SAR target recognition method and device based on electromagnetic characteristics and deep learning

Abstract

This application relates to a SAR target recognition method and device based on electromagnetic characteristics and deep learning. The method comprises: extracting SAR target scattering centers using an ASC model and constructing a target component model based on this; then, extracting feature vectors from the original target image and the target component image using a feature extraction backbone network, fusing the component features using a component attention unit, and then classifying and outputting the fused feature vectors using a convolutional layer to obtain the final classification result. This method, combined with the target electromagnetic characteristics, can produce a more robust target classification algorithm. Furthermore, the method of extracting target and component features using a component attention mechanism enables quantitative derivation of component importance for each target classification result, improving the interpretability of the deep learning algorithm.

Description

Technical Field

[0001] The present application relates to the technical field of SAR image target recognition, and in particular to a SAR target recognition method and device based on electromagnetic characteristics and deep learning. Background Art

[0002] SAR has advantages such as all-weather, long-range, and high-resolution imaging capabilities, and has strong applicability in both military and civilian fields. Due to its various advantages, SAR has been widely used in remote sensing image target detection, target recognition, terrain classification, and other fields. Automatic target recognition in SAR images has been an important research topic in recent years. Recognition methods are mainly divided into two categories: traditional machine learning methods and deep learning-based methods. For most traditional machine learning methods, target feature extraction and classification are mainly based on electromagnetic scattering characteristics, which are interpretable and stable. However, the process of extracting effective recognition features is often complex and computationally inefficient. Compared with traditional methods, deep learning methods can directly learn the high-dimensional features of the target, thereby achieving higher target recognition accuracy. However, due to the "black box" nature of deep learning, these algorithms have poor generalization performance and are difficult to interpret. Summary of the Invention

[0003] Based on this, it is necessary to provide a SAR target recognition method and device based on electromagnetic characteristics and deep learning that can interpret and analyze target recognition results to address the above technical problems.

[0004] A SAR target recognition method based on electromagnetic characteristics and deep learning, characterized in that the method includes:

[0005] Acquire a SAR target image to be used for target identification;

[0006] Extracting multiple single scattering centers of the SAR target image using an attribute scattering center model, and obtaining multiple target component images by calculation based on the multiple single scattering centers;

[0007] Inputting the SAR target image and each target component image into a feature extraction backbone network, and extracting corresponding target feature vectors and multiple component feature vectors respectively;

[0008] Using a component attention unit, the target feature vector and multiple component feature vectors are fused to obtain a fused feature matrix. When fusing the target feature vector and the multiple component feature vectors, the features of each channel of the target feature vector are weightedly calculated with the feature vectors of each component to obtain a fused feature vector corresponding to each channel. The fused feature vectors of all channels constitute the fused feature matrix.

[0009] The convolution layer is used to perform target classification on the fused feature matrix, and the target in the SAR target image is identified according to the classification result.

[0010] In one embodiment, the step of calculating a plurality of target component images based on the plurality of single scattering centers comprises:

[0011] Calculating according to multiple parameters of each of the single scattering centers to obtain a cluster center of a parameter set;

[0012] Dividing the plurality of single scattering centers to obtain different scattering center sets;

[0013] Calculating each of the scattering center sets according to the cluster centers to reconstruct a corresponding frequency domain image of the target component;

[0014] Then, a two-dimensional Fourier inverse transform is performed on the frequency domain image of the target component to obtain the target component images of each component in the target.

[0015] In one embodiment, the feature extraction backbone network adopts a fully convolutional network.

[0016] In one embodiment, the target feature vector and the plurality of component feature vectors are in the form of three-dimensional feature vectors, and before the feature fusion is performed using the component attention unit, the target feature vector and the plurality of component feature vectors are converted into the form of two-dimensional feature vectors;

[0017] The fused feature matrix output by the component attention unit is in the form of a two-dimensional feature vector, and is also converted into a three-dimensional feature vector before being input into the convolutional layer.

