A method, system, storage medium, and product for marine vessel surveillance based on density knowledge mining and quantitative sensing synthetic aperture radar remote sensing images.

Abstract

This invention discloses a method, system, storage medium, and product for monitoring marine vessels using quantity-aware synthetic aperture radar (SAR) remote sensing images based on density knowledge mining. Belonging to the field of computer vision and remote sensing image processing technology, it addresses the shortcomings of existing technologies in adapting to complex marine backgrounds, handling scenarios with extremely uneven vessel numbers, and limiting model generalization ability due to scarce samples in high-density scenarios. This invention acquires SAR remote sensing images to be detected; constructs and trains a marine vessel monitoring model; and inputs the SAR remote sensing images to be detected into the trained marine vessel monitoring model to obtain quantity-aware monitoring results of marine vessels. This invention is used for target monitoring and analysis applications of SAR remote sensing images.

Description

Technical Field

[0001] A method, system, storage medium, and product for monitoring marine vessels using quantity-aware synthetic aperture radar (SAR) remote sensing images based on density knowledge mining are disclosed. This invention relates to the fields of computer vision and remote sensing image processing technology and is used for target monitoring and analysis of SAR remote sensing images. Background Technology

[0002] Marine vessel surveillance has significant applications in transportation, international trade, fisheries production, and national defense, and is one of the key technologies in marine information acquisition and management. Compared with optical and infrared remote sensing sensors, satellite synthetic aperture radar (SAR) has all-weather, day-and-night imaging capabilities, is not limited by cloud cover or lighting conditions, and can stably acquire high-resolution image data in complex marine environments, thus possessing significant advantages in marine vessel surveillance missions. Marine vessel surveillance based on satellite SAR data has become an important research area in remote sensing and computer vision.

[0003] While existing SAR methods for marine vessel surveillance have made some progress in target detection accuracy, feature representation capabilities, and computational efficiency, they still largely rely on data-driven feature learning mechanisms and fail to fully consider the critical issue of significant imbalances in the number and spatial density of vessels within a single SAR image. In real-world marine scenarios, the spatial distribution of vessels is highly uneven, and these differences in number and density significantly increase the difficulty of surveillance tasks. However, existing methods typically employ uniform feature extraction, candidate generation, and query strategies, making it difficult to adapt to different density scenarios, thus leading to decreased surveillance performance in high-density or extreme environments. Existing methods still face the following technical challenges when detecting densely distributed ship groups: 1. Insufficient adaptability to complex marine backgrounds: Early SAR marine vessel surveillance methods were mainly based on classic signal processing techniques such as constant false alarm rate detection. These methods set detection thresholds by modeling the statistical distribution of sea clutter in order to distinguish vessel targets from the background. However, the marine environment is highly dynamic, and changes in sea state and unstable clutter distribution make accurate modeling difficult. In diverse scenarios, statistical modeling-based methods often struggle to maintain stable performance, and detection results are easily affected by environmental changes. 2. Insufficient ability to handle scenarios with extremely uneven ship numbers: With the development of remote sensing big data, deep learning methods have been widely introduced into SAR marine ship surveillance tasks. Existing technologies such as Faster R-CNN, YOLO, DETR, and their improved models have been successfully applied to ship detection and have achieved good performance on multiple public datasets. These methods typically improve detection accuracy through multi-scale feature fusion, feature decoupling, global context modeling, or Transformer architecture. However, these methods mainly focus on feature representation, scale changes, or computational efficiency, without explicitly modeling or utilizing information on the number and density distribution of ships in the image. In scenarios with highly dense or extremely varied ship numbers, the models lack the ability to perceive the number or density of ships, easily leading to missed or false detections. 3. The scarcity of high-density scene samples limits the generalization ability of the model: In the existing public SAR ship datasets, most samples contain only a small number of ships, and the proportion of high-density ship scenes is low. Although high-density scenes are more difficult to monitor, the relevant training samples are relatively scarce. Existing methods usually rely on large-scale data for feature learning. When the model fails to have the ability to perceive the number or density of ships, it is difficult to effectively mine potential structural information in high-density scenes, which leads to a decline in monitoring performance. This problem caused by the uneven distribution of data has not been systematically solved in the existing technology. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, storage medium, and product for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining, thereby addressing the problems of insufficient adaptability of existing technologies to complex marine backgrounds; insufficient processing capabilities for scenarios with extremely uneven vessel numbers; and limited model generalization ability due to the scarcity of samples in high-density scenarios.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining includes the following steps: Step 1: Acquire the synthetic aperture radar remote sensing image to be detected; Step 2: Construct and train a marine vessel surveillance model. Input the synthetic aperture radar remote sensing image to be detected into the trained marine vessel surveillance model to obtain the number of marine vessels perceived and monitored. The marine vessel surveillance model includes a density knowledge mining module that mines density knowledge from the synthetic aperture radar remote sensing image to be detected to obtain a density map; a density knowledge representation module that performs binarization and connected region merging on the density map to generate a deterministic logical binary mask; and a density knowledge representation module that performs fuzzy feature extraction and fuzzy logic processing on the density map to obtain an uncertain fuzzy attention map; and a four-level density knowledge guidance module that, based on the deterministic logical binary mask and the uncertain fuzzy attention map, realizes the perception and surveillance of the number of marine vessels in the synthetic aperture radar remote sensing image to be detected from four levels.

[0006] Furthermore, in step 2, the density knowledge mining module comprises four sequentially connected stages, each stage comprising sequentially connected... Feature extraction module and hypergraph topology sensing module.

[0007] Furthermore, the aforementioned The feature extraction module includes three sequentially connected modules. Module; The first The module includes a first normalization layer, a multi-head self-attention layer, a second normalization layer that adds the input of the first normalization layer and the output of the multi-head self-attention layer as input, and a first multilayer perceptron, which are connected in sequence. The multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the first normalization layer into four non-overlapping windows. Multi-head self-attention is applied to each window, and a visual window merging process is performed. The second The module includes a third normalization layer that adds the input of the second normalization layer and the output of the first multilayer perceptron as input, a vertically shifted multi-head self-attention layer, a fourth normalization layer that adds the input of the third normalization layer and the output of the vertically shifted multi-head self-attention layer as input, and a second multilayer perceptron, which are connected in sequence. The vertical shift multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the third normalization layer into four non-overlapping windows, and performs vertical translation partitioning. Specifically, a cyclic left shift operation is applied to the window along the horizontal direction, with a shift range of half the window size. This moves the original left half of the window area to the rightmost side, and recombines it with the original rightmost half of the window area to form a new window. Then, a masked multi-head self-attention is used. The new window uses an explicit mask to shield attention calculations from different areas. The third The module includes a fifth normalization layer (which adds the input of the fourth normalization layer and the output of the second multilayer perceptron as input), a horizontally shifted multi-head self-attention layer, a sixth normalization layer (which adds the input of the fifth normalization layer and the output of the horizontally shifted multi-head self-attention layer as input), and a third multilayer perceptron (which adds the input of the sixth normalization layer and the output of the vertically shifted multi-head self-attention layer as input). Features, i.e., obtaining feature maps; The horizontally shifted multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the third normalization layer into four non-overlapping windows, and performs horizontal translation partitioning. Specifically, an additional local cyclic translation is introduced in the horizontal direction, that is, the upper half of the window located on the top side and the lower half of the window located on the bottom side are locally cyclically translated in a vertical translation manner.

