A SAR image target recognition method and system integrating electromagnetic scattering features

Abstract

The present invention relates to a SAR image target recognition method and system integrating electromagnetic scattering features, wherein the SAR image target recognition method comprises the following steps: generating a first intermediate feature map based on a deep convolutional neural network for a SAR test image, inputting the SAR test image and the first intermediate feature map into an image classification network to obtain a recognition result; wherein, the method for obtaining the model parameters of the deep convolutional neural network comprises the following steps: generating an attribute scattering center image based on an attribute scattering center model for a SAR training image, generating a reconstructed attribute scattering center image based on a deep convolutional neural network for the SAR training image, and obtaining the model parameters of the deep convolutional neural network based on the attribute scattering center image and the reconstructed attribute scattering center image. The present invention innovatively adopts the idea of ​​a distillation model to extract electromagnetic scattering features, and adopts the idea of ​​global attention to fuse electromagnetic scattering features for SAR image target recognition.

Description

Technical Field

[0001] The present invention belongs to the field of target recognition in computer vision, and specifically relates to a supervised SAR image target recognition method using deep learning, and in particular to a SAR image target recognition method integrating electromagnetic scattering features. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging sensor with the characteristics of multi-mode, multi-view, multi-band, multi-polarization and high resolution. It is not affected by weather conditions when working and has a certain ability to penetrate ground objects, enabling it to observe the ground all day and all weather and obtain high-resolution images of target scenes. The images obtained reflect the scattering characteristics of the monitored targets.

[0003] SAR Automatic Target Recognition (SAR-ATR) aims to use computers to automatically implement corresponding algorithms for target recognition without human intervention. Since the imaging mechanism of synthetic aperture radar is different from that of optical images, the interpretation of SAR images cannot directly follow the research results of optical images. In addition, the sensitivity of synthetic aperture radar to imaging conditions greatly increases the technical difficulty of image understanding.

[0004] The electromagnetic scattering feature extraction methods of SAR image scattering centers are generally based on non-parametric methods of Fourier analysis and parametric methods based on scattering center parameterized models. The existing three-dimensional electromagnetic scattering models that describe the electromagnetic scattering characteristics of SAR targets include point scattering models, attenuation exponential models, attribute scattering center models, etc. These models assume that the backscattering of the target can be approximated as the sum of a series of independent scattering center responses. By estimating the parameters of the model, the relevant features of the SAR image target scattering points can be extracted. However, at present, more complex parameter optimization methods such as sparse representation, maximum likelihood estimation, and genetic algorithms are usually used in model parameter extraction, which takes a long time to calculate and is not conducive to the subsequent classification algorithm design based on the target scattering characteristics. Therefore, designing a SAR image electromagnetic scattering feature reconstruction network with fast inference speed has important engineering application value.

[0005] In addition, the existing SAR image target recognition method that integrates electromagnetic scattering features first uses traditional methods to extract electromagnetic scattering features, and then performs feature fusion by clustering attribute scattering centers, channel splicing, feature addition, etc. These methods do not achieve end-to-end detection, and are complex and time-consuming to calculate, and cannot effectively use electromagnetic scattering features to assist the target recognition network for detection. For example, in Chinese patent applications CN 116051994 A and CN 116740561A, in the test phase, it is necessary to first use traditional methods to extract electromagnetic scattering parameters, and then perform a two-dimensional Fourier inverse transform to obtain an attribute scattering center image. However, this method only uses traditional methods to extract attribute scattering center images in the training phase, and does not increase the reasoning time in the test phase. In addition, in Chinese patent application CN 116778349 A, the scattering coefficients extracted by the scattering center extraction module only include the position and amplitude of each scattering center, but the electromagnetic scattering features also involve geometric structure information such as frequency dependence and azimuth dependence. This method uses electromagnetic scattering images to more completely reflect the electromagnetic scattering characteristics of the target. Summary of the invention

[0006] The purpose of the present invention is to overcome the above problems. In a first aspect, a SAR image target recognition method integrating electromagnetic scattering features is provided. The SAR image target recognition method comprises the following steps: generating a first intermediate feature map based on a deep convolutional neural network for a SAR test image, inputting the SAR test image and the first intermediate feature map into an image classification network to obtain a recognition result;

[0007] Among them, the method for obtaining the model parameters of the deep convolutional neural network includes the following steps: generating an attribute scattering center image based on the attribute scattering center model from the SAR training image, generating a reconstructed attribute scattering center image based on the deep convolutional neural network from the SAR training image, and obtaining the model parameters of the deep convolutional neural network based on the attribute scattering center image and the reconstructed attribute scattering center image.

