SAR target intelligent identification method and device based on decision support

Abstract

The invention discloses an SAR target intelligent identification method and device based on decision support, and belongs to the technical field of computer vision. The method comprises the following steps: sequentially performing instance segmentation and detection clustering processing on a to-be-identified SAR image to obtain target features of a to-be-identified aircraft target; wherein the target feature comprises a geometric dimension and a topological structure; according to the target features and an expert interpretation knowledge base, constructing a target attribute network used for representing the correlation between the target features and the types; and training and adjusting a preset deep neural network based on the target attribute network, and generating a neural support decision recognition model for recognizing the SAR target. According to the invention, the recognition accuracy based on the interpretable target recognition model can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and apparatus for intelligent SAR target recognition based on decision support. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave imaging radar, is unaffected by weather factors such as light and fog, and has the advantage of all-weather, all-time observation capabilities. With the continuous improvement of data reception and processing capabilities, the amount of SAR data has increased significantly, making accurate and rapid automatic target identification technology a pressing issue in SAR applications.

[0003] In related technologies, common SAR target recognition applications mainly focus on traditional pattern recognition methods. These methods rely heavily on the quantity and quality of data and do not fully combine the advantages of target physical characteristics knowledge and expert interpretation knowledge, resulting in low reliability of recognition results.

[0004] Therefore, there is an urgent need for a decision support-based intelligent SAR target identification method and device to solve the above-mentioned technical problems. Summary of the Invention

[0005] This invention provides a decision support-based intelligent SAR target identification method and apparatus, which can effectively improve the identification accuracy of SAR targets. The technical solution is as follows:

[0006] On the one hand, a decision support-based intelligent target identification method for SAR is provided, the method comprising:

[0007] The SAR image to be identified is sequentially segmented into instances and clustered for detection to obtain the target features of the aircraft target to be identified; wherein, the target features include geometric dimensions and topological structure;

[0008] Based on the target features and the expert judgment knowledge base, a target attribute network is constructed to represent the correlation between the target features and the type.

[0009] The preset deep neural network is trained and adjusted based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets.

[0010] On the other hand, a decision support-based intelligent SAR target identification device is provided, the device comprising:

[0011] The processing module is used to sequentially perform instance segmentation and detection clustering processing on the SAR image to be identified to obtain the target features of the aircraft target to be identified; wherein, the target features include geometric dimensions and topological structure;

[0012] The modeling module is used to construct a target attribute network that represents the correlation between the target features and the type, based on the target features and the expert judgment knowledge base.

[0013] The training module is used to train and adjust a preset deep neural network based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets.

[0014] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing computer programs, and the processor for executing the computer programs stored in the memory to implement the steps of the decision support-based SAR target intelligent identification method described above.

[0015] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the above-described decision support-based SAR target intelligent identification method.

[0016] On the other hand, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the decision support-based SAR target intelligent identification method described above.

[0017] The technical solution provided by this invention can bring at least the following beneficial effects: First, it studies the characteristics of SAR ground targets for different aircraft, focusing on methods for automatically extracting target features such as geometric dimensions and topological structures. Next, it combines expert interpretation logic with attribute induction based on the correlation between target characteristics and target types, forming a target recognition attribute network based on expert knowledge and attribute-category correlation. Finally, it combines a discrimination mechanism based on the target attribute network to construct a decision tree model supported by a deep neural network to achieve interpretable intelligent SAR target recognition. This method deeply couples SAR target characteristic knowledge and the interpretation knowledge of professional interpreters with a data-driven intelligent model, improving the decision transparency of the recognition algorithm, enhancing the model's adaptability, and increasing user trust in the model. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a flowchart of a decision support-based intelligent SAR target identification method provided in an embodiment of the present invention;

[0020] Figure 2 This is a roadmap for constructing a target attribute network based on target features and expert interpretation knowledge, provided by an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the construction of a neural support decision recognition model provided in an embodiment of the present invention;

[0022] Figure 4 This is a diagram showing the SAR target intelligent recognition result provided by an embodiment of the present invention;

[0023] Figure 5 This is a structural diagram of a SAR target intelligent identification device based on decision support provided in an embodiment of the present invention;

[0024] Figure 6 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0025] 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 some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] As mentioned earlier, the decision-making process of existing SAR target intelligent identification algorithms has poor interpretability, making it difficult for users to determine whether the model has made incorrect predictions, thus making it difficult to guarantee the accuracy of the model's predictions.