[0008] Furthermore, the specific implementation steps of the hypergraph topology sensing module are as follows: Step 2.1, based on the corresponding stage The feature map obtained by the feature extraction module is expanded into vertex features in a hypergraph, resulting in a vertex set describing the density space. Global average pooling and global max pooling are then applied to the vertex features in this set, and the resulting sets are concatenated to form a global vertex set. , is represented as: in, Indicates global average pooling. Indicates global max pooling. Representing vertex features, Indicates splicing, Length is real-valued vectors, For channel dimensions; Step 2.2: Generate hyperedge prototypes based on the connections between pairwise vertex features, adaptively adjust the hyperedge prototypes, calculate the participation relationship between vertex features and hyperedge prototypes, and construct the vertex-hyperedge participation matrix. Specifically, this is done through a projection layer. For global vertices Project to obtain the global offset And on this basis, a global prototype is introduced. This forms a dynamic hyperedge prototype. , is represented as: in, Represents the number of superedges; Introduce another projection layer Vertex features Mapped to query vector : in, This indicates a query for the projection weight matrix; For query vector conduct Subspace decomposition yields , Indicates the first Channel characteristics in each subspace Indicates the first The first query vector decomposition yields the... Subspace; simultaneously, hyperedge prototype The same subspace processing is also performed to obtain... , Indicates the first The first hyperedge prototype decomposition yields the... Subspace; No. In each subspace, the th subspace is calculated by scaling the dot product. Query vector of vertex features With the A hyperedge prototype The similarity between the subspaces is calculated; simultaneously, the overall similarity is obtained by averaging the similarities across all subspaces. The formula is: Finally, regarding similarity Normalization is performed to obtain the contribution weights of vertex features to the hyperedge prototype. The vertex-hyperedge participation matrix is ​​composed of contribution weights. Construction, contribution weight The formula is: in, Represents the number of vertex features; Step 2.3: Perform hypergraph convolution operation based on vertex-hyperedge participation matrix, that is, aggregate vertex features through hyperedge prototype to realize vertex feature projection, and then backpropagate the aggregated hyperedge-level features to vertex features to realize bidirectional information propagation between vertices and hyperedges; Hypergraph convolution is represented as: in, Indicates the first The first in the layer Feature vectors of vertex features Indicates the first The first in the layer Feature vectors of vertex features This is the weight matrix. It is a hyperedge prototype diagonal matrix. For activation function, Vertex features For the hyper-edge prototype Contribution weight; The above formula can be simplified to the following matrix form: in, For hypergraph topological evolution operators, through the incidence matrix and its transpose This enables bidirectional transfer and aggregation of features between vertex features and hyperedge prototypes. This indicates the burden of proof for overstaffing. Indicates the first The hierarchical feature matrix is ​​used to carry the semantic information of ships under the current level after spatial evolution; Step 2.4: Normalize the results obtained from hypergraph convolution using symmetric normalization to output topological enhancement features containing higher-order topological relationships. The symmetric normalization formula is: in, The vertex degree matrix is ​​a diagonal matrix with diagonal elements. Representing the The degree of a vertex feature. The hyperedge degree matrix is ​​a diagonal matrix, where each element on the diagonal is a hyperedge degree matrix. Representing the The degree of the prototype of the hyperedge; Normalized and contribution weights After addition, the results of the corresponding stage are output through the nonlocal module. The nonlocal module follows the hypergraph convolution operation and is used to make up for the limitation of hypergraph convolution, which focuses on the logical topological relationship between discrete nodes and ignores the physical spatial continuity of the whole graph, by establishing a global semantic calibration mechanism.

[0009] Furthermore, in step 2, the specific implementation steps of the density knowledge representation module are as follows: Step 3.1: Calculate the local density accumulation value on the density map generated by hypergraph convolution using a sliding window method to generate a binary density mask. Specifically, the sliding window size is matched with the average target size in the training data. For each position, the cumulative value of pixel intensity within the sliding window is calculated and compared with a predefined density threshold. Sliding windows with a density lower than the density threshold are set to zero, while the remaining sliding windows are set to 1. Connected component decomposition is performed on the sliding windows set to 1. Each connected component is traversed in turn. If the area of ​​the connected component is less than the preset minimum connected component area threshold, the connected component is deleted. Otherwise, the connected component is retained. Finally, a binary density mask is obtained, which is the deterministic logical binary mask. Step 3.2: Extract fuzzy features from the density map using fuzzy logic and generate an uncertain fuzzy attention map, specifically: First, the density map is processed using a... The convolution operator projects it from a high-dimensional topological space to a single-channel physical space, generating a density response matrix; Subsequently, the density response matrix is ​​input into the fuzzy layer and the Gaussian membership function in fuzzy logic is used to map the continuous response values ​​in the density response moment to the membership degree within the fuzzy set. The fuzzy layer maps continuous response values ​​to membership degrees within a fuzzy set using a Gaussian membership function, as shown in the formula: in, The number of membership functions. Indicates the first The membership degree of a Gaussian membership function and Represent its center and radius, respectively. The density response matrix represents the first... The first spatial sampling point The eigenvector values, i.e., the response values ​​in the density response moments; Subsequently, the membership degree of the fuzzy layer output is then... The activation response is used to approximate the "AND" fuzzy logic, and the formula is: in, and These represent the learned weights and biases in the AND fuzzy logic, respectively. The negative exponential mapping mechanism is used to achieve feature fuzzification. It is from the above The membership vector formed, i.e. The first in spatial sampling points , It is the final fuzzy logic output value, representing the confidence level of the validity of the sampled point as a hypergraph node or participating in the construction of hyperedges; Finally, the result after fuzzy logic fusion is subjected to bilinear interpolation in the spatial dimension to align it with the density map scale, and an uncertain fuzzy attention map is output.

[0010] Furthermore, the specific implementation steps of the four-level density knowledge guidance module are as follows: Step 4.1: Execute input-level density knowledge guidance: Multiply the synthetic aperture radar remote sensing image to be detected with a deterministic logical binary mask to obtain densely clustered ship regions. Then, based on these densely clustered ship regions, multiple sub-regions are obtained, denoted as... And Gaussian scaling is introduced to adjust each sub-region; The specific steps for tailoring based on densely clustered ship regions are as follows: Isolate all active pixels in the densely clustered areas of ships, i.e., pixels with an isolation mask value of 1; After isolation, spatially adjacent pixels are clustered into connected regions; By extracting rectangular clippings using the bounding boxes of each connected region, we obtain the ship detection boxes corresponding to multiple sub-regions. A sub-region refers to the region corresponding to a connected region. The specific steps for adjusting each sub-region using Gaussian scaling are as follows: Traverse each sub-region and obtain scaling control parameters by random sampling based on preset Gaussian distribution parameters; The scaling factor is calculated based on the scaling control parameters, using the following formula: in, Indicates the scaling factor. Represents the original intensity of the density response matrix. Indicates scaling control parameters; The scale of the current sub-region is adjusted based on the scaling factor, and it is determined whether the adjusted sub-region is larger than the maximum value of the preset size range. If so, the sub-region is cropped based on the maximum value. Then, it is determined whether the adjusted sub-region is smaller than the minimum value of the preset range. If so, the adjusted sub-region is filled with the synthetic aperture radar remote sensing image to be detected based on the minimum value. Otherwise, the adjusted sub-region is obtained directly. Step 4.2: Perform feature-level density knowledge guidance: The adjusted sub-region and the synthetic aperture radar remote sensing image to be detected are fed into a backbone network with shared weights for feature extraction and mapping into a feature pyramid network. During the mapping process, the downsampling loss of the backbone network is compensated by coordinate scaling to ensure that the local detail features and global semantic features overlap perfectly in the spatial dimension. The features of each layer of the feature pyramid network are fused with the details of the sub-region and the background of the synthetic aperture radar remote sensing image to be detected. The backbone network includes ResNet. The uncertain fuzzy attention map is aligned with the spatial resolution of the feature pyramid network through multi-scale scaling. The aligned uncertain fuzzy attention map is then fused with the feature pyramid network in a weighted manner. After cross-scale information compensation through two top-down upsampling paths, a density pyramid network with global density perception and multi-scale discrimination capabilities is constructed. The multi-scale scaling refers to using a bilinear interpolation algorithm to perform proportional spatial resampling of the uncertain fuzzy attention map according to the downsampling ratio of each level of the feature pyramid network to generate an attention weight map that is aligned pixel by pixel with the spatial resolution of the feature pyramid, ensuring the consistency of spatial position of fuzzy knowledge during cross-scale fusion. Using the density pyramid network as a spatial weight operator, element-level multiplication mapping is performed on the features in the feature pyramid network layer by layer to achieve recalibration of the feature response. After calibration, a density attention feature pyramid with global density distribution awareness is generated through the density attention feature pyramid network. The density attention feature pyramid network consists of a self-attention layer, a spatial attention layer, and a channel attention layer connected in sequence; For the features obtained from each layer of the density attention feature pyramid, the self-attention layer uses three different learnable linear projections to obtain the corresponding query, key, and value representations, namely: in, This represents the features of each layer in the feature pyramid network. The embedding parameter matrix represents the features of each layer of the density pyramid network. They represent Query representation , key representation AND value representation Weights; Embedding parameter matrix They represent The query represents , key representation AND value representation The weights; The spatial attention layer consists of two linear attention operations and a convolutional spatial attention operation. The results of the two convolutional spatial attention operations are then added channel by channel to the results of the corresponding linear attention operations. Linear attention operations involve cross-branch interaction between the two obtained query representations. After the interaction, the two key representations are processed separately. After function processing, the result is multiplied by the corresponding value representation, and then multiplied channel-by-channel by the interactive query representation to obtain the linear attention operation result, as shown in the formula: in, Indicates the dimensions of the query vector and key vector after linear projection; Convolutional spatial attention operations involve performing depthwise separable convolutions on the two-valued representations and then performing the corresponding convolutions on the interactive query representation. Multiplication yields the result of the convolutional spatial attention operation, as shown in the formula: in, express Multiply; The channel attention layer concatenates the two channel-wise sums obtained from the spatial attention layer, then processes them sequentially through the channel shuffling convolution and compression activation modules, and adds the processing result to the channel concatenation result to obtain the final feature map, i.e., the density attention feature pyramid. The compression process of the compression excitation module is as follows: in, and For learnable weights used in adaptive compression and activation features, This represents the calculated channel attention scalar vector. The representation features are compressed into a reduced-dimensional space in the first fully connected layer to reduce the number of channels from Dimensional reduction ,Can The feature tensor to be processed that enters the compression excitation module, That is, the linear rectifier function, which is responsible for introducing nonlinear eigenmaps; The final feature map is as follows: ; Step 4.3: Perform candidate-level density knowledge guidance: Spatially align the feature maps of each layer of the density attention feature pyramid with the density response matrix; The candidate region generation network generates the candidate box with the highest density in the density map. The number of candidates at each location is And the position with the lowest density The number of candidates at position 0 is set to 0. Based on a linear mechanism, the number of candidates at position 1 is 0. Number of candidates for each position The calculation is as follows: in, Indicates the first Density at each location; Step 4.4: Perform query-level density knowledge guidance: Map candidate boxes to the uncertain fuzzy attention map and the aligned density attention feature pyramid. After mapping, the ship quantity level is estimated based on the query quantity estimation network and the mapped uncertain fuzzy attention map. The classification score vector is obtained based on the classification branch and the mapped density attention feature pyramid. The number of marine ships in the synthetic aperture radar remote sensing image to be detected is then perceived and monitored.

