Zero-sample space target micro component segmentation method based on spatial constraint guidance

Abstract

The invention discloses a zero-sample space target micro component segmentation method based on spatial constraint guidance, and solves the problems that in the prior art, spatial geometry and semantic information fusion is insufficient, interference of morphological similar structures is likely to occur, and cross-scale adaptability is poor. High-precision and high-robustness segmentation taking three-dimensional geometry as a core constraint and considering semantic interpretability is realized; the method comprises the following steps: preprocessing an input triangular network model to obtain a geometrically coherent 3D primitive cluster; extracting a micro component candidate area in the cluster and pre-segmenting the micro component candidate area to obtain a candidate set at the vertex side of the grid, and mapping the candidate set to the three-dimensional point cloud to obtain an initial segmentation result of the micro component at the point cloud side; enhancing and optimizing the semantic features of the three-dimensional point cloud through a visual language model, and generating an initial segmentation result; and performing micro part label resetting and non-micro part outlier filtering on the initial segmentation result to obtain a final 3D point cloud segmentation result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of 3D segmentation, in particular to a zero-shot space target micro-component segmentation method based on space constraint guidance. BACKGROUND

[0002] With the rise of space missions such as on-orbit service, space situation awareness and autonomous operation and maintenance, there is a higher requirement for the fine identification and segmentation of key components of satellites and other space targets. In these tasks, although the segmentation of main structures and large-scale components such as solar wings is relatively mature, the accurate segmentation of micro-functional components such as antennas, sensors and docking mechanisms is the key to mastering the fine working state, communication intention and on-orbit interaction capability of the target, and therefore has higher application value and more severe technical challenges. These micro-components are small in size and weak in geometric features, and are easily overwhelmed by noise or confused with the edge profiles of large components in three-dimensional data, making their segmentation much more difficult than that of conventional components.

[0003] The existing space target component segmentation methods mainly have the following challenges when dealing with this task. On the one hand, methods relying on two-dimensional image semantic segmentation lack strong constraints on real three-dimensional geometric shapes, making it difficult to accurately depict key structural attributes such as normal consistency, curvature and thickness of components. Under complex imaging conditions such as weak light, strong reflection or partial occlusion, large-area and approximately planar structures such as solar wings are often misjudged as elongated antennas. On the other hand, deep learning methods directly applied to three-dimensional point clouds or mesh models usually rely on a large number of labeled three-dimensional models for supervised training. However, in practical applications, the cost of obtaining such datasets is high, and the model has limited generalization ability when facing new configurations of satellites that have not appeared in the training set, making it difficult to meet the needs of zero-shot scenarios.

[0004] In addition, although the visual language model (VLM) provides a new idea for zero-shot segmentation, it lacks sufficient discrimination in distinguishing specific component categories. Without explicitly introducing three-dimensional geometric priors and interpretable filtering mechanisms, the generated semantic information is prone to noise and boundary drift, especially for micro-components such as antennas, which have poor segmentation result stability. At the same time, although traditional geometric segmentation algorithms do not require training, they are highly sensitive to threshold values, model scales and surface complexity. Fixed parameters are difficult to be universal for satellite models of different scales and configurations, and are prone to under-segmentation or over-segmentation problems, affecting the stability and reproducibility of the method.

[0005] In summary, the existing technology has the problems of insufficient fusion of spatial geometry and semantic information, susceptibility to similar structure interference and poor cross-scale adaptability in the zero-shot segmentation task of space target micro-components. SUMMARY

[0006] The application solves the problems of insufficient fusion of spatial geometry and semantic information, interference of similar structure and poor cross-scale adaptability in the segmentation task of space target micro components in the zero sample scene in the prior art, and realizes a segmentation method with three-dimensional geometry as the core constraint and considering semantic interpretability to realize high-precision and high-robustness target.

[0007] The application provides a zero sample space target micro component segmentation method based on spatial constraint guidance, comprising: Model preprocessing and geometric segmentation are performed on the input triangular network model to obtain a segmented 3D primitive cluster; wherein the segmented 3D primitive cluster is a geometrically continuous 3D primitive cluster; A candidate region of the micro component is extracted in the 3D primitive cluster and pre-segmented to obtain a micro component candidate set on the grid vertex side, and the micro component candidate set on the grid vertex side is mapped to a three-dimensional point cloud corresponding to the triangular network model to obtain an initial segmentation result of the micro component on the point cloud side; The three-dimensional point cloud is subjected to semantic feature enhancement and optimization by a two-dimensional visual language model to generate an initial semantic segmentation result of the three-dimensional point cloud; The initial semantic segmentation result of the three-dimensional point cloud is subjected to micro component category resetting, and the initial segmentation result of the micro component on the point cloud side is covered into the reset result to obtain a coverage result, and the coverage result is subjected to filtering of non-micro component outliers to obtain a final 3D point cloud segmentation result.

[0008] In a possible implementation manner, the model preprocessing and geometric segmentation of the input triangular network model to obtain a segmented 3D primitive cluster comprises: According to the vertex coordinates of each triangular face in the triangular network model, face unit normal vector estimation is performed to obtain the face unit normal vector of each triangular face , and according to all adjacent face unit normal vectors of the vertex , vertex normal estimation is performed to obtain the vertex normal corresponding to each vertex . According to the topological structure of the triangular network model, undirected graph construction is performed, the face unit normal vector and the vertex normal are used to assign edge weights to the undirected graph, and the segmented 3D primitive cluster is obtained by segmenting the weighted undirected graph.

[0009] In a possible implementation manner, the undirected graph construction according to the topological structure of the triangular network model comprises: All vertices of the triangular network model are used as the undirected graph vertex ; The edges formed by pairs of vertices sharing at least one triangular face in the triangular network model are considered as an undirected graph. edge .

[0010] In one possible implementation, the statement based on the face unit normal vector... and the vertex normal The undirected graph is weighted and then the weighted undirected graph is segmented to obtain a segmented 3D primitive cluster, including: Determine the edges in the undirected graph The first vertex corresponding to both ends Second vertex and determine the first vertex. The corresponding first vertex normal Second vertex The normal to the second vertex ; According to the normal of the first vertex and the normal of the second vertex Calculate the dot product between two vertices. and based on the dot product Calculate the initial edge weight of the corresponding edge. ; According to the first vertex Second vertex Calculate the unit direction vector Calculate the second vertex Corresponding vertex normal In the unit direction vector Projection on and according to the projection Determine the edge The convexity includes: concave transitions and convex transitions; The initial edge weight corresponding to the edge that is a convex transition in the undirected graph. Perform a square operation on the initial edge weight corresponding to the edge with a concave transition in the undirected graph. Without performing any operations, the final edge weight corresponding to each edge in the undirected graph is obtained; A threshold-based undirected graph partitioning algorithm is adopted to merge vertices in ascending order of edge weights and dynamically increase the merging threshold to obtain a preliminary primitive cluster. The initial primitive cluster is merged to obtain the segmented 3D primitive cluster.

[0011] In a possible implementation, the threshold-based undirected graph segmentation algorithm is used to obtain a preliminary primitive cluster by merging vertices in ascending order of edge weight and dynamically increasing the merging threshold, including: Each vertex of the undirected graph is taken as an independent primitive cluster, and the segmentation threshold control parameter is used to control the merging of the primitive clusters An initial internal threshold is set for each independent primitive cluster ; The edge weights of the edges in the undirected graph are sorted in ascending order, the sorted edges are traversed in a loop, and the following operations are performed: The first primitive cluster and the second primitive cluster corresponding to the two vertices of the current edge are determined, and the first initial internal threshold and the second initial internal threshold corresponding to the first primitive cluster and the second primitive cluster are determined; The first primitive cluster and the second primitive cluster are merged based on a judgment condition to obtain a new primitive cluster, wherein the judgment condition includes judging whether the first primitive cluster and the second primitive cluster belong to different primitive clusters and judging whether the edge weight corresponding to the current edge is less than or equal to the first initial internal threshold and the second initial internal threshold; Based on the new primitive cluster and the edge weight corresponding to the current edge, an internal threshold corresponding to the new primitive cluster is adaptively calculated to obtain a preliminary primitive cluster.