[0027] Based on this, the concept of this invention is to construct a discrimination logic graph based on expert knowledge and attribute-category correlation, and combine it with a discrimination mechanism based on target attributes to construct a decision tree model supported by a deep neural network to achieve interpretable intelligent SAR target recognition.

[0028] The specific implementation of the above concept is described below.

[0029] Please refer to Figure 1 This invention provides a SAR target intelligent identification method based on decision support, the method comprising:

[0030] Step 100: The SAR image to be identified is sequentially segmented into instances and clustered to obtain the target features of the aircraft target to be identified; wherein, the target features include geometric dimensions and topological structure;

[0031] Step 102: Based on the target features and the expert interpretation knowledge base, construct a target attribute network to represent the correlation between the target features and the type;

[0032] Step 104: Based on the target attribute network, train and adjust the preset deep neural network to generate a neural support decision recognition model for identifying SAR targets.

[0033] In this embodiment of the invention, firstly, the characteristics of automatic extraction methods for target features such as geometric extraction size and topological structure are studied for SAR ground targets of different aircraft. Then, combined with expert interpretation logic, attribute induction is performed based on the correlation between target characteristics and target type, forming a target recognition attribute network based on expert knowledge and attribute-category correlation. Finally, combined with a discrimination mechanism based on the target attribute network, a decision tree model supported by a deep neural network is constructed to achieve interpretable intelligent SAR target recognition. This method deeply couples SAR target characteristic knowledge and the interpretation knowledge of professional interpreters with a data-driven intelligent model, improving the decision transparency of the recognition algorithm, enhancing the model's adaptability, and increasing user trust in the model. Through the research on interpretable target recognition technology combining neural networks and attribute networks, the recognition accuracy based on the interpretable target recognition model reaches over 80%.

[0034] The following description Figure 1 The execution method for each step is shown.

[0035] First, for step 100, the SAR image to be identified is sequentially segmented and clustered to obtain the target features of the aircraft target to be identified.

[0036] In this embodiment of the invention, the target features of the aircraft target to be identified are extracted in the following manner:

[0037] The SAR image is segmented using a multi-layer convolutional neural network with cross-entropy as the loss function and SARBake baseline annotation as the label, to obtain the segmentation result of the aircraft target to be identified in the SAR image.

[0038] The size of the segmentation result is estimated based on the bounding box, and the coordinate information of the estimated result is adjusted using a segmentation mask to obtain the geometric size of the aircraft target to be identified.

[0039] The strong scattering points of the extracted aircraft target to be identified are clustered according to the preset clustering center points, and the clustering results are modeled using a Gaussian mixture model to generate the scattering structure features of the aircraft target to be identified.

[0040] Based on the scattering structure characteristics, a graph network structure is established with the cluster center as the topological node and the spatial relationship between different nodes as the edge;

[0041] The graph network structure is updated based on graph aggregation to obtain the topology of the aircraft target to be identified, which eliminates the spatial relationships between irrelevant nodes.

[0042] Specifically, different types of aircraft targets share common characteristics such as target size and aspect ratio. The geometric characteristics of SAR targets, such as their size, provide strong prior information for target identification. Instance segmentation of the target and shadow regions of SAR targets can extract these geometric features. This embodiment uses the SARBake method, employing a convolutional neural network to segment SAR target instances. Cross-entropy is used as the loss function, and SARBake baseline annotations are used as labels. A multi-layer convolutional network segments the samples. The network output is the probability of the center pixel of the input sample corresponding to (target / shadow / background). This yields the foreground and shadow segmentation results of the SAR target.

[0043] After obtaining the detection and instance segmentation results, the bounding boxes and segmentation masks can be used simultaneously to calculate the aspect ratio of the target ship and obtain its geometric features. First, bounding box information is obtained: this is bounding box extraction. Once a target is detected, the algorithm provides a bounding box, i.e., a rectangular box enclosing the target. The width and height of this box can be used to estimate the target's size. Then, for each target, a detection-segmentation collaborative algorithm is used to obtain mask information. Using the bounding box coordinates, the target's width and length are calculated, and then the aspect ratio is calculated. If the mask obtained by the segmentation algorithm provides more accurate target shape information, consider using the segmentation mask to adjust the bounding box calculation to obtain a more accurate aspect ratio; geometric information can also be intuitively obtained from the mask results.