[0018] In one embodiment, the features of each channel of the target feature vector are weighted together with the feature vectors of each component to obtain the fused feature vector corresponding to each channel using the following formula:

[0019]

[0020]

[0021] In the above formula, q n represents the features of the nth channel, (k p ,v p ) represents the pth component feature vector, n is the total number of target component images divided, d q,k Represents the component feature dimension, Softmax() represents the Softmax function, represents the component weight, b m Represents the fused feature vector of the mth channel.

[0022] In one embodiment, after the target in the SAR target image is identified according to the classification result, an interpretability analysis of the importance of each target component is performed based on the fused feature matrix and the classification result.

[0023] In one embodiment, analyzing the importance of each target component based on the fusion feature matrix and the classification result includes:

[0024] Derived based on the fusion feature matrix and the classification results, the result of the convolution calculation using the target features is used as an indicator to measure the importance of a certain component to the final target recognition result;

[0025] After calculating the importance of all components, the target decision map is obtained by multiplying the pixels of each component location area on the SAR target image by the corresponding importance and normalizing them. The target decision map shows the distribution map of pixels and areas that play a key role in the classification result.

[0026] Perform an interpretability analysis on the classification results according to the target decision diagram.

[0027] A SAR target recognition device based on electromagnetic characteristics and deep learning, the device comprising:

[0028] SAR target image acquisition module, used to acquire SAR target images to be used for target identification;

[0029] a target component image construction module, which extracts a plurality of single scattering centers of the SAR target image using an attribute scattering center model, and obtains a plurality of target component images by performing calculations based on the plurality of single scattering centers;

[0030] A feature vector extraction module is used to input the SAR target image and each target component image into a feature extraction backbone network to extract corresponding target feature vectors and multiple component feature vectors respectively;

[0031] a feature fusion module, configured to fuse the target feature vector and multiple component feature vectors using a component attention unit to obtain a fused feature matrix, wherein when fusing the target feature vector and multiple component feature vectors, the features of each channel of the target feature vector are weightedly calculated with the feature vectors of each component to obtain a fused feature vector corresponding to each channel, and the fused feature vectors of all channels constitute the fused feature matrix;

[0032] The target recognition module is used to perform target classification on the fused feature matrix using a convolutional layer, and to recognize targets in the SAR target image according to the classification results.

[0033] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0034] Acquire a SAR target image to be used for target identification;

[0035] Extracting multiple single scattering centers of the SAR target image using an attribute scattering center model, and obtaining multiple target component images by calculation based on the multiple single scattering centers;

[0036] Inputting the SAR target image and each target component image into a feature extraction backbone network, and extracting corresponding target feature vectors and multiple component feature vectors respectively;

[0037] Using a component attention unit, the target feature vector and multiple component feature vectors are fused to obtain a fused feature matrix. When fusing the target feature vector and the multiple component feature vectors, the features of each channel of the target feature vector are weightedly calculated with the feature vectors of each component to obtain a fused feature vector corresponding to each channel. The fused feature vectors of all channels constitute the fused feature matrix.

[0038] The convolution layer is used to perform target classification on the fused feature matrix, and the target in the SAR target image is identified according to the classification result.

[0039] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0040] Acquire a SAR target image to be used for target identification;

[0041] Extracting multiple single scattering centers of the SAR target image using an attribute scattering center model, and obtaining multiple target component images by calculation based on the multiple single scattering centers;

[0042] Inputting the SAR target image and each target component image into a feature extraction backbone network, and extracting corresponding target feature vectors and multiple component feature vectors respectively;

[0043] Using a component attention unit, the target feature vector and multiple component feature vectors are fused to obtain a fused feature matrix. When fusing the target feature vector and the multiple component feature vectors, the features of each channel of the target feature vector are weightedly calculated with the feature vectors of each component to obtain a fused feature vector corresponding to each channel. The fused feature vectors of all channels constitute the fused feature matrix.