[0008] The second aspect provides a SAR image target recognition system integrating electromagnetic scattering features, the system comprising at least one processor; and a memory storing instructions, which, when executed by at least one processor, implements the steps of the aforementioned method.

[0009] The beneficial effects of the present invention are as follows: the present invention innovatively adopts the idea of ​​distillation model to extract electromagnetic scattering features, and adopts the idea of ​​global attention to fuse electromagnetic scattering features for SAR image target recognition. The reconstructed attribute scattering center model proposed by the present invention can assist SAR target detection and recognition, etc., and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of a SAR image target recognition method integrating electromagnetic scattering features;

[0011] Figure 2 It is a schematic diagram of a SAR image target recognition training method integrating electromagnetic scattering features;

[0012] Figure 3 Schematic diagram of SAR image target recognition system integrating electromagnetic scattering features

[0013] Figure 4 is a flow chart of the identification method of the present invention;

[0014] Figure 5 It is a flow chart of generating a simulated SAR image in the recognition method of the present invention;

[0015] Figure 6 is a flow chart of a reconstruction module in the identification method of the present invention;

[0016] Figure 7 is a flow chart of a local residual module in the recognition method of the present invention;

[0017] Figure 8 is a flow chart of the enhanced channel attention module in the recognition method of the present invention;

[0018] Fig. 9 It is a flow chart of the feature fusion module in the recognition method of the present invention. DETAILED DESCRIPTION

[0019] The following examples further illustrate the content of the present invention, but should not be construed as limiting the present invention. Without departing from the spirit and substance of the present invention, modifications or substitutions made to the methods, steps or conditions of the present invention all fall within the scope of the present invention.

[0020] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and implementation modes of the present invention.

[0021] The purpose of the embodiments of the present invention is to solve the technical problems in the prior art and to provide a SAR image target recognition method and system integrating electromagnetic scattering features.

[0022] like Figure 1 As shown, in some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, and the SAR image target recognition method comprises the following steps:

[0023] S11: Generate a first intermediate feature map based on the deep convolutional neural network based on the SAR test image;

[0024] S12: Inputting the SAR test image and the first intermediate feature map into an image classification network to obtain a recognition result;

[0025] like Figure 2 As shown, the method for obtaining the model parameters of the deep convolutional neural network includes the following steps:

[0026] S21: Generate an attribute scattering center image based on the attribute scattering center model from the SAR training image;

[0027] S22: Generate a reconstructed attribute scattering center image based on a deep convolutional neural network from the SAR training image;

[0028] S23: Obtaining model parameters of the deep convolutional neural network based on the attribute scattering center image and the reconstructed attribute scattering center image.

[0029] In these embodiments of the present invention, the idea of ​​distillation model is used to extract electromagnetic scattering features, and the idea of ​​global attention is used to fuse electromagnetic scattering features for SAR image target recognition. First, through innovative feature fusion methods, the electromagnetic scattering features are fused with the features extracted by the image classification model to perform global information fusion, thereby improving the classification network's attention to the foreground, thereby increasing the accuracy of SAR image classification. Second, the maximum likelihood estimation optimization method is used only in the training phase to extract the attribute scattering center image, which does not increase the inference time in the test phase.

[0030] The term "fused electromagnetic scattering features" refers to the use of an Attributed Scattering Center (ASC) model in the training phase. Based on the concept of the present invention, the existing Attributed Scattering Center model is also within the selection scope of the present invention.

[0031] The term "deep convolutional neural network" refers to a convolutional neural network (CNN) that includes convolution calculations and has a deep structure. Based on the concept of the present invention, existing CNNs are within the selection scope of the present invention.

