Radar clutter region division method based on multi-stage optimization

Abstract

The invention discloses a radar clutter region division method based on multi-stage optimization. The method comprises the following steps: acquiring a pixel-level initial radar clutter classification map; abstracting macroscopic clutter regions in the initial radar clutter classification map as nodes, abstracting an overlapping relation between the macroscopic clutter regions as edges, and constructing a scene map in a map structure form; determining an optimal label of the node by taking the minimum global energy function as an optimization target; and taking the optimal tag as a final tag of the corresponding macroscopic clutter region to obtain a final radar clutter classification map. According to the method, multi-stage processing of graph structuring and global topological optimization is carried out on a pixel-level initial radar clutter classification graph, so that an original classification result graph full of noise can be reconstructed into a radar environment partition graph which is correct in macrostructure and accurate in key boundary and contains high-level semantic information such as a shore-sea junction region, and the method is suitable for a radar environment classification algorithm. The accuracy, robustness and interpretability of radar environment perception are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of radar signal / data processing technology, and particularly relates to a radar clutter region division method based on multi-stage optimization. Background Technology

[0002] As the core of modern detection systems, radar is increasingly widely used in aerospace, maritime navigation, and meteorological monitoring. The performance of a radar system largely depends on the accurate analysis of echoes from complex environments. Radar clutter, as a non-target echo signal, is a key interference factor affecting target detection and environmental perception. Therefore, processing the data received by the radar to generate a clutter classification map that accurately reflects the spatial distribution of different clutter types is crucial for subsequent tasks such as clutter suppression, target tracking, and scene understanding.

[0003] With the development of deep learning, methods such as convolutional neural networks and graph neural networks have been successfully applied to generate pixel-level radar clutter classification maps. These methods can automatically learn complex features from raw radar echo data to achieve accurate classification of various categories such as ground clutter and sea clutter.

[0004] However, existing pixel-level clutter classification methods have fundamental limitations: First, they lack macroscopic structural understanding, resulting in a large number of noise points and fragmented areas in the output image, making it impossible to form a coherent representation of large-scale environments such as land clutter areas and sea clutter areas; second, severe regional fragmentation distorts the real topological relationships, making it difficult to accurately maintain the global consistency of key geographical structures such as the shore-sea boundary, inland areas, and water areas; and third, the output only stays at the low-level clutter type discrimination level, failing to extract higher-level semantic information with greater decision-making value, such as the shore-sea boundary transition zone and the clean and detectable area.

[0005] Therefore, how to start from the noisy, fragmented, and topologically distorted pixel-level classification results to achieve macroscopic structure perception, global topology restoration, and reconstruction of advanced composite semantics, thereby generating a radar environment zoning map with complete structure, accurate topology, and rich semantics, remains a core technical challenge that needs to be solved in the current field of radar environment perception. Summary of the Invention

[0006] The purpose of this invention is to provide a radar clutter region segmentation method based on multi-stage optimization, so as to optimize the initial radar clutter classification map and improve the classification accuracy of the radar clutter classification map.

[0007] This invention adopts the following technical solution: a radar clutter region division method based on multi-stage optimization, comprising the following steps:

[0008] Obtain pixel-level initial radar clutter classification map;

[0009] The macroscopic clutter regions in the initial radar clutter classification map are abstracted into nodes, and the overlapping relationships between macroscopic clutter regions are abstracted into edges, thus constructing a scene graph in the form of a graph structure; wherein, the feature vector of the node includes the geometric features, physical statistical features, and texture features of the macroscopic clutter region;

[0010] Construct a global energy function based on nodes and their corresponding clutter labels, and determine the optimal label of a node with the goal of minimizing the global energy function;

[0011] The optimal label is used as the final label for the corresponding macroscopic clutter region to obtain the final radar clutter classification map.

[0012] The beneficial effects of this invention are: by performing multi-stage processing of the pixel-level initial radar clutter classification map through graph structuring and global topology optimization, this invention can reconstruct the noisy original classification result map into a radar environment partition map with correct macroscopic structure, accurate key boundaries, and high-level semantic information such as the coastal-sea boundary area, which greatly improves the accuracy, robustness and interpretability of radar environment perception. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the overall steps of the method proposed in this invention.

