A segmented planar tank reconstruction method based on iterative graph cut guided optimization

Abstract

This invention discloses a piecewise planar tank reconstruction method based on iterative graph cut-guided optimization. By embedding high-level geometric priors into the energy function of the graph cut, the selection of facets gains semantic understanding capabilities, and the graph structure and candidate face set are dynamically updated. This allows the graph cut results to drive the update of geometric parameters, and the new parameters, in turn, feed back into the graph model reconstruction, forming a closed-loop refinement. The realized "graph structure-geometric model" joint iterative mechanism can eliminate invalid assumptions round by round and continuously focus on effective facets, upgrading from static local optimization to dynamic global guidance, significantly improving the applicability of graph cut in complex 3D reconstruction. Finally, lightweight topological constraints are applied to the highly refined candidate subset to ensure the watertightness and manifold of the output mesh. This invention significantly reduces the dependence on subsequent topology guarantee modules and the computational burden, making it particularly suitable for robust reconstruction of regular objects in industrial scenarios, such as tanks, pipes, and boxes.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer graphics, 3D vision and industrial tank inspection, and specifically relates to a segmented planar tank reconstruction method based on iterative graph cut guided optimization. Background Technology

[0002] Existing 3D surface reconstruction techniques mainly include Poisson reconstruction, advancing front, and scale-space reconstruction. These methods perform well when dealing with smooth surfaces, but they have significant limitations when handling man-made structures in industrial settings (industrial storage tanks).

[0003] 1. The model is too complex: it generates a mesh containing a large number of redundant triangles, making it difficult to use for subsequent analysis or rendering;

[0004] 2. Unable to retain sharp features: Smooths key geometric features such as right angles and edges;

[0005] 3. Sensitive to noise and missing data: Prone to holes, self-intersections, or non-manifold structures;

[0006] Therefore, there is an urgent need for a reconstruction framework that combines robustness, efficiency, and topological correctness, capable of automatically eliminating invalid assumptions in complex noisy environments and focusing on high-quality candidate surfaces that conform to the overall geometric laws of the object. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies and achieve the goal of eliminating invalid assumptions in tank surface reconstruction round by round and continuously focusing on effective areas, this invention adopts the following technical solution:

[0008] A segmented planar tank reconstruction method based on iterative graph cut-guided optimization includes the following process:

[0009] Obtain the original point cloud data of the tank surface, and calculate the normal of the point based on the local surface where the point is located;

[0010] To determine a planar model, obtain at least three points whose normals are not collinear. Compare the consistency between the interior points of the planar model and the planar model itself, and update the planar model to obtain the optimal planar model with the most consistent interior points, i.e., the largest planar structure. Construct initial candidate facets based on the interior points of the optimal planar model. Continue to construct the next initial candidate facet from the remaining points, i.e., remove the original point cloud data belonging to the current optimal planar model, continue to search for the largest planar structure in the remaining point cloud data, and then construct the next initial candidate facet using the interior points of the largest planar structure, finally obtaining the initial candidate facet set.

[0011] Using candidate facets as nodes and the adjacency relationships between nodes as edges, a geometric perception graph model is constructed iteratively. The minimum graph cut is then solved based on the conformity between the geometric perception graph model and the global geometric model of the tank, so as to retain candidate facets that conform to the global geometric structure of the tank.

[0012] For the candidate face set optimized by graph cut, the mixed integer linear programming (MILP) method is used to solve the global topology to ensure that the final result conforms to the geometric constraints and topological relationships of the object, so as to extract the final set of retained faces for tank surface reconstruction.

[0013] Furthermore, the calculation of the normal is achieved by searching a set of neighboring points of a point, constructing the covariance matrix of each point in the neighborhood based on the centroid of the neighboring point set, performing eigenvalue decomposition on the covariance matrix to obtain a set of eigenvalues ​​and their corresponding eigenvectors, and using the eigenvector corresponding to the smallest eigenvalue as the normal vector of the point. The normal direction is either the line viewpoint direction or the direction with an angle less than the angle threshold (90 degrees) with the viewpoint direction.

[0014] Furthermore, the process of generating the initial candidate facets is as follows:

[0015] Three points with non-collinear normals are randomly selected to determine a unique plane model, and the parameters of the plane model are calculated based on the three points.