[0012] In a possible implementation, the preliminary primitive cluster is merged to obtain a segmented 3D primitive cluster, including: A minimum component size parameter is set; The edge set of the undirected graph is traversed, and the following operations are performed: The primitive clusters to which the two vertices connected by the current edge belong are determined, and if the two primitive clusters are different and at least one of the primitive clusters includes a number of vertices less than the minimum component size parameter, the two primitive clusters are merged into a new primitive cluster; The edge set is repeatedly traversed and the above merging operation is performed until there is no primitive cluster that meets the merging condition; The finally obtained primitive cluster is taken as the segmented 3D primitive cluster.

[0013] In a possible implementation, the candidate region of the micro-component in the 3D primitive cluster is extracted and pre-segmented to obtain a micro-component candidate set on the grid vertex side, including: The normalized mean normal, convergence degree, and normal variance of each primitive cluster in the 3D primitive cluster are calculated; The corresponding micro-component score is constructed according to the convergence degree and the normal variance of each primitive cluster, and four thresholds are set to screen the primitive clusters to obtain a first screening set; Refilter the first filtered set by using structure prior knowledge to obtain a candidate set of micro components on the side of the grid vertex.

[0014] In a possible implementation, the mapping of the candidate set of micro components on the side of the grid vertex to the three-dimensional point cloud corresponding to the triangular network model to obtain an initial segmentation result of micro components on the side of the point cloud comprises: constructing a KD tree spatial index of the candidate set of micro components on the side of the grid vertex, calculating an average nearest neighbor distance of the three-dimensional point cloud, and defining an initial matching tolerance based on the average nearest neighbor distance; using the KD tree spatial index, performing the following nearest neighbor query on each point in the three-dimensional point cloud; wherein the nearest neighbor query comprises: calculating the distance between the current point cloud and the nearest neighbor grid vertex, and including the point with a distance less than the initial matching tolerance in a matching point set; calculating the matching coverage of the matching point set, and if the matching coverage is lower than a first threshold, resetting the first threshold to perform the nearest neighbor query to obtain a final matching point set; adding a micro component label to the point cloud in the final matching point set to obtain an initial segmentation result of micro components on the side of the point cloud.

[0015] In a possible implementation, the semantic feature enhancement and optimization of the three-dimensional point cloud by the two-dimensional visual language model to generate an initial semantic segmentation result of the three-dimensional point cloud comprises: for each category in the satellite component category set, generating a diversified text description set by a large language model, and inputting the text description set into a text encoder to obtain text features; projecting the three-dimensional point cloud to a multi-view image set, and extracting visual features of the multi-view image set by a visual encoder; projecting the text features and the visual features to each three-dimensional point of the three-dimensional point cloud to obtain an initial semantic feature of each point; based on the initial semantic feature and the combined feature , constructing a cost matrix for measuring points and primitives, and updating the segmented 3D primitive cluster based on the cost matrix, performing a hierarchical aggregation operation of local smoothing and global propagation on each primitive in the updated segmented 3D primitive cluster to obtain refined visual features; wherein the combined feature is a combination of a description sub of each point in the point cloud coinciding with the vertex of the triangular network model and the vertex normal; The dot product of the refined visual features and the text features is calculated, and initial segmentation log probabilities of each point belonging to various categories are obtained by SoftMax normalization, and then initial semantic segmentation results of the three-dimensional point cloud are obtained.

[0016] In a possible implementation, the initial semantic segmentation result of the three-dimensional point cloud is reset to a micro-component category, including: For each point labeled as a micro-component in the initial segmentation result, the following operations are performed: The KD tree is used to retrieve K nearest neighbor points of the current micro-component point in the point cloud, to form a set of neighbor points The number K is an integer greater than or equal to 1. The occurrence frequency of the category to which all points in the set of neighbor points belong is counted, the category with the highest occurrence frequency is determined as the majority category in the neighborhood, and the category of the current micro-component point itself is updated to the majority category The loop ends until the reset result is obtained.

[0017] The one or more technical solutions provided in the application have at least the following technical effects or advantages: ​​​​The present application obtains a segmented 3D primitive cluster by model preprocessing and geometric segmentation on the input triangular network model. The preprocessing removes noise and redundant data, improving the neatness and processability of the model. The geometric segmentation decomposes the complex model into a coherent 3D primitive cluster, helping to maintain the integrity and consistency of the geometric structure and laying the foundation for subsequent analysis. Then, the candidate regions of micro components are extracted and pre-segmented in the 3D primitive cluster to obtain a micro component candidate set on the grid vertex side. The micro component candidate set on the grid vertex side is mapped to the three-dimensional point cloud corresponding to the triangular network model to obtain the initial segmentation result of the micro component on the point cloud side. This can accurately locate micro components such as antennas and avoid missing details. Pre-segmentation provides preliminary identification candidates, and mapping to the point cloud ensures data consistency, achieving smooth conversion from the grid to the point cloud and improving the accuracy and efficiency of the initial segmentation. Then, based on the initial segmentation result of the micro component on the point cloud side, the three-dimensional point cloud is subjected to semantic feature enhancement and optimization by a two-dimensional visual language model to generate a global semantic segmentation result of the three-dimensional point cloud. The visual language model introduces semantic understanding to enhance the feature representation of the point cloud and optimize the segmentation boundary and class recognition, thereby generating a more intelligent and comprehensive global semantic segmentation and improving the semantic accuracy and robustness of the segmentation result. Finally, the global semantic segmentation result is subjected to micro component class resetting and filtering of non-micro component outliers to obtain the final 3D point cloud segmentation result. This ensures the accuracy of the micro component label by resetting the class, filters outliers to remove noise and misclassification, refines the segmentation result, and improves the purity, reliability, and practicality of the final output, meeting the actual application requirements. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A zero-shot space target micro component segmentation method based on spatial constraint guidance is provided for the embodiments of the present application. Figure 2 A set primitive cluster segmentation flowchart based on a network model is provided for the embodiments of the present application. Figure 3 An overall process framework diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0020] The present application provides a zero-shot space target micro component segmentation method based on spatial constraint guidance, as shown in Figure 1comprising steps S101-S104.

[0021] S101, performing model preprocessing and geometric segmentation on the input triangular mesh model to obtain a segmented 3D primitive cluster; wherein the segmented 3D primitive cluster is a geometrically coherent 3D primitive cluster; Specifically, in step S101, the model preprocessing and geometric segmentation are performed on the input triangular mesh model to obtain a segmented 3D primitive cluster, comprising steps S1011-S1013.

[0022] S1011, according to the vertex coordinates of each triangular face in the triangular mesh model, performing face normal vector estimation to obtain the face unit normal vector of each triangular face , and according to all adjacent face unit normal vectors of the vertex performing vertex normal estimation to obtain the vertex normal corresponding to each vertex . S1012, according to the topological structure of the triangular mesh model, performing undirected graph construction, assigning edge weights to the undirected graph according to the face unit normal vector and the vertex normal , and segmenting the weighted undirected graph to obtain the segmented 3D primitive cluster.

[0023] Here, according to the topological structure of the triangular mesh model, the undirected graph construction includes: (1) taking all vertices of the triangular mesh model as vertices of the undirected graph ; (2) taking the edges formed by vertex pairs sharing at least one triangular face in the triangular mesh model as edges of the undirected graph .