[0011] Furthermore, the specific steps of step 4.4 are as follows: Step 4.41: Guide the generated candidate boxes with candidate-level density knowledge through... The operation is projected onto the density attention feature pyramid, and the feature subsets corresponding to the candidate boxes are extracted to form a feature database; Step 4.42: Flatten the feature database to obtain the initial query, which also serves as the key and value in the decoding process; Step 4.43: Process the mapped uncertain fuzzy attention map using a query quantity estimation network to obtain dynamic query data. Calculate the classification score vector of each candidate box in the mapped density attention feature pyramid using a classification branch. The operators are sorted from highest to lowest based on the foreground confidence in the classification score vector, and selected from the initial query based on dynamic query data. The most valuable query is used to obtain the final query, which is then used as the final decoded query. in, Indicates by the first An adaptive query set consisting of query vectors, that is, estimating the output level value of the network based on the number of queries. from Before the middle cut The feature vector with the highest score, Calculated by the query volume estimation network. Let represent the classification score vector derived from the class confidence prediction. express Operator; The query quantity estimation network includes a global pooling layer and a multilayer perceptron connected in sequence. The global pooling layer is used to extract macroscopic statistical information of the whole-image density features in the uncertain fuzzy attention map, while the multilayer perceptron triggers specific query data from a preset quantity level based on the macroscopic statistical information through confidence mapping. The classification branch includes at least two convolutional modules connected in series and a 1×1 prediction convolutional layer for class probability mapping. The convolutional module includes a 3×3 convolutional layer, a normalization layer, and a non-linear activation layer connected in sequence. Step 4.44 The decoder's feedforward network and multi-head self-attention complete the final query decoding and ocean vessel search, that is, to perceive and monitor the number of ocean vessels in the synthetic aperture radar remote sensing image to be detected. The calculation process is as follows: in, Indicates the number of queries The first in Each gear It includes four quantity levels, each corresponding to a specific number of queries. This setting is based on data statistical analysis and experimental experience. Represents query vector, Represents value vector, Indicates the transpose of the key vector. This indicates the results of multi-head self-attention searching for ocean vessel data.

[0012] A quantitative sensing synthetic aperture radar remote sensing image marine vessel surveillance system based on density knowledge mining includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the quantitative sensing synthetic aperture radar remote sensing image marine vessel surveillance method based on density knowledge mining.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for monitoring marine vessels using density-aware synthetic aperture radar remote sensing images based on density knowledge mining.

[0014] A computer program product includes a computer program that, when executed by a processor, implements the steps of a method for monitoring marine vessels using density-aware synthetic aperture radar remote sensing images based on density knowledge mining.

[0015] Compared with the prior art, the advantages of the present invention are as follows: I. This invention addresses the extreme imbalance characteristics of marine vessels in SAR images, characterized by "extremely dense local areas and sparse global distribution." By introducing a quantity perception mechanism into the marine vessel surveillance model, it breaks through the limitations of the fixed detection paradigm of traditional algorithms. This mechanism enables the surveillance system to adjust the focus area of ​​feature extraction and the allocation weight of computing resources in real time according to the dynamic changes in vessel density within the scene. This ensures that the model maintains a stable and high-precision identification capability for both dense clusters and isolated targets under complex and variable sea conditions, significantly improving the generalization performance of the surveillance system. II. This invention is achieved through The deep coupling between the feature extraction module and the hypergraph topology perception module systematically uncovers the deep topological relationships between ship targets. By generating deterministic logical binary masks and uncertain fuzzy attention maps, it can not only achieve accurate localization of target clustering areas through hard discrimination, but also effectively characterize the fuzziness of ship density edges through probabilistic modeling of continuous values. This characterization of high-order topological relationships and spatial evolution trends significantly enhances the model's analytical accuracy for nearshore disturbances and complex community arrangements. Third, the four-level density knowledge guidance module in this invention covers a collaborative guidance architecture of input level, feature level, candidate level and query level, and integrates density knowledge throughout the entire life cycle from image preprocessing to target refinement. Among them, the density guidance at the input level and feature level enhances the marine vessel surveillance model's ability to focus on vessel clustering areas, and at the candidate level, it generates density-adaptive candidate boxes based on density information, so that dense areas generate more candidates and sparse areas generate fewer candidates, thereby having quantity perception ability in the pre-detection stage, effectively balancing missed detections and false detections, and improving overall surveillance performance. IV. This invention utilizes a query quantity estimation network to transform global density knowledge into discrete query control signals. Through a dynamic query selection mechanism, it achieves real-time matching between query size and scenario complexity, compared to traditional methods. The redundant fixed query method in the detector is replaced by a solution that provides a sufficient search window to ensure accuracy in dense target scenes and significantly reduces unnecessary computational overhead in sparse scenes. This ensures the accuracy of large-scale SAR image monitoring while significantly improving the utilization efficiency of hardware resources and the system response speed. V. This invention fully utilizes the hypergraph topology sensing mechanism in the hypergraph topology sensing module to mine the spatial correlation between targets. Since typical interferences in SAR images (such as sea clutter, isolated islands and reefs, speckle noise, etc.) usually exhibit random, isolated distributions and lack the community-like topological characteristics of ship targets, this invention, by establishing high-order topological consistency constraints, can effectively identify and eliminate false targets that do not conform to spatial distribution patterns. This feature screening based on topological priors significantly reduces the false alarm rate in near-shore and complex sea conditions, improving the purity of surveillance missions. VI. The density knowledge representation module in this invention transforms density knowledge into a deterministic logical binary mask and an uncertain fuzzy attention map, and introduces it into the training process as supervision information for the intermediate layer. Compared with traditional detection algorithms that rely solely on bounding box regression, this density supervision mechanism provides the network with a more guided feature learning direction, enabling the model to quickly converge to key regions in the early stages of training and avoiding ineffective searches in massive background features. This not only shortens the model training cycle but also enhances the algorithm's transfer learning robustness when facing SAR images from different sensors and with different polarization modes. VII. This invention addresses the detection challenges in marine surveillance caused by the coexistence of "large-scale ships accompanying small ships" and "very large oil tankers and small near-shore boats." This invention employs feature-level density guidance and a four-level collaborative architecture to perform density-weighted compensation on the feature responses of small targets. Utilizing the spatial prior of the density map, the marine vessel surveillance model can effectively identify dense areas of small targets that are easily overlooked by traditional feature extraction networks, mitigating the masking effect of large target features on small targets in traditional algorithms. Through cross-level transfer of density knowledge, the model ensures extremely high capture sensitivity for ship targets across the entire scale range. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the framework structure of the present invention; Figure 3 This is a schematic diagram of the cyclic shifting operation in the vertical and horizontal directions in this invention; Figure 4 The three cascaded structures in this invention Module structure diagram; Figure 5 This is a schematic diagram of the network structure of the hypergraph topology sensing module in this invention; Figure 6 This is a schematic diagram of the density mask generation method in this invention; Figure 7 This is a schematic diagram of the Gaussian scale adjustment module in this invention; Figure 8 A schematic diagram of the cross-density guided fusion structure of a single layer of features in a density attention feature pyramid; Figure 9 This is a comparison of the accuracy of the experimental results of the present invention and existing technologies on the SSDD and HRSID datasets; Figure 10 This is a schematic diagram of the structure of a computer device according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] A method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining includes the following steps: Step 1: Obtain the synthetic aperture radar remote sensing image to be detected, such as inputting a synthetic aperture radar remote sensing image from the database. ,in, and These represent the height and width of the image, respectively. Step 2: Construct and train a marine vessel surveillance model. Input the synthetic aperture radar remote sensing image to be detected into the trained marine vessel surveillance model to obtain the number of marine vessels perceived and monitored. like Figure 2 As shown, the marine vessel surveillance model includes a density knowledge mining module that mines density knowledge from the synthetic aperture radar remote sensing image to be detected to obtain a density map; a density knowledge representation module that performs binarization and connected region merging on the density map to generate a deterministic logical binary mask; and a density knowledge representation module that performs fuzzy feature extraction and fuzzy logic processing on the density map to obtain an uncertain fuzzy attention map; and a four-level density knowledge guidance module that, based on the deterministic logical binary mask and the uncertain fuzzy attention map, realizes the perception and surveillance of the number of marine vessels in the synthetic aperture radar remote sensing image to be detected from four levels.

