SAR (Synthetic Aperture Radar) target identification method based on graph neural network and position information fusion

Abstract

The invention relates to an SAR (Synthetic Aperture Radar) target identification method based on a graph neural network and position information fusion. The method comprises the following steps: acquiring a main body pixel point set of a target image of the SAR; wherein each main body pixel point feature is composed of attribute feature information of the main body pixel point and position information of the main body pixel point in the target image. And constructing a position information fused target map structure according to the pixel point vectors. And constructing a collaborative target recognition model based on the graph neural network, wherein the collaborative target recognition model comprises a collaborative graph attention layer and a relative position embedding layer. And respectively inputting the main body pixel point set into a collaborative graph attention layer and a relative position embedding layer to carry out iterative training so as to extract corresponding information features. And adding position information to the relative position embedding layer in a residual structure manner to perform fusion of the attribute information features and the position information features, and outputting fusion features. By adopting the method, the recognition performance of the graph neural network can be efficiently and quickly improved in a lightweight manner.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and particularly to a SAR target recognition method based on the fusion of graph neural network and position information. Background Art

[0002] Synthetic Aperture Radar (SAR) is a coherent imaging radar operating in the microwave band and is an active remote sensing sensor. Compared with other sensors such as optical and infrared sensors, SAR imaging is not restricted by weather, lighting and other conditions, and can observe targets of interest all day and all weather. SAR Automatic Target Recognition (ATR) plays an important role in battlefield reconnaissance, situation awareness, detection and guidance and other fields. Existing image classification networks (convolutional neural network and Transformer network) have achieved excellent results in the field of SAR recognition and have higher recognition accuracy compared with traditional SAR recognition algorithms.

[0003] After investigation, currently SAR is mostly carried on some resource-constrained mobile platforms (edge devices), such as satellites, airplanes, ships and ground vehicles. The energy, processor computing resources and storage space of these platforms are limited. However, the current SAR target recognition models based on deep learning require high computing power and storage space, because:

[0004] (1) These models usually have a large number of network layers and parameters. These parameters need to be learned and optimized during the training process and are used during the inference process, which requires a large amount of storage space. During the image recognition process, forward propagation, backward propagation, and loss calculation all require a large number of matrix multiplications and non-linear activation function calculations, which require a large amount of computing resources. In addition, batch processing further exacerbates the requirements of the model for computing volume and storage. For large networks, these parameters may occupy dozens or even hundreds of megabytes of storage space.

[0005] (2) Image recognition tasks usually involve processing a large number of high-resolution images, which contain a large number of pixels, including target areas and background areas. The value of each pixel needs to be processed by a neural network, which involves a large amount of computing and space storage capabilities.

[0006] Therefore, the current SAR target recognition algorithms based on deep learning are generally carried on data center devices and are difficult to be carried on resource-constrained platforms. If the data is collected and transmitted back to the data center for processing, the delay is long, which requires the device communication ability and is difficult to meet the real-time processing conditions at the edge. Summary of the Invention

[0007] Based on this, in view of the above technical problems, it is necessary to provide a SAR target recognition method based on the fusion of graph neural network and position information, which can be carried on the edge platform and improve the target recognition efficiency.

[0008] A SAR target recognition method based on the fusion of graph neural network and position information, the method comprising:

[0009] Obtain the set of main pixel points of the target image of the SAR. Wherein, each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image.

[0010] Construct a target graph structure with position information fusion according to the pixel point vector.

[0011] Construct a collaborative target recognition model based on the graph neural network, the collaborative target recognition model comprising: a collaborative graph attention layer and a relative position embedding layer.

[0012] Input the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information feature.

[0013] According to the target graph structure, the updated attribute feature information, and the updated position information feature, add a relative position information embedding layer to the collaborative graph attention layer in the form of a residual structure, obtain the fusion feature through aggregation, after the fusion feature passes through the global pooling layer, use the cross-entropy loss function for iterative training of the structural features, and output the target image.

[0014] A SAR target recognition device based on the fusion of graph neural network and position information, the device comprising:

[0015] A main pixel point set acquisition module, configured to acquire the set of main pixel points of the target image of the SAR; wherein, each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image;

[0016] A graph structure construction module, configured to construct a target graph structure with position information fusion according to the pixel point vector.

[0017] A model construction module, configured to construct a collaborative target recognition model based on the graph neural network, the collaborative target recognition model comprising: a collaborative graph attention layer and a relative position embedding layer.

[0018] A lightweight feature optimization module, configured to input the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information feature.