[0024] Here, according to the face unit normal vector and the vertex normal , the edge weight of the undirected graph is assigned, and the weighted undirected graph is segmented to obtain the segmented 3D primitive cluster, including: (1) determining the first vertex and the second vertex corresponding to the edge in the undirected graph, and determining the first vertex normal corresponding to the first vertex and the second vertex normal corresponding to the second vertex ; (2) calculating the dot product between the first vertex normal and the second vertex normal between the two vertices according to the first vertex normal and the second vertex normal .And based on the dot product Calculate the initial edge weight of the corresponding edge. ; (3) Based on the first vertex Second vertex Calculate the unit direction vector Calculate the second vertex Corresponding vertex normal unit direction vector Projection on And according to the projection Determine the edge The convexity includes concave transitions and convex transitions. (4) Initial edge weights for edges that are convex transitions in an undirected graph. Perform squaring to initialize the edge weights of edges that are concave in the undirected graph. Without performing any operations, we obtain the final edge weight corresponding to each edge in the undirected graph; (5) A threshold-based undirected graph partitioning algorithm is adopted to merge vertices in ascending order of edge weight and dynamically increase the merging threshold to obtain a preliminary primitive cluster. Here, a threshold-based undirected graph partitioning algorithm is adopted. By merging vertices in ascending order of edge weights and dynamically increasing the merging threshold, a preliminary primitive cluster is obtained, including: (5.1) Treat each vertex of the undirected graph as an independent primitive cluster, and control the parameters according to the segmentation threshold. Set a corresponding initial internal threshold for each individual primitive cluster. ; (5.2) Sort the edge weights of the undirected graph in ascending order, traverse the sorted edges, and perform the following operations: (5.3) Determine the first primitive cluster and the second primitive cluster corresponding to the two vertices of the current edge, and determine the first initial internal threshold and the second initial internal threshold corresponding to the first primitive cluster and the second primitive cluster respectively; (5.4) Based on the judgment conditions, the first primitive cluster and the second primitive cluster are merged to obtain a new primitive cluster; wherein, the judgment conditions include: judging whether the first primitive cluster and the second primitive cluster belong to different primitive clusters, and judging whether the edge weight corresponding to the current edge is less than or equal to the first initial internal threshold and the second initial internal threshold. (5.5) Based on the new primitive cluster and the edge weight corresponding to the current edge, the internal threshold corresponding to the new primitive cluster is adaptively calculated to obtain the preliminary primitive cluster.

[0025] (6) Merge the initial primitive clusters to obtain the segmented 3D primitive clusters. Here, merging the initial primitive clusters to obtain the segmented 3D primitive clusters includes: (6.1) setting a minimum component size parameter; (6.2) traversing the edge set of the undirected graph, performing the following operations: (6.2.1) determining the two primitive clusters to which the two vertices connected by the current edge belong, if the two primitive clusters are different, and at least one of the two primitive clusters includes a number of vertices less than the minimum component size parameter, merging the two primitive clusters into a new primitive cluster; (6.2.2) repeating the traversal of the edge set and performing the above merging operation until there is no primitive cluster that meets the merging condition; (6.3) taking the final set of primitive clusters as the segmented 3D primitive clusters.

[0026] For example, see Figure 2 , Figure 2 The geometric primitive cluster segmentation flowchart based on the grid model provided by the embodiment of the application. Here, it needs to be made clear that in this example, the antenna is taken as a specific example of a micro-component.

[0027] The purpose of this step is to use the topological and geometric information of the grid model to generate a three-dimensional primitive cluster with geometric continuity and stable boundaries through normal estimation, graph segmentation and small cluster merging, laying a foundation for subsequent antenna candidate screening.

[0028] S1.1, geometric feature extraction. In order to enhance the geometric representation ability of the point cloud, on the basis of normal estimation, the local geometric descriptor of the point cloud is further extracted. Specifically, for each point in the point cloud coinciding with the grid vertex , its fast point feature histogram (FPFH) descriptor is calculated.

[0029] FPFH is a local surface geometric descriptor used to capture the curvature, normal distribution and other geometric statistical properties of the point cloud in the neighborhood of the point. In this embodiment, the FPFH feature is calculated by using the Open3D library, and the feature vector dimension is 33.

[0030] For each point , its complete combined feature is composed of the FPFH descriptor and the vertex normal : ; The combined feature is used to calculate the geometric feature distance between the point and the primitive in the semantic smoothing step of S3, enhancing the geometric consistency of semantic smoothing.

[0031] S1.2, surface normal estimation.

[0032] For each triangle in the mesh model , suppose its three vertices are , , , to obtain the unit normal vector of the triangle, first calculate two edge vectors and , then get a vector perpendicular to the triangle through the cross product operation, and finally normalize the processing. The calculation formula of the surface normal is: ; Specifically, in the code implementation, the cross product operation adopts the standard formula of three-dimensional vector cross product: for vector and vector , the components of the cross product result are: ; ; ; Then calculate the length of the cross product vector , and normalize the components: , , , thus obtaining the unit normal vector .

[0033] This step ensures that each triangle has a clear normal direction, providing a geometric basis for subsequent vertex normal smoothing and edge weight calculation. Accurate estimation of the surface normal is crucial for distinguishing convex and concave transitions, and identifying flat areas and curved surfaces.

[0034] S1.3, vertex normal estimation. To obtain the normal vector at the vertex of the mesh, it is necessary to integrate the normal information of all triangles adjacent to the vertex. Let be the set of triangles adjacent to vertex , denote the number of elements in the set (i.e. the number of adjacent faces), then the normal of vertex uses the arithmetic mean of the adjacent face normals: ; where represents the sum of the face normals of all triangles in the set .

[0035] Further, in order to ensure smooth transition of vertex normal when adjacent faces change, an incremental average strategy is adopted in actual calculation: when traversing adjacent triangle faces of a vertex , let the current accumulated normal be , the newly added face normal be , and the number of visited faces be , then the update formula is: ; where is a linear interpolation function defined as . Through this incremental update method, the current accumulated normal and the new normal are fused according to the weight of each time a new face normal is added, and finally converge to the arithmetic mean of all adjacent face normals.

[0036] This smooth vertex normal estimation method can effectively suppress noise and discontinuity in the mesh, making the subsequent edge weight calculation based on normal difference more robust and avoiding fragmentation caused by abnormal normal of a single triangle.

[0037] S1.4, graph construction, edge weight definition and segmentation.

[0038] S1.4.1, graph construction and edge weight definition based on normal difference and containing convexity adjustment.

[0039] According to the topological structure of the mesh model, an undirected graph is constructed. The vertex set corresponds to all vertices of the mesh, and the edge set corresponds to vertex pairs that share at least one triangle. Specifically, for each triangle in the mesh, three edges are formed between its three vertices, and the complete edge set can be obtained by traversing all triangles.

[0040] In order to reflect the geometric difference between adjacent vertices and guide graph segmentation, the weight of each edge needs to be defined. The design of edge weight follows the principle of "larger normal difference, higher edge weight (tending to segmentation), consistent normal, lower edge weight (tending to preservation)", and special processing is performed for convex corner regions to avoid excessive segmentation.

[0041] For two vertices and connected by an edge , let their normals be and , then the calculation of edge weight includes the following steps: First step: calculate the cosine difference of normals. Calculate the dot product of the two vertex normals​ , the dot product value ranges from , where 1 means the normals are exactly the same, -1 means the normals are exactly opposite. Define the initial edge weight as the difference of the normals cosine: ; The edge weight is close to 0 when the normals are close to the same (dot close to 1), and close to 2 when the normals are quite different (dot close to -1), which meets the requirement of "larger difference, higher weight" for segmentation.

[0042] Second step: convexity adjustment. To determine whether the region where the edge is located is a convex turn, the unit direction vector : ; Then calculate the projection of the normal on the direction . When , it indicates that the normal of the vertex is directed to the extension direction of the edge, and the region is a convex turn; otherwise, it is a concave turn.

[0043] Third step: convexity weight amplification. For convex turn regions, since the antenna and other protruding structures often have convex surfaces, in order to avoid excessive segmentation in these regions, the edge weight needs to be squared to reduce its relative size, so that the edge is less likely to be segmented: If , then .

[0044] Through the above convexity adjustment, the edge weight of the convex region is compressed twice, so that even if there is a certain difference in the normal, the edge still tends to be retained in the same component, thereby maintaining the integrity of the convex structure such as the antenna; while the edge weight of the concave region remains linear and normally participates in the segmentation judgment.

[0045] In summary, the maximum weight of the edge is determined by the following formula: ; If , then .

[0046] Where , are the vertex normals, and is the unit direction vector.

[0047] S1.4.2, threshold-based graph segmentation and threshold adaptive update.

[0048] After completing the edge weight calculation, a threshold-based graph segmentation algorithm is used to segment the graph​ The segmentation is performed to generate a preliminary set of geometric primitive clusters. The core idea of the algorithm is that for each edge, if its weight is less than the current threshold of the components to which its two ends belong, then the two components are merged; after the merging, the threshold of the new component is updated to adaptively increase with the component size, so as to keep the stability of large components while allowing flexible merging of small components.