[0020] The density knowledge mining module consists of four interconnected stages, each stage comprising a series of interconnected components. Feature extraction module and hypergraph topology sensing module.

[0021] The The feature extraction module includes three sequentially connected modules. Module; The first The module includes a first normalization layer, a multi-head self-attention layer, a second normalization layer that adds the input of the first normalization layer and the output of the multi-head self-attention layer as input, and a first multilayer perceptron, which are connected in sequence. The multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the first normalization layer into four non-overlapping windows. Multi-head self-attention is applied to each window, and a visual window merging process is performed. The second The module includes a third normalization layer that adds the input of the second normalization layer and the output of the first multilayer perceptron as input, a vertically shifted multi-head self-attention layer, a fourth normalization layer that adds the input of the third normalization layer and the output of the vertically shifted multi-head self-attention layer as input, and a second multilayer perceptron, which are connected in sequence. The vertically shifted multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the third normalization layer into four non-overlapping windows, and performs vertical translation partitioning. Specifically, a cyclic left shift operation is applied to the window along the horizontal direction, with a shift magnitude of half the window size. This moves the original left half of the window region to the rightmost side, and then recombines it with the original rightmost half of the window region to form a new window. Then, a masked multi-head self-attention layer is used. The new window uses an explicit mask to shield attention calculations from different regions. Figure 3 As shown, regions A and B are moved to the far right and recombined with regions C and D respectively to form new windows. Since A and C, or B and D, are not spatially adjacent in the feature map, their corresponding image patches cannot be directly used to calculate attention. Therefore, attention calculations from A or B with C or D must be masked using an explicit mask to prevent invalid cross-region attention. Cyclic translation effectively mitigates the increased computational complexity introduced by the small sub-windows, reorganizing the sub-windows into a standard window structure to ensure that the number of multi-head self-attention calculations remains at 4.

[0022] The third The module includes a fifth normalization layer (which adds the input of the fourth normalization layer and the output of the second multilayer perceptron as input), a horizontally shifted multi-head self-attention layer, a sixth normalization layer (which adds the input of the fifth normalization layer and the output of the horizontally shifted multi-head self-attention layer as input), and a third multilayer perceptron (which adds the input of the sixth normalization layer and the output of the vertically shifted multi-head self-attention layer as input). Features, i.e., obtaining feature maps; The horizontally shifted multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the third normalization layer into four non-overlapping windows, and performs horizontal translation partitioning. Specifically, an additional local cyclic translation is introduced in the horizontal direction, that is, the upper half of the window located on the top side and the lower half of the window located on the bottom side are locally cyclically translated in a vertical translation manner.

[0023] Improve computational efficiency by leveraging the cyclic translation window mechanism, using the corresponding The module performs feature extraction.

[0024] Among them, variables , , , These represent feature outputs at different depths, while , , This characterizes the intermediate topological states of the multi-head self-attention layer, the vertically shifted multi-head self-attention layer, and the horizontally shifted multi-head self-attention layer under the corresponding self-attention mechanisms; specifically: It's the first one. The input to the module, serving as the starting information for hierarchical feature modeling, carries the original ocean scattering characteristics that have not been processed by the current contiguous block, reflecting the model's initial capture of pixel-level information within a small receptive field; It is a feature The product after processing by the first normalization layer LN and the multi-head self-attention layer (MSA) represents the initial topological association between ship targets in the local sea area established by the model in the initial window partitioning stage. Together with its residual path, it constitutes the intermediate representation of this layer. It's the first one. The output of the module represents the feature mapping that has been passed from the shallow network to the current level and has already acquired preliminary semantic information. It is the "raw material" for the subsequent deep feature evolution. Characterize the second The intermediate state features within a module are input features. The state after modeling through the third normalization layer and the vertical shift multi-head self-attention layer represents the complex topological relationships between ship targets and between targets and near-shore interference objects that are further established within a specific receptive field. It is the core intermediate state for the transformation from isolated pixels to structured target relationships. For the second The final output characteristic of the module is an intermediate state. The finalized features, enhanced by nonlinear projection through a feedforward network and stabilized by residual connections, are simultaneously passed down as drivers for subsequent deeper layers. Then it represents the third The intermediate state features of the module in a higher-order semantic space, through cross-window information interaction, further break the fixed boundary constraints and realize the abstract extraction of features of large-scale ship communities. Then it is the third one. The deep features ultimately produced by the module, which are the result of the evolution of multi-layer window attention and nonlinear mapping, represent the model's high-order understanding of the ship distribution status in complex scenarios. Through the aforementioned variables... The chain-like evolution and hierarchical iteration ensure that the model can fully extract knowledge of ship distribution in synthetic aperture radar images from micro-pixels to macro-density.

[0025] based on The feature extraction results show that the hypergraph topology perception refines features through adaptive hyperedge generation and hyperedge convolution to obtain density knowledge and gain a more comprehensive spatial perception capability. Among them, the adaptive hyperedge generation can dynamically infer the membership probability of vertices in different hyperedges, thereby obtaining stronger computational adaptability. Hyperedge convolution is responsible for updating features. Each hyperedge refines the features of all vertices it connects, generates a hyperedge-level representation through linear projection, and propagates it back to the vertices to complete the representation update.

[0026] like Figure 5 As shown, the specific implementation steps of the hypergraph topology sensing module are as follows: Step 2.1, based on the corresponding stage The feature map obtained by the feature extraction module is expanded into vertex features in a hypergraph, resulting in a vertex set describing the density space. Global average pooling and global max pooling are then applied to the vertex features in this set, and the resulting sets are concatenated to form a global vertex set. , is represented as: in, Indicates global average pooling. Indicates global max pooling. Representing vertex features, Indicates splicing, Length is real-valued vectors, For channel dimensions; Step 2.2: Generate hyperedge prototypes based on the connections between pairwise vertex features, adaptively adjust the hyperedge prototypes, calculate the participation relationship between vertex features and hyperedge prototypes, and construct the vertex-hyperedge participation matrix. Specifically, this is done through a projection layer. For global vertices Project to obtain the global offset And on this basis, a global prototype is introduced. This forms a dynamic hyperedge prototype. , is represented as: in, Represents the number of superedges; Introduce another projection layer Vertex features Mapped to query vector : in, This indicates a query for the projection weight matrix; For query vector conduct Subspace decomposition yields , Indicates the first Channel characteristics in each subspace Indicates the first The first query vector decomposition yields the... Subspace; simultaneously, hyperedge prototype The same subspace processing is also performed to obtain... , Indicates the first The first hyperedge prototype decomposition yields the... Subspace; No. In each subspace, the th subspace is calculated by scaling the dot product. Query vector of vertex features With the A hyperedge prototype The similarity between the subspaces is calculated; simultaneously, the overall similarity is obtained by averaging the similarities across all subspaces. The formula is: Finally, regarding similarity Normalization is performed to obtain the contribution weights of vertex features to the hyperedge prototype. The vertex-hyperedge participation matrix is ​​composed of contribution weights. Construction, contribution weight The formula is: in, Represents the number of vertex features; Step 2.3: Perform hypergraph convolution operation based on vertex-hyperedge participation matrix, that is, aggregate vertex features through hyperedge prototype to realize vertex feature projection, and then backpropagate the aggregated hyperedge-level features to vertex features to realize bidirectional information propagation between vertices and hyperedges; Hypergraph convolution is responsible for updating features. Each hyperedge refines the features of all the vertices it connects to, generates a hyperedge-level representation through linear projection, and then propagates it back to the vertices to complete the representation update, such as... Figure 6 As shown. This bidirectional process ensures efficient modeling of high-order relationships.

[0027] Existing empirical studies have shown that vertices sharing membership in high-weighted hyperedges should exhibit stronger information propagation capabilities because these weighted connections demonstrate higher reliability in knowledge transfer.