[0016] Perform consistency verification on the interior points of the planar model; calculate the distance from the interior point to the planar model and the angle between the interior point normal and the planar model normal, and count the consistent interior points and the number of consistent interior points that simultaneously satisfy the distance and the angle less than the corresponding threshold.

[0017] The number of consistent interior points is compared with the number of consistent interior points of the historical best plane model. If the number of consistent interior points is greater than that of the historical best plane model, the current plane model is taken as the historical best plane model, and the number of consistent interior points and the number of consistent interior points are updated.

[0018] The selection of execution points and the updating of the planar model are performed repeatedly until the number of consistent internal points is maximized and / or the number of iterations is reached, thus obtaining the optimal planar model (the largest planar structure).

[0019] Furthermore, based on the finally determined optimal planar model, global interior point detection is performed to obtain all point cloud data belonging to the optimal planar model, so as to form an initial candidate patch;

[0020] After removing (or marking as processed) the points belonging to the optimal planar model from the original point cloud dataset to prevent them from participating in the subsequent planar fitting process, the loop iterates back to continue searching for the next optimal planar model (the largest planar structure) in the remaining point cloud data to construct the next initial candidate patch, until the remaining point cloud is insufficient to constitute a valid planar model and / or no new planar model that meets the threshold can be found.

[0021] Furthermore, the geometric perception graph model constructs a data item energy function based on the point support of the initial candidate facets and their conformity with the global geometric model of the tank.

[0022] By solving for the minimum value of the energy function, we obtain the initial candidate faceplate rejection / rejection labels for the current round. Based on these labels, we update the initial candidate faceplate set, remove the initial candidate faceplates with the rejection label, and obtain the retained candidate faceplate set.

[0023] Furthermore, the energy function also includes a smoothing term energy function constructed based on the consistency of the rejection or retention labels of adjacent initial candidate patches and the deviation between the included angle of adjacent initial candidate patches and the expected included angle.

[0024] Furthermore, based on the currently retained set of candidate facets, the global geometry model of the tank is refitted to calculate the deviation between the remaining candidate facets and the global geometry model of the tank, so as to remove candidate facets whose deviation exceeds a threshold.

[0025] Reconstruct the ensemble perception graph model until the change in the number of candidate patches is less than a threshold or the change in the energy function is less than a threshold.

[0026] Furthermore, the global topology solution involves extracting the geometric features of the candidate patch set and calculating the energy of the data items for each patch; setting geometric and topological constraints according to the tank type; and using the branch and bound method to perform a global optimal solution to obtain a patch set that satisfies both the patch geometric constraints and the overall topology.

[0027] Furthermore, the global topology solution also includes calculating the energy of the smoothing term by constructing an adjacency graph.

[0028] Furthermore, the constraints include geometric constraints of planar consistency constraints and distance constraints, as well as topological constraints of connectivity constraints and boundary constraints;

[0029] The plane consistency constraint means that the difference between the normal vectors of adjacent facets is less than the maximum allowable angle.

[0030] The distance constraint means that the distance between the retained patches is less than the maximum allowed distance;

[0031] The connectivity constraint refers to the retention of faces to form a connected overall structure;

[0032] The boundary constraints refer to the overall size and shape constraints of the tank.

[0033] The advantages and beneficial effects of this invention are as follows:

[0034] This invention can reconstruct compact, watertight, and geometrically clear segmented planar meshes from noisy point clouds in industrial sites. Specifically designed for various industrial storage tanks, it employs a graph cut optimization algorithm as the core selection tool, moving beyond simple smoothing to a decision engine incorporating geometric knowledge. The invention significantly enhances noise resistance and robustness, handling industrial point clouds with up to 30% outliers. It improves reconstruction efficiency; the graph cut algorithm itself is polynomial-time, with convergence typically achieved in 3-5 iterations. This invention supports multiple geometric priors, adapting to tanks, buildings, and mechanical parts by changing the M-model. The final result maintains watertightness, meeting industrial standards through lightweight post-processing. This invention is particularly suitable for high-precision 3D modeling of rotationally symmetric tanks, irregularly shaped storage tanks with combined planar / curved / cylindrical surface structures, vertical oil storage tanks, and horizontal pressure vessels. It can be widely applied in digital twins throughout the tank lifecycle, BIM modeling of sites, detection of tank corrosion and deformation defects, volume calibration, in-service monitoring, and reverse engineering. Attached Figure Description