[0014] Figure 2 This is a comparative diagram showing the processing effects of the initial preprocessing module in an embodiment of the present invention;

[0015] Figure 3 This is a schematic diagram of the scene graph construction module in an embodiment of the present invention;

[0016] Figure 4 This is a schematic diagram illustrating the processing effect of the critical boundary geometry flow refining module in an embodiment of the present invention;

[0017] Figure 5 This is a schematic diagram of the high-level semantic region generated by the final partition map layered synthesis module in this embodiment of the invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this invention provides a radar clutter region partitioning method based on multi-stage optimization. The method receives a pixel-level initial radar clutter classification map and clutter multi-dimensional feature information corresponding to the clutter map as input, and reconstructs it into a final partition map with complete structure, correct topology and rich semantics through a series of multi-stage optimization steps.

[0020] Specifically, the present invention provides a radar clutter region segmentation method based on multi-stage optimization, comprising the following steps: obtaining a pixel-level initial radar clutter classification map; abstracting the macroscopic clutter regions in the initial radar clutter classification map into nodes and the overlapping relationships between macroscopic clutter regions into edges, constructing a scene graph in graph structure form; wherein, the feature vector of a node includes the geometric features, physical statistical features, and texture features of the macroscopic clutter region; constructing a global energy function based on the node and its corresponding clutter label, and determining the optimal label of the node with the minimum of the global energy function as the optimization objective; using the optimal label as the final label of the corresponding macroscopic clutter region to obtain the final radar clutter classification map.

[0021] This invention, through multi-stage processing of pixel-level classification maps, including preprocessing, graph structuring, global topology optimization, and local boundary refinement, can reconstruct a noisy original classification result map into a radar environment zoning map with correct macroscopic structure, accurate key boundaries, and high-level semantic information such as the coastal-sea boundary area. This greatly improves the accuracy, robustness, and interpretability of radar environment perception.

[0022] Step 1: Preprocessing of the initial clutter classification map.

[0023] This section transforms the initial clutter classification map, filled with pixel-level noise and fragmented regions, into a simplified map composed of several macroscopic, stable regions, providing stable and clean input for subsequent structured analysis. In other words, before constructing the scene map in graph structure form, the process includes: performing noise filtering based on connected component analysis on the initial radar clutter classification map; and performing morphological closure operations on the macroscopic regions in the noise-filtered initial radar clutter classification map to obtain the macroscopic clutter regions.

[0024] (1.1) Isolated noise area.

[0025] This step aims to filter out isolated noise points or tiny artifacts that are physically insignificant and consist of only a very small number of pixels. (See reference...) Figure 2 From the initial observation map ( Figure 2 (Left) to the initial denoised image ( Figure 2 The transformation process (in the original text) involves performing noise filtering based on connected component analysis on the initial clutter classification map of the input. The process is as follows:

[0026] (1.1.1) Definition and Partition: Let M be a set of pixels whose binary mask represents a specific clutter category to be processed (in this invention, the category is the label of the clutter). Using the connected component labeling algorithm, set M is divided into N disjoint subsets C1, C2…C… N ,Right now:

[0027]

[0028] In the formula, each subset C i (i.e., connected components) are all independent connected components, i.e., C i Any two points within C can be connected by a line that is completely contained within C. i The paths within are connected, i∈N, j∈N, and φ represents the empty set.

[0029] (1.1.2) Area calculation and filtering: Define each connected region C i The area of ​​a set is the number of pixels it contains, which is the cardinality S(C) of its set. i )=|C i |

[0030] Set an area threshold T a (In this embodiment, T is set) a =50 pixels). This threshold is used to filter all connected regions, retaining all regions with an area not less than T. a The region is used to form a new filtered set M. filtered Its definition is:

[0031]

[0032] The final set M filtered This refers to the effective region of this clutter category after all isolated noise has been removed.

[0033] (1.2) Macro-regional integration and filling:

[0034] Even after removing isolated noise, the image still contains macroscopic regions that should belong to the same whole but are fragmented into multiple neighboring pieces, or misclassified holes within these regions. This step addresses this structural incompleteness issue. (Refer to...) Figure 2 From the initial denoised image ( Figure 2 From the middle to the final simplified diagram ( Figure 2 The conversion process (right) performs morphological closure operations for clutter categories such as ground clutter and sea clutter, which are usually physically presented as large continuous areas.

[0035] Operating principle: By combining expansion followed by erosion in a sequential manner, adjacent areas that have been broken due to noise are connected, and tiny pores within the macroscopic area are filled. The expansion process widens the boundary of the area, bridging the separated areas; the subsequent erosion process shrinks the boundary, so that the integrated area maintains connectivity while roughly restoring its original size and outline.

