] Among them, IS k (p) is the kth distance neighbor NN of object p k (p) and the inverse k-th distance neighborhood RNN k (p), den(p) refers to the local density of object p, which is expressed as the inverse of the k-distance of object p, as shown in Equations (7) and (8).
[0076] IS k (p) = NN k (p)∪RNN k (p) (7)
[0077]
[0078] By classifying the data in the core set and the candidate set according to the mutual neighborhood degree, the outlier factor LOF and the comprehensive outlier factor INFLO are calculated for the data in the divided core set and the candidate set. The relationship between the outlier factor threshold t is judged, and the outlier degree greater than the threshold t is an outlier point.
[0079] Step 3: Use the elbow method to calculate the number of clustering divisions c for the core set in step 2, propose a method to determine the initial cluster center position based on the mutual neighborhood degree, and then use the fuzzy C clustering method based on the Mahalanobis distance to achieve denoising and smoothing, including:
[0080] Use the clustering method to select the appropriate cluster center by optimizing the number of cluster centers and the location of the center points, and avoid the problem that the algorithm easily falls into the local optimal solution when randomly selecting the initial cluster center.
[0081] Starting from the determined optimal initial value of cluster center, fuzzy C clustering based on Mahalanobis distance is used to iteratively cluster the data and calculate the membership matrix. According to the membership, the small-scale noise points are smoothed and denoised.
[0082] Step 3.1 Use the improved elbow method to determine the optimal number of clustering partitions c and the set V of these c cluster centers, where V = {v1, v2, …, v c}.
[0083] The core concentrated data is used as the point that may become the data center, and the number of clusters c is iteratively calculated. The value range of the number of clusters is from 1 to θ.
[0084] in, n a is the number of data points in the core set, and m is the number of data points in the core set and the alternative set as a whole.
[0085] Sort the mutual neighborhood degrees of the core points from large to small, select the reference cluster centers in order, and use them as the selection order of the reference points.
[0086] Calculate the sum of the squares of the distances from each sample point to the reference cluster center in the nearest core set to which it belongs. Through iteration, the sum of the squared errors (SSE) corresponding to each cluster number c is obtained, and the quantitative relationship between the number c and SSE is determined. The coordinates of each point are recorded as (x i ,y i ).
[0087] The optimal value of c is determined by the change in the slope of the line graph of SSE decreasing with the increase of c. As the value of c increases, the degree of aggregation of the cluster represented by SSE will gradually increase, first decreasing sharply and then slowing down.
[0088] When c=i and c=i+1, the slope between the two points is After obtaining the slope between adjacent points in the iterative process, the ratio r of the slope k between two adjacent points is calculated. The calculation method is shown in formula (9):
[0089]
[0090] If the slope ratio r i When the value is less than the threshold ε or c = θ in the iterative calculation, the iteration is stopped and r is taken from the calculated result. i The point with the largest value of is the elbow point, and the optimal number of clusters is obtained from it.
[0091] Step 3.2 selects points as initial cluster centers in the core set based on mutual neighborhood values.
[0092] Take the core set as the selected sample set, sort the mutual neighborhood degree MND values obtained in step 3.1 from large to small, and select the maximum value point as the first cluster center c1 in this order.
[0093] In order from large to small, select points p with smaller MND one by one, and screen them according to the conditions determined by the cluster center. The data points that meet the conditions are included in the cluster center V. After c points that meet the conditions are selected, the iteration process ends. The screening conditions include:
[0094] a) Determine the data point p and the i cluster centers v1 to v1 that have been determined i The Euclidean distance of the coordinates of each point in the cluster. When the distance between point p and other cluster centers is greater than the preset threshold α, condition b is judged. Otherwise, the point is skipped to avoid the initial cluster center selection being too concentrated.
[0095] b) Point p to be determined i The local density is greater than pi The local density of the mutual neighboring points is the cluster center point of the load condition, as shown in formula (10):
[0096]
[0097] The initial cluster center V is determined by the above method of determining the optimal cluster number c and selecting the location of the cluster center.