[0035] Figure 1 This is a flowchart of a segmented planar tank reconstruction method based on iterative graph cut guided optimization in an embodiment of the present invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0037] like Figure 1 As shown, a segmented planar tank reconstruction method based on iterative graph cut-guided optimization includes the following steps:

[0038] Step S1: Input Data Preparation; Input data preparation is the foundation of the entire algorithm process. Its core task is to preprocess the raw point cloud data and calculate the key geometric attribute of each point—the normal—to provide data support for subsequent plane fitting and geometric structure analysis. This step mainly includes the following two key aspects:

[0039] Step S1.1: Data Preprocessing; First, load the original 3D point cloud dataset. To improve the efficiency and accuracy of subsequent calculations, the point cloud usually needs to be preprocessed, including removing outliers (such as isolated points caused by sensor noise) and performing voxel grid downsampling. Downsampling can reduce the amount of data and speed up the calculation while preserving the main geometric features of the object.

[0040] Step S1.2: Normal Calculation; The normal is one of the most important attributes describing the local geometry of a point cloud, and it is perpendicular to the local surface where the point is located. Accurate normal information is crucial for determining surface orientation and performing plane segmentation. This step uses a local neighborhood fitting method based on Principal Component Analysis (PCA) to calculate the normal for each point; the specific algorithm flow is as follows:

[0041] Step S1.2.1: Neighborhood search; for each point to be processed Search for the set of neighboring points in the point cloud. There are two common search strategies:

[0042] (1) k-Nearest Neighbor (k-NN): Select the nearest neighbor to the point One point.

[0043] (2) Radius Search: Select the radius search function. Centered on, with radius All points within the spherical neighborhood.

[0044] Step S1.2.2: Construct the covariance matrix; let the neighborhood point set be... The center of mass is Then the covariance matrix It can be represented as:

[0045]

[0046] in, It is the number of points in the neighborhood.

[0047] Step S1.2.3: Eigenvalue decomposition; for the covariance matrix Eigenvalue decomposition yields three eigenvalues. (Assuming) ) and their corresponding eigenvectors .

[0048] Step S1.2.4: Determine the normal direction; minimum eigenvalue Corresponding feature vector This represents the direction in which the point cloud changes least within that local region, i.e., the direction perpendicular to the local surface; therefore, it is used as a point. normal vector .

[0049] Step S1.2.5: Normal Orientation; Since the normal calculated by PCA has bidirectionality (i.e. and (These are all mathematical solutions). To ensure the consistency of the normal direction throughout the point cloud (e.g., all pointing outwards from the object or all pointing towards the sensor), normal orientation processing is required. Typically, based on the viewpoint position, the normal direction is adjusted to point towards the viewpoint or at an angle less than 90 degrees to it.

[0050] After the above processing, the original point cloud data is transformed into a point cloud with normal attributes, providing the necessary geometric information for subsequent steps.

[0051] Step S2: Generate candidate faces; To generate a comprehensive and non-redundant set of candidate face patches, the Random Sample Consensus (RANSAC) algorithm is typically used, which includes the following steps:

[0052] Step S2.1: Initialize parameters; Before starting the iteration, the following key parameters need to be set:

[0053] Distance threshold ( ): Used to determine whether a point belongs to a plane, usually set to 2cm;

[0054] Normal angle threshold ( ): Used to determine whether the normal direction of a point is consistent with the plane normal, usually set to 15°;

[0055] Maximum number of iterations ( ): Controls the convergence speed and accuracy of the algorithm;

[0056] Confidence parameter: used to dynamically adjust the number of iterations.

[0057] Step S2.2: Iterative Loop; Each iteration includes the following core steps:

[0058] Step S2.2.1: Random sampling; From the current unsegmented point cloud set, randomly select three non-collinear points. These three points uniquely determine a planar model M;

[0059] Step S2.2.2: Model parameter calculation; Calculate the mathematical parameters of the candidate plane based on three sampling points:

[0060] Calculate the unit normal vector ;

[0061] Calculate the constant term of the plane equation ;

[0062] The complete plane equation is expressed as: ;

[0063] Step S2.2.3: Consistency check (internal point detection); Verify for all remaining points:

[0064] Distance calculation: Calculate the distance for each point Distance to plane M ;

[0065] Normal consistency: Calculate the normal of the point. The angle between the plane and the normal n ;

[0066] Interior point determination condition: simultaneously satisfying and Count the set of all interior points that satisfy the conditions. and their quantity .