[0019] A feature fusion module, which is used to add a relative position information embedding layer to the collaborative graph attention layer in the form of a residual structure according to the target graph structure, the updated attribute feature information, and the updated position information feature, obtain a fused feature through aggregation, and after the fused feature passes through a global pooling layer, use a cross-entropy loss function to perform iterative training on the structural features and output the target image.

[0020] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0021] Obtain a set of main pixel points of the target image of the SAR. Each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image.

[0022] Construct a target graph structure for position information fusion according to the pixel point vector.

[0023] Construct a collaborative target recognition model based on a graph neural network. The collaborative target recognition model includes: a collaborative graph attention layer and a relative position embedding layer.

[0024] Input the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information feature.

[0025] According to the target graph structure, the updated attribute feature information, and the updated position information feature, add a relative position information embedding layer to the collaborative graph attention layer in the form of a residual structure, obtain a fused feature through aggregation, and after the fused feature passes through a global pooling layer, use a cross-entropy loss function to perform iterative training on the structural features and output the target image.

[0026] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0027] Obtain a set of main pixel points of the target image of the SAR. Each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image.

[0028] Construct a target graph structure for position information fusion according to the pixel point vector.

[0029] Construct a collaborative target recognition model based on a graph neural network. The collaborative target recognition model includes: a collaborative graph attention layer and a relative position embedding layer.

[0030] Input the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information feature.

[0031] According to the target graph structure, the updated attribute feature information, and the updated position information feature, a relative position information embedding layer is added in the collaborative graph attention layer in the form of a residual structure. After obtaining the fused features through aggregation, the fused features pass through the global pooling layer, and then the cross-entropy loss function is used for iterative training of the structural features to output the target image.

[0032] The above SAR target recognition method based on the fusion of graph neural network and position information, first, extracts the set of main pixel points of the SAR target image, including attribute feature information and position information, and constructs a pixel-level target graph structure to combine the feature information with the spatial relationship, thereby enhancing the accurate modeling of the target. Secondly, the collaborative graph attention layer is adopted to use the iterative training mechanism to extract the feature attention scores between the main pixel points and their neighbor pixel points, strengthen the key features in an adaptive manner, and at the same time reduce the influence of irrelevant information on the model calculation. Furthermore, the relative position embedding layer effectively extracts the position information of the pixel points, so that the position information is effectively retained and extracted during the iterative training of the features, improving the network's perception ability of spatial relationships and enhancing the comprehensive expression ability of the local and global features of the target. This solution optimizes the computational overhead and reduces the parameter scale through a lightweight graph neural network architecture. At the same time, combined with an efficient feature learning mechanism, the model can perform efficient calculations on resource-constrained edge devices without relying on the high computing power support of the data center, thereby avoiding the high latency and communication burden brought by data transmission and realizing the real-time recognition ability of SAR targets. Finally, this solution successfully improves the recognition performance of the graph neural network, while meeting the requirements of edge devices for lightweight and high efficiency, providing an innovative solution for SAR target recognition in resource-constrained environments. Brief Description of the Drawings

[0033] Figure 1 It is a flowchart of the SAR target recognition method based on the fusion of graph neural network and position information in one embodiment;

[0034] Figure 2 It is a schematic diagram of the information update of the collaborative graph attention layer in one embodiment;

[0035] Figure 3 It is a schematic diagram of the information update of the relative position embedding layer in one embodiment;

[0036] Figure 4 It is an overall framework diagram of the SAR target recognition network in one embodiment;

[0037] Figure 5 It is a schematic diagram of the confusion matrix in one embodiment;

[0038] Figure 6The structural block diagram of an SAR target recognition device based on the fusion of graph neural network and position information in an embodiment;

[0039] Figure 7 The internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0040] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] In one embodiment, as Figure 1 shown, a SAR target recognition method based on the fusion of graph neural network and position information is provided, including the following steps:

[0042] Step 102, obtaining a set of main pixel points of the target image of the SAR.

[0043] Wherein, each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image.

[0044] Specifically, the Otsu method is used to perform threshold segmentation on the image to extract the main part of the target. For sample N m , m ∈ {1,..., N}, N is the total number of samples, the main part of the target extracted from this sample is defined as the set P = {p1, p2,..., p K}, K is the number of pixel points of the main body extracted from this sample, and each pixel point p i can be defined as p i = (s i , pos i ), where i ∈ {1,..., K}. In this definition, s i is the feature vector describing the attribute of this point, pos i is the position coordinate of this point in the image. Specifically: Among them, is the intensity value of the scattering point at this position; pos i = (x i , y i ), where (x i , y i ) is the position coordinate of this scattering point in the image coordinate system.

[0045] Step 104, constructing a target graph structure with position information fusion according to the pixel point vectors.