[0098] Step 3.3 constructs the objective function of the fuzzy C clustering method based on Mahalanobis distance, calculates the cluster center through continuous iteration and replaces the cluster center of the previous iteration with the new cluster center until the change of the cluster center is less than the threshold ω.
[0099] At this time, the sample set X={x1,x2,…,x m}, there are m samples in total, x i ∈R 3 The initial cluster center V = {v1,v2,…,v c}. X represents an m×3 input matrix. Step 3.2 determines c subsets and divides X into subsets S1, S2, …, S i ,…S c .
[0100] The objective function determined by the FCM algorithm combined with the Mahalanobis distance is shown in formula (11):
[0101]
[0102] where u ij Represents sample data x j For S i The membership degree of m is used to limit the fuzziness of the membership matrix. The larger the m, the higher the fuzzy classification degree. Σ represents the covariance matrix of the sample. ij The value of must meet the following constraints as shown in equations (12) and (13):
[0103] u ij ≥0(1≤i≤c,1≤j≤n) (12)
[0104]
[0105] The Lagrangian function is constructed by combining the constraints as shown in formula (14):
[0106]
[0107] Take the first-order partial derivative of the variables in the Lagrangian function and set it to zero, so as to determine the membership u in the condition for F to take the extreme value ij and cluster center v iAs shown in formula (15) and (16):
[0108]
[0109]
[0110] Initialize the cluster center matrix V with the position of the cluster center V in the calculation result (i) , when ||V (i+1) -V (i) Stop the iteration when ||≤ω, otherwise repeat step 3.3.
[0111] Step 3.4 smoothes the small-scale noise according to the degree of membership. The filtering formula is shown in formula (17):
[0112]
[0113] Where p' is the point after smoothing and denoising; λ is the smoothing coefficient, and its value range is λ∈[0,1]. When the smoothing coefficient is larger, the filtering process is smoother, and when the value is smaller, the model will not change much; n p is the direction of the normal vector at point p; u i is the cluster center v of point p i The membership value between i Take the cluster center with the maximum membership degree in V corresponding to point p.
[0114] The above preferred examples are only used to illustrate the technical solutions of the present invention rather than to limit the present invention. Although the present invention has been described in detail through the above specific examples, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.
Claims
1. Annular forging point cloud denoising method based on local density and improved fuzzy C-means, Features: The following steps are involved: Step 1: Use the parameter-adaptive radius filter denoising method to denoise outliers that deviate from the point cloud model structure and obtain the point cloud data after the initial denoising; Step 2: Using the density-based point cloud denoising method, the data set is divided into a core set and an alternative set based on the value of mutual neighborhood degree; different outlier factor calculation methods are used for the data characteristics of the two data subsets, and large-scale outliers are screened out by outlier degree; Step 3: Use the elbow method to calculate the number of clustering divisions c for the core set in step 2, propose a method to determine the initial cluster center position based on the size of the mutual neighborhood degree, and then use the fuzzy C clustering method based on the Mahalanobis distance to achieve denoising and smoothing; The radius filtering denoising method using parameter adaptation in step 1 includes the following steps: Get the point cloud data set D of the laser scanner, and calculate the appropriate radius filter parameters according to the overall distribution of the point cloud data; where the number of data points is n; the number of neighboring points of data point p within the specified radius is q; the k points closest to p in the data set are the nearest neighbor points NN k (p); Calculate any data point p i The average distance between points in the k nearest neighbors The average k-neighborhood distance of each point is then averaged to obtain the average neighborhood distance of each point in the overall point cloud data. Traverse each data point, within the filter radius Within the range of , if the number of neighboring points in its radius neighborhood reaches the specified minimum number of neighboring points Q, then the point is retained, otherwise the point is recorded as an isolated point and deleted; In step 2, the data set is divided into a core set and an alternative set, including the following steps: Based on the k-neighborhood data of all objects counted in step 1, the farthest distance between p and o is determined as k-dist(p) among the k-neighboring