[0067] Step S2.3: Model evaluation and update; after each iteration:

[0068] Compare the number of interior points in the current model Compared with the historical best model;

[0069] If the current model has more interior points, then update the optimal model to the current model;

[0070] At the same time, update the best set of interior points to the current set of interior points.

[0071] Step S2.4: Iteration Termination Condition; The iteration terminates when any of the following conditions are met:

[0072] The preset maximum number of iterations has been reached;

[0073] Find models where the number of interior points exceeds a certain proportion;

[0074] Planar segmentation and reconstruction.

[0075] Step S2.5: Extract facets; Based on the finally determined optimal planar model, perform a global interior point detection to obtain all point cloud data belonging to the plane. These points constitute an initial candidate facet.

[0076] Step S2.6: Point cloud update; Remove (or mark as processed) points belonging to the plane from the original point cloud dataset to prevent them from participating in the subsequent plane fitting process.

[0077] Step S2.7: Repeat the detection; return to the iteration step and continue searching for the next largest planar structure in the remaining point cloud until one of the following termination conditions is met:

[0078] The remaining point cloud is too small to form an effective plane.

[0079] Unable to find a new plane that meets the preset threshold

[0080] The RANSAC process described above can efficiently separate multiple independent planar candidate patches from messy point cloud data. These patches constitute the initial candidate set. This serves as the input data for the subsequent S3 graph cut optimization algorithm.

[0081] Step S3: Iterative Graph Cut; This step is the core optimization process. By iteratively constructing a geometry-aware graph model and solving for the minimum cut, erroneous candidate faces are gradually eliminated, focusing on faces that conform to the object's geometry. Specifically, it includes the following steps:

[0082] Step S3.1: Construct a geometric perception graph model;

[0083] Node definition: The current set of candidate faces Each facet in Defined as a graph A node .

[0084] Edge definition: Based on the adjacency table, add edges between nodes sharing a boundary. .

[0085] Constructing the energy function: Defining the energy function of the graph cut. ,in Represents a node The label (1 means keep, 0 means remove).

[0086] Unary Term Definition: Measured by a single facet The quality consists of two parts:

[0087] (1) Point support: Calculate how many point cloud data points support the patch, that is, the distance from the point to the plane where the patch is located is less than the threshold. The number of points is denoted as ;

[0088] (2) Geometric prior conformity: Calculate the surface patch The degree of conformity with the currently estimated global geometric model of the tank (such as the cylindrical axis and the dominant direction of the box); for example, if the object is a cylinder, calculate the deviation between the normal vector of the surface patch and the normal vector field of the cylindrical side surface. The energy of the data term is defined as:

[0089]

[0090] in, Indicates the total number of points. Indicates the current geometric model, This represents two weighted terms. Indicates measurement mode With a certain reference model The deviation between them.

[0091] Smoothness Term Definition: Used to constrain the label consistency of adjacent faces, preventing isolated faces from being incorrectly removed or retained. Defines adjacent faces... and The energy of the smoothing term is:

[0092]

[0093] in, The weight of the smoothing term controls the constraint strength for the consistency of labels between adjacent facets; For the label difference function, when hour, ,otherwise ; Adjacent face and The included angle; The desired angle between adjacent facets (as in a planar structure) (indicating coplanarity). The standard deviation of the Gaussian kernel controls the tolerance for angular deviation.

[0094] Step S3.2: Perform graph cut optimization;

[0095] Finding the minimum cut: Solving the energy function using the maximum flow / minimum cut algorithm. The minimum value is used to obtain the binary label assignment for the current round. .

[0096] Update candidate sets: based on tags Update candidate face sheets Set, remove all faces with a label of 0.

[0097] Step S3.3: Dynamically update the geometric prior and candidate set;

[0098] (1) Geometric model re-estimation: based on the currently retained set of high-confidence patches Refit the global geometric model of the object For example, the principal axis direction of an object can be updated by analyzing the distribution of normal vectors of facets, or the radius of a cylinder can be updated by fitting a contour.