[0028] Hypergraph convolution is represented as: in, Indicates the first The first in the layer Feature vectors of vertex features Indicates the first The first in the layer Feature vectors of vertex features This is the weight matrix. It is a hyperedge prototype diagonal matrix. For activation function, Vertex features For the hyper-edge prototype Contribution weight; The above formula can be simplified to the following matrix form: in, For hypergraph topological evolution operators, through the incidence matrix and its transpose This enables bidirectional transfer and aggregation of features between vertex features and hyperedge prototypes. This indicates the burden of proof for overstaffing. Indicates the first The hierarchical feature matrix is ​​used to carry the semantic information of ships under the current level after spatial evolution; Step 2.4: Symmetric normalization is applied to normalize the results obtained from hypergraph convolution, outputting topological enhancement features containing higher-order topological relationships. These features are used for subsequent density knowledge modeling and object perception tasks. When hypergraph convolutional layers are repeatedly stacked in a deep architecture, this feature may introduce numerical instability during training, thereby increasing the possibility of gradient explosion or gradient vanishing. Therefore, appropriate normalization must be applied to ensure numerical stability.

[0029] The symmetric normalization formula is: in, The vertex degree matrix is ​​a diagonal matrix with diagonal elements. Representing the The degree of a vertex feature. The hyperedge degree matrix is ​​a diagonal matrix, where each element on the diagonal is a hyperedge degree matrix. Representing the The degree of the prototype of the hyperedge; Normalized and contribution weights After addition, the results of the corresponding stage are output through the nonlocal module. The nonlocal module follows the hypergraph convolution operation and is used to make up for the limitation of hypergraph convolution, which focuses on the logical topological relationship between discrete nodes and ignores the physical spatial continuity of the whole graph, by establishing a global semantic calibration mechanism.

[0030] To represent the density knowledge obtained from feature extraction and hypergraph topology perception, the threshold density map is first binarized and connected regions are merged to generate a density mask. Then, fuzzy feature extraction and fuzzy logic processing are performed on the density map to obtain an uncertain fuzzy attention map. Finally, the two density knowledge representations, deterministic logic binary mask and uncertain fuzzy attention map, are passed to the fourth-level density knowledge guide to improve ship surveillance performance.

[0031] like Figure 6 As shown, the density threshold τ serves as a noise control mechanism for the predicted density map and also constrains the number of ships in each density clipping. A higher threshold introduces irregular boundaries, making edge pixels easier to filter, thus increasing the proportion of clipping regions containing only sparse ships. Furthermore, a minimum connected region threshold is introduced to avoid unreasonable extreme magnification caused by generating excessively small clipping regions. The specific implementation steps of the density knowledge representation module are as follows: Step 3.1: Calculate the local density accumulation value on the density map generated by hypergraph convolution using a sliding window method to generate a binary density mask. Specifically, the sliding window size is matched with the average target size in the training data. For each position, the cumulative value of pixel intensity within the sliding window is calculated and compared with a predefined density threshold. Sliding windows with a density lower than the density threshold are set to zero, while the remaining sliding windows are set to 1. Connected component decomposition is performed on the sliding windows set to 1. Each connected component is traversed in turn. If the area of ​​the connected component is less than the preset minimum connected component area threshold, the connected component is deleted. Otherwise, the connected component is retained. Finally, a binary density mask is obtained, which is the deterministic logical binary mask. Step 3.2: Use fuzzy logic to extract fuzzy features from the density map and generate an uncertain fuzzy attention map to characterize the uncertain regions in the density distribution and supplement the expressive power of the deterministic mask. Because density boundaries often lack clear definition, they are difficult to define precisely even for human perception. Therefore, fuzzy feature extraction and fuzzy logic techniques are used to further process the learned density map to generate an uncertain fuzzy attention map. .

[0032] Specifically: First, the density map is processed using a... The convolution operator projects it from a high-dimensional topological space to a single-channel physical space, generating a density response matrix; Subsequently, the density response matrix is ​​input into the fuzzy layer and the Gaussian membership function in fuzzy logic is used to map the continuous response values ​​in the density response moment to the membership degree within the fuzzy set. The fuzzy layer maps continuous response values ​​to membership degrees within a fuzzy set using a Gaussian membership function, as shown in the formula: in, The number of membership functions. Indicates the first The membership degree of a Gaussian membership function and Represent its center and radius, respectively. The density response matrix represents the first... The first spatial sampling point The eigenvector values, i.e., the response values ​​in the density response moments; Subsequently, the membership degree of the fuzzy layer output is then... The activation response is used to approximate the "AND" fuzzy logic, as shown in the formula: in, and These represent the learned weights and biases in the AND fuzzy logic, respectively. The negative exponential mapping mechanism is used to achieve feature fuzzification. It is from the above The membership vector formed, i.e. The first in spatial sampling points , It is the final fuzzy logic output value, representing the confidence level of the validity of the sampled point as a hypergraph node or participating in the construction of hyperedges; Finally, the result after fuzzy logic fusion is subjected to bilinear interpolation in the spatial dimension to align it with the density map scale, and an uncertain fuzzy attention map is output.

[0033] By utilizing deterministic logical binary masks and uncertain fuzzy attention maps, quantity-aware synthetic aperture radar (SAR) surveillance of marine vessels is achieved at four levels: input-level density knowledge guidance, feature-level density knowledge guidance, candidate-level density knowledge guidance, and query-level density knowledge guidance. The specific implementation steps of the four-level density knowledge guidance module are as follows: Step 4.1: Execute input-level density knowledge guidance: Multiply the synthetic aperture radar remote sensing image to be detected with a deterministic logical binary mask to obtain densely clustered ship regions. Then, based on these densely clustered ship regions, multiple sub-regions are obtained, denoted as... And Gaussian scaling is introduced to adjust each sub-region; The specific steps for tailoring based on densely clustered ship regions are as follows: Isolate all active pixels in the densely clustered areas of ships, i.e., pixels with an isolation mask value of 1; After isolation, spatially adjacent pixels (8-connectivity) are clustered into connected regions; By extracting rectangular clippings using the bounding boxes of each connected region, we obtain the ship detection boxes corresponding to multiple sub-regions. A sub-region refers to the region corresponding to a connected region. The specific steps for adjusting each sub-region using Gaussian scaling are as follows: Traverse each sub-region and obtain scaling control parameters by random sampling based on preset Gaussian distribution parameters; The scaling factor is calculated based on the scaling control parameters, using the following formula: in, Indicates the scaling factor. Represents the original intensity of the density response matrix. Indicates scaling control parameters; The scale of the current sub-region is adjusted based on the scaling factor, and it is determined whether the adjusted sub-region is larger than the maximum value of the preset size range. If so, the sub-region is cropped based on the maximum value. Then, it is determined whether the adjusted sub-region is smaller than the minimum value of the preset range. If so, the adjusted sub-region is filled with the synthetic aperture radar remote sensing image to be detected based on the minimum value to reduce missed detections. Otherwise, the adjusted sub-region is obtained directly.

[0034] Furthermore, directly performing clipping and scaling operations may inadvertently introduce critical dimensional information about the ship, leading to conflicting training results. To address this issue, density clipping adjustment (… In this algorithm, we use a Gaussian scaling distribution to adjust the ship sizes within different scale sub-regions, thereby ensuring that the ships maintain appropriate proportions. The algorithm process is as follows: Figure 7 As shown. Used to scale factor The region Adjust to prediction box . Used to define the region Cut evenly into two regions. Used to scale by a given factor The region Fill in the original detected synthetic aperture radar remote sensing image. An average value suitable for the receptive field of the backbone network will be selected. Then, based on... Determination of standard deviation of medium-sized ship dimensions Therefore, the mean and standard deviation of ships are learned in advance, which can narrow down large areas and expand small areas.

[0035] Step 4.2: Perform feature-level density knowledge guidance: The adjusted sub-region and the synthetic aperture radar remote sensing image to be detected are fed into a backbone network with shared weights for feature extraction and mapping into a feature pyramid network. During the mapping process, the downsampling loss of the backbone network is compensated by coordinate scaling to ensure that the local detail features and global semantic features overlap perfectly in the spatial dimension. The features of each layer of the feature pyramid network are fused with the details of the sub-region and the background of the synthetic aperture radar remote sensing image to be detected. The backbone network includes ResNet. Feature-level density knowledge guides the construction of density pyramid networks based on uncertain fuzzy attention maps. ), and with the feature pyramid network ( This process involves fusing features to achieve feature-level attention modeling, forming a density attention feature pyramid, and constructing the density attention feature pyramid (…). By utilizing the density knowledge obtained from the mining process described earlier, we can achieve... and Alignment of hierarchical levels. For example... Figure 8 As shown, where, express A single level in the middle, express A single level in the output express A single level within. It consists of self-attention, spatial attention, and channel attention.

