Method for removing multipath noise from underwater sonar images based on geometric probability
By applying geometric probability method in underwater sonar images, combined with SCUNet, DBSCAN and K-means++ algorithms, multipath effect noise is removed, and the problem of multipath noise removal in underwater sonar is solved, improving the accuracy of three-dimensional reconstruction and stability.
Patent Information
- Application Number
- CN202410671471.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-28
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-05-28
AI Technical Summary
The prior art lacks efficient and adaptive robust methods to remove multipath effect noise in underwater imaging sonar, affecting the accuracy of three-dimensional object reconstruction of the NeRF method of neural radiation field.
Using a geometric probability-based method, the sonar image is preprocessed, clustered and multi-path echo effect point removal, and the multi-path effect point is detected and removed by using SCUNet, DBSCAN and K-means++ algorithms to detect and remove multi-path effect points to adapt to different perspectives and Underwater environment.
It significantly improves the quality of sonar images, reduces the error of three-dimensional reconstruction, enhances the accuracy of reconstruction results, reduces interference from multipath effect, and avoids perforation and reconstruction of false objects.
Smart Images

Figure CN118537250B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of sonar image noise removal, and in particular to a method for removing multipath effect noise from underwater sonar images based on geometric probability. Background Art
[0002] With the advancement of sensor, energy and communication technologies, the performance of underwater robots in underwater perception has been greatly improved, and they are widely used in the fields of seabed topography mapping, marine ecological monitoring, underwater resource exploration, etc. Imaging sonar plays a vital role in the navigation and positioning of underwater robots because it is not limited by light and has better penetration and directionality.
[0003] In recent years, neural radiance field (NeRF) technology has achieved remarkable success in 3D reconstruction of objects using imaging sonar, especially showing powerful capabilities in new view synthesis and surface reconstruction.
[0004] However, in real-world scenarios, noise in 3D reconstructions comes from various sources, including reverberation and inaccurate sensor readings, most notably multipath reflections from objects or structures.
[0005] Efficient and accurate noise removal can greatly help the reconstruction of 3D objects by NeRF method, making the reconstructed 3D objects more accurate. However, if the noise of the dataset is not removed, the inaccurate dataset will have a great adverse effect on the precise structure of the modeled object, which may endanger the authenticity of the reconstructed 3D structure.
[0006] According to relevant information, there have been successful studies on mitigating radar multipath effects. However, their methods are not applicable to underwater imaging sonar scenarios. According to the information, there is currently no efficient and adaptive robust method to remove multipath effect noise in imaging sonar. Summary of the invention
[0007] In view of the above problems in the prior art, the present invention proposes a method for directly detecting and removing multipath effect points on a single two-dimensional sonar image based on the geometric principle of multipath effect imaging, that is, a method for removing multipath effect noise from underwater sonar images based on geometric probability.
[0008] The technical solution adopted by the present invention is as follows:
[0009] A method for removing multipath effect noise from underwater sonar images based on geometric probability comprises the following steps:
[0010] S1: Preprocess and denoise the sonar image;
[0011] S2: clustering the echo points of the pre-processed and denoised image;
[0012] S3: Remove multipath echo effect points from the clustered image.
[0013] Preferably, S1 specifically comprises the following steps:
[0014] S1.1: Preprocess sonar images using the open source SCUNet denoising model;
[0015] S1.2: The sonar image denoised in S1.1 is subjected to secondary denoising, i.e., a threshold operation is applied after normalizing the sonar echo values, and echoes below the predefined threshold are systematically set to zero;
[0016] S1.3: Reset the values of the area far away from the sonar equipment in the sonar image after the secondary denoising in 1.2 to zero.
[0017] Preferably, the SCUNet denoising model in S1.1 selects the grayscale images-25 denoising mode.
[0018] Preferably, S2 specifically includes the following steps:
[0019] S2.1: Map the sonar image dataset stored in the rectangular view onto a fan-shaped view plane and establish the coordinate grid of the fan-shaped view by the following formula
[0020] r′=H′δr+r min
[0021]
[0022] Among them, r is the radius of the fan view, Indicates the angle range of the fan view, r min is the minimum distance, θ represents the angular range of the sonar horizontal angle, δr represents the resolution of the sonar in the r direction, It represents the resolution of the sonar in the horizontal angular direction, and (H', W') is the horizontal and vertical range of the rectangular view.