[0032] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, wherein the deep convolutional neural network further includes a deep convolutional neural network baseline model and several layers of reconstruction modules, wherein:

[0033] The reconstruction module is configured to input the deep convolutional neural network baseline model or the feature map output by the reconstruction module of the previous layer, and after being processed by the deconvolution layer, normalization function and activation function, output or output the feature map or the first intermediate feature map to the reconstruction module of the next layer; in some preferred embodiments, the reconstruction module has three layers; in some further preferred embodiments, the reconstruction module includes 1 to 3 3×3 deconvolution layers, 1 batch normalization function and 1 ReLU function; in some further preferred embodiments, the deep convolutional neural network baseline model includes but is not limited to the ResNet18 network.

[0034] The term “deep convolutional neural network baseline model” includes but is not limited to the ResNet18 baseline model.

[0035] In these embodiments of the present invention, for electromagnetic scattering feature extraction, a deep convolutional neural network is used to reconstruct an attribute scattering center model to solve the problems of complex electromagnetic scattering feature extraction and time-consuming calculation of SAR targets in the prior art. For example, the target recognition network based on the ResNet18 baseline model is improved, and a reconstruction module is added to reconstruct the electromagnetic scattering features, optimize the extraction method, and improve the calculation efficiency. In addition. For SAR target recognition, an image classification model that fuses electromagnetic scattering features is used to improve target recognition accuracy. For example, the feature fusion module fuses electromagnetic scattering features to improve its recognition performance while maintaining a fast inference speed. Compared with other SAR image target recognition methods, the embodiments of the present invention can reduce the training time, realize end-to-end fusion of electromagnetic scattering features for recognition, and improve the model recognition accuracy.

[0036] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, wherein the image classification network includes a plurality of residual modules, a feature fusion module, and a global average pooling module, wherein:

[0037] The residual module is used to input the SAR test image or the first-size SAR image, and output the second-size SAR image after being processed by the deconvolution layer, the normalization function, the activation function, and the enhanced channel attention module;

[0038] The feature fusion module is used to input the second-size SAR image and the feature map output by the reconstruction module, and output a fused SAR feature map after convolution kernel, dimensional expansion, normalization, transposition and expansion rearrangement processing; in some preferred embodiments, the feature map output by the reconstruction module of the second-size SAR image and the fused SAR feature map have the same size;

[0039] The global average pooling module is used to input the second-size SAR image and output a final feature output vector.

[0040] The specific values ​​of the terms "first size" and "second size" are not limited. For example, they can all be 100×100, 64×50×50, 128×13×13, and 256×7×7. When used as the input and output of the same residual module, the first size and the second size cannot have the same value.

[0041] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, wherein the enhanced channel attention module is configured to: input a second intermediate feature map, process it through global average pooling, a fully connected layer, and a sigmoid function, and output a weighted feature map, wherein the weighted feature map is the second-size SAR image.

[0042] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, and a method for generating an attribute scattering center image based on an attribute scattering center model from a SAR training image is as follows:

[0043] Resizing the SAR training image to obtain a preprocessed SAR training image;

[0044] The preprocessed SAR training image is subjected to image segmentation, model structure selection, parameter initialization, parameter optimization and image generation to obtain an attribute scattering center parameter estimation set;

[0045] The attribute scattering center image is obtained by reconstructing the scattering center parameter estimation set.

[0046] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, and the method for obtaining the model parameters of the deep convolutional neural network based on the attribute scattering center image and the reconstructed attribute scattering center image comprises the following steps:

[0047] The mean absolute error loss between the attribute scattering center image and the reconstructed attribute scattering center image is calculated to obtain the model parameters of the deep convolutional neural network.

[0048] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, wherein the deep convolutional neural network and the image classification network constitute a joint model, and a method for obtaining model parameters of the joint model comprises the following steps:

[0049] Calculating the mean absolute error loss between the attribute scatter center image and the reconstructed attribute scatter center image;

[0050] Calculating a cross entropy loss based on the final feature output vector and the SAR image label;

[0051] The weighted sum of the mean absolute error loss and the cross entropy loss is used to obtain a total loss function;

[0052] The model parameters of the joint model are obtained according to the total loss function.