[0036] The uncertain fuzzy attention map is aligned with the spatial resolution of the feature pyramid network through multi-scale scaling. The aligned uncertain fuzzy attention map is then fused with the feature pyramid network in a weighted manner. After cross-scale information compensation through two top-down upsampling paths, a density pyramid network with global density perception and multi-scale discrimination capabilities is constructed. The multi-scale scaling refers to using a bilinear interpolation algorithm to perform proportional spatial resampling of the uncertain fuzzy attention map according to the downsampling ratio of each level of the feature pyramid network to generate an attention weight map that is aligned pixel by pixel with the spatial resolution of the feature pyramid, ensuring the consistency of spatial position of fuzzy knowledge during cross-scale fusion. Using the density pyramid network as a spatial weight operator, element-level multiplication mapping is performed on the features in the feature pyramid network layer by layer to achieve recalibration of the feature response. After calibration, a density attention feature pyramid with global density distribution awareness is generated through the density attention feature pyramid network. The density attention feature pyramid network consists of a self-attention layer, a spatial attention layer, and a channel attention layer connected in sequence; For the features obtained from each layer of the density attention feature pyramid, the self-attention layer uses three different learnable linear projections to obtain the corresponding query, key, and value representations, namely: in, This represents the features of each layer in the feature pyramid network. The embedding parameter matrix represents the features of each layer of the density pyramid network. They represent Query representation , key representation AND value representation Weights; Embedding parameter matrix They represent The query represents , key representation AND value representation The weights; The spatial attention layer consists of two linear attention operations and a convolutional spatial attention operation. The results of the two convolutional spatial attention operations are then added channel by channel to the results of the corresponding linear attention operations. Linear attention operations involve cross-branch interaction between the two obtained query representations. After the interaction, the two key representations are processed separately. After function processing, the result is multiplied by the corresponding value representation, and then multiplied channel-by-channel by the interactive query representation to obtain the linear attention operation result, as shown in the formula: in, Indicates the dimensions of the query vector and key vector after linear projection; Convolutional spatial attention operations involve performing depthwise separable convolutions on the two-valued representations and then performing the corresponding convolutions on the interactive query representation. Multiplication yields the result of the convolutional spatial attention operation, as shown in the formula: in, express Multiply; The channel attention layer concatenates the channel-wise sums of the two results obtained from the spatial attention layer, then processes them sequentially through a channel-shuffling convolution and a compression activation module. The processed result is then added to the channel-concatenated result to obtain the final feature map, i.e., the density attention feature pyramid. Essentially, the outputs of the two branches are concatenated for channel activation. To break channel dependencies, the channels of the feature map are first shuffled and then processed through a convolutional layer. (Compression excitation module) to activate more important channels while suppressing redundant channels; among them, fully connected ( The layer compresses and excites key channel features, with a compression ratio of [value missing]. ; The compression process of the compression excitation module is as follows: in, and For learnable weights used in adaptive compression and activation features, This represents the calculated channel attention scalar vector. The representation features are compressed into a reduced-dimensional space in the first fully connected layer to reduce the number of channels from Dimensional reduction ,Can The feature tensor to be processed that enters the compression excitation module, That is, the linear rectifier function, which is responsible for introducing nonlinear eigenmaps; The final feature map is as follows: ; Step 4.3: Perform candidate-level density knowledge guidance: Spatially align the feature maps of each layer of the density attention feature pyramid with the density response matrix; The candidate region generation network generates the candidate box with the highest density in the density map. The number of candidates at each location is And the position with the lowest density The number of candidates at position 0 is set to 0. Based on a linear mechanism, the number of candidates at position 1 is 0. Number of candidates for each position The calculation is as follows: in, Indicates the first Density at each location; Step 4.4: Perform query-level density knowledge guidance: Map candidate boxes to the uncertain fuzzy attention map and the aligned density attention feature pyramid. After mapping, the ship quantity level is estimated based on the query quantity estimation network and the mapped uncertain fuzzy attention map. The classification score vector is obtained based on the classification branch and the mapped density attention feature pyramid. The number of marine ships in the synthetic aperture radar remote sensing image to be detected is then perceived and monitored.

[0037] The previously generated candidate boxes are projected onto the density attention feature pyramid feature map through the ROIAlign operation, and feature subsets are extracted to form a feature database. A query quantity estimation network is designed under query-level density knowledge guidance to predict the number of queries selected from the feature database, enabling dynamic query selection and thus enhancing the ability to represent the location information of ship queries. It includes a query quantity estimation network (…). ) and a tool for optimizing from the feature database Dynamic query selection of the selected location (in the query) This module is composed of Feature subset extraction is provided.

[0038] The specific steps are as follows: Step 4.41: Guide the generated candidate boxes with candidate-level density knowledge through... The operation is projected onto the density attention feature pyramid, and the feature subsets corresponding to the candidate boxes are extracted to form a feature database; Step 4.42: Flatten the feature database to obtain the initial query, which also serves as the key and value in the decoding process; Step 4.43: Process the mapped uncertain fuzzy attention map using a query quantity estimation network to obtain dynamic query data. Calculate the classification score vector of each candidate box in the mapped density attention feature pyramid using a classification branch. The operators are sorted from highest to lowest based on the foreground confidence in the classification score vector, and selected from the initial query based on dynamic query data. The most valuable query is used to obtain the final query, which serves as the final decoded query; the formula is: in, Indicates by the first An adaptive query set consisting of query vectors, that is, estimating the output level value of the network based on the number of queries. from Before the middle cut The feature vector with the highest score, Calculated by the query volume estimation network. Let represent the classification score vector derived from the class confidence prediction. express Operator; The query quantity estimation network includes a global pooling layer and a multilayer perceptron connected in sequence. The global pooling layer is used to extract macroscopic statistical information of the whole-image density features in the uncertain fuzzy attention map, while the multilayer perceptron triggers specific query data from a preset quantity level based on the macroscopic statistical information through confidence mapping. The classification branch includes at least two convolutional modules connected in series and a 1×1 prediction convolutional layer for class probability mapping. The convolutional module includes a 3×3 convolutional layer, a normalization layer and a non-linear activation layer connected in sequence. That is, it performs spatial feature mapping on the mapped density attention feature pyramid and extracts the probability information of each feature point (or candidate box) belonging to "ship".

[0039] The query quantity estimation network uses an uncertain fuzzy attention graph through... It is converted into a feature vector, and then passed through a 4-neuron... Layers, and through Activate mapping to , , , The number range, where N>50 is used for ablation analysis, and the corresponding results are... Values ​​include 150, 300, 500, and 800. The value should be much larger than the actual number of ships, because negative queries are also needed. This adaptive query selection mechanism can increase the number of decoding queries when encountering dense ship traffic, thereby reducing missed detections.

[0040] Step 4.44 The decoder's feedforward network and multi-head self-attention complete the final query decoding and ocean vessel search, that is, to perceive and monitor the number of ocean vessels in the synthetic aperture radar remote sensing image to be detected. The calculation process is as follows: in, Indicates the number of queries The first in Each gear Includes four quantity levels , , and The corresponding query counts are 150, 300, 500, and 800 respectively. This setting is based on data statistical analysis and experimental experience. Represents query vector, Represents value vector, Indicates the transpose of the key vector. This indicates the results of multi-head self-attention searching for ocean vessel data.

[0041] like Figure 9 As shown, experiments were conducted using the SSDD and HRSID datasets. SSDD contains 1,160 SAR images and 2,551 ships. HRSID contains 5,604 SAR images and 3,642 ships. The training-test split was consistent with previous reports.

[0042] Comparing the DKMG-Net (Marine Vessel Surveillance Model) of this invention with several mainstream state-of-the-art methods, DKMG-Net achieves superior performance, surpassing the second-best methods with average accuracies of 3.1% and 3.8% on SSDD and HRSID, respectively. It also achieves the highest scores across all evaluation metrics, reaching performance levels unattainable by alternative methods. These advantages undoubtedly stem from the rational and effective utilization of density knowledge at four levels: Input-level density knowledge guidance (i-DKG) enables focusing on high-density regions; Feature-level density knowledge guidance (f-DKG) implements density feature self-attention; Candidate-level density knowledge guidance (p-DKG) generates density-adaptive candidate boxes; and Query-level density knowledge guidance (q-DKG) guides the selection of better queries for density knowledge learning. In summary, these factors collectively establish the competitive advantage of DKMG-Net.

[0043] like Figure 9 As shown, density knowledge guidance at each level is effective. With the gradual introduction of new strategies, detection accuracy also improves progressively. Compared to baseline methods, density knowledge guidance (DKG) achieves an overall average accuracy improvement of approximately 7% on SSDD and HRSID, which is an encouraging result. Overall, region focusing from i-DKG, discriminative features from f-DKG, adaptive differential sparse candidate boxes from p-DKG, and optimal monitoring queries from q-DKG collectively drive the respective accuracy improvements. By quantifying the spatial distribution density of ships, density knowledge can significantly improve ship monitoring performance, thereby enhancing the positioning accuracy in high-density areas. Furthermore, i-DKG supports region focusing to enhance the search capability for small ships, achieving robust and efficient detection results in complex scenarios.