[0023] Map the valid points in the rectangular view to the fan view and calculate their coordinates (x, y) in the fan view
[0024]
[0025] S2.2: Apply DBSCAN algorithm to cluster valid echo points;
[0026] S2.3: Apply the K-means++ algorithm to each main cluster {L1,L2,L3,...,L m , ...}.
[0027] Preferably, S2.2 is specifically:
[0028] S2.2.1: From the large class C in DBSCAN i Randomly select a center among the data points;
[0029] S2.2.2: For each data point P i , calculate D(P i ), namely P i The distance to the closest center that has already been chosen;
[0030] S2.2.3: Randomly select a new data point as the new center using a weighted probability distribution, where the probability of selecting point P is the same as D(P) 2 is proportional to;
[0031] S2.2.4: Repeat steps S2.2.2 and S2.2.3 until K centers are selected;
[0032] S2.2.5: After determining K center points according to the above steps, select these K points as the cluster centers for initial aggregation;
[0033] S2.2.6: Calculate the Euclidean distance from each sample point to the K cluster cores, find the cluster core closest to the point, and assign it to the corresponding cluster;
[0034] S2.2.7: After all points are assigned to clusters, the M points are divided into K clusters; then the centroid or average distance center of each cluster is recalculated and set as the new cluster core;
[0035] S2.2.8: Repeat steps S2.2.6-S2.2.7 until a termination condition is reached. The termination condition for this experiment is 200 iterations of the S2.2.6-S2.2.7 process.
[0036] After the clustering operation is completed, all valuable points are clustered into small clusters {c1, c2, c3, ..., c i}.
[0037] Preferably, the valid echo point in S2.2 means that the multipath effect echo must meet the following three conditions:
[0038] a) The distance to the echo generated by the second reflection must be smaller than the distance between multipath effect points;
[0039] b) The echo from the multipath effect and the echo from the second reflection are in the same direction as the sonar;
[0040] c) The distance from the multipath effect point to the sonar origin is within the range of the total distance traveled during the second reflection.
[0041] Preferably, a minimum distance constraint is introduced for the valid echo points in S2.2, that is, a minimum distance judgment is added between the two classes found, and only those exceeding the minimum distance threshold are considered to be multipath effect pairs. This constraint is used to reduce the misjudgment of multipath effect points.
[0042] Preferably, S2.3 is specifically:
[0043] The initial number of clusters for the K-means++ clustering algorithm is pre-specified, denoted as k; the number of clusters specified is proportional to the number of echoes in the main cluster;
[0044] The algorithm is as follows,
[0045] 1) Initialization: The main data points L obtained from DBSCAN m The first cluster E1 is randomly selected;
[0046] 2) For i=2 to k:
[0047] a. Calculate each data point x i The square distance D(x i ) 2 , where D(x i ) is from x i The distance to the nearest cluster center;
[0048] b. Select a new cluster center E according to the algorithm described in S2.2.5 i ;
[0049] 3) Run the standard K-means algorithm using the selected cluster centers for initialization;
[0050] 4) In each iteration, perform the following steps:
[0051] a. For each sample x i , calculate its distance to each cluster center and assign it to the cluster with the closest center
[0052]
[0053] Among them, E i is the sample x i The cluster to which it belongs, μ j is the center of cluster j;
[0054] b. For each cluster j, recalculate its center μj As the average value of all samples in the cluster
[0055]
[0056] Among them, |E j | represents the number of samples in cluster j;
[0057] c. Repeat steps a and b until the cluster assignment result is stable or reaches the specified termination criterion or number of iterations;
[0058] After the clustering process is completed, all valid points are grouped into separate small clusters {C1, C2, C3, ..., C m ,…}.