[0049] Specifically, the segmentation process includes the following steps: Initialization: each vertex in the graph is initialized as an independent component comp, and the initial threshold of each component is set as wherein is a segmentation threshold control parameter. This parameter controls the granularity of segmentation: The larger it is, the coarser the segmentation (tending to generate larger components); The smaller it is, the finer the segmentation (tending to generate smaller components).

[0050] Edge sorting: all edges in the edge set are sorted in ascending order of the weight . In this way, edges with high geometric similarity can be processed first, and the local coherent region is gradually expanded to the global segmentation result.

[0051] Edge-by-edge merging judgment: each edge is processed in turn according to the sorting order. Let the two vertices connected by the edge belong to the representative elements (obtained by the find operation of the union-find set) of the components to which they currently belong be and . If the vertices do not belong to the same component, the following judgment is performed: If and , then the merging is performed: merge and into a new component, and perform the join operation of the union-find set; then find the representative element of the merged component ; finally, update the threshold of the new component as: ; wherein represents the vertex size of the new component after the merging.

[0052] The significance of the threshold update formula is that the threshold of the new component is composed of the "weight of the currently merged edge " and the "scale decay term ". As the component size increases, the decay term Gradually reduce, so that the large component threshold growth slows down, thus avoiding excessive merging caused by the boundary blur; and small components of the threshold remains higher growth potential, easy to absorb the adjacent fragmentation area.

[0053] Data structure: component management uses disjoint-set / union-find data structure, supporting efficient find (find representative) and join (merge component) operations. Each component records its rank, representative (p) and size, and adopts the rank-based merging strategy to maintain the balance of the disjoint-set tree during merging, and adopts path compression during finding to speed up subsequent operations.

[0054] After the above threshold-based graph segmentation process, the grid is segmented into several preliminary geometric primitive clusters , each primitive cluster The internal vertex normal is relatively consistent, and the normal difference at the boundary between clusters is significant, thereby realizing the preliminary segmentation of geometric continuity.

[0055] Segmentation threshold control parameters The optimal value is automatically selected through the subsequent parameter adaptive determination step S1.5.

[0056] S1.4.3, small cluster merging.

[0057] After the graph segmentation of S1.4.2, a certain number of small-scale components (fragmented areas) will be generated. These small components often correspond to noise, isolated vertices or local geometric mutations in the grid, which will affect the accuracy and stability of the subsequent antenna candidate selection if not handled. Therefore, it is necessary to perform merging operation on the components with a size less than the preset threshold, and merge them into adjacent larger components to eliminate segmentation fragmentation.

[0058] Specifically, set the minimum component size parameter , for all components obtained by preliminary segmentation, execute the following small cluster merging process: Traverse all edges: traverse all edges in the edge set . .

[0059] Check component size: for the edge The two vertices connected by the edge find the representative elements of the components to which they currently belong through the disjoint-set find operation And .

[0060] Perform merging: if , and meet any of the following conditions: Or , perform merging operation: merge And Merge into a new component.

[0061] This process continues until all edges adjacent to the smaller component have been processed. Since each merge increases the component size, after several iterations, the vast majority of smaller components will be absorbed into adjacent larger components. Ultimately, the size of all components will be no less than [a certain value]. This allows us to obtain a set of geometric primitive clusters with stable size and clear boundaries. .

[0062] The purpose of the small cluster merging operation is: (1) Eliminate noise and isolated vertices to improve the robustness of segmentation results; (2) Reduce fragmentation and lower the computational burden of subsequent processing; (3) Ensure that each primitive cluster has sufficient vertex support to provide a reliable sample basis for subsequent statistical calculations (normal mean, variance).

[0063] Minimum component size parameter The optimal value is automatically selected through the subsequent parameter adaptive determination step in S1.5. In typical satellite model segmentation tasks, the common value range for this parameter is 50 to 300, with the specific value depending on the number of vertices, geometric complexity, and model scale. The lower bound of the adjusted parameter is fixed at 50 to ensure that the primitive clusters have sufficient statistical reliability.

[0064] S1.5, parameters are determined adaptively.

[0065] To improve the robustness and generalization ability of the method under different satellite model scales and geometric complexities, feature analysis is performed on the input mesh model before performing graph segmentation in S1.4, and the segmentation threshold control parameters are adaptively determined based on the model features. With minimum component size parameter The optimal value.

[0066] S1.5.1, Model Feature Extraction.

[0067] Extract the following four types of geometric features from the input mesh model: Vertex count: Directly counts the total number of vertices in the grid. This reflects the discretization accuracy of the model; Model Scale: Calculate the 3D scale of the mesh axis-aligned bounding box. Define the model scale as the diagonal length: ; Geometric complexity: Geometric complexity is estimated using vertex density and is defined as follows: ; Surface variation degree: the distance between the top vertex pair of the sampled part The average edge length is calculated as the average distance between adjacent samples , normalized to the model scale: .

[0068] S1.5.2, Candidate parameter generation.

[0069] Segmented generation according to the number of model vertices and the initial candidate set of : 1) If , generate: , ; 2) If , generate: , ; 3) If , generate: , .

[0070] Based on geometric complexity adjustment: if , multiply each value in the candidate set by 0.7 to accommodate the need for more subdivision for high complexity models; Based on model scale adjustment: multiply each value in the candidate set by the normalized model scale and ensure that the adjusted value is not less than 50. Let the adjusted be , then: If , take downward rounding; If , take .

[0071] Combination generation: perform Cartesian product on the adjusted candidate set and the candidate set to obtain a number of candidate pairs.

[0072] S1.5.3, Candidate evaluation and optimization.

[0073] For each group of candidate parameters , the following evaluation process is performed: ​Run S1 and S2: Run the graph segmentation process of S1.1 to S1.4.3 and the antenna detection process of S2 with the set of parameters to obtain the segmentation result; Parse the antenna detection result: Extract the number of antenna candidates and the average antenna score from the output file ; ; Calculate the confidence score: Define the comprehensive confidence score as ; The formula combines the antenna quality and the reasonableness of the number, encouraging parameter combinations with both high quality and appropriate number of antennas; Traverse the evaluation results of all candidate parameters, select the parameter combination with the highest confidence score as the optimal parameter combination, and re-execute the complete S1 graph segmentation process with the set of parameters to obtain the final primitive cluster .

[0074] After the complete process of S1.1 to S1.5.3, a three-dimensional primitive cluster with geometric continuity and stable size is obtained , the vertex normal inside each primitive cluster is relatively consistent, the surface morphology is continuous, the normal difference at the boundary between clusters is significant, and the geometric transition is obvious. These primitive clusters provide high-quality geometric input for the subsequent candidate selection of micro components in the S2 step, ensuring the accuracy and stability of micro component (such as antenna) segmentation.

[0075] S102, extract the candidate region of the micro component in the 3D primitive cluster and perform pre-segmentation to obtain the micro component candidate set on the grid vertex side, and map the micro component candidate set on the grid vertex side to the three-dimensional point cloud corresponding to the triangular network model to obtain the initial segmentation result of the micro component on the point cloud side; Specifically, in step S102, the candidate region of the micro component in the 3D primitive cluster is extracted and pre-segmented to obtain the micro component candidate set on the grid vertex side, including the following steps S1021 to S1023.

[0076] S1021, calculate the normalized mean normal, convergence degree and normal variance of each primitive cluster in the 3D primitive cluster; S1022, construct the corresponding micro component score according to the convergence degree and normal variance corresponding to each primitive cluster, and set four thresholds to screen the primitive clusters to obtain a first screening set; S1023, use structure prior knowledge to further screen the first screening set to obtain the micro component candidate set on the grid vertex side.

[0077] ​Specifically, in step S102, the candidate set of small parts on the vertex side of the mesh is mapped onto the 3D point cloud corresponding to the triangular network model to obtain the initial segmentation result of the small parts on the point cloud side, including: (1) Construct a KD tree spatial index for the candidate set of small parts on the vertex side of the mesh, calculate the average nearest neighbor distance of the 3D point cloud, and define the initial matching tolerance based on the average nearest neighbor distance; (2) Using the KD-tree spatial index, perform the following nearest neighbor query for each point in the 3D point cloud; where the nearest neighbor query includes: (2.1) Calculate the distance between the current point cloud and the nearest neighbor grid vertex, and include points whose distance is less than the initial matching tolerance into the matching point set; (2.2) Calculate the matching coverage of the matching point set. If the matching coverage is lower than the first threshold, reset the first threshold and execute the nearest neighbor query to obtain the final matching point set. (2.3) Add tiny component labels to the point cloud in the final matching point set to obtain the initial segmentation result of the tiny components on the point cloud side.