[0046] Specifically, the target main body extracted is used to define the graph structure. For sample Nm , define the graph structure as G(N m ) = {V m , E m , POS m}, where V m is composed of a finite and non-empty set of vertices, denoted as V m = {s1, s2, s3,..., s K}, E m represents the set of edges between vertices, denoted as E = {(u, v)|u, v ∈ |1,..., K|}, and the set of vertex positions is denoted as POS = {pos1, pos2,..., pos K}. The construction of edges is based on the adjacency relationship between nodes. Specifically, for node p i , if there exists another point p j such that x i ∈ (x j - 1, x j + 1), y i ∈ (y j - 1, y j + 1), then the adjacency relationship between p i and p j holds, and the edge between p i and p j is connected by an undirected edge. In the graph structure, an undirected edge is usually represented by a bidirectional edge, so the geometric edge set E includes (i, j) and (j, i), indicating that there is an undirected connection between nodes i and j.

[0047] Step 106, construct a collaborative target recognition model based on the graph neural network.

[0048] The collaborative target recognition model includes: a collaborative graph attention layer and a relative position embedding layer.

[0049] Step 108, input the set of main body pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information features.

[0050] Specifically, the current mainstream graph neural networks only focus on vertex information, and the contribution degree of each neighbor vertex to vertex update is the same, ignoring the graph structure information. For this reason, the graph attention network was proposed. A node attention mechanism is introduced in the graph attention layer to assign different weights to neighbor nodes, enabling the model to effectively learn the dependency relationship between nodes, thereby more effectively capturing the local and global structure information of graph data. The update flow chart of the collaborative graph attention layer is as Figure 2As shown in the figure, the attention of neighbor nodes is calculated through the co-attention score calculation formula, and then the aggregated function is used for updating to obtain the updated node features. In the co-graph attention layer, for a certain node p i , there exists a neighbor node p j . First, calculate the attention score of the neighbor node:

[0051]

[0052] where a T is a linear weight vector, both a and W are parameters to be learned, and t is the number of iterations. In a graph neural network, one iteration usually refers to the process of passing through one layer of the graph neural network. and are the node information corresponding to p i and p j at the iteration number t, and || represents vector concatenation. Then, the attention score is normalized through softmax among all neighbors :

[0053]

[0054] where is defined as the set of neighbor nodes including the self-node, and the node feature s i of node p i is calculated by the weighted average of the transformations of its neighbor nodes:

[0055]

[0056] In the co-graph attention layer, the attention score can be calculated through the interaction between the query node feature s i and the neighbor node feature s j to capture richer non-linear relationships between node features, provide a finer adjustment space, and enhance the representation ability of the attention score.

[0057] Step 110: According to the target graph structure, the updated attribute feature information, and the updated position information feature, add a relative position information embedding layer in the co-graph attention layer in the form of a residual structure, obtain the fused feature through aggregation, and after the fused feature passes through the global pooling layer, use the cross-entropy loss function for iterative training of the structural features and output the target image.

[0058] Specifically, to improve the recognition ability of the network, the position information of the nodes is introduced into the feature representation, and a feature extraction layer dedicated to processing the node position information is designed to more fully explore the potential information of the position features and improve the recognition effect.

[0059] The main target graph structure is input into the relative position embedding layer to extract the position information features of the main body. The information aggregation formula of the relative position embedding layer is as follows:

[0060]

[0061] where t is the number of iterations, and the function is used to calculate the features between two nodes. The function ρ(*) is an aggregation operation used to aggregate the features of all nodes connected to node p i . The function g t (*) processes the aggregated neighbor node information and updates the node feature s i of node p i . Among them, the calculation formula of is as follows:

[0062]

[0063] pos i -pos j is the relative position between node p i and node p j , which is used to avoid the problem that the coordinate differences of the same type of targets at different positions and angles are relatively large. To solve the problem of lack of translational invariance and rotational invariance in SAR images, the variance Δpos is introduced. A MLP (Multilayer Perceptron) network is used to predict the position deviation of the node using the feature s i of p i t , which is defined as:

[0064]

[0065] In the final design, two different MLP networks are used to model f t and g t , and the Mean function is used for aggregation to mitigate the information impact caused by the loss of local area nodes, making the network more robust. Finally, we obtain the aggregation formula of this network layer:

[0066]

[0067] The overall flowchart for updating the relative position embedding layer is as shown in Figure 3 . The relative position information is used for the feature representation of the graph structure. The node information of the query node is used to calculate Δpos to offset the image changes caused by the perspective change in the SAR image to a certain extent. Finally, the node information is updated using the position information and the neighbor node information.