[0044] The convolution layer is used to perform target classification on the fused feature matrix, and the target in the SAR target image is identified according to the classification result.

[0045] The above-mentioned SAR target recognition method and device based on electromagnetic characteristics and deep learning utilizes an ASC model to extract SAR target scattering centers and constructs a target component model based on this. Next, a feature extraction backbone network is used to extract feature vectors for both the original target image and the target component image. A component attention module is then used to fuse the component features. Finally, a convolutional layer is used to classify and output the fused feature vectors, yielding the final classification result. This method, combined with target electromagnetic characteristics, yields a more robust target classification algorithm. Furthermore, the component attention mechanism allows for the quantitative derivation of component importance for each target classification result, improving the interpretability of deep learning algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 1 is a flow chart of a SAR target recognition method based on electromagnetic characteristics and deep learning in one embodiment;

[0047] Figure 2 Schematic diagram of the calculation process in the component attention unit in one embodiment;

[0048] Figure 3 Schematic diagram of a calculation process for classifying and outputting fused features in one embodiment;

[0049] Figure 4 is a flowchart of a SAR target recognition method based on electromagnetic characteristics and deep learning in another embodiment;

[0050] Figure 5 1 is a block diagram of a SAR target recognition device based on electromagnetic characteristics and deep learning in one embodiment;

[0051] Figure 6 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0053] like Figure 1 As shown, a SAR target recognition method based on electromagnetic characteristics and deep learning is provided, comprising the following steps:

[0054] Step S100, obtaining a SAR target image to be identified;

[0055] Step S110, extracting multiple single scattering centers of the SAR target image using an attribute scattering center model, and obtaining multiple target component images by calculation based on the multiple single scattering centers;

[0056] Step S120, inputting the SAR target image and each target component image into a feature extraction backbone network, and extracting corresponding target feature vectors and multiple component feature vectors respectively;

[0057] Step S130: Using the component attention unit, the target feature vector and the multiple component feature vectors are fused to obtain a fused feature matrix. When fusing the target feature vector and the multiple component feature vectors, the features of each channel of the target feature vector are weighted together with the feature vectors of each component to obtain a fused feature vector corresponding to each channel. The fused feature vectors of all channels form a fused feature matrix.

[0058] Step S140: Using a convolutional layer to perform target classification on the fused feature matrix, and identifying the target in the SAR target image according to the classification result.

[0059] In this example, a target recognition method based on target electromagnetic characteristics and a deep learning network is proposed, enabling interpretable analysis of SAR target recognition results. Target component models are constructed using the Attributed Scattering Center (ASC) model. A classification network is trained based on component features. A component attention mechanism is employed to extract and fuse target and component features. The fused features are then used to train a classification network for target recognition. This method, incorporating target electromagnetic characteristics for target recognition, improves the robustness of SAR target recognition results. Furthermore, the importance of individual components can be quantified and derived based on the target recognition results and component features, making the target recognition results interpretable.

[0060] In step S110, calculating based on the multiple single scattering centers to obtain multiple target component images includes: calculating based on multiple parameters of each single scattering center to obtain a cluster center of a parameter set, dividing the multiple single scattering centers to obtain different scattering center sets, and calculating for each scattering center set based on the cluster center to reconstruct a corresponding target component frequency domain image, and then performing a two-dimensional inverse Fourier transform on the target component frequency domain image to obtain a target component image of each component in the target.

[0061] Specifically, in this step, the ASC model is used to extract the SAR target scattering center and the target component model is constructed based on this.