[0059] Preferably, S3 specifically includes the following steps:
[0060] S3.1: Randomly select a completed clustering operation cluster, denoted as C i , whose cluster center coordinates are represented by X i ; Then, the corresponding line segment OX is obtained by connecting it with the sonar image origin O i ;
[0061] S3.2: Traverse all other clusters and calculate the number of nodes from the cluster center X j To line segment OX i The distance is calculated and the distance threshold a is set; if the distance is within the threshold a, the intersecting clusters X are marked j , and classify it into category D i ;
[0062] S3.3: First, determine point P k Is it related to line segment OX? i intersect;
[0063] Then, in the intersecting cluster D i All points P in each cluster k Traverse on; any point P that meets the intersection condition k are grouped into set S; if none of the potentially intersecting clusters meets the intersection criteria, meaning S is an empty set, the algorithm proceeds to the next cluster C i for subsequent iterations;
[0064] In addition, if the distance between the centers of two clusters is less than the specified threshold c, they are considered to be in contact, thus preventing misjudgment as a multipath effect;
[0065] S3.4: Using the exhaustive enumeration method, iterate the remaining clusters and calculate the distance OX from their cluster centers to the sonar image origin O m ; Calculate OXi Cluster X that intersects with the labels in S2.3 j The centroid coordinates of all intersection points in the j , use E to represent X m and Y j The potential distance between m and Y j In the actual environment, the position will be anywhere within the elevation range of the sonar scan, so the range of E is determined by X m and Y j The total possible distance C covered by the sound wave through secondary reflection is calculated as OX m +OY j +E, get the calculated range [C min ,C max ];
[0066] S3.5: Let G represent OX i twice the length of G if it falls on [C min ,C max ] range, it is considered as a potential point affected by the multipath effect; if it is outside the range, the offset value distance is calculated from the nearest end, and the value is substituted into the Gaussian function to calculate the probability p that it is a multipath affected point. If the probability p is greater than the set threshold, it is considered as a potential multipath affected point;
[0067] Finally, in the rectangular view, all echo values corresponding to points in the cluster identified as being affected by multipath effects are set to zero.
[0068] Preferably, the determination point P in S3.3 k Is it related to line segment OX? i Intersection is determined by the following two criteria:
[0069] 1) Judge by the intersection coefficient e
[0070]
[0071] Among them, the intersection coefficient e indicates point P k Relative to OX i The relative position of the projection on
[0072] Since point P k In the line segment OX i The projection on must be between O and X i So the range of e is from 0 to 1;
[0073] 2) Set the angle threshold b and calculate the vector OP k and OX iThe angle between them is calculated and the intersection is confirmed when the angle is below a specified threshold b for collinearity.
[0074] The present invention has the following characteristics and advantages:
[0075] 1) In terms of the processing results of sonar image data: pre-processing improves image quality through denoising, and clustering helps to distinguish physical object echoes from multipath effect echoes; the algorithm can adapt to data sets with different perspective information, providing a flexible solution to cope with different underwater environments and data conditions. It not only retains the accuracy of physical pixels, but also plays a role in significantly reducing multipath effects.
[0076] 2) In 3D reconstruction, using processed data sets can effectively reduce the error at the final convergence time.
[0077] 3) In terms of the results of three-dimensional reconstruction, without removing the multipath effect, the reconstructed results are interfered by the multipath effect noise, resulting in perforations in the original graphics and the reconstruction of objects that do not exist in the actual environment; however, after using the multipath effect processing algorithm of the present invention, the reconstruction results are greatly improved compared to the original results, especially when processing challenging multipath effect data. BRIEF DESCRIPTION OF THE DRAWINGS
[0078] Figure 1 is an algorithm flow chart of an embodiment of the present invention;
[0079] Figure 2 It is a schematic diagram of mapping points on a rectangular view onto a fan-shaped view plane;
[0080] Figure 3 It is a schematic diagram of the method steps of the present invention;
[0081] Figure 4 is a comparison diagram of the original and processed two-dimensional sonar images of data set A obtained by the algorithm of the embodiment of the present invention;
[0082] Figure 5 is a comparison diagram of the two-dimensional sonar image pre-processing and post-processing of the B data set obtained by the algorithm of the embodiment of the present invention;
[0083] Figure 6 is the average intensity loss-iteration number graph of the 3D reconstruction of the unprocessed data;
[0084] Figure 7 is a graph of average intensity loss-iteration number of three-dimensional reconstruction after data is processed by the algorithm of the embodiment of the present invention;
[0085] Figure 8 is a schematic diagram of 3D reconstruction of unprocessed data and processed data of data set A obtained by the algorithm of an embodiment of the present invention;
[0086] Fig. 9 It is a schematic diagram of 3D reconstruction of unprocessed data and processed data of the B data set obtained by the algorithm of the embodiment of the present invention. DETAILED DESCRIPTION
[0087] Depend on Figure 1 and Figure 3 As shown in the figure, the experiment was conducted on a simulated dataset. In dataset A, the feature is that only the area close to the 3D volume is sampled, and not enough sampling exploration is performed on the surrounding environment, which means that there may be missing angles in some environments. In dataset B, the feature is that in addition to sampling the reconstructed 3D volume at sufficient angles, we also intentionally sample the environment within a certain range at sufficient angles. In other words, the perspective is richer.