[0068] Furthermore, based on the proposed graph neural network layer above, the overall network structure is designed. In this network, the co-attention layer is used to more effectively extract node features, and the position information of nodes is injected into the network in the form of a residual structure. Through the fusion of intensity information and position information, more efficient target recognition performance is achieved. The specific network framework is as follows Figure 4 shown as:

[0069] The input graph structure first passes through a simple GCN (Graph Convolutional Network) layer, and the node features are updated through the graph convolutional layer as:

[0070]

[0071] where \(W_1\) and \(W_2\) are learnable weight matrix parameters. In addition, the aggregated features are normalized (Batch Normalization, BN) and activated by a non-linear activation function (activation layer). The main purpose of the GCN layer is to perform an information aggregation so that the nodes contain the information of one-hop neighbor nodes, with certain structure and local information, to reduce the error of the computational variance \(\Delta_{pos}\) of the first relative position embedding layer. The backbone of the network processes the node information of the graph structure through the co-graph attention layer + BN layer + dropout layer + activation layer. The processing and aggregation of position information are mainly carried out through the relative position embedding layer + activation layer for feature processing and feature extraction, and the position information is injected into the network in the form of a residual structure, enabling the network to dynamically handle the dependencies between node information and position information. Such a structure is looped three times to further extract high-order features. Then, the GMP (Global Max Pooling) layer is used to globally pool the extracted features to extract global information from the node-level features. The output tensor of GMP corresponds to the maximum value in each corresponding dimension of each node in the graph:

[0072]

[0073] where \(d\) is the dimension of node mapping. Finally, the globally pooled features are input into a four-layer MLP to obtain the final classification result

[0074] In the design of the loss function, the cross-entropy loss is used to calculate the loss function:

[0075]

[0076] where \(C\) is the number of sample categories, \(y\) i,c is the true label of the \(i\)-th sample in category \(c\), is the predicted probability.

[0077] In the above SAR target recognition method based on the fusion of graph neural network and position information, first, the set of main pixel points of the SAR target image is extracted, including attribute feature information and position information, and a pixel-level target graph structure is constructed to combine the feature information with the spatial relationship, thereby enhancing the accurate modeling of the target. Secondly, a collaborative graph attention layer is adopted, and an iterative training mechanism is used to extract the feature attention scores between the main pixel points and their neighboring pixel points, strengthen the key features in an adaptive manner, and reduce the influence of irrelevant information on the model calculation. Furthermore, in the relative position embedding layer, the position information of the pixel points is effectively extracted, so that the position information is effectively retained and extracted during the feature iterative training process, improving the network's perception ability of spatial relationships and enhancing the comprehensive expression ability of the local and global features of the target. This solution optimizes the computational overhead and reduces the parameter scale through a lightweight graph neural network architecture, and at the same time combines an efficient feature learning mechanism, enabling the model to perform efficient calculations on resource-constrained edge devices without relying on the high computing power support of the data center, thereby avoiding the high latency and communication burden brought by data transmission and realizing the real-time recognition ability of SAR targets. Finally, this solution successfully improves the recognition performance of the graph neural network, meets the requirements of edge devices for lightweight and high efficiency, and provides an innovative solution for SAR target recognition in resource-constrained environments.

[0078] In one embodiment, the target image captured by SAR is obtained, and the Otsu method is used to perform threshold segmentation on the target image to obtain the set of main pixel points.

[0079] In one embodiment, according to the adjacent position relationship between any nodes in the pixel point vector, if there is an adjacent node of the first node, the first node and the adjacent node are connected by an undirected edge to construct a target graph structure with fused position information:

[0080] G(N m )={V m ,E m ,POS m}

[0081] Wherein, V m is a finite and non-empty vertex set, E m is the edge set between vertices, POS m is the position information set of vertices, N m is a sample of m pixel nodes.

[0082] In one embodiment, in the collaborative graph attention layer, the attention score of the adjacent node of the first node in the pixel node vector is calculated:

[0083]

[0084] Wherein, aT is a linear weight vector, W is a parameter to be learned, t is the number of iterations, is the node p i The node information corresponding to the node p at the iteration number t, is the node p j The node information corresponding to the node p at the iteration number t, || is vector concatenation;

[0085] The attention score is normalized among all adjacent nodes through softmax:

[0086]

[0087] where N(i) is the set of adjacent nodes including the node itself, softmax j mainly performs the normalization operation for the adjacent node p j is, α i,j is, e(p i , p j ) is the attention score of the adjacent node p of the first node j of the first node, w and j are both adjacent nodes, w ∈ N(i), e(p i , p w ) is the attention score of the adjacent node p w for the first node, mainly sums the attention scores for each adjacent node of the first node.