[0036] Definition and Implementation: The morphological closure operation operates on a binary mask A corresponding to a set of binary regions consisting of pixel coordinates, and is defined by a structuring element B, which is also a set of pixel coordinates. Its mathematical expression is:

[0037]

[0038] In the formula, B is a structuring element (a form of mask). This represents a morphological dilation operation. This represents a morphological erosion operation. In this embodiment, to better adapt to the natural, non-rectangular contour of the clutter region, structural element B is selected as an elliptical core with a size of 11×11, so as to maintain its original shape to the greatest extent while achieving effective region integration.

[0039] Step 2: Scene graph construction and feature generation module.

[0040] This section transforms the simplified graph output from step 1, which resides in a two-dimensional image space, into a structured scene graph described in graph theory space. Please refer to [link to relevant documentation]. Figure 3 This figure is a schematic diagram of the scene graph construction module in an embodiment of the present invention, which intuitively shows the separation of macroscopic clutter regions (such as...). Figure 3 The left region (i.e., the region after erosion in step 1) is abstracted into a graph structure composed of nodes and edges (e.g., ...). Figure 3 The process (right). This transformation allows it to move away from pixel-level local computation and instead utilize mature graph optimization algorithms to perform global reasoning on the macroscopic topological relationships of the scene at a higher level.

[0041] (2.1) Graph structure generation.

[0042] This step aims to build the skeleton of the scene graph, namely nodes and edges.

[0043] (2.1.1) Graph Node Generation: In the simplified graph output from step 1, each retained, independent macroscopic clutter region is mapped to a node in the graph. That is, each macroscopic region is abstracted as a node in the graph. This elevates the basic unit of analysis from pixels to regions, which is the foundation for achieving macroscopic semantic understanding.

[0044] (b) Edge Generation: Edges between nodes are used to represent the spatial adjacency relationships between macroscopic regions. To determine the connectivity between nodes, this embodiment employs a neighborhood detection method based on morphological expansion. Specifically, for any two macroscopic regions C... i With C j (i≠j), if the following first condition is satisfied, then in the node v representing them... i With v j Establish an edge between them:

[0045]

[0046] In the formula, C i Let C represent the i-th macroscopic clutter region. jB represents the j-th macroscopic clutter region. adj Indicates C i The structuring element for morphological dilation operations, where φ represents the empty set.

[0047] This condition means that when region C i After the boundary expands outward, if it merges with region C... j If they overlap, they are determined to be adjacent.

[0048] (2.2) Extraction of multimodal feature vectors of nodes.

[0049] To ensure that each node in the scene graph possesses sufficient information to support subsequent classification and optimization, this step extracts and associates a high-dimensional, multimodal feature vector with each node. This feature vector aims to provide a comprehensive quantitative description of the macroscopic region represented by the node from multiple perspectives, including geometric shape, physical properties, and internal texture.

[0050] For the representative region C i node v i Its eigenvector f i This can be represented as the concatenation of multiple sub-feature vectors:

[0051]

[0052] In the formula, These represent geometric features, physical statistical features, and texture feature vectors, respectively. The specific extraction process is as follows:

[0053] (2.2.1) Geometric feature extraction: used to describe region C i The outline and shape, including the following information about the macroscopic clutter region:

[0054] Area S(C) i ), that is, the total number of pixels contained in the region;

[0055] Perimeter P(C) i ), which is the pixel length of the region boundary;

[0056] Solidity Sol(C) i The ratio of the area of ​​a region to the area of ​​its convex hull reflects the regularity of the boundary.

[0057]

[0058] In the formula, H(C) i ) is region C i The set of pixels of the convex hull.

[0059] Eccentricity Ecc(C i), used to describe the degree of extension of the shape, is calculated by the second central moment of the region, and has a value range of [0,1], where 0 is a circle and 1 is a line.

[0060] (2.2.2) Extraction of physical statistical features, including the mean, standard deviation and variance of key physical channels in the macro clutter region.

[0061] These characteristics reflect region C i The inherent physical properties of the internal radar echo signal. This embodiment selects K=4 key physical channels (i.e., amplitude, phase, power, and SNR) from the multidimensional clutter feature information. For each channel k, let its feature map be I. k Extract the channel in region C i The three core statistics include:

[0062] mean The average level reflecting signal strength:

[0063]

[0064] Where p represents the index of a pixel in the macro clutter region.