[0088] A method for removing multipath effect noise from underwater sonar images based on geometric probability comprises the following steps:
[0089] S1: Preprocess and denoise the sonar image;
[0090] S2: clustering the echo points of the pre-processed and denoised image;
[0091] S3: Remove multipath echo effect points from the clustered image.
[0092] Preferably, S1 specifically comprises the following steps:
[0093] S1.1: Preprocess sonar images using the open source SCUNet denoising model;
[0094] S1.2: The sonar image denoised in S1.1 is subjected to secondary denoising, i.e., a threshold operation is applied after normalizing the sonar echo values, and echoes below the predefined threshold are systematically set to zero;
[0095] S1.3: Reset the values of the area far away from the sonar equipment in the sonar image after the secondary denoising in 1.2 to zero.
[0096] Preferably, the SCUNet denoising model in S1.1 selects the grayscale images-25 denoising mode.
[0097] Preferably, S2 specifically includes the following steps:
[0098] S2.1: Map the sonar image dataset stored in the rectangular view onto a fan-shaped view plane and establish the coordinate grid of the fan-shaped view by the following formula
[0099] r′=H′δr+r min
[0100]
[0101] Among them, r is the radius of the fan view, Indicates the angle range of the fan view, r min is the minimum distance, θ represents the angular range of the sonar horizontal angle, δr represents the resolution of the sonar in the r direction, It represents the resolution of the sonar in the horizontal angular direction, and (H', W') is the horizontal and vertical range of the rectangular view.
[0102] Map the valid points in the rectangular view to the fan view and calculate their coordinates (x, y) in the fan view
[0103]
[0104] S2.2: Apply DBSCAN algorithm to cluster valid echo points;
[0105] S2.3: Apply the K-means++ algorithm to each main cluster {L1,L2,L3,...,L m , ...}.
[0106] Preferably, S2.2 is specifically:
[0107] S2.2.1: From the large class C in DBSCAN i Randomly select a center among the data points;
[0108] S2.2.2: For each data point P i , calculate D(P i ), namely P i The distance to the closest center that has already been chosen;
[0109] S2.2.3: Randomly select a new data point as the new center using a weighted probability distribution, where the probability of selecting point P is the same as D(P) 2 is proportional to;
[0110] S2.2.4: Repeat steps S2.2.2 and S2.2.3 until K centers are selected;
[0111] S2.2.5: After determining K center points according to the above steps, select these K points as the cluster centers for initial aggregation;
[0112] S2.2.6: Calculate the Euclidean distance from each sample point to the K cluster cores, find the cluster core closest to the point, and assign it to the corresponding cluster;
[0113] S2.2.7: After all points are assigned to clusters, the M points are divided into K clusters; then the centroid or average distance center of each cluster is recalculated and set as the new cluster core;
[0114] S2.2.8: Repeat steps S2.2.6-S2.2.7 until a termination condition is reached. The termination condition for this experiment is 200 iterations of the S2.2.6-S2.2.7 process.
[0115] After the clustering operation is completed, all valuable points are clustered into small clusters {c1, c2, c3, ..., c i}.
[0116] Preferably, the valid echo point in S2.2 means that the multipath effect echo must meet the following three conditions:
[0117] a) The distance to the echo generated by the second reflection must be smaller than the distance between multipath effect points;
[0118] b) The echo from the multipath effect and the echo from the second reflection are in the same direction as the sonar;
[0119] c) The distance from the multipath effect point to the sonar origin is within the range of the total distance traveled during the second reflection.