[0078] For example, S2, convex isolation and candidate filtering. Here, it should be clarified that in this example, the antenna is illustrated as a specific instance of a tiny component.

[0079] The purpose of this step is to obtain the geometric primitive cluster from S1. The algorithm identifies and extracts candidate regions for small components (such as antennas). Through normal statistics calculation, combined scoring, multiple threshold filtering, and structural prior exclusion, it achieves pre-segmentation of small components (such as antennas), providing high-quality initial candidates for subsequent semantic smoothing. The entire S2 step includes four core steps: cluster-level normal statistics calculation, combined scoring and multiple threshold filtering, forced exclusion of non-antenna structures, and matching of mesh vertices and point clouds.

[0080] S2.1, Calculation of cluster-level normal statistics.

[0081] To characterize each primitive cluster To determine the internal normal distribution characteristics, three key statistics need to be calculated: normalized mean normal, convergence, and normal variance. These statistics can effectively characterize the geometric features of convex structures such as antennas, which exhibit "uniform normal pointing and low dispersion."

[0082] S2.1.1, Normalized mean normal direction.

[0083] For primitive clusters Let the set of vertex normals it contains be... Cluster size is First, sum the vectors of all normals within the cluster, then normalize them to obtain the main direction of the cluster: ; where, denotes vector sum, denotes Euclidean norm of vector. The normalized mean normal represents the dominant direction of normals within the cluster, which is the reference for subsequent statistical quantity calculation.

[0084] In actual calculation, in order to avoid the normal sum being a zero vector, it is necessary to check the length of the normal sum first: let , if , then perform normalization; otherwise, set as a zero vector or skip the processing of the cluster.

[0085] S2.1.2, convergence degree statistical quantity.

[0086] Convergence degree is used to measure the consistency degree of the normals within the cluster with the dominant direction. Specifically, calculate the dot product of each normal with the normalized mean normal , take the positive part, and then average all the vertices: ; where, denotes max operation, which is used to filter out the normals with an angle greater than 90 degrees with the dominant direction, and only keep the components pointing to the dominant direction. The value range of is [0, 1], where 1 indicates that all normals within the cluster completely point to the dominant direction (perfect convergence), and 0 indicates no convergence.

[0087] Antennas and other convex structures usually have a high convergence degree (close to 1) because their surface normals tend to point to the same focus or direction; while flat surfaces or randomly distributed areas have a lower convergence degree.

[0088] S2.1.3, normal variance statistical quantity.

[0089] Normal variance is used to measure the dispersion degree of the normals within the cluster. Specifically, calculate the squared Euclidean distance of each normal from the normalized mean normal , and then average all the vertices: ; where, denotes squared Euclidean distance of vector. The value range of is

[0090] The convex structure such as antenna usually has a low variance due to its surface normal relatively concentrated; while the irregular surface or the region containing multiple curved surfaces has a high variance.

[0091] By the above three statistics, , , , the normal distribution characteristics of each primitive cluster can be comprehensively described, and quantitative criteria are provided for subsequent antenna candidate screening.

[0092] S2.2, combined score and multiple threshold screening.

[0093] In order to comprehensively consider the two aspects of "high consistency and low dispersion", a linear weighted combined score is constructed, and multiple thresholds are used to perform candidate screening.

[0094] S2.2.1, combined score construction.

[0095] For the primitive cluster , the antenna score is defined as a linear combination of convergence and normalized variance score: ; Wherein, and are weighting coefficients, used to balance the relative importance of convergence and variance in the score; is a normalization constant, used to map the variance to a reasonable numerical range. The term represents the variance score: when , , the term is 1 (optimal), and when , the term is 0.

[0096] In this embodiment, is preferably 0.7, is preferably 0.3, is preferably 0.5. This set of parameter configurations makes the convergence the main factor and the variance the auxiliary factor, which is consistent with the geometric characteristics of the antenna structure.

[0097] S2.2.2, multiple threshold screening.

[0098] In order to ensure that the candidate cluster meets multiple necessary conditions at the same time, the following four threshold constraints are set: (1) Convergence threshold: , which requires that the normals in the cluster have high enough consistency of principal direction; in this embodiment, is preferably 0.85.

[0099] (2) Variance threshold: , requiring the normal dispersion within the cluster to be low enough. In this embodiment, is preferably 0.15.

[0100] (3) Scale constraint: , requiring the cluster scale to be within a reasonable range. Herein, is the lower limit of the cluster scale, used to filter out noise clusters or fragmented regions with too few vertices; is the upper limit of the cluster scale, used to exclude satellite body structures with too many vertices. In this embodiment, is preferably 100, is preferably 800.

[0101] (4) Comprehensive score threshold: , requiring the combined score to meet the minimum requirement, wherein is the comprehensive decision threshold. In this embodiment, is preferably 0.80.

[0102] Only the primitive clusters that meet the above four conditions can enter the next step of structure prior verification; otherwise, they are directly excluded. This set of thresholds has been experimentally verified to effectively suppress false antennas while maintaining a high recall rate.

[0103] S2.3, Forced exclusion of non-antenna structures.

[0104] After multiple threshold screenings in S2.2, some non-antenna structures (such as solar wing large planes, satellite body ring components, and low-profile surface features) may be misjudged as antenna candidates due to similar normal distribution characteristics. Therefore, targeted structure priors need to be introduced for forced exclusion to improve the accuracy and robustness of antenna identification.

[0105] S2.3.1, Solar wing exclusion.

[0106] Solar wings usually appear as large-area flat sheets with extremely low normal variance, normal height alignment, and extremely thin thickness. To distinguish them from antennas, the following three criteria are set. When at least two of them are true, the cluster is determined to be a solar wing and is excluded: Criterion (a): Low plane variance. If , it is considered that the normal dispersion within the cluster is extremely low, showing a planar feature, preferably 0.005.

[0107] Criterion (b): Normal height alignment. Count the number of normals within the cluster that satisfy the absolute value of the dot product between the normal and the normalized mean normal greater than the threshold , and define the alignment ratio as: ; If , the cluster is considered to have its normals almost perfectly aligned with the principal directions. Preferably, 0.99, 0.95.

[0108] Criterion (c): Extreme thinness ratio. Calculate the axis-aligned bounding box of the cluster with its three dimension scales , , , define the thickness ratio: ; If , the cluster is considered to be extremely thin in one dimension, presenting a sheet-like feature; preferably 0.05.

[0109] When at least two of the above three criteria are satisfied, the cluster is determined to be a solar wing and excluded from the subsequent micro-component candidate set.

[0110] S2.3.2, Ring-shaped structure exclusion.

[0111] Some components of the satellite body (such as docking rings, thruster rings) present a ring-shaped geometry, whose bounding box scales are close in two dimensions and small in the third, and the point distribution presents a ring-like pattern. To exclude such structures, the following criteria are set: Geometric prior judgment: Calculate the axis-aligned bounding box of the cluster with its three dimension scales , and sort them in size as . If both of the following are satisfied: ; ; the cluster is considered to meet the ring-shaped geometric prior. Preferably, 0.30, 0.60.

[0112] Ring consistency test: Calculate the set of distances from all vertices in the cluster to the center of the bounding box , and calculate the coefficient of variation (CV) of the distances: ; where denotes the variance of the distances, denotes the mean of the distances. The coefficient of variation measures the degree of dispersion of the distance distribution: The smaller it is, the more uniformly the points are distributed on the circumference of a circle at a fixed radius from the center, and the more it meets the ring-shaped feature.

[0113] If , the cluster is determined as a ring structure and excluded. is preferably 0.30.

[0114] S2.3.3, Low-profile surface feature exclusion.

[0115] Some small components on satellite surface (such as heat sink, sensor window) have concentrated normal distribution, but they are almost not convex relative to the satellite body, presenting low-profile features. In order to exclude such structures, the following criterion is set: Height distribution analysis: with the cluster center as the reference point, the distance of each vertex to the center is defined as: ; The maximum value and the minimum value of the height set are calculated, and the height variation range is defined as: ; The reference height is defined as the midpoint of the maximum and minimum height: ; If and are met at the same time, the cluster is determined as a low-profile surface feature and excluded. is preferably 0.2, is preferably 0.1.