[0088] After performing weighted averaging on the normalized adjacent nodes, the node feature of the first node is obtained:

[0089]

[0090] where, is the node feature of the first node pi at the iteration number t + 1, α i,w is the normalized attention score of the adjacent node p w is, is the node feature of the adjacent node p w at the iteration number t.

[0091] In one embodiment, the relative position information is used to describe the node feature of the first node:

[0092]

[0093] where, is the aggregated node feature, and three explicit mapping networks are respectively used to complete the corresponding aggregation work, is to complete the aggregation of adjacent nodes and update the feature, is to calculate the feature between two nodes, To complete the prediction of the position deviation, pos i is the position information of the first node p i pos j is the position information of the neighbor node p j is the node feature of node p j at the iteration number t, where j is the adjacent node number;

[0094] According to the node features of the first node, use the MLP linear network to predict the variance of the node features, that is, use the node features of the first node at the t-th iteration to estimate the change in the SAR main structure information caused by changes such as angles, and add a linear network to predict this variance:

[0095]

[0096] where Δpos is the position deviation, is the node feature of the first node p i at the iteration number t.

[0097] In one embodiment, input the set of main body pixel points into the collaborative graph attention layer for iterative training. To effectively ensure the required predicted position deviation Δpos, the prediction requires aggregating the adjacent node information. Before the relative position embedding layer, input the set of main body pixel points into a layer of graph convolutional layer to update the node features of the current graph convolutional layer:

[0098]

[0099] where s i is the updated node feature, W1 and W2 are the first weight matrix parameter and the second weight matrix parameter respectively, s j is the node feature of the adjacent node, and N(i) is the set of adjacent nodes including the self-node. The updated node features include: the updated attribute feature information and the updated position information feature.

[0100] In one embodiment, according to the target graph structure, the updated attribute feature information, and the updated position information feature, inject the position deviation of each node in the target graph structure into the relative position embedding layer as a residual structure, and use the residual structure method to fuse the attribute features extracted by the collaborative graph attention and the position information extracted by the relative position embedding layer to output the fused feature. After the fused feature passes through the global pooling layer, use the cross-entropy loss function for iterative training of the structure features to output the target image.

[0101] ​It should be noted that in the establishment of the network model, the existing convolutional neural network and Transformer network are discarded. Based on the graph neural network, a collaborative graph attention layer and a relative position embedding layer are proposed, and a special SAR graph neural network model is designed. The features obtained after one iteration are respectively input into the collaborative graph attention layer and the relative position embedding layer for iterative training, and the attribute features extracted by the collaborative graph attention and the position information extracted by the relative position embedding layer are fused in the way of a residual structure to output the fused features. After the fused features pass through the global pooling layer, the cross-entropy loss function is used for iterative training of the structural features to output the recognition result. On the premise of ensuring light weight, the fused information improves the recognition performance of the graph neural network. In lightweight applications, the above method uses lower computational costs and storage costs to meet the lightweight requirements of edge application scenarios.

[0102] In one embodiment, as Figure 2 shown, a collaborative graph attention layer information update is provided, which uses node information to calculate the collaborative attention scores of the central node and neighbor nodes in sequence, and updates the central node features using the collaborative attention scores and neighbor node features.

[0103] In one embodiment, as Figure 3 shown, a relative position embedding layer information update process is provided, which uses the relative position information of the central node and neighbor nodes to update the central node features, and uses the central node features obtained from the previous update to calculate Δpos to eliminate the graph structure differences caused by angle changes to a certain extent.

[0104] In one embodiment, as Figure 4 shown, an overall framework of the SAR target recognition network is provided. The graph structure first passes through a layer of GCN, then through a fusion structure of three layers of collaborative attention layers and relative position embedding layers, and finally uses the global pooling layer and a four-layer MLP to complete the recognition task and output the recognition result. The specific contents are as follows:

[0105] Dataset Preparation and Network Setting: For SAR data, the SAR-AIRcraft-1.0 dataset is used. This dataset is a publicly available SAR aircraft target dataset for large-scale complex scenarios. The data is collected from the Gaofen-3 satellite, with a single-polarization mode, a spatial resolution of 1m, and a spotlight imaging mode. The dataset mainly comes from three civil airports: Shanghai Hongqiao Airport, Beijing Capital Airport, and Taoyuan Airport in Taiwan, with complex scenarios. The dataset covers 7 fine-grained aircraft categories, namely A220, A320 / 321, A330, ARJ21, Boeing737, Boeing787, and other, with a total of 16,463 aircraft target instances. The instance targets are cropped according to the annotation boxes in the SAR-AIRcraft-1.0 dataset to obtain a series of instance samples, and the corresponding specific quantities are shown in Table 1. In the experiment, half of the training set is used as the training data, that is, 6,754 samples are selected for training, and 2,950 samples are used for testing. During the training process, the batch size is set to 64, the learning rate is 0.001, the dropout rate is 0.3, the training epochs of the network are 100, and the hidden layer size is 128.