[0044] After obtaining common target features, this embodiment uses the Harris-Laplace detection algorithm to extract strong scattering points of the aircraft target. Nine cluster centers are used to cluster the scattering points, and then a Gaussian mixture model is used to model the extracted strong scattering points and their pixel intensity distribution, thereby obtaining the scattering structure features of the aircraft target. (To address the sparsity of SAR images, the OMP algorithm is used to estimate the set of reflection center parameters, such as amplitude, geometric coordinates, and frequency dependence. Through block processing and iterative optimization techniques, accurate modeling of multiple scattering centers in large-size SAR images is achieved.)

[0045] Furthermore, after extracting the scattering cluster centers and corresponding scattering features of each aircraft target, each scattering cluster center is defined as a topological node, and the spatial relationship between different nodes is modeled as an edge. A graph neural network is then used to convert the scattering information and spatial relationship into a discriminative object representation.

[0046] Furthermore, in constructing the aircraft target topology, the cluster centers of each scattering point of the aircraft target are first used as node features in the graph neural network. Then, the graph network and aggregated features are updated based on graph aggregation, eliminating irrelevant spatial relationships between nodes to obtain the expected aircraft target scattering topology. After obtaining the topology with aircraft category information, it can be used as a strong physical prior information to assist the network in identifying aircraft types in downstream SAR aircraft target-related tasks. In addition, since the SAR aircraft target topology reflects the aircraft target contour information to some extent, the nose and tail dimensions and wing dimensions of the aircraft target can be obtained using the aircraft target topology and detection results, serving as one of the prior information for determining the aircraft target type and used for subsequent aircraft target attribute information.

[0047] Then, for step 102, based on the target features and the expert interpretation knowledge base, a target attribute network is constructed to characterize the correlation between the target features and the type.

[0048] In this embodiment of the invention, the target attribute network is established in the following manner: the target features are summarized according to their category source and extraction method to obtain various attribute information containing textual descriptions and representation vectors; the attribute information and aircraft categories are summarized based on a preset neural network to obtain the association between attributes and categories; an initial attribute network is established based on an expert judgment knowledge base, with key elements as nodes and expert knowledge as edges; the representation vectors are substituted into the nodes, and the information of the edges is calculated according to the association to generate a visually interpretable target attribute network.

[0049] like Figure 2 As shown, the target features are first categorized into different attributes based on their source and extraction method, and each attribute is explained in text. To allocate decision nodes based on attributes and construct the correlation between different attributes and categories, the different attributes need to be unified into a vector representation form. That is, the attributes categorized from the target characteristics are unified into attribute explanations primarily based on text descriptions and attribute representations primarily based on vector representations. For aircraft topology, this study names the attribute "scattering point topology map." Attribute explanations may include: Y-type, V-type, T-type, double-point type, specific size data, etc., with the scattering point topology map serving as the attribute representation vector.

[0050] In this embodiment of the invention, the following attribute information can be summarized according to different target features: when the target feature is the geometric size, an attribute representation vector corresponding to the geometric size is generated based on the obtained size information; when the target feature is the scattering characteristics of the topology, an attribute representation vector corresponding to the scattering characteristics is generated based on the obtained scattering parameters; when the target feature is the scattering point distribution characteristics of the topology, the scattering point topology map is determined as the attribute representation vector corresponding to the scattering point distribution characteristics.

[0051] In other words, for geometric size features, the length, width, and height dimensions can be obtained from the detection box, target mask, and shadow mask. Attribute representation vectors for each attribute are given according to the attribute category (e.g., aspect ratio: L / W, target size: [L, W, H]). For attribute scattering features, the scatterer type attribute can be determined based on parameters such as frequency correlation α and scattering center length. For scattering point distribution features, the scattering point topology map can be directly used as the attribute representation vector. The attribute representation vectors obtained for each attribute interpretation can be used as labels for that interpreted attribute. A multi-task learning approach is used to train the attribute prediction model, thereby obtaining a unified attribute representation form in the same size feature space.