[0053] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, and the step of inputting the SAR test image and the first intermediate feature map into an image classification network to obtain a recognition result includes:

[0054] Inputting the SAR test image and the first intermediate feature map into the joint model to obtain the final feature output vector;

[0055] The index of the maximum value of the final feature output vector is obtained as the recognition result.

[0056] In some embodiments of the present invention, a SAR image target recognition method integrating electromagnetic scattering features is provided, wherein the SAR image target recognition method comprises one or more of the following steps:

[0057] The attribute scattering center model is a three-dimensional electromagnetic scattering parameterized model derived by structural calculation of some typical scatterers through simulation;

[0058] The image segmentation is a watershed threshold method;

[0059] The model structure selection is to determine the local scattering center and the distributed scattering center;

[0060] The parameter initialization is to obtain the initial value of the parameter based on the constructor;

[0061] The parameter optimization is to obtain the optimal value of the parameter based on the initial value of the parameter;

[0062] The image generation is to obtain a simulated SAR image based on the optimal value of the parameter;

[0063] Repeating the model structure selection, parameter initialization, parameter optimization and image generation steps to obtain the attribute scattering center parameter estimation set;

[0064] Based on the attribute scattering center parameter estimation set, obtaining the reconstructed attribute scattering center image through inverse Fourier transform;

[0065] The reconstructed attribute scattering center image is normalized.

[0066] like Figure 3As shown, in some other embodiments of the present invention, a SAR image target recognition system integrating electromagnetic scattering features is also provided, wherein the system 1 includes at least one processor 2; and a memory 3 storing instructions, which, when executed by at least one processor 2, implements the steps of the aforementioned method.

[0067] The present invention is further described below in conjunction with specific implementation methods. Figures 4 to 9 As shown, it includes the process of reconstructing attribute scattering centers using the maximum likelihood parameter estimation method, the joint training process of the attribute scattering center reconstruction model and the image classification model, and the testing process of using the model to perform SAR image target recognition tasks.

[0068] Example 1 Model Training

[0069] A SAR image target recognition method integrating electromagnetic scattering features includes a training process of a SAR image classification model integrating electromagnetic scattering features, specifically including the following steps:

[0070] Step 1: Preprocess the raw SAR images for training by image resizing.

[0071] Step 2: The preprocessed image is subjected to image segmentation, model structure selection, parameter initialization, parameter optimization and image generation to obtain the attribute scattering center parameter estimation set.

[0072] Step 3: Use a reconstruction method to obtain an attribute scattering center image for the scattering center parameter estimation set and perform normalization processing.

[0073] Step 4: Preprocess the raw SAR image pairs by image resizing, random horizontal flipping, and normalization.

[0074] Step 5: Use the ResNet18 network as the baseline model and add an image reconstruction module after the baseline model to form an attribute scattering center reconstruction network. Input the SAR image preprocessed in step 4 into the attribute scattering center reconstruction network to obtain a reconstructed attribute scattering center image.

[0075] Step 6: Calculate the mean absolute error loss between the reconstructed attribute scattering center image and the attribute scattering center image obtained in step 3.

[0076] Step 7: Input the image preprocessed in step 4 and the intermediate feature map of the attribute scattering center reconstruction network into the image classification network for forward propagation.

[0077] Step 8: Send the output of the fourth layer of the image classification network to the global average pooling module to obtain the final feature output vector, and calculate the cross entropy loss in combination with the SAR image label. The weighted sum of the mean absolute error loss and the cross entropy loss is used as the total loss function.

[0078] Step 9: Use the total loss function for back propagation, update the network weights, and continue the next training.

[0079] Step 10: Repeat steps 6 to 9 until the loss approaches zero and the network finally converges. Save the model parameters of the last iteration.

[0080] Example 2 Target Identification

[0081] The process of using the SAR image classification model integrated with electromagnetic scattering features obtained in Example 1 for SAR image target recognition specifically includes the following steps:

[0082] Step 11: Resize the original SAR image for testing and normalize it.

[0083] Step 12: Input the test image into the trained model to obtain the final output feature vector.

[0084] Step 13: Get the index of the maximum value of the final feature output vector as the recognition result.