[0044] Based on the same inventive concept, this application also provides a system (or device) for implementing the aforementioned target detection method for sensing, retrieving, and locating synthetic aperture radar remote sensing images. It can be a server or a terminal, and its internal structure diagram can be as follows. Figure 10 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The system's (or device's) database stores target detection data for sensing, retrieving, and locating targets based on synthetic aperture radar (SAR) remote sensing images. The computer device's I / O interfaces facilitate information exchange between the processor and external devices. The computer device's communication interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a sensing, retrieving, and locating detection method for SAR remote sensing images.

[0045] Those skilled in the art will understand that Figure 10 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0046] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0047] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0048] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0049] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

Claims

1. A method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining, characterized in that, Includes the following steps: Step 1: Acquire the synthetic aperture radar remote sensing image to be detected; Step 2: Construct and train a marine vessel surveillance model. Input the synthetic aperture radar remote sensing image to be detected into the trained marine vessel surveillance model to obtain the number of marine vessels perceived and monitored. The marine vessel surveillance model includes a density knowledge mining module that mines density knowledge from the synthetic aperture radar remote sensing image to be detected to obtain a density map; a density knowledge representation module that performs binarization and connected region merging on the density map to generate a deterministic logical binary mask; and a density knowledge representation module that performs fuzzy feature extraction and fuzzy logic processing on the density map to obtain an uncertain fuzzy attention map; and a four-level density knowledge guidance module that, based on the deterministic logical binary mask and the uncertain fuzzy attention map, realizes the perception and surveillance of the number of marine vessels in the synthetic aperture radar remote sensing image to be detected from four levels.

2. The method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining according to claim 1, characterized in that, In step 2, the density knowledge mining module includes four sequentially connected stages, each stage comprising sequentially connected... Feature extraction module and hypergraph topology sensing module.

3. The method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining, as described in claim 5, is characterized in that: The The feature extraction module includes three sequentially connected modules. Module; The first The module includes a first normalization layer, a multi-head self-attention layer, a second normalization layer that adds the input of the first normalization layer and the output of the multi-head self-attention layer as input, and a first multilayer perceptron, which are connected in sequence. The multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the first normalization layer into four non-overlapping windows. Multi-head self-attention is applied to each window, and a visual window merging process is performed. The second The module includes a third normalization layer that adds the input of the second normalization layer and the output of the first multilayer perceptron as input, a vertically shifted multi-head self-attention layer, a fourth normalization layer that adds the input of the third normalization layer and the output of the vertically shifted multi-head self-attention layer as input, and a second multilayer perceptron, which are connected in sequence. The vertical shift multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the third normalization layer into four non-overlapping windows, and performs vertical translation partitioning. Specifically, a cyclic left shift operation is applied to the window along the horizontal direction, with a shift range of half the window size. This moves the original left half of the window area to the rightmost side, and recombines it with the original rightmost half of the window area to form a new window. Then, a masked multi-head self-attention is used. The new window uses an explicit mask to shield attention calculations from different areas. The third The module includes a fifth normalization layer (which adds the input of the fourth normalization layer and the output of the second multilayer perceptron as input), a horizontally shifted multi-head self-attention layer, a sixth normalization layer (which adds the input of the fifth normalization layer and the output of the horizontally shifted multi-head self-attention layer as input), and a third multilayer perceptron (which adds the input of the sixth normalization layer and the output of the vertically shifted multi-head self-attention layer as input). Features, i.e., obtaining feature maps; The horizontally shifted multi-head self-attention layer uses a standard window partitioning method to systematically divide the output of the third normalization layer into four non-overlapping windows, and performs horizontal translation partitioning. Specifically, an additional local cyclic translation is introduced in the horizontal direction, that is, the upper half of the window located on the top side and the lower half of the window located on the bottom side are locally cyclically translated in a vertical translation manner.

4. The method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining according to claim 3, characterized in that, The specific implementation steps of the hypergraph topology sensing module are as follows: Step 2.1, based on the corresponding stage The feature map obtained by the feature extraction module is expanded into vertex features in a hypergraph, resulting in a vertex set describing the density space. Global average pooling and global max pooling are then applied to the vertex features in this set, and the resulting sets are concatenated to form a global vertex set. , is represented as: in, Indicates global average pooling. Indicates global max pooling. Representing vertex features, Indicates splicing, Length is real-valued vectors, For channel dimensions; Step 2.2: Generate hyperedge prototypes based on the connections between pairwise vertex features, adaptively adjust the hyperedge prototypes, calculate the participation relationship between vertex features and hyperedge prototypes, and construct the vertex-hyperedge participation matrix. Specifically, this is done through a projection layer. For global vertices Project to obtain the global offset And on this basis, a global prototype is introduced. This forms a dynamic hyperedge prototype. , is represented as: in, Represents the number of superedges; Introduce another projection layer Vertex features Mapped to query vector : in, This indicates a query for the projection weight matrix; For query vector conduct Subspace decomposition yields , Indicates the first Channel characteristics in each subspace Indicates the first The first query vector decomposition yields the... Subspace; simultaneously, hyperedge prototype The same subspace processing is also performed to obtain... , Indicates the first The first hyperedge prototype decomposition yields the... Subspace; No. In each subspace, the th subspace is calculated by scaling the dot product. Query vector of vertex features With the A hyperedge prototype The similarity between the subspaces is calculated; simultaneously, the overall similarity is obtained by averaging the similarities across all subspaces. The formula is: Finally, regarding similarity Normalization is performed to obtain the contribution weights of vertex features to the hyperedge prototype. The vertex-hyperedge participation matrix is ​​composed of contribution weights. Construction, contribution weight The formula is: in, Represents the number of vertex features; Step 2.3: Perform hypergraph convolution operation based on vertex-hyperedge participation matrix, that is, aggregate vertex features through hyperedge prototype to realize vertex feature projection, and then backpropagate the aggregated hyperedge-level features to vertex features to realize bidirectional information propagation between vertices and hyperedges; Hypergraph convolution is represented as: in, Indicates the first The first in the layer Feature vectors of vertex features Indicates the first The first in the layer Feature vectors of vertex features This is the weight matrix. It is a hyperedge prototype diagonal matrix. For activation function, Vertex features For the hyper-edge prototype Contribution weight; The above formula can be simplified to the following matrix form: in, For hypergraph topological evolution operators, through the incidence matrix and its transpose This enables bidirectional transfer and aggregation of features between vertex features and hyperedge prototypes. This indicates the burden of proof for overstaffing. Indicates the first The hierarchical feature matrix is ​​used to carry the semantic information of ships under the current level after spatial evolution; Step 2.4: Normalize the results obtained from hypergraph convolution using symmetric normalization to output topological enhancement features containing higher-order topological relationships. The symmetric normalization formula is: in, The vertex degree matrix is ​​a diagonal matrix with diagonal elements. Representing the The degree of a vertex feature. The hyperedge degree matrix is ​​a diagonal matrix, where each element on the diagonal is a hyperedge degree matrix. Representing the The degree of the prototype of the hyperedge; Normalized and contribution weights After addition, the results of the corresponding stage are output through the nonlocal module. The nonlocal module follows the hypergraph convolution operation and is used to make up for the limitation of hypergraph convolution, which focuses on the logical topological relationship between discrete nodes and ignores the physical spatial continuity of the whole graph, by establishing a global semantic calibration mechanism.

5. The method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining according to claim 4, characterized in that, In step 2, the specific implementation steps of the density knowledge representation module are as follows: Step 3.1: Calculate the local density accumulation value on the density map generated by hypergraph convolution using a sliding window method to generate a binary density mask. Specifically, the sliding window size is matched with the average target size in the training data. For each position, the cumulative value of pixel intensity within the sliding window is calculated and compared with a predefined density threshold. Sliding windows with a density lower than the density threshold are set to zero, while the remaining sliding windows are set to 1. Connected component decomposition is performed on the sliding windows set to 1. Each connected component is traversed in turn. If the area of ​​the connected component is less than the preset minimum connected component area threshold, the connected component is deleted. Otherwise, the connected component is retained. Finally, a binary density mask is obtained, which is the deterministic logical binary mask. Step 3.2: Extract fuzzy features from the density map using fuzzy logic and generate an uncertain fuzzy attention map, specifically: First, the density map is processed using a... The convolution operator projects it from a high-dimensional topological space to a single-channel physical space, generating a density response matrix; Subsequently, the density response matrix is ​​input into the fuzzy layer and the Gaussian membership function in fuzzy logic is used to map the continuous response values ​​in the density response moment to the membership degree within the fuzzy set. The fuzzy layer maps continuous response values ​​to membership degrees within a fuzzy set using a Gaussian membership function, as shown in the formula: in, The number of membership functions. Indicates the first The membership degree of a Gaussian membership function and Represent its center and radius, respectively. The density response matrix represents the first... The first spatial sampling point The eigenvector values, i.e., the response values ​​in the density response moments; Subsequently, the membership degree of the fuzzy layer output is then... The activation response is used to approximate the "AND" fuzzy logic, and the formula is: in, and These represent the learned weights and biases in the "AND" fuzzy logic, respectively. The negative exponential mapping mechanism is used to achieve feature fuzzification. It is composed of the above The membership vector formed, i.e. The first in spatial sampling points , It is the final fuzzy logic output value, representing the confidence level of the validity of the sampled point as a hypergraph node or participating in the construction of hyperedges; Finally, the result after fuzzy logic fusion is subjected to bilinear interpolation in the spatial dimension to align it with the density map scale, and an uncertain fuzzy attention map is output.