[0062] The ASC model has been widely used in the field of target recognition. The model assumes that the backscatter of the target can be well approximated as the sum of the responses of each scattering center, as shown below:

[0063]

[0064] In formula (1), q represents the total number of single scattering centers. For a single scattering center, the backscattered field can be parameterized as the frequency f and the azimuth angle Function:

[0065]

[0066] In formula (2) and formula (1), f c represents the radar center frequency, C represents the propagation speed of electromagnetic waves. Θ represents the ASC model parameter set, Among them, (x i ,y i ) represents the spatial position, A i Represents the relative amplitude, α i represents the dependence of the ASC model on frequency and set structure, L i and denote the length and angle dependence of the distributed scattering centers, γ i The azimuthal dependence of the localized scattering center is represented by . The seven parameters in the ASC model parameter set can be estimated using the parameter optimization method.

[0067] In this embodiment, the K-means clustering method is used to calculate the cluster centers of the model parameter set based on the seven parameters of the ASC model. Then, multiple single scattering centers are divided into different scattering center sets. For each scattering center set, the reconstructed local image, that is, the frequency domain image of the target component, is calculated based on the cluster center using formula (3):

[0068]

[0069] In formula (3), E R Represents the reconstructed data of the model parameters in a cluster combined in the frequency domain. Θ R represents the estimated model parameter set, Θ R ={θ Ri}(i=1,2,…,p), p represents the number of scattering centers in a cluster. θ Ri represents the parameter set of the i-th scattering center in the R-th cluster.

[0070] Then, a two-dimensional inverse Fourier transform is performed on the reconstructed frequency domain image of the target component to obtain a reconstructed target component image.

[0071] In step S120, a feature extraction backbone network is used to extract feature vectors for each of the SAR target image and the target component image. After obtaining the target component image, it and the SAR target image are input into the feature extraction backbone network. This network can employ various conventional classification network structures. In this embodiment, a fully convolutional network structure is used as an example.

[0072] The SAR target image input to the fully convolutional network is 128×128, and the target component image is 88×88. After passing through four convolutional layers with 16, 32, 64, and 128 channels respectively, three-dimensional feature matrices of 128×4×4 and 128×1×1 are obtained.

[0073] Before using the component attention unit for feature fusion, the target feature vector and multiple component feature vectors are converted into two-dimensional feature vector form, that is, the 128×4×4 feature matrix of the SAR target image is flattened to 128×16, and the 128×1×1 feature matrix of each target component image is flattened to 128×1.

[0074] In step S130, the features of each channel of the target feature vector are weighted with the feature vectors of each component to obtain the fusion feature vectors of each channel. Here, the 128-dimensional features of the first channel in the feature vector of the 128×16 SAR target image are used as an example to illustrate. Figure 2 The component attention calculation process in Calculate the weight of each target component feature in the final fusion feature output and the fusion feature result. The component attention calculation process is expressed as the following formula:

[0075]

[0076] In formula (4-6), q n represents the features of the nth channel, (k p ,v p ) represents the pth component feature vector, n is the total number of target component images divided, d q,k Represents the component feature dimension, Softmax() represents the Softmax function, represents the component weight, b m Represents the fused feature vector of the mth channel.

[0077] Taking the first channel as an example, q1 represents the feature of the first channel, (k p ,v p ) represents the pth component feature vector, n is the total number of target component images divided, d q,k Indicates the component feature dimension, which is 128 here. Softmax() represents the Softmax function. represents the component weight, and b1 represents the fused feature vector of the first channel.

[0078] First, multiply the first channel feature by each target component feature to obtain α 1,p , and then calculated by the Softmax function That is the component weight, and finally the final output feature is calculated by weighting the features of each component, which is b1.

[0079] The above calculation is performed on each channel of the 128×16 feature vector of the SAR target image, and finally the 128×16 fused feature, that is, the fused feature matrix, is obtained, and it is converted back to a 128×4×4 two-dimensional vector form.

[0080] In step S140, the fused feature matrix is ​​finally classified and outputted using a convolutional layer to obtain the final classification result.