[0065] Standard deviation Measuring the dispersion of signal strength:

[0066]

[0067] variance The square of the standard deviation.

[0068] Flatten the 12 statistics of K=4 channels to form a 12-dimensional physical feature vector.

[0069] (2.2.3) Texture feature extraction: Texture features include the contrast, dissimilarity, homogeneity, and energy of the gray-level co-occurrence matrix of the amplitude channel in the macroscopic clutter region. These features quantify the region C. i Spatial distribution pattern of internal signal strength. In this embodiment, this feature is based on feature map I of the amplitude channel. amp The calculation yielded the result.

[0070] First, region C i Feature map I of the amplitude channel amp The feature values ​​in the image are normalized to the integer grayscale range of [0, 255] to obtain the grayscale image I′. amp Then, within the two-dimensional bounding box of this region, calculate I′. amp The gray-level co-occurrence matrix. Let the normalized gray-level co-occurrence matrix be G(m,n) (where m and n represent the row and column indices of the gray-level co-occurrence matrix, respectively), and extract four features, including:

[0071] Contrast Ratio:

[0072]

[0073] Dissimilarity:

[0074]

[0075] Homogeneity:

[0076]

[0077] energy:

[0078]

[0079] Finally, the above 4-dimensional geometric features, 12-dimensional physical features, and 4-dimensional texture features are concatenated to form a 20-dimensional feature vector f. i This vector comprehensively describes node v i The various attributes of the macro-region it represents provide sufficient data support for the global topology optimization in step 3.

[0080] Step 3: Global topology optimization module based on regional Markov random fields.

[0081] This step performs a global, context-aware optimization of the scene graph constructed in Step 2. Although the nodes in Step 2 already possess rich features, their initial category labels (derived from Step 1) are based only on local, pixel-level information and may contain errors (for example, a plot of land with vague appearance features might be misclassified as sea clutter). This step corrects the category labels of all nodes by solving a Markov Random Field (MRF) model over the entire scene graph, utilizing the topological adjacency relationships and feature similarities between regions, thereby outputting a more reasonable and accurate optimized graph in terms of macroscopic topology.

[0082] (3.1) Construction of global energy function.

[0083] The core of this step lies in constructing a global energy function E(L) that describes the state of the entire scene graph. The goal of this energy function is to find an optimal class label assignment scheme L = {l} for each node in the node set V of the graph. v |v∈V},l v Let the label of node v be the label that minimizes the total energy E(L). A low-energy configuration indicates that the label allocation scheme achieves the optimal balance between data authenticity and spatial smoothness.

[0084] According to Markov random field theory, the global energy function E(L) can be defined as the sum of two parts:

[0085]

[0086] In the formula, E(L) represents the global energy function based on the label assignment scheme L, V represents the node set, Edges is the set of all edges (i.e., adjacency relationships) in the graph, (u, v) represents the edge between node u and node v, and U(v, l) represents the edge between node u and node v. v This is called a unary potential, which means assigning a label l to a node v. v The cost. This cost depends only on the characteristics of node v itself, P(l u ,l v This is called a binary potential, representing a pair of adjacent nodes u and v, each assigned a label l. u and l v The cost. This cost is used to introduce context constraints. λ is a weighting coefficient used to balance the contributions of data items and smoothing items to the total energy.

[0087] (3.2) Definition and calculation of univariate potential.

[0088] Univariate potential U(v,l) v The multimodal feature vector f of node v was quantized. v (Results from Step 2) and candidate labels v The degree of mismatch between them. In this embodiment, the unary potential is defined by the probability output of the random forest classifier.

[0089] (3.2.1) Classifier Training: First, extract the feature vector X = {f} of all nodes in the scene graph from step 2. v |v∈V} and their initial category labels Use this data to train a random forest classifier.

[0090] (3.2.2) Dynamic Category Determination: To improve robustness and efficiency, optimization is performed only on clutter categories actually existing in the current radar scenario. A clutter category list C is obtained using a trained classifier. active .

[0091] (3.2.3) Probability Calculation and Cost Definition: Input the feature vectors of all nodes back into the trained classifier to obtain the posterior probability Prob(l|f) of each node v belonging to each candidate class l. v The univariate potential is defined as the negative log-likelihood of that probability:

[0092] U(v,l v ) = -log(Prob(l v |f v (14)

[0093] This definition indicates that the higher the probability predicted by the classifier, the lower the corresponding univariate cost. To ensure numerical stability, the case where the probability value is 0 can be replaced by a very small positive number.