[0120] Preferably, a minimum distance constraint is introduced for the valid echo points in S2.2, that is, a minimum distance judgment is added between the two classes found, and only those exceeding the minimum distance threshold are considered to be multipath effect pairs. This constraint is used to reduce the misjudgment of multipath effect points.
[0121] Preferably, S2.3 is specifically:
[0122] The initial number of clusters for the K-means++ clustering algorithm is pre-specified, denoted as k; the number of clusters specified is proportional to the number of echoes in the main cluster;
[0123] The algorithm is as follows,
[0124] 1) Initialization: The main data points L obtained from DBSCAN m The first cluster E1 is randomly selected;
[0125] 2) For i=2 to k:
[0126] a. Calculate each data point x i The square distance D(x i ) 2 , where D(x i ) is from x i The distance to the nearest cluster center;
[0127] b. Select a new cluster center E according to the algorithm described in S2.2.5 i ;
[0128] 3) Run the standard K-means algorithm using the selected cluster centers for initialization;
[0129] 4) In each iteration, perform the following steps:
[0130] a. For each sample x i , calculate its distance to each cluster center and assign it to the cluster with the closest center
[0131]
[0132] Among them, E i is the sample x i The cluster to which it belongs, μ j is the center of cluster j;
[0133] b. For each cluster j, recalculate its center μ j As the average value of all samples in the cluster
[0134]
[0135] Among them, |E j | represents the number of samples in cluster j;
[0136] c. Repeat steps a and b until the cluster assignment result is stable or reaches the specified termination criterion or number of iterations;
[0137] After the clustering process is completed, all valid points are grouped into separate small clusters {C1, C2, C3, ..., C m ,…}.
[0138] Preferably, S3 specifically includes the following steps:
[0139] S3.1: Randomly select a completed clustering operation cluster, denoted as C i , whose cluster center coordinates are represented by X i ; Then, the corresponding line segment OX is obtained by connecting it with the sonar image origin O i ;
[0140] S3.2: Traverse all other clusters and calculate the number of nodes from the cluster center X j To line segment OX i The distance is calculated and the distance threshold a is set; if the distance is within the threshold a, the intersecting clusters X are marked j , and classify it into category D i ;
[0141] S3.3: First, determine point P k Is it related to line segment OX? i intersect;
[0142] Then, in the intersecting cluster D i All points P in each cluster k Traverse on; any point P that meets the intersection condition k are grouped into set S; if none of the potentially intersecting clusters meets the intersection criteria, meaning S is an empty set, the algorithm proceeds to the next cluster C i for subsequent iterations;
[0143] In addition, if the distance between the centers of two clusters is less than the specified threshold c, they are considered to be in contact, thus preventing misjudgment as a multipath effect;
[0144] S3.4: Using the exhaustive enumeration method, iterate the remaining clusters and calculate the distance OX from their cluster centers to the sonar image origin O m ; Calculate OX i Cluster X that intersects with the labels in S2.3 j The centroid coordinates of all intersection points in the j , use E to represent X m and Y j The potential distance between m and Y j In the actual environment, the position will be anywhere within the elevation range of the sonar scan, so the range of E is determined by X m and Y j The total possible distance C covered by the sound wave through secondary reflection is calculated as OX m +OY j +E, get the calculated range [C min ,C max ];
[0145] S3.5: Let G represent OX i twice the length of G if it falls on [C min ,C max ] range, it is considered as a potential point affected by the multipath effect; if it is outside the range, the offset value distance is calculated from the nearest end, and the value is substituted into the Gaussian function to calculate the probability p that it is a multipath affected point. If the probability p is greater than the set threshold, it is considered as a potential multipath affected point;
[0146] Finally, in the rectangular view, all echo values corresponding to points in the cluster identified as being affected by multipath effects are set to zero.
[0147] Preferably, the determination point P in S3.3 k Is it related to line segment OX? i Intersection is determined by the following two criteria:
[0148] 1) Judge by the intersection coefficient e
[0149]
[0150] Among them, the intersection coefficient e indicates point P k Relative to OX i The relative position of the projection on
[0151] Since point P k In the line segment OX i The projection on must be between O and X i So the range of e is from 0 to 1;
[0152] 2) Set the angle threshold b and calculate the vector OP k and OX i The angle between them is calculated and the intersection is confirmed when the angle is below a specified threshold b for collinearity.