[0081] like Figure 3 As shown in the figure, a 128×4×4 fusion feature matrix is ​​convolved with C convolution kernels, where N represents the number of final classifications, to obtain a C×1×1 feature vector. Finally, the target category probability vector is calculated through Batch Normalization (BN) and Softmax activation function. The category with the highest probability is the final classification result of the target.

[0082] like Figure 4 As shown, when performing target recognition according to this method, the feature extraction backbone network, component attention units, and convolutional layers described above can be combined into a target recognition network. This network is then iteratively trained using training data. A loss function is calculated after each iteration, and the parameters of the feature extraction backbone network, component attention units, and convolutional layers are adjusted based on the loss function results until the loss function converges. This results in a trained target recognition network. This trained target recognition network can then be applied to target recognition tasks in SAR images.

[0083] In this embodiment, after the target in the SAR target image is identified, the importance of each target component can be analyzed for interpretability according to the fusion feature matrix and the classification results, including deduction based on the fusion feature matrix and the classification results, using conv k (V p ) is used as an indicator to measure the importance of a certain component to the final target recognition result. Then, after calculating the importance of all components, the target decision diagram is obtained by multiplying the pixels of the component position area on the SAR target image by the corresponding importance and normalizing them. The target decision diagram represents the target pixel and area distribution map that plays a key role in the classification result. Finally, the classification result is analyzed for interpretability based on the target decision diagram.

[0084] Specifically, the fusion feature output by the component attention unit is F, f ij and Figure 2 The corresponding way of the 16 channel features output in can be expressed as: ij =b (i-1)×4+j .

[0085]

[0086] Let t = (i-1) × 4 + j, then:

[0087]

[0088] Substitute F into the convolution calculation formula, and for each convolution kernel k, record its parameters as It can be deduced as follows:

[0089]

[0090] The above derivation shows that the output of convolution can be regarded as the weighted convolution calculation of each component feature and then the sum. For the final BN and Softmax operations, they mainly transform the value range of the final output, and the relative size of the output of different components does not change. Therefore, conv can be used k (V p ) is used as an indicator to measure the importance of a certain component of the target to the final target recognition result.

[0091] After calculating the importance of all components, the decision map of the target can be obtained by multiplying the corresponding importance of each component position pixel on the target image and normalizing it. This is a distribution map of pixels and regions in the target that play a key role in the classification results, thereby performing interpretability analysis of the classification results.

[0092] In this SAR target recognition method based on electromagnetic properties and deep learning, the ASC model is first used to extract SAR target scattering centers. Based on this, a target component model is constructed. Feature vectors are extracted from both the SAR target image and the target component image using a feature extraction backbone network. Component features are then fused using a component attention unit. The fused feature vectors are then classified and output using a convolutional layer to produce the final classification result. Based on this target recognition method, component importance is derived based on the component attention and convolution classification output calculation processes, and interpretable analysis is performed in conjunction with the network classification results. This method, combined with target electromagnetic properties, yields a more robust target classification algorithm. After obtaining the deep learning-based target classification results, the component attention mechanism is then used to quantify the component importance of each target classification result, improving the interpretability of the deep learning algorithm. Furthermore, this interpretability analysis method based on the component attention mechanism is highly versatile and can be flexibly embedded into various existing deep learning algorithm frameworks, making it highly practical.

[0093] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0094] In one embodiment, Figure 5 As shown, a SAR target recognition device based on electromagnetic characteristics and deep learning is provided, including: a SAR target image acquisition module 200, a target component image construction module 210, a feature vector extraction module 230, a feature fusion module 240 and a target recognition module 250, wherein:

[0095] SAR target image acquisition module 200, used to acquire SAR target images to be used for target identification;

[0096] A target component image construction module 210 extracts a plurality of single scattering centers 220 of the SAR target image using an attribute scattering center model, and calculates a plurality of target component images based on the plurality of single scattering centers;

[0097] A feature vector extraction module 230 is configured to input the SAR target image and each target component image into a feature extraction backbone network to extract corresponding target feature vectors and multiple component feature vectors respectively;

[0098] A feature fusion module 240 is configured to fuse the target feature vector and multiple component feature vectors using a component attention unit to obtain a fused feature matrix. When fusing the target feature vector and multiple component feature vectors, a weighted calculation is performed on the features of each channel of the target feature vector and each component feature vector to obtain a fused feature vector corresponding to each channel. The fused feature vectors of all channels constitute the fused feature matrix.