[0106] Table 1 Quantities of Instance Targets in Different Categories in the SAR-AIRcraft-1.0 Dataset

[0107]

[0108] Next, two types of quantization criteria are introduced, including recognition performance metrics and resource efficiency metrics. Among them, the recognition performance metrics include accuracy, precision, recall, and F1-score; the resource efficiency metrics include training time, number of parameters, floating-point operation count, inference time, accuracy / FLOPS, and accuracy / number of parameters.

[0109] The specific introduction is as follows:

[0110] (1) Accuracy: Accuracy represents the proportion of correctly classified samples in all predictions. The calculation formula is:

[0111]

[0112] where TP is the number of correct predictions for true positive samples, TN is the number of correct predictions for true negative samples, FP is the number of predictions for false positive samples, and FN is the number of predictions for false negative samples.

[0113] (2) Precision: Precision measures the proportion of samples that are actually positive among the samples predicted as positive by the model. The calculation formula is:

[0114]

[0115] (3) Recall Rate: The recall rate reflects the ability of the model to identify actual positive class samples. The calculation formula is:

[0116]

[0117] (4) F1-Score: The harmonic mean of precision and recall, comprehensively considering the balance between precision and recall. The calculation formula is:

[0118]

[0119] (5) Training Time: The training time measures the total time required for the model to start training until it is completed. This metric reflects the computational efficiency of the model during training.

[0120] (6) Number of Parameters: The number of parameters is the total number of trainable parameters in the model. A smaller number of parameters not only reduces the storage requirements of the model but also helps to reduce the computational overhead and accelerate the inference process. It is one of the key metrics for lightweight models.

[0121] (7) Floating Point Operations (FLOPs): FLOPs represent the number of floating point operations required by the model during one inference process. This metric directly measures the computational complexity of the model. The fewer FLOPs, the higher the computational efficiency of the model.

[0122] (8) VRAM Occupancy: VRAM occupancy refers to the size of the VRAM occupied by the model during inference or training. VRAM occupancy is a key factor in evaluating the resource consumption of the model when running on a GPU (Graphics Processing Unit). A model with a smaller VRAM occupancy is more easily deployable on resource-constrained devices.

[0123] (9) Inference Time: The inference time refers to the time required for the model to make a prediction for a single input sample. For applications that require real-time response, the inference time is the core metric for measuring the efficiency of the model, usually requiring it to be as low as possible to ensure a good user experience.

[0124] (10) Accuracy / FLOPs: This ratio measures the classification performance that can be achieved per unit of computational effort (FLOPs). Through this metric, the performance efficiency of the model under limited computational resources can be evaluated. A higher accuracy / FLOPs ratio indicates that the model can achieve higher classification performance with fewer computational resources.

[0125] (11) Accuracy / Number of Parameters: This ratio represents the classification performance brought by each parameter. A higher accuracy / number of parameters ratio means that the model can achieve high classification accuracy while maintaining a small number of parameters, which is an important indicator of lightweight models.

[0126] The experimental platform is an Intel(R) Core(TM) i7-13700 CPU, with a Windows 11 Pro operating system, a GeForce RTX 4090 24GB GPU, and the neural network deployment platform is Python 3.7 and Pytorch 1.9.

[0127] Tables 2 and 3 present the quantitative results obtained from the SAR image recognition experiment using the network structure proposed in the present invention. Figure 5 The confusion matrix is given. Table 2 presents the precision, recall, and F1-score corresponding to each category. Since the accuracy is a global metric, it is placed in Table 3. Additionally, Table 3 also presents the training time, number of parameters, FLOPS, video memory occupancy, inference time, and the ratios of accuracy / FLOPS and accuracy / number of parameters. All results are the averages of ten experiments.