[0116] After the forced exclusion of the above three types of structures (solar wing, ring structure, low-profile surface), the remaining primitive clusters are the candidate set of micro-components that pass all geometric and statistical tests, which are used for subsequent semantic smoothing and label transfer.

[0117] S2.4, Matching of mesh vertices and point cloud.

[0118] Under the premise that the mesh model and the point cloud have been aligned to the same coordinate system, in order to map the antenna labels obtained on the mesh vertex side through S2 screening to the point cloud, to complete the label transfer and assist the subsequent point cloud segmentation, the nearest neighbor matching based on geometric distance needs to be performed. This matching process uses KD tree to accelerate the nearest neighbor search, and through adaptive tolerance and back-off mechanism to improve the robustness of the matching.

[0119] S2.4.1, Construction of KD tree index.

[0120] Let the antenna vertex set obtained on the mesh vertex side after S2 screening be , where each vertex is a three-dimensional coordinate. In order to efficiently perform the nearest neighbor query from the point cloud to the mesh vertices, a KD-tree is constructed to build the KD-tree spatial index.

[0121] The KD-tree is a kind of spatial partitioning data structure, which can complete a single nearest neighbor query in time complexity. In the present embodiment, the KD-tree construction and query can be implemented by using existing libraries, without limitation to specific implementation libraries.

[0122] S2.4.2, calculate the average nearest neighbor distance and define a small tolerance.

[0123] Let the point cloud be , where each point . In order to set the matching tolerance adaptively, it is necessary to estimate the local density characteristics of the point cloud first. Specifically, for each point in the point cloud , the distance to the nearest neighbor point is calculated, and then the average nearest neighbor distance is obtained by averaging the nearest neighbor distances of all points: ; wherein denotes the nearest neighbor point of the point , and denotes the Euclidean distance.

[0124] Based on the average nearest neighbor distance , a small tolerance is defined: ; wherein is a tolerance proportion coefficient. Preferably, can be fixed as or preferably 1.0. The small tolerance is used for the first matching, so as to ensure that only the point cloud points that are close enough to the mesh antenna vertices are marked as the antenna.

[0125] S2.4.3, perform the nearest neighbor matching and calculate the matching coverage rate.

[0126] For each point in the point cloud , the KD-tree is used to query the nearest neighbor point and the corresponding distance in the antenna vertex set : ; If , it is considered that the point cloud point is successfully matched to the mesh antenna, and it is added to the matching set .

[0127] statistical matching set , and calculate the matching coverage : ; The matching coverage represents the proportion of the number of points in the point cloud that are successfully matched to the mesh antenna.

[0128] The total number of mesh antenna vertices, used to evaluate the degree of coverage of the mesh antenna area by the point cloud. When the point cloud density is high and well aligned, may be greater than 1; when the point cloud density is low, less than 1. If is high, it indicates that the point cloud is well aligned with the mesh and the point cloud density is sufficient, and the small tolerance is sufficient to complete the matching; if is low, the fallback mechanism needs to be enabled.

[0129] S2.4.4, adaptive tolerance fallback.

[0130] When , where is the lower limit threshold of the matching coverage, it indicates that the small tolerance is too strict, resulting in too few point cloud points successfully matched. At this time, the fallback mechanism is enabled: a larger tolerance is used to re-execute the matching.

[0131] Define the large tolerance : ; where, is the large tolerance proportion coefficient. Preferably, is 2.5, is 0.10.

[0132] Re-execute the nearest neighbor matching of S2.4.3 with to obtain a new matching set . Since is significantly greater than , the size of is usually significantly increased, thereby improving the matching coverage.

[0133] S2.4.5, label assignment.

[0134] For point cloud points that meet the distance threshold (i.e. or ), the antenna label is assigned. For unmatched point cloud points, their original label is retained or a non-antenna label is assigned.

[0135] Through the above matching process, the vertex side (mesh) side antenna labeling is mapped to the geometry of the point cloud, providing an initial label for subsequent point cloud semantic segmentation, and can be used to verify the accuracy of the S2 screening result.

[0136] S103, performing semantic feature enhancement and optimization on the three-dimensional point cloud through a two-dimensional visual language model to generate an initial semantic segmentation result of the three-dimensional point cloud; Specifically, in step S103, the semantic feature enhancement and optimization are performed on the three-dimensional point cloud through a two-dimensional visual language model to generate an initial semantic segmentation result of the three-dimensional point cloud, including steps S1031 to S1035.

[0137] S1031, for each category in the satellite component category set, generating a diversified text description set through a large language model, and inputting the text description set into a text encoder to obtain text features; S1032, projecting the three-dimensional point cloud to a multi-view image set, and extracting visual features from the multi-view image set through a visual encoder; S1033, projecting the text features and the visual features to each three-dimensional point of the three-dimensional point cloud to obtain initial semantic features of each point; S1034, based on the initial semantic features and the combined features , constructing a cost matrix for measuring points and primitives, and updating the segmented 3D primitive cluster based on the cost matrix, performing local smoothing and global propagation hierarchical aggregation operations on each primitive in the updated segmented 3D primitive cluster to obtain refined visual features; wherein the combined features are obtained by combining the description sub of each point in the point cloud where the vertices of the triangular network model coincide and the vertex normal; S1035, calculating the dot product of the refined visual features and the text features, and obtaining the initial segmentation log probability of each point belonging to each category through SoftMax normalization, and then obtaining the initial semantic segmentation result of the three-dimensional point cloud.

[0138] For example, S3: please refer to Figure 3 , Figure 3 is the overall flow framework diagram provided by the embodiment of the application. The flow principle of S3 and S4 is included. Here, it needs to be clear that in this example, the antenna is taken as a specific example of a micro component.

[0139] With the help of the knowledge possessed by the two-dimensional visual language model (VLM), semantic features are given to the three-dimensional point cloud. Specifically, a generative pre-training transformation model 4 (GPT-4) is used to generate text features, a contrastive language-image pre-training (CLIP) is used to generate visual features, and the two types of features are associated with each point in the three-dimensional point cloud.

[0140] S3.1, text feature acquisition for each category of satellite components , 10 queries are initiated to GPT-4 to obtain a set of diversified text descriptions .

[0141] During the query process, a unified prompt template is used for all satellite component categories, which is expressed as: "This is the [category] of the satellite. Please describe its detailed appearance for segmentation, including contour boundaries, shape, relative position, and any unique structural features."

[0142] The obtained text descriptions are input to the text encoder of CLIP to obtain corresponding text features ; Visual feature acquisition According to the general method in existing research, the three-dimensional point cloud P is projected to multiple pre-set perspectives to generate an image set .

[0143] The generated image set is input to the visual encoder of CLIP, and after processing, a dense visual feature is generated .

[0144] Feature projection projects the text features and visual features obtained through the above steps into two-dimensional features, and finally obtains the initial two-dimensional VLM feature representation of each point in the three-dimensional point cloud.

[0145] S3.2, module purpose This module is used to solve the problem of local semantic noise caused by occlusion of the perspective, projection error, and not considering the inherent geometric relevance of the three-dimensional point cloud when projecting directly from two-dimensional images to three-dimensional point clouds. By denoising the initial VLM feature representation and strengthening semantic consistency, the accuracy of three-dimensional point cloud semantic segmentation is improved.

[0146] S3.2.1, cost function optimization updates 3D primitives Using the three-dimensional geometric features obtained earlier, the 3D primitives are updated by optimizing the cost function, which can balance spatial proximity, geometric consistency, and semantic consistency.

[0147] The assignment problem of points to primitives is converted into an optimal transmission problem, and a cost matrix is defined , whose expression is: ; In the above formula, and represent the coordinates of the point and the primitive ; and represent the geometric features of the point and the primitive , which combines the FPFH feature and the normal vector ; and correspond to the semantic features of the point and the primitive ; , , are hyperparameters, respectively used to balance the contributions of the coordinate distance , the geometric feature distance and the semantic feature distance in the cost matrix.

[0148] Two-step hierarchical aggregation optimizes local smoothing of features: local smoothing is performed within the neighborhood of the primitive, and the embedding features of points with similar geometric and semantic features are fused to eliminate local semantic noise.