points o of p, and the points in o that are k-neighboring points with p are taken as the mutual neighboring points of p, which are recorded as the mutual neighboring point set MNS of p, as shown in the following formula: MNS k (p)={o|o∈NN k (p)∧p∈NN k (q)} The proportion of mutual neighboring points of object p in the k nearest neighboring points is defined as the mutual neighboring degree When the value of MND is greater than the set threshold μ, the data is marked as the core set D of point cloud data D a , the number is recorded as n a ; Otherwise, mark it as candidate set D b , the number is n b , as shown below: In step 2, calculating the outlier factors of the core set and the candidate set includes the following steps: After dividing the point cloud data D into a core set and an alternative set, different methods are used to denoise the two data sets according to their different data characteristics; The local outlier factor (LOF) method is used to calculate the outlier degree of each data point on the core set data. k The higher the value of (p), the more likely p is an outlier, and the calculation method is as follows: When calculating the outlier factor for the candidate set data, not only the nearest neighbors of the data point should be considered, but also the anti-nearest neighbor RNN k (p) data, calculate the comprehensive outlier factor value INFLO k (p), as shown in the following formula: Among them, IS k (p) is the kth distance neighbor NN of object p k (p) and the inverse k-th distance neighborhood RNN k (p), den(p) refers to the local density of object p, expressed as the inverse of the k-distance of object p; In step 3, the specific process of using the improved elbow method to determine the optimal number c of clustering divisions and the set V of the c cluster centers includes the following steps: Sort the mutual neighborhood degrees of the core points from large to small, select the reference cluster centers in order, and use them as the selection order of the reference points; Calculate the sum of the squares of the distances from each sample point to the reference cluster center in the nearest core set to which it belongs; obtain the sum of squares of errors SSE corresponding to each cluster quantity c through iteration, and determine the quantitative relationship between the quantity c and SSE; obtain the slope between adjacent points during the iteration process, and then calculate the ratio r of the slope k between two adjacent points. The calculation method is shown in the following formula: If the slope ratio ri is less than the threshold ε or c = θ in the iterative calculation, the iteration is stopped and r is taken from the calculated result. i The point with the largest value of is the elbow point, and the optimal number of clusters is obtained from it; Then, sort the points from large to small by the mutual neighborhood degree MND value, and select the maximum value point as the first cluster center c1 in this order; select the points p with smaller MND one by one in order from large to small, and screen them according to the conditions determined by the cluster center set below a) Determine the data point p and the i cluster center points v that have been determined 1 to v i The Euclidean distance of the coordinates of each point in the cluster. When the distance between point p and other cluster centers is greater than the preset threshold α, condition b is judged. Otherwise, the point is skipped to avoid the initial cluster center selection being too concentrated. b) Point p to be determined i The local density is greater than p i The local density of the mutual neighboring points is the cluster center point of the load condition, as shown in the following formula (10): p(p i )≥ρ(p j ), In step 3, the small-scale noise is smoothed using a fuzzy C clustering algorithm based on Mahalanobis distance, including the following steps: Based on the obtained initial number of cluster centers and cluster center coordinates, the FCM algorithm of Mahalanobis distance is constructed to determine the objective function as shown in the following formula (11): u in the formula ij The value must conform to in ij ≥0(1≤i≤c,1≤j≤n) Then the Lagrange multiplier method is used to obtain the membership of each point relative to the cluster center, and the obtained membership is used to smooth the small-scale noise. The filter function is shown as follows: Where p' is the point after smoothing and denoising; λ is the smoothing coefficient, and its value range is λ∈[0,1]. When the smoothing coefficient is larger, the filtering process is smoother, and when the value is smaller, the model will not change much; n p is the normal vector direction of point p; u i is the cluster center v of point p i The membership value between i Take the cluster center with the maximum membership degree in V corresponding to point p.