[0052] Furthermore, the association between attributes and categories is determined through the following process: the target attribute information is input into the association recognition network to determine the correlation strength between the target attribute information and the aircraft target category; if the correct recognition rate of the output target category is greater than a preset threshold, then the target attribute information and the target category are determined to be strongly correlated; if the target category cannot be correctly identified and the correct recognition rate for other categories is less than the threshold, then the target attribute information and the target category are determined to be uncorrelated.

[0053] Specifically, a certain attribute representation of the target is used as input to the network to directly learn the target type. If this attribute cannot distinguish between two target types, it indicates that the two target types have similar attributes. For example, this study feeds the features output by the encoder into a small attribute-category correlation decoding network. The network is trained through supervision between the input features and category labels. The training aims to determine the ease with which the feature correctly identifies the target and distinguishes other target categories under attribute i. For instance, under attribute i (i belongs to I, and there are I attributes in total), if the probability of feature λ correctly identifying class p (p belongs to 1 to n, and there are n categories in total) is close to 1, it proves that this class has a strong correlation with attribute i. Conversely, if feature λ cannot correctly identify class p, but shows a relatively balanced probability on some categories, it proves that these categories likely all possess attribute i simultaneously. Through this training process, the correlation between different categories and different attributes can be obtained.

[0054] Furthermore, the key elements identified from expert knowledge are represented as nodes in an attribute network. Each node is interpreted as a specific attribute meaning, and an attribute representation vector is used as the node's representation. Using the knowledge provided by experts, edges connecting these nodes are established to represent the relationships between attributes and between attributes and target types. The inner product between attribute vectors or the score between an attribute vector and the target type can represent the information of the edges. These edges are then visualized using tools such as Matplotlib and Graphviz to obtain an interpretable target attribute network.

[0055] For step 103, the preset deep neural network is trained and adjusted based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets.

[0056] In this embodiment of the invention, the neural support decision recognition model is generated through the following process: the weights of the pre-trained model are processed according to the agglomerative hierarchical clustering method to obtain a decision tree with a hierarchical structure; the decision tree is embedded into the final linear layer of the preset recognition network to generate an initial decision tree model; the initial decision tree model is adjusted according to the tree supervision loss function and the target attribute network to generate a neural support decision recognition model for recognizing SAR targets.

[0057] like Figure 3 As shown, constructing a neural support decision tree requires initialization using a high-performance pre-trained model, which facilitates rapid convergence and improved final recognition performance in target recognition tasks. Therefore, this study uses the WideResNet28x10 model as the baseline model to train a high-performance pre-trained model for the target recognition task.

[0058] Furthermore, the neural network and decision tree are co-modeled by constructing an induced tree structure. The neural-supported decision tree uses path probabilities for reasoning to tolerate highly uncertain intermediate decisions, constructs a hierarchical structure from pre-trained model weights to reduce overfitting, and is trained with hierarchical loss to significantly improve the learning of high-level decisions. This embodiment employs agglomerative hierarchical clustering to build a hierarchical structure based on model weights, thereby constructing an induced tree, and uses average links to measure the distance between two clusters. Given two clusters C... i and C j Their distance is calculated as follows:

[0059]

[0060] Where n1 and n2 represent clusters C i and C j The number of data points in the middle.

[0061] Next, the decision tree is embedded into the neural network: mimicking the forward pass of a neural network, the root node of the decision tree receives an input vector, and its K child nodes each have their own representative vectors. During inference, the output of each node is the inner product of the input vector and its representative vector. Each child node in the tree can be viewed as a node in the neural network, and the combination of the representative vectors of each node can be considered as the parameter matrix of the neural network. After obtaining the output of the child nodes, softmax is used to probabilize the output, and weighted calculations are performed based on the node probabilities during forward inference. After constructing the embedded network structure, the network can perform inference in this way.

[0062] The model is then fine-tuned. In this embodiment, the model is first fine-tuned using tree-supervised loss, soft inference is performed using soft decision trees, and optimization is performed where there may be "slight redundancy" to achieve the best results.