[0085] Example 3 Model Training

[0086] A SAR image target recognition method based on electromagnetic scattering features. Figure 4 As shown in FIG. 1 , the training process of the SAR image classification model integrating electromagnetic scattering features specifically includes the following steps:

[0087] Take the MSTAR dataset collected by the Sandia National Laboratory SAR sensor platform as an example, which contains ten types of ground targets: BMP2, BTR70, T72, T62, BRDM2, BTR60, ZSU23 / 4, D7, ZIL131 and 2S1. These images are acquired by an X-band SAR system with a resolution of 0.3m×0.3m and a size of about 128×128 pixels.

[0088] Step 1: Read the original SAR image for training and resize the image to 100×100.

[0089] Step 2: The preprocessed image is subjected to image segmentation, model structure selection, parameter initialization, parameter optimization and image generation to obtain the attribute scattering center parameter estimation set. The specific method is as follows:

[0090] 1) The property scattering center model is a three-dimensional electromagnetic scattering parameterized model derived by simulation based on the structure calculation of some typical scatterers, as shown in formula (1).

[0091]

[0092] In formula (1), f represents the radar center frequency, C represents the propagation speed of electromagnetic waves, represents the azimuth, is the model parameter set, θ p =x p ,y p ,α p ,γ p ,L p ,A p ] T represents the parameter of the pth scattering center, Represents the relative amplitude, x p ,y p represents the coordinates of the scattering center in the range and azimuth directions relative to the zero phase reference point, γ p represents the azimuthal dependence of the scattering amplitude at the local scattering center, α p represents frequency dependence, L p represents the length of the distribution scattering center, It represents the ratio of the azimuth angle of the distribution scattering center to the maximum aperture.

[0093] 2) Use the watershed threshold method for image segmentation. Find the maximum peak amplitude based on the original SAR data, and calculate the -3dB and -20dB of the peak. It is considered that the pixels above -20dB of the maximum peak belong to the scattering center response area corresponding to the maximum value, and the pixel area above -3dB of the maximum peak is the peak area.

[0094] 3) The model structure can be divided into local scattering centers and distributed scattering centers. If the peak area contains only one peak point, the -3dB width of the peak point in the image domain azimuth profile is used to determine the type of scattering center. If the width is greater than the set threshold of 0.5, it is determined to be a distributed scattering center, and L p Initialize to the obtained -3dB width, otherwise it is determined as a local scattering center; if it contains multiple peak points, calculate the relative distance coordinate difference between other peak points and the maximum peak point, and consider that the peak points with an absolute value of the difference less than 1 may form a distribution scattering center with the maximum peak point, which is called the candidate peak point, and then calculate the relative azimuth coordinate difference between the candidate peak point and the maximum peak point, select the peak point with the smallest absolute value of the azimuth coordinate difference, and consider that the peak point and the maximum peak point form a distribution scattering center and L p Initialized to the absolute value of the azimuth coordinate difference.

[0095] 4) For parameter initialization, select the peak area center coordinate x p ,y p As initial coordinates; Initialized to 0; γ p Initialized to 0. Constructor:

[0096] J(θ)=[abs(d R )-abs(e R )] T [abs(d R )-abs(e R )] (2)

[0097] where d R represents the original image data corresponding to the scattering center area, e R Indicates that the initial value θ p =[x p ,y p ,α p ,γ p ,L p ,A p ] T Substitute the data estimated by the attribute scattering center model.

[0098] i. In x p ,y p ,γ p ,L p In the initialization, let A p =1, and α p =[0,1 / 2,1] are substituted into J(θ) respectively, and the α corresponding to the minimum value of J(θ) is calculated. p The value is considered to be α p Initial value.

[0099] ii. In x p ,y p ,α p ,γ p ,L p In the case of initialization, Where S R Indicates that the initial value θ p =[x p ,y p ,α p ,γ p ,L p ,A p =1] T Substitute the values ​​obtained in (1).

[0100] 5) Set θ p =[x p ,y p ,α p ,γ p ,L p ,A p ] T Substitute into (2), optimize the parameters of the function J(θ) and find θp =[x p ,y p ,α p ,γ p ,L p ,A p ] T The optimal value of .