[0128] Table 2 Quantitative Results of SAR Recognition by the Network Proposed in the Present Invention 1

[0129] Category Precision Recall F1-score A330 92.984±0.66 96.413±0.326 94.665±0.185 A320 / A321 65.194±2.374 95.193±0.962 77.375±1.990 A220 92.394±4.158 93.548±3.225 92.821±0.513 ARJ21 98.387±1.011 91713±0.553 94.927±0.175 Boeing737 95.274±1.050 95.091±0.182 95.180±0.615 Boeing787 96.604±0.676 93.943±0.551 95.255±0.612 Other 95.378±0.029 95.058±0.624 95.216±0.328 Average 90.888±0.556 94.423±0.437 92.206±0.010

[0130] Table 3 Quantitative Results of SAR Recognition by the Network Proposed in the Present Invention 2

[0131]

[0132] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown sequentially in the direction of the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover,

[0133] In one embodiment, as Figure 6 shown, a SAR target recognition device based on the fusion of graph neural network and location information is provided, including: a main pixel point set acquisition module 602, a graph structure construction module 604, a model construction module 606, a lightweight feature optimization module 608, and a feature fusion module 610, where:

[0134] The main pixel point set acquisition module 602 is configured to acquire the main pixel point set of the target image of the SAR; wherein, each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image;

[0135] The graph structure construction module 604 is configured to construct a target graph structure with fused position information according to the pixel point vectors.

[0136] The model construction module 606 is configured to construct a collaborative target recognition model based on a graph neural network. The collaborative target recognition model includes: a collaborative graph attention layer and a relative position embedding layer.

[0137] The lightweight feature optimization module 608 is configured to input the main pixel point set into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training, so as to extract the updated attribute feature information and position information features.

[0138] The feature fusion module 610 is configured to add a relative position information embedding layer to the collaborative graph attention layer in the form of a residual structure according to the target graph structure, the updated attribute feature information, and the updated position information features, and obtain a fused feature through aggregation. After the fused feature passes through the global pooling layer, the cross-entropy loss function is used for iterative training of the structural features, and the target image is output.

[0139] For the specific limitations of the SAR target recognition device based on the graph neural network and position information fusion, reference can be made to the limitations of the SAR target recognition method based on the graph neural network and position information fusion in the above text, which will not be elaborated here. Each module in the above SAR target recognition device based on the graph neural network and position information fusion can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor in the computer device in hardware form or independent of it, or stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0140] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 7As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a SAR target recognition method based on the fusion of graph neural networks and position information. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad set on the shell of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0141] Those skilled in the art can understand that Figure 6 - 7 the structure shown in the figure is only a block diagram of some structures related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0142] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the following steps are implemented:

[0143] Obtain the set of main pixel points of the target image of SAR. Among them, each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image.

[0144] Construct a target graph structure with position information fusion according to the pixel point vector.

[0145] Construct a collaborative target recognition model based on a graph neural network. The collaborative target recognition model includes: a collaborative graph attention layer and a relative position embedding layer.

[0146] Input the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information features.

[0147] According to the target graph structure, the updated attribute feature information, and the updated position information features, add a relative position information embedding layer to the collaborative graph attention layer in the form of a residual structure. Through aggregation, the fusion feature is obtained. After the fusion feature passes through the global pooling layer, the cross-entropy loss function is used for iterative training of the structural features, and the target image is output.

[0148] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0149] Obtain a set of main pixel points of the target image of the SAR. Each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image.

[0150] Construct a target graph structure with fused position information according to the pixel point vectors.

[0151] Construct a collaborative target recognition model based on a graph neural network. The collaborative target recognition model includes: a collaborative graph attention layer and a relative position embedding layer.

[0152] Input the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information features.

[0153] According to the target graph structure, the updated attribute feature information, and the updated position information features, add a relative position information embedding layer to the collaborative graph attention layer in the form of a residual structure. After obtaining the fused features through aggregation, the fused features pass through a global pooling layer, and then the cross-entropy loss function is used for iterative training of the structural features to output the target image.

[0154] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0155] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0156] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the invention patent should be subject to the appended claims.

Claims

1. A SAR target recognition method based on the fusion of graph neural network and position information, characterized in that, The method includes: Obtaining a set of main pixel points of the target image of SAR; wherein, each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image; Constructing a target graph structure with fused position information according to the pixel point vectors; Constructing a collaborative target recognition model based on a graph neural network, the collaborative target recognition model including: a collaborative graph attention layer and a relative position embedding layer; Inputting the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information features; According to the target graph structure, the updated attribute feature information, and the updated position information features, adding the relative position information embedding layer to the collaborative graph attention layer in a residual structure manner, obtaining a fused feature through aggregation, and after the fused feature passes through a global pooling layer, using a cross-entropy loss function for iterative training of the structural features to output the target image.

2. The method according to claim 1, wherein Obtaining a set of main pixel points of the target image of SAR includes: Obtaining the target image captured by SAR, and performing threshold segmentation on the target image using the Otsu method to obtain a set of main pixel points.