[0063] L y =CE(D 预测 D 标签 )

[0064] L soft =CE(D nbdt D 标签 )

[0065] L=βL y +γL soft +ωL attr

[0066] The overall loss function L of the network is the class cross-entropy loss L. y With tree supervision loss L soft and attribute representation constraint loss L attr The weighted result. As shown in the formula above, β, γ, and ω are the weights corresponding to the three loss functions. The standard cross-entropy loss can effectively separate each leaf node, but it is not trained to separate each internal node. To correct this, a tree-supervised loss—the cross-entropy loss of the path probability class distribution—is added.

[0067] Tree-supervised loss requires a predefined hierarchical structure. Tree-supervised loss may impede learning speed in the early stages of training, when leaf weights are meaningless. Therefore, the weights of the tree supervision need to be adjusted according to different datasets. L-trees should only be used when the accuracy of the original model is not reproducible. soft Fine-tuning is performed. Unlike hierarchical softmax, the path probability cross-entropy loss L... soft This disproportionately increases the weight of earlier decisions in the hierarchical structure, encouraging accurate high-level decision-making.

[0068] Furthermore, to enable the attribute network to effectively guide the decision tree, this study adds attribute representation constraints, denoted as L.attr After establishing the node correspondence between the induced tree and the attribute network, for two nodes in the same layer of the attribute network (e.g., both being aspect ratio attributes), the angle between their attribute representation vectors represents that the interpretations of these two attributes should be different. Similarly, for two nodes in the induced tree, the vectors in the induced tree also need to be constrained to have a certain degree of difference. Therefore, the constraint of the vector angle is added to the loss function.

[0069] In summary, replacing the final linear layer of the network with a decision tree allows NBDT to use path probabilities for inference, tolerating highly uncertain intermediate decisions. It constructs a hierarchical structure from pre-trained model weights to reduce overfitting, and uses hierarchical loss to significantly improve the learning of high-level decisions. The inference process of the algorithm is shown below:

[0070] First, NBDT uses a neural network backbone to represent each sample; the backbone consists of all neural network layers preceding the final linear layer. Second, this embodiment runs the final fully connected layer as a skewed decision tree. However, classic decision trees cannot recover from errors in the early stages of the hierarchy, and running classic decision trees only on neural features significantly reduces accuracy, exceeding 10% as demonstrated in existing experiments. Therefore, this embodiment proposes three modified decision rules:

[0071] First, the weights of the skewed decision rule are seeded using neural network weights: skewed decision trees only support binary decisions, and each decision uses a hyperplane. In contrast, this embodiment associates the weight vector with each node. Second, the node probability is calculated, i.e., the sub-probability is given by the inner product of softmax. Third, a leaf node is selected using the path probability, and the final category is calculated based on the leaf node.

[0072] The above steps yield a neural support decision recognition model for identifying SAR targets. To improve the model's accuracy, it is necessary to randomly select samples from the dataset to form training and testing sets, and then train and test the model algorithm. Specifically:

[0073] For the training task, the training process includes decision node localization based on the target recognition attribute network, mapping the target attributes to the nodes of the induced tree, and then fine-tuning the model based on the tree supervision loss.

[0074] For reasoning tasks, the reasoning process first uses a pre-trained model to extract features from the input image, then inputs them into an inductive tree to obtain the model's decision sequence, and simultaneously outputs the decision results and the decision process for target recognition, resulting in an interpretable recognition result—the target recognition accuracy of the interpretable model reaches over 90%.

[0075] During the testing phase of the model, M SAR target images were selected from the test samples. The same prediction results were obtained on the SAR target recognition model based on convolutional neural networks and the SAR target recognition model based on neural support decision trees studied in this project.

[0076] This application, through research on interpretable target recognition technology combining neural networks and attribute networks, achieves a recognition accuracy of over 80% based on the interpretable target recognition model. Figure 4 It can be clearly seen that the method of the present invention can effectively identify targets in SAR images.

[0077] Please refer to Figure 5 This invention provides a decision support-based intelligent SAR target identification device, which includes:

[0078] The processing module 500 is used to sequentially perform instance segmentation and detection clustering processing on the SAR image to be identified to obtain the target features of the aircraft target to be identified; wherein, the target features include geometric dimensions and topological structure;

[0079] Modeling module 502 is used to construct a target attribute network that represents the correlation between the target features and the type based on the target features and the expert judgment knowledge base;

[0080] The training module 504 is used to train and adjust a preset deep neural network based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets.