[0094] In other words, U(v,l) v The calculation method is as follows: train a classifier using the feature vectors of all nodes in the scene graph and the initial labels; use the trained classifier to obtain the posterior probability of node v belonging to each label, and use the negative log-likelihood of the posterior probability as U(v,l) v ).

[0095] (3.3) Definition and calculation of binary potential.

[0096] Binary potential P(l) u ,l v Spatial topological constraints are introduced. This embodiment uses an improved Pott model based on prior knowledge for definition. Let P be a cost matrix, whose elements P ij Defines when a pair of adjacent nodes are each assigned a label l i and l j The cost of time.

[0097] (3.3.1) Zero penalty for the same category: encourages adjacent areas to be of the same category.

[0098] (3.3.2) Default penalty for different categories: By default, adjacent areas are discouraged from being of different categories.

[0099] P ij =ω default for all i≠j (15)

[0100] In the formula, ω default A higher penalty value (i.e., the first threshold) is applied.

[0101] (3.3.3) Weight Adjustment Based on Prior Knowledge: The adjacency cost between specific categories is adjusted based on the physical priors of radar clutter. For example, the adjacency between ground clutter and sea clutter (i.e., the shore-sea boundary) is reasonable and common in the physical world. Therefore, if both ground clutter and sea clutter are in the category subset of the current scene, their corresponding penalty terms are specifically adjusted:

[0102] P 地,海 =ω coast (16)

[0103] Where, ω coast It is significantly lower than ω default The low penalty value (i.e., the second threshold) allows the model to form class boundaries at reasonable locations (such as between land and sea) at a very small cost while maintaining overall smoothness.

[0104] In other words, P(l) u ,l v The calculation method for label l is as follows: u and tag l v When they are the same, P(l) u ,l v ) is 0; when label l u and tag l v At different times, P(l) u ,l v ) is the first threshold; when label l u and tag l v When it is a preset combination, P(l) u ,l v The second threshold is less than the first threshold.

[0105] (3.4) Global energy minimization reasoning.

[0106] After defining the complete energy function E(L), the goal of this step is to find a globally optimal label allocation scheme L. * This minimizes the total energy.

[0107] L * =argmin L E(L) (17)

[0108] (3.4.1) Robustness handling: If the system determines that the current scene contains only a single clutter category (i.e., fewer than 2 clutter categories), then there is no need for binary optimization, and this step can be skipped, directly using the optimal solution of the unary potential. If the scene graph does not contain any adjacency relationships (i.e., the set of edges is empty), then the binary potential term is zero, and the optimization problem degenerates into solving the minimum value of the unary potential independently for all nodes.

[0109] In other words, when the initial radar clutter classification map contains only a single clutter label or the set of edges in the scene graph is empty, based on U(v,l) v Determine the optimal label for the node.

[0110] (3.4.2) Graph Cut Solution: Under normal circumstances, this combinatorial optimization problem can be solved by an efficient graph cut algorithm to obtain a powerful approximate optimal solution. The algorithm takes the costs of nodes, edges, unary potentials, and binary potentials as inputs and solves iteratively using the minimum cut principle until the energy function converges.

[0111] (3.4.3) Result Mapping: Finally, the algorithm outputs the optimal label set. Generate a blank image of the same size as the initial radar clutter classification image, with data type integer. Then, iterate through all nodes v in the scene image and assign the optimal label to all pixels covered by their corresponding macroscopic regions. This image shows the result of global topology optimization. It corrects the unreasonable region categories in the original image, making the overall scene division more topologically consistent.

[0112] Step 4: Key boundary refinement and semantic partition generation.

[0113] To address the limitations of the topology optimization map output in step 3 regarding boundary accuracy, a final partitioning map containing advanced composite semantics is synthesized through semantic reconstruction. While the output of step 3 ensures the correct categories and topological relationships of macroscopic regions, its boundaries are still constrained by the blocky contours of the macroscopic regions in step 1. This step refines the key boundaries by introducing a geometric model based on partial differential equations and designs a hierarchical synthesis logic, ultimately outputting a radar environment partitioning map that combines macroscopic topological correctness, microscopic boundary accuracy, and rich advanced semantics.