3. The method according to claim 1, wherein Constructing a target graph structure with fused position information according to the pixel point vectors includes: According to the adjacent position relationship between any nodes in the pixel point vectors, if there is an adjacent node of the first node, then the first node and the adjacent node are connected by an undirected edge to construct a target graph structure with fused position information: G(N m ) = {V m , E m , POS m} Among them, V m is a finite and non-empty set of vertices, E m is a set of edges between vertices, POS m is a set of position information of vertices, N m is m samples, m ∈ {1, …, N}, and N is the total number of samples.

4. The method according to any one of claims 1 to 3, characterized in that, In the collaborative graph attention layer, calculating the attention score of the adjacent node of the first node in the pixel node vectors: where a T is a linear weight vector, W is a parameter to be learned, t is the number of iterations, is the node information of the first node p i corresponding to the number of iterations t, is the node information of the adjacent node p j corresponding to the number of iterations t, and || is vector concatenation; The attention score is normalized among all the adjacent nodes through softmax: Among them, N(i) is the set of adjacent nodes including the self-node, and softmax j performs a normalization operation on the adjacent node p j , and α i,j is the normalized attention score of the first node p i with respect to the adjacent node p corresponding to the first node j , e(p i , p j ) is the attention score of the adjacent node p j of the first node. w and j are both adjacent nodes. is the sum of the attention scores for each adjacent node of the first node. e(p i , p w ) is the attention score of the adjacent node p w for the first node; After performing weighted averaging on the normalized adjacent nodes, obtaining the node feature of the first node: Among them, is the node feature of the first node p i at the iteration number t + 1, and α i,w is the adjacent node p w the normalized attention score, is the adjacent node p w the node feature at the iteration number t.

5. The method according to any one of claims 1 to 3, characterized in that In the relative position embedding layer, using relative position information to describe the node feature of the first node: Among them, is the aggregated node feature, and three explicit mapping networks are used to complete the corresponding aggregation work respectively, is to complete the aggregation of adjacent nodes and update the features, is to calculate the features between two nodes, is to complete the prediction of the position deviation, pos i is the position information of the first node p i pos j is the position information of the neighbor node p j is the node feature of node p j at the iteration number t, and j is the adjacent node number;​ Predicting the variance of the node feature according to the node feature of the first node using an MLP linear network: where Δpos is the position deviation, is the node feature of the first node p i at the iteration number t.

6. The method according to claim 5, wherein Inputting the set of main pixel points into the collaborative graph attention layer and the relative position embedding layer respectively for iterative training to extract the updated attribute feature information and position information features includes: Inputting the set of main pixel points into the collaborative graph attention layer for iterative training, and before the relative position embedding layer, inputting the set of main pixel points into a graph convolutional layer to update the node features of the current graph convolutional layer: Among them, s i is the updated node feature, W1 and W2 are the first weight matrix parameter and the second weight matrix parameter respectively, s j is the node feature of adjacent nodes, and N(i) is the set of adjacent nodes including the node itself; The updated node features include: updated attribute feature information and updated position information features.

7. The method according to claim 6, characterized in that, According to the target graph structure, the updated attribute feature information, and the updated position information features, adding the relative position information embedding layer to the collaborative graph attention layer in a residual structure manner, obtaining a fused feature through aggregation, and after the fused feature passes through a global pooling layer, using a cross-entropy loss function for iterative training of the structural features to output the target image includes: According to the target graph structure, the updated attribute feature information, and the updated position information feature, the position deviation of each node in the target graph structure is injected into the relative position embedding layer as a residual structure, and the attribute features extracted by fusing the collaborative graph attention in the way of the residual structure and the position information extracted by the relative position embedding layer are fused to output a fused feature. After the fused feature passes through the global pooling layer, the cross-entropy loss function is used for iterative training of the structural features to output the target image.

8. A SAR target recognition device based on the fusion of graph neural network and position information, characterized in that The device includes: A main pixel point set acquisition module, configured to acquire a main pixel point set of the target image of the SAR; wherein, each main pixel point feature is composed of the attribute feature information of the main pixel point and the position information of the main pixel point in the target image; A graph structure construction module, configured to construct a target graph structure for fusing position information according to the pixel point vector; A model construction module, configured to construct a collaborative target recognition model based on a graph neural network, where the collaborative target recognition model includes: a collaborative graph attention layer and a relative position embedding layer; A lightweight feature optimization module, configured to respectively input the main pixel point set into the collaborative graph attention layer and the relative position embedding layer for iterative training to extract the updated attribute feature information and position information feature; A feature fusion module, configured to add the relative position information embedding layer to the collaborative graph attention layer in the way of a residual structure according to the target graph structure, the updated attribute feature information, and the updated position information feature, and obtain a fused feature through aggregation. After the fused feature passes through the global pooling layer, the cross-entropy loss function is used for iterative training of the structural features to output the target image.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

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

Images