[0149] Global propagation: enables contextually similar primitives to exchange context information, refines the primitive-level features through graph convolution, and finally obtains the refined visual feature set .

[0150] S3.3: Initial segmentation generation.

[0151] The initial segmentation log probability is calculated by the dot product of the smoothed VLM visual feature representation and all n-class text features, and the calculation result is input into the SoftMax normalization function for processing to obtain the initial segmentation log probability , whose mathematical expression is: .

[0152] S104, the initial semantic segmentation result of the three-dimensional point cloud is reset to the micro component class, and the initial segmentation result of the micro component of the point cloud side is covered to the reset result to obtain the covering result. The covering result is filtered to obtain the final 3D point cloud segmentation result.

[0153] Here, the initial semantic segmentation result of the three-dimensional point cloud is reset for the micro-component class, including: For each point labeled as a micro-component in the initial segmentation result, the following operations are performed: (1) By KD tree retrieval, the nearest neighbor points of the current micro-component point in the point cloud are retrieved to form a set of nearest neighbor points ; (2) The occurrence frequency of the class to which all points in the set of nearest neighbor points belong is counted, and the class with the highest occurrence frequency is determined as the majority class in the neighborhood , and the class of the current micro-component point itself is updated to the majority class ; (3) Until the loop ends, the reset result is obtained.

[0154] S4: semantic filtering processing. Here, it needs to be clear that in this example, the antenna is taken as a specific example of a micro-component.

[0155] The initial segmentation result is refined, and the final three-dimensional point cloud segmentation result is obtained through antenna area reset and outlier filtering.

[0156] The step S4 of the embodiment specifically includes: S4.1, since the visual language model (VLM) has a large noise in understanding the class semantics of micro-components (such as antennas), which will have a negative impact on the segmentation of other classes. In order to eliminate this interference, first, the neighborhood reset processing is performed on the points labeled as antennas in the initial segmentation result .

[0157] Specifically, for the initial segmentation result , where represents the three-dimensional coordinates of the i-th point, represents its semantic label, and N is the total number of points. Define the set of antenna points as: ; For each antenna point , its set of nearest neighbor points is retrieved by KD tree: .

[0158] The occurrence frequency of each class label in the neighborhood is counted, and the label with the highest occurrence frequency is selected as the majority label : ; wherein ​is the indicator function. Update the point label to the majority label, i.e. . This process implements the neighborhood reset for antenna noise prediction, and the reset point cloud is .

[0159] S4.2: To recover the real antenna region, the antenna point cloud segmented by the geometric-aware antenna isolation module is overlaid to the result after reset . .

[0160] This overlay operation ensures that the antenna region is segmented based on reliable geometric features and is not affected by VLM semantic noise, and the antenna- overlaid point cloud is obtained .

[0161] For the points in that are not in the antenna category, further outlier removal is performed to eliminate local segmentation abnormalities. Define the set of non-antenna points: ; For each non-antenna point , retrieve its nearest neighbors , calculate the majority label within the neighborhood (calculation method same as S41). If the point label is inconsistent with the neighborhood majority label, i.e. , it is determined to be an outlier and its label is updated: ; This operation is based on the assumption of spatial continuity, i.e. geometrically adjacent points should have the same semantic label. Through the neighborhood consistency constraint, isolated false segmentation points are effectively eliminated, and the final refined three-dimensional segmentation result is obtained , where is the final label after the complete filtering process.

[0162] This embodiment verifies the effect of the spatial target recognition method based on dual-mode fusion described in Embodiment One through simulation experiments.

[0163] The present embodiment is evaluated and verified on the NASA-2 / 3D-Part multi-modal satellite part segmentation dataset. NASA-2 / 3D-Part is a multi-modal benchmark dataset for satellite part segmentation based on NASA 3D resources, which contains 109 satellite models such as Active Cavity Radiometer Irradiance Monitor Satellite (AcrimSat), Satélite de Aplicaciones Científicas-D (Aquarius), Cloud Satellite (CloudSat), and Tracking and Data Relay Satellite (TDRS).

[0164] The dataset contains original models and augmented samples, which are generated by splitting three core components of the satellite and reassembling them in various configurations, greatly improving the structural diversity of the models. The data types cover high-quality 3D triangular meshes and surface-sampled 3D point clouds. The 3D point cloud of each model is generated by uniformly sampling 2048 points from the mesh surface using the Trimesh library. The corresponding part ground truth labels are also provided. The label of the 3D point cloud is encoded as the fourth attribute in addition to the spatial coordinates in PLY format, which includes three types of core structural elements: body, solar wing, and antenna.

[0165] Evaluation criteria: The existing segmentation methods and the present embodiment are compared in the 3D point cloud segmentation task. Single-class accuracy (calculated for Body, Antenna, and Wing), overall accuracy (Allacc), and mean intersection over union (mIoU) are used as evaluation indicators. Among them, the existing segmentation methods are 3D zero-shot understanding methods (PointCLIPV2, GeoZe). The evaluation results are shown in Table 1.

[0166] Table 1: Accuracy of space target part segmentation

[0167] In Table 1, Point Clip V2 represents powerful 3D open-world learning through prompt CLIP and GPT. The full name is: PointCLIP V2: Prompting CLIP and GPT for Powerful 3D Open-world Learning; GeoZe represents geometrically-driven aggregation for zero-shot 3D point cloud understanding.

[0168] Table 1 shows the performance comparison results of the method of the present application and the current mainstream zero-shot three-dimensional point cloud understanding method (including PointCLIP V2 and GeoZe) on the NASA-3D-Part dataset. The evaluation indicators include the recognition accuracy of each component - main body, antenna, solar wing, and the overall recognition accuracy and mean intersection over union (mIoU).

[0169] The key performance is as follows: In the main body component segmentation task, the present application achieves a recognition accuracy of 81.72%, which is slightly higher than the 81.21% of the GeoZe method, and significantly better than the 45.31% of the PointCLIP V2 method, indicating that the present application has more accurate recognition ability for satellite core structures.

[0170] In terms of antenna component segmentation, the present application achieves an accuracy of 48.71%, which is significantly higher than 16.47% of GeoZe and 8.11% of PointCLIP V2, effectively alleviating the problem that curved structure components are easily disturbed by noise in the segmentation process.

[0171] In the solar wing component recognition, the present application achieves an accuracy of 70.68%, which is slightly higher than the GeoZe method, and significantly ahead of the 36.12% of PointCLIP V2, verifying the effectiveness of the strategy of the present application to fuse geometric features and semantic information in large-area component recognition.

[0172] In terms of overall performance, the present application performs best in both overall recognition accuracy of 75.78% and mIoU of 68.26%, with an increase of 0.66% and 0.84% respectively compared with GeoZe, and a more significant increase compared with PointCLIP V2, fully embodying the technical advantages of the present application in three-dimensional semantic prior optimization and cross-modal knowledge transfer architecture.

[0173] In summary, by introducing the spatially guided semantic smoothing mechanism and the geometrically aware antenna isolation module, the present application significantly improves the accuracy and robustness of satellite component point cloud segmentation, especially in the recognition task of small structures such as antennas, providing a more advanced technical solution for satellite three-dimensional point cloud segmentation under zero-shot conditions.

[0174] The various embodiments in this specification are described in a progressive manner, and the same or similar parts among the various embodiments can be referred to each other, and each embodiment focuses on the difference from other embodiments. The whole or part of the present application can be used in a variety of general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, mobile communication terminals, multi-processor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like.

[0175] The above examples are only used to illustrate the technical solutions of the present application, and are not limited to the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the present application.

Claims

1. A zero-sample space target micro-component segmentation method based on space constraint guidance, characterized in that, include: The input triangular network model is preprocessed and geometrically segmented to obtain a segmented 3D primitive cluster; wherein the segmented 3D primitive cluster is a geometrically coherent 3D primitive cluster. Candidate regions of small components are extracted from the 3D primitive cluster and pre-segmented to obtain a set of candidate small components on the vertex side of the mesh. The set of candidate small components on the vertex side of the mesh is then mapped onto the 3D point cloud corresponding to the triangular network model to obtain the initial segmentation result of the small components on the point cloud side. The semantic features of the three-dimensional point cloud are enhanced and optimized using a two-dimensional visual language model to generate the initial semantic segmentation result of the three-dimensional point cloud. The initial semantic segmentation result of the 3D point cloud is reclassified into a small component category, and the initial segmentation result of the small components on the point cloud side is overlaid into the reclassified result to obtain an overlay result. The overlay result is then filtered for outliers that are not small components to obtain the final 3D point cloud segmentation result.