[0081] In this embodiment of the invention, when the processing module 500 performs instance segmentation and detection clustering processing on the SAR image to be identified to obtain the target features of the aircraft target to be identified, it specifically performs the following operations: Instance segmentation of the SAR image is performed using a multi-layer convolutional neural network with cross-entropy as the loss function and SARBake baseline annotation as the label to obtain the segmentation result of the aircraft target to be identified in the SAR image; size estimation is performed on the segmentation result based on the bounding box, and the coordinate information of the estimated result is adjusted using a segmentation mask to obtain the geometric size of the aircraft target to be identified; strong scattering points of the extracted aircraft target to be identified are clustered according to preset cluster center points, and the clustering result is modeled using a Gaussian mixture model to generate the scattering structure features of the aircraft target to be identified; a graph network structure is established based on the scattering structure features, with the cluster center as the topological node and the spatial relationship between different nodes as the edge; the graph network structure is updated based on graph aggregation to obtain the topological structure of the aircraft target to be identified after eliminating the spatial relationship between irrelevant nodes.

[0082] In this embodiment of the invention, when the modeling module 502 constructs a target attribute network to represent the correlation between the target features and types based on the target features and the expert interpretation knowledge base, it specifically performs the following operations: It performs inductive processing on the target features according to the category source and extraction method of the target features to obtain various attribute information containing textual descriptions and representation vectors; it performs inductive processing on the attribute information and aircraft categories based on a preset neural network to obtain the association relationship between attributes and categories; it establishes an initial attribute network with key elements as nodes and expert knowledge as edges based on the expert interpretation knowledge base; it substitutes the representation vectors into the nodes and calculates the edge information according to the association relationship to generate a visually interpretable target attribute network.

[0083] In this embodiment of the invention, the step of summarizing the target features according to the category source and extraction method of the target features to obtain various attribute information containing text descriptions and representation vectors includes: when the target feature is the geometric size, generating an attribute representation vector corresponding to the geometric size based on the obtained size information; when the target feature is the scattering characteristics of the topology, generating an attribute representation vector corresponding to the scattering characteristics based on the obtained scattering parameters; and when the target feature is the scattering point distribution characteristics of the topology, determining the scattering point topology map as the attribute representation vector corresponding to the scattering point distribution characteristics.

[0084] In this embodiment of the invention, the inductive processing of the attribute information and aircraft category based on a preset neural network to obtain the association relationship between the attribute and the category includes: inputting the target attribute information into the association recognition network to determine the correlation strength between the target attribute information and the aircraft target category; if the correct recognition rate of the output target category is greater than a preset threshold, then the target attribute information and the target category are determined to have a strong correlation; if the target category cannot be correctly identified and the correct recognition rate for other categories is less than the threshold, then the target attribute information and the target category are determined to have no correlation.

[0085] In this embodiment of the invention, when the training module 504 performs training and adjustment of a preset deep neural network based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets, it specifically performs the following operations: processing the weights of the pre-trained model according to the agglomerative hierarchical clustering method to obtain a decision tree with a hierarchical structure; embedding the decision tree into the final linear layer of the preset recognition network to generate an initial decision tree model; adjusting the initial decision tree model according to the preset loss function and the target attribute network to generate a neural support decision recognition model for identifying SAR targets; wherein, the loss function includes tree supervision loss, category cross-entropy loss, and attribute representation constraint loss.

[0086] It should be noted that the SAR target intelligent recognition device based on decision support provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the SAR target intelligent recognition device based on decision support provided in the above embodiments and the SAR target intelligent recognition method embodiments based on decision support belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0087] Embodiments of this application also provide a computer device, please refer to... Figure 6 The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, at least one program, code set or instruction set being loaded and executed by the processor to implement the decision support-based SAR target intelligent recognition method provided in the above-described method embodiments.

[0088] The embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the decision support-based SAR target intelligent identification method provided in the above-described method embodiments.

[0089] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform any of the decision support-based SAR target intelligent identification methods described in the above embodiments.