[0114] In one embodiment, after using the optimal label as the final label for the corresponding macroscopic clutter region, the method further includes: constructing a local region of interest based on the boundary to be refined in the radar clutter classification map; constructing an edge stopping function based on the edge information of the amplitude channel of the original radar feature map to optimize the local region of interest; and iteratively evolving the boundary in the optimized local region of interest based on the geodesic active contour model.

[0115] (4.1) Geometric refinement of critical boundaries.

[0116] This section precisely aligns the topological boundaries obtained from graph optimization to the actual physical boundaries that exist in the original radar feature map.

[0117] Figure 4 This is a schematic diagram illustrating the processing effect of the critical boundary geometry flow refining module in an embodiment of the present invention. This embodiment uses the most representative shore-sea boundary (i.e., the boundary between the land clutter region and the sea clutter region) as an example for explanation.

[0118] (4.1.1) Construction of Local Region of Interest (ROI): To significantly improve computational efficiency and avoid solving computationally intensive partial differential equations on the entire image, this step first defines a Local Region of Interest (ROI) around the boundary to be refined (from the output of step 3). This ROI can be constructed by performing morphological dilation on the land and sea binary masks respectively, and then finding the intersection of the dilated regions, forming a region surrounding the initial boundary (e.g., ...). Figure 5 (a)) The surrounding strip-shaped calculation area.

[0119] (4.1.2) Edge Stopping Function (ESF) Construction: The driving force for boundary refinement comes from the edge information in the original radar feature map (amplitude channel). Therefore, an edge stopping function g(|▽I|) needs to be constructed. This function approaches 0 in regions with large image gradients |▽I| (i.e., at physical edges), and approaches 1 in flat regions with smaller gradients. In this embodiment, the original amplitude feature map I is first... amp Gaussian smoothing is performed to suppress noise, and then its gradient is calculated. The edge stopping function g is:

[0120]

[0121] In the formula, is a Gaussian kernel, and k is the gradient sensitivity coefficient.

[0122] (4.1.3) Boundary Evolution Based on Geodesic Active Profile (GAC): This embodiment employs a geodesic active profile model based on the level-set method to iteratively evolve the boundary. This model implicitly represents the boundary curve as the zero level set (i.e., the position where β = 0) of a high-dimensional level-set function β(x,y,t) (x,y are the x and y coordinates of the pixel, and t is the evolution time). The evolution of this function β is controlled by a partial differential equation (PDE):

[0123]

[0124] In the formula, Let β be the evolution rate of β with time t. This is the edge stopping function, which stops the evolution when the curve reaches the physical edge (g≈0). ν is the curvature term of the level set β, used to maintain the smoothness of the boundary curve. ν is a constant force, which can be set to a negative value in this embodiment to apply a continuous inward contraction force, allowing the boundary to adhere to the potential energy trough defined by the edge stopping function g.

[0125] (4.2) Generation of the final semantic partitioning map.

[0126] Based on the aforementioned optimization results, layered fusion is performed according to a preset priority, and new composite semantic regions with advanced application value are generated in the process.

[0127] Figure 5 This is a schematic diagram illustrating the generation of high-level semantic regions by the final partition map layered synthesis module in this embodiment of the invention. The synthesis logic of this embodiment is as follows:

[0128] (4.2.1) Basic partition drawing: Using the topology optimization map output in step 3 as the base map, the meteorological clutter area and the clean area are drawn into the final semantic partition map.

[0129] (4.2.2) Refined partitioning: Using the refined high-precision ground clutter mask obtained in step 1, the ground clutter region it represents is drawn onto the semantic partitioning map. This layer will cover the old land and sea areas with rough boundaries that may exist in the previous step.

[0130] (4.2.3) Semantic generation: The coastal boundary area is generated algorithmically. First, the high-precision boundary line refined in step 1 is obtained. Second, the boundary line is subjected to morphological operation with a preset width to form a boundary zone.

[0131] Next, the boundary ring is intersected with the sea clutter region determined in step 3; finally, this intersection region is defined as a new shore-sea boundary region with high-level semantics and drawn onto the final partition map with the highest priority.

[0132] Through the above layered synthesis, the final output radar environment zoning map not only includes basic clutter zones such as land, sea, and meteorology with complete structure and correct topology, but also provides high-precision shore-sea physical boundaries guaranteed by the geometric flow model, and offers advanced composite semantic information on the shore-sea boundary area, greatly improving the accuracy and application value of radar environment perception. This completes the radar clutter region delineation.