2. The zero-sample space target micro-component segmentation method based on spatial constraint guidance according to claim 1, characterized in that, The process of preprocessing and geometrically segmenting the input triangular network model to obtain the segmented 3D primitive clusters includes: Based on the vertex coordinates of each triangle face in the triangular network model, the face normal vector is estimated to obtain the unit normal vector of each triangle face. And based on the face unit normal vectors of all adjacent vertices of the vertex. Perform vertex normal estimation to obtain the vertex normal corresponding to each vertex. ; Based on the topology of the triangular network model, an undirected graph is constructed, and the surface unit normal vector is used as the basis for the graph. and the vertex normal The undirected graph is weighted and the weighted undirected graph is segmented to obtain a segmented 3D primitive cluster.

3. The zero-sample space target micro-component segmentation method based on spatial constraints guided according to claim 2, characterized in that, The construction of the undirected graph based on the topology of the triangular network model includes: Treat all vertices of the triangular network model as an undirected graph. vertex ; The edges formed by pairs of vertices sharing at least one triangular face in the triangular network model are considered as an undirected graph. edge .

4. The zero-sample space target micro-component segmentation method based on spatial constraints guided according to claim 2, characterized in that, The basis of the surface unit normal vector and the vertex normal The undirected graph is weighted and then the weighted undirected graph is segmented to obtain a segmented 3D primitive cluster, including: Determine the edges in the undirected graph The first vertex corresponding to both ends Second vertex and determine the first vertex. The corresponding first vertex normal Second vertex The corresponding second vertex normal ; According to the normal of the first vertex and the normal of the second vertex Calculate the dot product between two vertices. and based on the dot product Calculate the initial edge weight of the corresponding edge. ; According to the first vertex Second vertex Calculate the unit direction vector Calculate the second vertex Corresponding vertex normal In the unit direction vector Projection on and according to the projection Determine the edge The convexity includes: concave transitions and convex transitions; The initial edge weight corresponding to the edge that is a convex transition in the undirected graph. Perform a square operation on the initial edge weight corresponding to the edge with a concave transition in the undirected graph. Without performing any operations, the final edge weight corresponding to each edge in the undirected graph is obtained; A threshold-based undirected graph partitioning algorithm is adopted to merge vertices in ascending order of edge weights and dynamically increase the merging threshold to obtain a preliminary primitive cluster. The initial primitive cluster is merged to obtain the segmented 3D primitive cluster.

5. The zero-sample space target micro-component segmentation method based on spatial constraints guided according to claim 4, characterized in that, The threshold-based undirected graph partitioning algorithm merges vertices in ascending order of edge weights and dynamically increases the merging threshold to obtain a preliminary primitive cluster, including: Each vertex of the undirected graph is treated as an independent primitive cluster, and the segmentation threshold is controlled according to the parameters. Set a corresponding initial internal threshold for each individual primitive cluster. ; Sort the edge weights of the undirected graph in ascending order, traverse the sorted edges, and perform the following operations: Determine the first primitive cluster and the second primitive cluster corresponding to the two vertices of the current edge, and determine the first initial internal threshold and the second initial internal threshold corresponding to the first primitive cluster and the second primitive cluster, respectively; Based on the judgment conditions, the first primitive cluster and the second primitive cluster are merged to obtain a new primitive cluster; wherein, the judgment conditions include: judging whether the first primitive cluster and the second primitive cluster belong to different primitive clusters, and judging whether the edge weight corresponding to the current edge is less than or equal to the first initial internal threshold and the second initial internal threshold. Based on the new primitive cluster and the edge weight corresponding to the current edge, the internal threshold corresponding to the new primitive cluster is adaptively calculated to obtain the preliminary primitive cluster.

6. The zero-sample space target micro-component segmentation method based on spatial constraint guidance according to claim 4, characterized in that, The step of merging the initial primitive clusters to obtain the segmented 3D primitive clusters includes: Set the minimum component size parameter; Traverse the edge set of the undirected graph and perform the following operation: Determine the primitive clusters to which the two vertices connected by the current edge belong. If the two primitive clusters are different, and at least one of the primitive clusters contains fewer vertices than the minimum component size parameter, then merge the two primitive clusters into a new primitive cluster. Repeatedly traverse the edge set and perform the above merging operation until there are no more primitive clusters that meet the merging conditions; The resulting primitive cluster is used as the segmented 3D primitive cluster.

7. The zero-sample space target micro-component segmentation method based on spatial constraint guidance according to claim 1, characterized in that, The process involves extracting candidate regions for micro-components from the 3D primitive cluster and pre-segmenting them to obtain a candidate set of micro-components at the mesh vertex side, including: Calculate the normalized mean normal, convergence, and normal variance of each primitive cluster in the 3D primitive cluster; Based on the convergence and normal variance of each primitive cluster, a corresponding micro-component score is constructed, and multiple thresholds are set to filter the primitive clusters to obtain the first filter set. The first filter set is further filtered using prior structural knowledge to obtain a candidate set of tiny components on the vertex side of the mesh.

8. The zero-sample space target micro-component segmentation method based on spatial constraint guidance according to claim 1, characterized in that, The step of mapping the candidate set of small components on the vertex side of the mesh onto the 3D point cloud corresponding to the triangular network model to obtain the initial segmentation result of the small components on the point cloud side includes: Construct a KD tree spatial index for the candidate set of small parts on the vertex side of the mesh, calculate the average nearest neighbor distance of the 3D point cloud, and define an initial matching tolerance based on the average nearest neighbor distance; Using the KD-tree spatial index, perform the following nearest neighbor query for each point in the 3D point cloud; wherein, the nearest neighbor query includes: Calculate the distance between the current point cloud and the nearest neighbor grid vertex, and include points whose distance is less than the initial matching tolerance into the matching point set; Calculate the matching coverage rate of the matching point set. If the matching coverage rate is lower than a first threshold, reset the first threshold and perform a nearest neighbor query to obtain the final matching point set. Add tiny component labels to the point cloud in the final matching point set to obtain the initial segmentation result of the tiny components on the point cloud side.

9. The zero-sample space target micro-component segmentation method based on spatial constraint guidance according to claim 2, characterized in that, The step of enhancing and optimizing the semantic features of the 3D point cloud using a 2D visual language model to generate an initial semantic segmentation result for the 3D point cloud includes: For each category in the satellite component category set, a diverse set of text descriptions is generated using a large language model, and the set of text descriptions is input into a text encoder to obtain text features; The 3D point cloud is projected onto a multi-view image set, and visual features are extracted from the multi-view image set using a visual encoder. The text features and the visual features are projected onto each 3D point of the 3D point cloud to obtain the initial semantic features of each point; Based on the initial semantic features and combined features A cost matrix is ​​constructed to measure the relationship between points and primitives. Based on this cost matrix, the segmented 3D primitive cluster is optimized and updated. Hierarchical aggregation operations involving local smoothing and global propagation are then performed on each primitive in the updated segmented 3D primitive cluster to obtain refined visual features. The combined features... It is a descriptor of each point in the point cloud where the vertices of the triangular network model coincide. It is obtained by combining the vertex normal; The dot product of the refined visual features and the text features is calculated, and the initial segmentation log probability of each point belonging to each category is obtained by SoftMax normalization, thereby obtaining the initial semantic segmentation result of the 3D point cloud.

10. The zero-sample space target micro-component segmentation method based on spatial constraint guidance according to claim 1, characterized in that, The process of reclassifying the initial semantic segmentation results of the 3D point cloud into minor component categories includes: For each point in the initial segmentation result that is marked as a tiny component, perform the following operations: Retrieve the current tiny component point in the point cloud using a KD-tree. The nearest neighbor points constitute the nearest neighbor set. ; Statistics on the nearest neighbor set The frequency of each point's category is used to determine the most frequent category within its neighborhood. And update the category of the current micro-part point itself to the majority category. ; The process continues until the loop ends, at which point the reset result is obtained.

Images