[0090] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0091] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0092] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0093] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A SAR target intelligent identification method based on decision support, characterized in that, The method includes: The SAR image to be identified is sequentially segmented into instances and clustered for detection to obtain the target features of the aircraft target to be identified; wherein, the target features include geometric dimensions and topological structure; Based on the target features and the expert judgment knowledge base, a target attribute network is constructed to represent the correlation between the target features and the type. The preset deep neural network is trained and adjusted based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets.

2. The method as described in claim 1, characterized in that, The SAR image to be identified is sequentially segmented into instances and clustered for detection to obtain the target features of the aircraft target to be identified, including: The SAR image is segmented using a multi-layer convolutional neural network with cross-entropy as the loss function and SARBake baseline annotation as the label, to obtain the segmentation result of the aircraft target to be identified in the SAR image. The size of the segmentation result is estimated based on the bounding box, and the coordinate information of the estimated result is adjusted using a segmentation mask to obtain the geometric size of the aircraft target to be identified. The strong scattering points of the extracted aircraft target to be identified are clustered according to the preset clustering center points, and the clustering results are modeled using a Gaussian mixture model to generate the scattering structure features of the aircraft target to be identified. Based on the scattering structure characteristics, a graph network structure is established with the cluster center as the topological node and the spatial relationship between different nodes as the edge; The graph network structure is updated based on graph aggregation to obtain the topology of the aircraft target to be identified, which eliminates the spatial relationships between irrelevant nodes.

3. The method as described in claim 1, characterized in that, The step of constructing a target attribute network to represent the correlation between the target features and the type based on the target features and the expert judgment knowledge base includes: Based on the category source and extraction method of the target features, the target features are summarized and processed to obtain various attribute information containing textual descriptions and representation vectors; The attribute information and aircraft category are summarized and processed based on a preset neural network to obtain the correlation between attributes and categories; An initial attribute network is established based on the expert judgment knowledge base, with key elements as nodes and expert knowledge as edges; The representation vector is substituted into the node, and the edge information is calculated based on the association relationship to generate a visually interpretable target attribute network.

4. The method as described in claim 3, characterized in that, The target features are then processed inductively based on their category source and extraction method to obtain various attribute information containing textual descriptions and representation vectors, including: When the target feature is the geometric dimension, an attribute representation vector corresponding to the geometric dimension is generated based on the acquired dimension information; When the target feature is the scattering characteristic of the topology, an attribute representation vector corresponding to the scattering characteristic is generated based on the obtained scattering parameters; When the target feature is the scattering point distribution characteristic of the topology, the scattering point topology map is determined as the attribute representation vector corresponding to the scattering point distribution characteristic.

5. The method as described in claim 3, characterized in that, The method of summarizing the attribute information and aircraft category based on a preset neural network to obtain the association relationship between attribute and category includes: The target attribute information is input into the correlation recognition network to determine the correlation strength between the target attribute information and the aircraft target category; If the correct recognition rate of the target category is greater than a preset threshold, then the target attribute information and the target category are determined to be strongly correlated. If the target category cannot be correctly identified and the correct identification rate for the remaining categories is less than the threshold, then the target attribute information and the target category are determined to be unrelated.

6. The method as described in claim 1, characterized in that, The step of training and adjusting a preset deep neural network based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets includes: By processing the weights of the pre-trained model using agglomerative hierarchical clustering, a decision tree with a hierarchical structure is obtained. The decision tree is embedded into the final linear layer of a preset recognition network to generate an initial decision tree model; The initial decision tree model is adjusted according to the preset loss function and the target attribute network to generate a neural support decision recognition model for identifying SAR targets; wherein, the loss function includes tree supervision loss, category cross-entropy loss and attribute representation constraint loss.

7. A SAR target intelligent identification device based on decision support, characterized in that, The device includes: The processing module is used to sequentially perform instance segmentation and detection clustering processing on the SAR image to be identified to obtain the target features of the aircraft target to be identified; wherein, the target features include geometric dimensions and topological structure; The modeling module is used to construct a target attribute network that represents the correlation between the target features and the type, based on the target features and the expert judgment knowledge base. The training module is used to train and adjust a preset deep neural network based on the target attribute network to generate a neural support decision recognition model for identifying SAR targets.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images