[0133] The present invention also discloses a radar clutter region division method based on multi-stage optimization, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the above-described method when executing the computer program.

[0134] The present invention also discloses an embodiment that provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0135] The present invention also provides a computer program product that, when run on a data storage device, enables the data storage device to implement the steps in the above-described method embodiments.

[0136] If the integrated unit module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0137] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0138] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

Claims

1. A radar clutter region partitioning method based on multi-stage optimization, characterized in that, Includes the following steps: Obtain pixel-level initial radar clutter classification map; The macroscopic clutter regions in the initial radar clutter classification map are abstracted into nodes, and the overlapping relationships between macroscopic clutter regions are abstracted into edges, thus constructing a scene graph in the form of a graph structure; wherein, the feature vector of the node includes the geometric features, physical statistical features, and texture features of the macroscopic clutter region; Construct a global energy function based on nodes and their corresponding clutter labels, and determine the optimal label of the node with the minimum of the global energy function as the optimization objective; The optimal label is used as the final label for the corresponding macroscopic clutter region to obtain the final radar clutter classification map.

2. The radar clutter region division method based on multi-stage optimization as described in claim 1, characterized in that, The global energy function is: Where E(L) represents the global energy function based on the label allocation scheme L, v represents a node, V represents the node set, and l v U(v,l) represents the label of node v. v This indicates that the node v is assigned the label l. v The cost, λ represents the weight coefficient, Edges represents the set of edges in the scene graph, (u, v) represents the edge between node u and node v, P(l u ,l v This indicates that adjacent nodes u and v are each assigned a label l. u and tag l v The cost.

3. The radar clutter region division method based on multi-stage optimization as described in claim 2, characterized in that, U(v,l v The calculation method for ) is as follows: A classifier is trained using the feature vectors of all nodes in the scene graph and the initial labels. The trained classifier is used to obtain the posterior probability of node v belonging to each label, and the negative log-likelihood of the posterior probability is taken as U(v,l). v ).

4. The radar clutter region division method based on multi-stage optimization as described in claim 3, characterized in that, P(l u ,l v The calculation method for ) is as follows: When label l u and tag l v When they are the same, P(l) u ,l v ) is 0; When label l u and tag l v At different times, P(l) u ,l v ) is the first threshold; When label l u and tag l v When it is a preset combination, P(l) u ,l v The second threshold is less than the first threshold.

5. A radar clutter region partitioning method based on multi-stage optimization as described in claim 3 or 4, characterized in that, The geometric features include the area, perimeter, solidity, and eccentricity of the macroscopic clutter region; The physical statistical characteristics include the mean, standard deviation, and variance of the key physical channels in the macroscopic clutter region; wherein, the key physical channels include the amplitude channel, phase channel, power channel, and SNR channel; The texture features include the contrast, dissimilarity, homogeneity, and energy of the gray-level co-occurrence matrix of the amplitude channel in the macroscopic clutter region.

6. The radar clutter region division method based on multi-stage optimization as described in claim 5, characterized in that, When the first condition is met, the overlapping relationship between macroscopic clutter regions is abstracted as an edge; The first condition is: Among them, C i Let C represent the i-th macroscopic clutter region. j This represents the j-th macroscopic clutter region. B represents the morphological dilation operation. adj Indicates C i The structuring element for morphological dilation operations, where φ represents the empty set.

7. The radar clutter region division method based on multi-stage optimization as described in claim 6, characterized in that, Before constructing a scene graph in graph structure form, the following are also included: Noise filtering based on connected component analysis is performed on the initial radar clutter classification map; A morphological closure operation is performed on the macroscopic region in the initial radar clutter classification map after noise filtering to obtain the macroscopic clutter region.

8. A radar clutter region partitioning method based on multi-stage optimization as described in claim 6 or 7, characterized in that, After using the optimal label as the final label for the corresponding macroscopic clutter region, the following is also included: Construct a local region of interest based on the boundary to be refined in the radar clutter classification image; An edge stopping function is constructed based on the edge information of the amplitude channel of the original radar feature map to optimize the local region of interest; The boundary is iteratively evolved based on the geodesic active profile model in the optimized local region of interest.

9. A radar clutter region partitioning method based on multi-stage optimization as described in claim 2, characterized in that, When the initial radar clutter classification map contains only a single clutter label or the set of edges in the scene map is empty, based on U(v,l) v Determine the optimal label for the node.

10. A radar clutter region partitioning method based on multi-stage optimization, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-9.

Images