6. The method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining according to claim 5, characterized in that, The specific implementation steps of the four-level density knowledge guidance module are as follows: Step 4.1: Execute input-level density knowledge guidance: Multiply the synthetic aperture radar remote sensing image to be detected with a deterministic logical binary mask to obtain densely clustered ship regions. Then, based on these densely clustered ship regions, multiple sub-regions are obtained, denoted as... And Gaussian scaling is introduced to adjust each sub-region; The specific steps for tailoring based on densely clustered ship regions are as follows: Isolate all active pixels in the densely clustered areas of ships, i.e., pixels with an isolation mask value of 1; After isolation, spatially adjacent pixels are clustered into connected regions; By extracting rectangular clippings from the bounding boxes of each connected region, we obtain the ship detection boxes corresponding to multiple sub-regions. A sub-region refers to the region corresponding to a connected region. The specific steps for adjusting each sub-region using Gaussian scaling are as follows: Traverse each sub-region and obtain scaling control parameters by random sampling based on preset Gaussian distribution parameters; The scaling factor is calculated based on the scaling control parameters, using the following formula: in, Indicates the scaling factor. Represents the original intensity of the density response matrix. Indicates scaling control parameters; The scale of the current sub-region is adjusted based on the scaling factor, and it is determined whether the adjusted sub-region is larger than the maximum value of the preset size range. If so, the sub-region is cropped based on the maximum value. Then, it is determined whether the adjusted sub-region is smaller than the minimum value of the preset range. If so, the adjusted sub-region is filled with the synthetic aperture radar remote sensing image to be detected based on the minimum value. Otherwise, the adjusted sub-region is obtained directly. Step 4.2: Perform feature-level density knowledge guidance: The adjusted sub-region and the synthetic aperture radar remote sensing image to be detected are fed into a backbone network with shared weights for feature extraction and mapping into a feature pyramid network. During the mapping process, the downsampling loss of the backbone network is compensated by coordinate scaling to ensure that the local detail features and global semantic features overlap perfectly in the spatial dimension. The features of each layer of the feature pyramid network are fused with the details of the sub-region and the background of the synthetic aperture radar remote sensing image to be detected. The backbone network includes ResNet. The uncertain fuzzy attention map is aligned with the spatial resolution of the feature pyramid network through multi-scale scaling. The aligned uncertain fuzzy attention map is then fused with the feature pyramid network in a weighted manner. After cross-scale information compensation using two top-down upsampling paths, a density pyramid network with global density perception and multi-scale discrimination capabilities is constructed. The multi-scale scaling refers to using a bilinear interpolation algorithm to perform proportional spatial resampling of the uncertain fuzzy attention map according to the downsampling ratio of each level of the feature pyramid network to generate an attention weight map that is aligned pixel by pixel with the spatial resolution of the feature pyramid, ensuring the consistency of spatial location of fuzzy knowledge during cross-scale fusion. Using the density pyramid network as a spatial weight operator, element-level multiplication mapping is performed on the features in the feature pyramid network layer by layer to achieve recalibration of the feature response. After calibration, a density attention feature pyramid with global density distribution awareness is generated through the density attention feature pyramid network. The density attention feature pyramid network consists of a self-attention layer, a spatial attention layer, and a channel attention layer connected in sequence; The self-attention layer uses three different learnable linear projections to obtain the corresponding query, key, and value representations for the features in each layer of the density attention feature pyramid. in, The features of each layer of the feature pyramid network are represented. The embedding parameter matrix represents the features of each layer of the density pyramid network. They represent Query representation , key representation AND value representation Weights; Embedding parameter matrix They represent The query represents , key representation AND value representation The weights; The spatial attention layer consists of two linear attention operations and a convolutional spatial attention operation. The results of the two convolutional spatial attention operations are then added channel by channel to the results of the corresponding linear attention operations. Linear attention operations involve cross-branch interaction between the two obtained query representations. After the interaction, the two key representations are processed separately. After function processing, the result is multiplied by the corresponding value representation, and then multiplied channel-by-channel by the interactive query representation to obtain the linear attention operation result, as shown in the formula: in, Indicates the dimensions of the query vector and key vector after linear projection; Convolutional spatial attention operations involve performing depthwise separable convolutions on the two-valued representations and then performing the corresponding convolutions on the interactive query representation. Multiplication yields the result of the convolutional spatial attention operation, as shown in the formula: in, express Multiply; The channel attention layer concatenates the two channel-wise sums obtained from the spatial attention layer, then processes them sequentially through the channel shuffling convolution and compression activation modules, and adds the processing result to the channel concatenation result to obtain the final feature map, i.e., the density attention feature pyramid. The compression process of the compression excitation module is as follows: in, and For learnable weights used in adaptive compression and activation features, This represents the calculated channel attention scalar vector. The representation features are compressed into a reduced-dimensional space in the first fully connected layer to reduce the number of channels from Dimensional reduction ,Can The feature tensor to be processed that enters the compression excitation module, That is, the linear rectifier function, which is responsible for introducing nonlinear eigenmaps; The final feature map is as follows: ； Step 4.3: Perform candidate-level density knowledge guidance: Spatially align the feature maps of each layer of the density attention feature pyramid with the density response matrix; The candidate region generation network generates the candidate box with the highest density in the density map. The number of candidates at each location is And the position with the lowest density The number of candidates at position 0 is set to 0. Based on a linear mechanism, the number of candidates at position 1 is 0. Number of candidates for each position The calculation is as follows: in, Indicates the first Density at each location; Step 4.4: Perform query-level density knowledge guidance: Map candidate boxes to the uncertain fuzzy attention map and the aligned density attention feature pyramid. After mapping, the ship quantity level is estimated based on the query quantity estimation network and the mapped uncertain fuzzy attention map. The classification score vector is obtained based on the classification branch and the mapped density attention feature pyramid. The number of marine ships in the synthetic aperture radar remote sensing image to be detected is then perceived and monitored.

7. The method for monitoring marine vessels using quantity-aware synthetic aperture radar remote sensing images based on density knowledge mining according to claim 6, characterized in that, The specific steps of step 4.4 are as follows: Step 4.41: Guide the generated candidate boxes with candidate-level density knowledge through... The operation is projected onto the density attention feature pyramid, and the feature subsets corresponding to the candidate boxes are extracted to form a feature database; Step 4.42: Flatten the feature database to obtain the initial query, which also serves as the key and value in the decoding process; Step 4.43: Process the mapped uncertain fuzzy attention map using a query quantity estimation network to obtain dynamic query data. Calculate the classification score vector of each candidate box in the mapped density attention feature pyramid using a classification branch. The operators are sorted from highest to lowest based on the foreground confidence in the classification score vector, and selected from the initial query based on dynamic query data. The most valuable query is used to obtain the final query, which serves as the final decoded query; the formula is: in, Indicates by the first An adaptive query set consisting of query vectors, that is, estimating the output level value of the network based on the number of queries. from Before the middle cut The feature vector with the highest score, Calculated by the query volume estimation network. Let represent the classification score vector derived from the class confidence prediction. express Operator; The query quantity estimation network includes a global pooling layer and a multilayer perceptron connected in sequence. The global pooling layer is used to extract macroscopic statistical information of the whole-image density features in the uncertain fuzzy attention map, while the multilayer perceptron triggers specific query data from a preset quantity level based on the macroscopic statistical information through confidence mapping. The classification branch includes at least two convolutional modules connected in series and a 1×1 prediction convolutional layer for class probability mapping. The convolutional module includes a 3×3 convolutional layer, a normalization layer, and a non-linear activation layer connected in sequence. Step 4.44 The decoder's feedforward network and multi-head self-attention complete the final query decoding and ocean vessel search, that is, to perceive and monitor the number of ocean vessels in the synthetic aperture radar remote sensing image to be detected. The calculation process is as follows: in, Indicates the number of queries The first in Each gear It includes four quantity levels, each corresponding to a specific number of queries. This setting is based on data statistical analysis and experimental experience. Represents query vector, Represents value vector, Indicates the transpose of the key vector. This indicates the results of multi-head self-attention searching for ocean vessel data.

8. A marine vessel surveillance system based on density knowledge mining using quantity-aware synthetic aperture radar remote sensing images, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

Images