[0099] The target recognition module 250 is configured to perform target classification on the fused feature matrix using a convolutional layer, and recognize targets in the SAR target image according to the classification results.

[0100] Regarding the specific definition of the SAR target recognition device based on electromagnetic characteristics and deep learning, please refer to the definition of the SAR target recognition method based on electromagnetic characteristics and deep learning above, and will not be repeated here. The various modules in the above-mentioned SAR target recognition device based on electromagnetic characteristics and deep learning can be implemented in whole or in part by software, hardware, or a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0101] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. 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 a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a SAR target recognition method based on electromagnetic characteristics and deep learning is implemented. The display screen of the computer device can be a liquid crystal display or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0102] Those skilled in the art will understand that Figure 6The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0103] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0104] Acquire a SAR target image to be used for target identification;

[0105] Extracting multiple single scattering centers of the SAR target image using an attribute scattering center model, and obtaining multiple target component images by calculation based on the multiple single scattering centers;

[0106] Inputting the SAR target image and each target component image into a feature extraction backbone network, and extracting corresponding target feature vectors and multiple component feature vectors respectively;

[0107] Using a component attention unit, the target feature vector and multiple component feature vectors are fused to obtain a fused feature matrix. When fusing the target feature vector and the multiple component feature vectors, the features of each channel of the target feature vector are weightedly calculated with the feature vectors of each component to obtain a fused feature vector corresponding to each channel. The fused feature vectors of all channels constitute the fused feature matrix.

[0108] The convolution layer is used to perform target classification on the fused feature matrix, and the target in the SAR target image is identified according to the classification result.

[0109] 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:

[0110] Acquire a SAR target image to be used for target identification;

[0111] Extracting multiple single scattering centers of the SAR target image using an attribute scattering center model, and obtaining multiple target component images by calculation based on the multiple single scattering centers;

[0112] Inputting the SAR target image and each target component image into a feature extraction backbone network, and extracting corresponding target feature vectors and multiple component feature vectors respectively;

[0113] Using a component attention unit, the target feature vector and multiple component feature vectors are fused to obtain a fused feature matrix. When fusing the target feature vector and the multiple component feature vectors, the features of each channel of the target feature vector are weightedly calculated with the feature vectors of each component to obtain a fused feature vector corresponding to each channel. The fused feature vectors of all channels constitute the fused feature matrix.

[0114] The convolution layer is used to perform target classification on the fused feature matrix, and the target in the SAR target image is identified according to the classification result.

[0115] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the 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-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory 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 (DDRSDRAM), 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).

[0116] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

[0117] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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, they should be considered to be within the scope of this specification.

[0118] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A SAR target recognition method based on electromagnetic characteristics and deep learning, characterized in that: The method comprises: Acquire a SAR target image to be used for target identification; Extracting multiple single scattering centers from the SAR target image using an attribute scattering center model, and performing calculations based on the multiple single scattering centers to obtain multiple target component images. Specifically, performing calculations based on multiple parameters of each of the single scattering centers to obtain cluster centers of a parameter set, dividing the multiple single scattering centers to obtain different scattering center sets, performing calculations based on the cluster centers for each of the scattering center sets to reconstruct a corresponding target component frequency domain image, and then performing a two-dimensional inverse Fourier transform on the target component frequency domain image to obtain a target component image of each component in the target. Inputting the SAR target image and each target component image into a feature extraction backbone network, and extracting corresponding target feature vectors and multiple component feature vectors respectively; The target feature vector and multiple component feature vectors are fused using a component attention unit to obtain a fused feature matrix, wherein when the target feature vector and multiple component feature vectors are fused, the features of each channel of the target feature vector are weighted with the feature vectors of each component to obtain a fused feature vector corresponding to each channel, and the fused feature matrix is ​​formed by the fused feature vectors of all channels, wherein the features of each channel of the target feature vector are weighted with the feature vectors of each component to obtain a fused feature vector corresponding to each channel using the following formula: In the above formula, represents the features of the nth channel, Indicates the component feature vectors, is the total number of target component images divided, Represents the component feature dimension, represents the Softmax function, represents the component weight, Represents the fused feature vector of the mth channel; The convolution layer is used to perform target classification on the fused feature matrix, and the target in the SAR target image is identified according to the classification result.