[0099] (2) Candidate set reconstruction: Calculate the remaining candidate patches and the new model The deviation is considered, and patches with deviations exceeding the threshold are removed.

[0100] (3) Iterative judgment: Check the set of candidate patches Whether convergence has occurred (the change in the number of patches retained in two consecutive rounds is less than a threshold, or the change in the energy function is less than a threshold). If convergence has not occurred, return to step S3.1 to rebuild the graph model; if convergence has occurred, proceed to step S4.

[0101] Step S4: Topology Solving; This step uses the Mixed Integer Linear Programming (MILP) method to solve the global topology of the candidate face set after graph cut optimization, ensuring that the final result conforms to the geometric constraints and topological relationships of the object; specifically, it includes the following steps:

[0102] Step S4.1: MILP problem modeling.

[0103] (1) Definition of decision variables:

[0104] binary variables : Indicates candidate facet Whether it is ultimately retained (1 indicates retention, 0 indicates rejection);

[0105] Continuous variables : indicates a piece of dough and dough sheets The strength of the connection between them.

[0106] (2) Objective function:

[0107] Minimize the total energy function:

[0108]

[0109] in, For dough Data item weights, For dough and The weights of the smoothing terms between them.

[0110] Step S4.2: Constraint construction.

[0111] Geometric constraints:

[0112] (1) Plane consistency constraint:

[0113]

[0114] The difference in normal vectors between adjacent preserved facets should be less than the maximum permissible angle.

[0115] (2) Distance constraints:

[0116]

[0117] The distance between the reserved facets should be less than the maximum allowable distance.

[0118] Topological constraints:

[0119] (1) Connectivity constraint: Ensures that the retained patches form a connected overall structure:

[0120]

[0121] in, Represents a piece of dough The set of adjacent faces,

[0122] (2) Boundary constraints: control the overall size and shape of the object.

[0123]

[0124] in, For dough The area.

[0125] Step S4.3: The MILP solution process is as follows:

[0126] Step S4.3.1: Problem instantiation.

[0127] (1) Parameter extraction:

[0128] From the candidate face set Extract geometric features from them;

[0129] Calculate the energy of the data items for each facet ;

[0130] Construct an adjacency graph and calculate the energy of the smoothing term. .

[0131] (2) Constraint Construction:

[0132] Set appropriate geometric constraints based on the object type (such as box, cylinder, etc.);

[0133] Define variable boundaries and integer constraints.

[0134] Step S4.3.2: Solve and execute.

[0135] (1) Initialize the solver:

[0136] Choose a suitable MILP solver;

[0137] Set the solution accuracy and time limits;

[0138] (2) Solution process:

[0139] Use the branch and bound method to find the global optimum;

[0140] Monitor the solution progress and the lower bound of the objective function.

[0141] Step S4.3: Post-processing of results.

[0142] (1) Extraction:

[0143] Obtain the optimal binary variable solution ;

[0144] Extract the final set of retained face patches ;

[0145] (2) Topology verification:

[0146] Check whether the results meet all constraints.

[0147] Verify the rationality of the overall topological structure of the object

[0148] Step S4.4: Algorithm output.

[0149] Step S5: Iterate through step S4 until the candidate face set converges;

[0150] The final set of facets obtained by MILP solution It possesses the characteristics of globally optimal topology, facet combinations that conform to object geometric constraints, and stable boundaries and connectivity.

[0151] This result, as the final output of the entire algorithm process, can be used for subsequent applications such as 3D reconstruction and object recognition.

[0152] Step S6: Topology-guaranteed post-processing; Apply manifold hard constraints v to the final refined candidate surfaces and output a watertight mesh.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization, characterized in that: Obtain the original point cloud data of the tank surface, and calculate the normal of the point based on the local surface where the point is located; Obtain at least three points whose normals are not collinear to determine a plane model. Compare the consistency between the interior points of the plane model and the plane model. Update the plane model to obtain the optimal plane model with the most consistent interior points. Construct initial candidate patches based on the interior points of the optimal plane model. Continue constructing the next initial candidate facet from the remaining points, eventually obtaining the initial candidate facet set; Using candidate facets as nodes and the adjacency relationships between nodes as edges, a geometric perception graph model is constructed iteratively. The minimum graph cut is then solved based on the conformity between the geometric perception graph model and the global geometric model of the tank, so as to retain candidate facets that conform to the global geometric structure of the tank. The global topology of the candidate face set after graph cut optimization is solved to extract the final set of face sets for tank surface reconstruction.

2. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 1, characterized in that: The calculation of the normal is achieved by searching a set of neighboring points of a point, constructing the covariance matrix of each point in the neighborhood based on the centroid of the neighboring point set, performing eigenvalue decomposition on the covariance matrix to obtain a set of eigenvalues ​​and their corresponding eigenvectors, and using the eigenvector corresponding to the smallest eigenvalue as the normal vector of the point. The normal direction is either the line viewpoint direction or the direction with an angle less than a threshold with the viewpoint direction.

3. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 1, characterized in that: The process of generating the initial candidate facets is as follows: Three points with non-collinear normals are randomly selected to determine a unique plane model, and the parameters of the plane model are calculated based on the three points. Perform consistency verification on the interior points of the planar model; Calculate the distance from the interior point to the plane model and the angle between the interior point normal and the plane model normal, and count the number of consistent interior points that simultaneously satisfy the distance and the angle less than the corresponding threshold. The number of consistent interior points is compared with the number of consistent interior points of the historical best plane model. If the number of consistent interior points is greater than that of the historical best plane model, the current plane model is taken as the historical best plane model, and the number of consistent interior points and the number of consistent interior points are updated. The selection of execution points and the updating of the planar model are performed repeatedly until the number of consistent internal points is maximized and / or the number of iterations is reached, thus obtaining the optimal planar model.

4. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 3, characterized in that: Based on the finally determined optimal planar model, global interior point detection is performed to obtain all point cloud data belonging to the optimal planar model, so as to form an initial candidate patch. After extracting the points belonging to the optimal planar model from the original point cloud dataset, return to the loop iteration and continue to search for the next optimal planar model in the remaining point cloud data to construct the next initial candidate patch, until the remaining point cloud is insufficient to constitute an effective planar model and / or a new planar model that meets the threshold cannot be found.

5. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 1, characterized in that: The geometric perception graph model constructs a data item energy function based on the point support of the initial candidate facets and their conformity with the global geometric model of the tank. By solving for the minimum value of the energy function, we obtain the initial candidate faceplate rejection / rejection labels for the current round. Based on these labels, we update the initial candidate faceplate set, remove the initial candidate faceplates with the rejection label, and obtain the retained candidate faceplate set.

6. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 5, characterized in that: The energy function also includes a smoothing term energy function constructed based on the consistency of the rejection / retention labels of adjacent initial candidate patches and the deviation between the included angle of adjacent initial candidate patches and the expected included angle.

7. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 5, characterized in that: Based on the currently retained set of candidate faces, the global geometry model of the tank is refitted to calculate the deviation between the remaining candidate faces and the global geometry model of the tank, so as to remove candidate faces whose deviation exceeds the threshold. Reconstruct the ensemble perception graph model until the change in the number of candidate patches is less than a threshold or the change in the energy function is less than a threshold.

8. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 1, characterized in that: The global topology solution is achieved by extracting the geometric features of the candidate patch set and calculating the energy of the data items for each patch; setting geometric and topological constraints according to the tank type; and using the branch and bound method to perform a global optimal solution to obtain a patch set that satisfies the patch geometric constraints and has a reasonable overall topology.

9. The method for reconstructing a segmented planar tank based on iterative graph cut-guided optimization according to claim 8, characterized in that: The global topology solution also includes calculating the energy of the smoothing term by constructing an adjacency graph.

10. A segmented planar tank reconstruction method based on iterative graph cut guided optimization according to claim 8, characterized in that: The constraints include geometric constraints such as planar consistency constraints and distance constraints, as well as topological constraints such as connectivity constraints and boundary constraints. The plane consistency constraint means that the difference between the normal vectors of adjacent facets is less than the maximum allowable angle. The distance constraint means that the distance between the retained patches is less than the maximum allowed distance; The connectivity constraint refers to the retention of faces to form a connected overall structure; The boundary constraints refer to the overall size and shape constraints of the tank.

Images