[0153] Combination Figure 4 and Figure 5 ,The results of processing and analyzing the sonar image data show that this algorithm has an excellent ability to remove multipath effects. In the A and B data sets, the method based on geometric probability to remove multipath effects only has a large number of false deletions when there are multiple true echoes at a horizontal angle at certain angles. In most cases, even if there are multiple true echoes at the same horizontal angle, the number of false deletions is small, which not only retains the accuracy of the physical pixels, but also plays a role in significantly reducing the multipath effect.
[0154] Combination Figure 6 and Figure 7 , which shows that compared with the unprocessed data set, using the processed data set can effectively reduce the error at the time of final convergence.
[0155] Combination Figure 8 and Fig. 9 , which shows that the dataset processed by the algorithm based on geometric probability to remove multipath effects is different from the dataset without considering removing multipath effects. Without removing multipath effects, the reconstruction results are disturbed by multipath effect noise, resulting in holes in the original graphics and the reconstruction of objects that do not exist in the actual environment. The reconstruction results obtained by our multipath effect processing algorithm are greatly improved compared with the original results.
[0156] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A method for removing multipath noise from underwater sonar images based on geometric probability, characterized in that: The following steps are involved: S1: Preprocess and denoise the sonar image; S2: clustering the echo points of the pre-processed and denoised image; S3: removing multipath echo effect points from the clustered image; S2 specifically includes the following steps: S2.1: Map the sonar image dataset stored in the rectangular view onto a fan-shaped view plane and establish the coordinate grid of the fan-shaped view by the following formula ; ; in, r is the radius of the fan view, φ Indicates the angle range of the fan view. is the minimum distance, θ Indicates the angular range of the sonar horizontal angle, represents the resolution of the sonar in the r direction, Indicates the resolution of the sonar in the horizontal angle direction. is the horizontal extent of the rectangular view, is the vertical extent of the rectangular view; Map the valid points in the rectangular view to the fan view and calculate their coordinates in the fan view ( x , y ) ; ; S2.2: Apply DBSCAN algorithm to cluster valid echo points; S2.3: Apply the K-means++ algorithm to each major cluster obtained by the DBSCAN algorithm { L 1 , L 2 , L 3 ,..., L m }; S3 specifically includes the following steps: S3.1: Randomly select a completed cluster, represented as C i , and its cluster center coordinates are expressed as X i ; Then, by comparing it with the sonar image origin O Connect to get the corresponding line segment OX i ; S3.2: Traverse all other clusters and calculate the number of nodes from the cluster center X j To Segment OX i distance and set the distance threshold a ; If the distance is within the threshold a If the clusters are within the same group, then the intersecting clusters X are marked. j , and classify them into clusters D i ; S3.3: First, determine the point P k Is it related to the line segment OX i intersect; Then, in the intersecting clusters D i All points in each cluster in P k Traverse on; any point that meets the intersection condition P k are grouped into set S; if none of the potentially intersecting clusters meets the intersection criteria, meaning S is an empty set, the algorithm proceeds to the next cluster C i for subsequent iterations; In addition, if the distance between the centers of two clusters is less than the specified threshold c, they are considered to be in contact, thus preventing misjudgment as a multipath effect; S3.4: Using the exhaustive enumeration method, iterate the remaining clusters and calculate the distance from their cluster centers to the sonar image origin. O Distance OX m ;calculate OX i Cluster X that intersects with the labels in S3.2 j The centroid coordinates of all intersection points in the Y j , represented by E X m and Y j The potential distance between m and Y j In actual environment, it will be anywhere within the elevation range of sonar scanning, so E The range of X m and Y j The total possible distance covered by the sound wave through secondary reflections is determined by the position and maximum elevation angle of the C Calculated as OX m + OY j + E , and the calculated range is [ C min , C max ]; S3.5: Set G express OX i Twice the length of G Falling in C min , C max ], it is considered as a potential point affected by the multipath effect; if it is outside this range, the offset value distance is calculated from the nearest end and the value is substituted into the Gaussian function to calculate the probability that it is a multipath affected point p , if the probability p If it is greater than the set threshold, it is considered as a potential multipath impact point; Finally, in the rectangular view, all echo values corresponding to points in the cluster identified as being affected by multipath effects are set to zero.