2. The SAR target recognition method according to claim 1, characterized in that: The feature extraction backbone network adopts a fully convolutional network.

3. The SAR target recognition method according to claim 2, characterized in that: The target feature vector and the plurality of component feature vectors are in the form of three-dimensional feature vectors, and before the component attention unit is used to perform feature fusion, the target feature vector and the plurality of component feature vectors are also converted into the form of two-dimensional feature vectors; The fused feature matrix output by the component attention unit is in the form of a two-dimensional feature vector, and is also converted into a three-dimensional feature vector before being input into the convolutional layer.

4. The SAR target recognition method according to claim 3, characterized in that: After identifying the target in the SAR target image according to the classification result, an interpretability analysis is performed on the importance of each target component according to the fusion feature matrix and the classification result.

5. The SAR target recognition method according to claim 4, characterized in that: The analyzing the importance of each target component according to the fusion feature matrix and the classification result includes: Derived based on the fusion feature matrix and the classification results, the result of the convolution calculation using the target features is used as an indicator to measure the importance of a certain component to the final target recognition result; After calculating the importance of all components, the target decision map is obtained by multiplying the pixels of each component location area on the SAR target image by the corresponding importance and normalizing them. The target decision map shows the distribution map of pixels and areas that play a key role in the classification result. Perform an interpretability analysis on the classification results according to the target decision diagram.

6. SAR target recognition device based on electromagnetic characteristics and deep learning, characterized in that: The device comprises: SAR target image acquisition module, used to acquire SAR target images to be used for target identification; a target component image construction module, which extracts a plurality of single scattering centers of the SAR target image using an attribute scattering center model, and obtains a plurality of target component images by performing calculations based on the plurality of single scattering centers; a feature vector extraction module, configured to input the SAR target image and each target component image into a feature extraction backbone network, extract corresponding target feature vectors and multiple component feature vectors, specifically, perform calculations based on multiple parameters of each single scattering center to obtain cluster centers of a parameter set, divide the multiple single scattering centers to obtain different scattering center sets, perform calculations based on the cluster centers for each scattering center set to reconstruct a corresponding target component frequency domain image, and then perform a two-dimensional inverse Fourier transform on the target component frequency domain image to obtain target component images of each component in the target; The feature fusion module is used to use the component attention unit to fuse the target feature vector and multiple component feature vectors to obtain a fused feature matrix. When fusing the target feature vector and multiple component feature vectors, the features of each channel of the target feature vector are weighted with the feature vectors of each component to obtain the fused feature vectors corresponding to each channel. The fused feature matrix is ​​composed of the fused feature vectors of all channels. The features of each channel of the target feature vector are weighted with the feature vectors of each component to obtain the fused feature vectors corresponding to each channel using the following formula: In the above formula, represents the features of the nth channel, Indicates the component feature vectors, is the total number of target component images divided, Represents the component feature dimension, represents the Softmax function, represents the component weight, Represents the fused feature vector of the mth channel; The target recognition module is used to perform target classification on the fused feature matrix using a convolutional layer, and to recognize targets in the SAR target image according to the classification results.

Images