2. The method according to claim 1, characterized in that S1 specifically includes the following steps: S1.1: Preprocess sonar images using the open source SCUNet denoising model; S1.2: The sonar image denoised in S1.1 is subjected to secondary denoising, i.e., a threshold operation is applied after normalizing the sonar echo values, and echoes below the predefined threshold are systematically set to zero; S1.3: reset the values of the area far away from the sonar equipment in the sonar image after secondary denoising in S1.2 to zero.
3. The method according to claim 2, characterized in that The SCUNet denoising model in S1.1 selects the grayscale images-25 denoising mode.
4. The method according to claim 1, characterized in that: S2.2 specifically: S2.2.1: From the large categories in DBSCAN Randomly select a center among the data points; S2.2.2: For each data point ,calculate ),Right now The distance to the closest center that has already been chosen; S2.2.3: Use the weighted probability distribution to randomly select a new data point as the new center, where the point is selected P The probability and is proportional to; S2.2.4: Repeat steps S2.2.2 and S2.2.3 until you select K center points; S2.2.5: Follow the above steps to confirm K After the center point is selected, K The center points are used as the cluster cores for initial aggregation; S2.2.6: Calculate each sample point to K The Euclidean distance of the cluster cores is used to find the cluster core closest to the point and assign it to the corresponding cluster. S2.2.7: After all points are assigned to clusters, M The points are divided into K clusters; then recalculate the centroid or average distance center of each cluster and set it as the new cluster core; S2.2.8: Repeat steps S2.2.6-S2.2.7 until a termination condition is reached. The termination condition for this experiment is 200 iterations of the S2.2.6-S2.2.7 process. After the clustering operation is completed, all valuable points are clustered into small clusters one by one. .
5. The method according to claim 1, characterized in that: The effective echo point in S2.2 means that the multipath effect echo must meet the following three conditions: 1) The distance to the echo generated by the second reflection must be smaller than the distance between the multipath effect points; 2) The multipath effect echo and the echo produced by the second reflection are located in the same direction as the sonar; 3) The distance from the multipath effect point to the sonar origin is within the total distance traveled during the second reflection.
6. The method according to claim 1, characterized in that A minimum distance constraint is introduced for the valid echo points in S2.2, that is, a minimum distance judgment is added between the two classes found. Only those exceeding the minimum distance threshold are considered to be multipath effect pairs. This constraint is used to reduce the misjudgment of multipath effect points.
7. The method according to claim 1, characterized in that S2.3 specifically: Pre-specify the initial number of clusters for the K-means++ clustering algorithm, expressed as k ;The number of clusters specified is proportional to the number of echoes in the main cluster; The algorithm is as follows, 1) Initialization: Main clusters obtained from DBSCAN L m Randomly select the first cluster in E 1 ; 2) For i =2 to k : t1. Calculate each data point x i The square distance { D ( x i )} 2 ,in D ( x i ) is from x i The distance to the nearest cluster center; t2. Select new cluster centers according to the algorithm described in S2.2.5 E i ; 3) Run the standard K-means algorithm using the selected cluster centers for initialization; 4) In each iteration, perform the following steps: z1. For each data point x i , calculate its distance to each cluster center and assign it to the cluster with the closest center ; in, E i It is a data point x i The cluster to which it belongs, μ j It is clustering j the center of; z2. For each cluster j , recalculate its center μ j As the average value of all samples in the cluster ; in, |E j | Representation clustering j The number of samples in ; z3. Repeat steps z1 and z2 until the clustering assignment result is stable or reaches the specified termination criterion or number of iterations; After completing the clustering process, all valid points are grouped into separate small clusters { C 1 , C 2 , C 3 ,..., C n }.
8. The method according to claim 1, characterized in that The determination point in S3.3 P k Is it related to the line segment OX i Intersection is determined by the following two criteria: 1) Through the intersection coefficient e To judge ; Among them, the intersection coefficient e Indication Point P k Relative to OX i The relative position of the projection on P k In line segment OX i The projection on must be located at O and X i Between, so e The range is from 0 to 1; 2) Set the angle threshold b and calculate the vector OP k and OX i The angle between them is calculated and the intersection is confirmed when the angle is below a specified threshold b for collinearity.