[0101] 6) Combination Figure 5 As shown, the optimal parameter values ​​are substituted into formula (1) to generate a simulated image, and the simulated image is subtracted from the original image. Specifically, the optimal model parameter values ​​are substituted into formula (1) to generate frequency domain data, two-dimensional interpolation is performed, and the data is multiplied by the Taylor window function. After zero padding, the simulated SAR image is obtained by two-dimensional inverse Fourier transform.

[0102] 7) Repeat 3)-6) until the number of extracted scattering centers reaches 20 or the peak value of the scattering center to be processed is less than -3dB of the maximum peak value of the image, and obtain the attribute scattering center parameter estimation set.

[0103] Step 3: Input the obtained parameter estimation set into equation (1) to obtain the frequency domain image, and obtain the attribute scattering center image through inverse Fourier transform. The image size is (20×13×13), and then normalized.

[0104] Step 4: The original SAR images for training are resized to 100×100, and the image pairs are preprocessed using random horizontal flipping and normalization with a probability of p=0.5. A training batch includes 32 SAR images.

[0105] Step 5: Use the ResNet18 network as the baseline model, and add an image reconstruction module after the baseline model to form an attribute scattering center reconstruction network. Input the SAR image preprocessed in step 4 into the attribute scattering center reconstruction network to obtain a reconstructed attribute scattering image. Combined with the attribute scattering center image obtained in step 3, calculate the mean absolute error loss. Figure 6 As shown, the specific method is:

[0106] 1) The original SAR image of size (100×100) is input into the ResNet18 baseline model, and the output is a feature map of size (512×4×4).

[0107] 2) Input the features of size (512×4×4) to the first reconstruction module. Specifically, it first passes through two 3×3 deconvolution layers, each of which is followed by a batch normalization function and a ReLU function, and outputs a feature map of size (256×11×11).

[0108] 3) Input the features of size (256×11×11) to the second-layer reconstruction module. Specifically, it first passes through a 3×3 deconvolution layer, a batch normalization function, and a ReLU function, and finally outputs a feature map of size (128×13×13).

[0109] 4) Input the features of size (512×4×4) to the third reconstruction module. Specifically, it first passes through three 3×3 convolutional layers, each of which is followed by a batch normalization function and a ReLU function, and outputs a feature map of size (20×13×13).

[0110] Step 6: Use the third layer reconstruction module to output an image of size (20×13×13) The attribute scattering center image I obtained in step 3 ASC , calculate the mean absolute error loss

[0111] Step 7: Input the image preprocessed in step 4 and the intermediate feature map of the attribute scattering center reconstruction network into the image classification network for forward propagation. Figure 7 As shown, the specific method is:

[0112] 1) The original SAR image of size (100×100) is input into the first residual module, and a feature map of size (64×50×50) is output. In a more specific embodiment, two 3×3 deconvolution layers are firstly performed, each of which is followed by a batch normalization function and a ReLU function, and the feature map extracted by the last convolution layer is added to the original input feature map to obtain an intermediate feature map, and then the intermediate feature map is input into the enhanced channel attention module.

[0113] 2) Input the feature map of size (64×50×50) to the second residual module, and output the feature map of size (128×13×13). In a more specific embodiment, firstly, two 3×3 deconvolution layers are passed, each of which is followed by a batch normalization function and a ReLU function, and the feature map extracted by the last convolution layer is added to the original input feature map to obtain an intermediate feature map, and then the intermediate feature map is input to the enhanced channel attention module.

[0114] 3) The feature map I with size (128×13×13) CLS , and the feature map I with a size of (128×13×13) output by the second layer reconstruction module in step 5 Rec_ They are input into the feature fusion module together, and the output feature map size is still (128×13×13). The specific method is:

[0115] i. will I CLSAfter two 1×1 convolution kernels, the number of channels is reduced to half of the original number. Then the h and w dimensions are expanded and the channels are rearranged to obtain two feature vectors of size (196×64) respectively.

[0116] ii. will I Rec_ After a 1×1 convolution kernel, the number of channels is reduced to half of the original number, and then the h and w dimensions are expanded to obtain a feature vector of size (64×196)

[0117] iii. and Multiply and normalize with softmax, and then normalize the feature vector Multiplying together, the resulting feature vector size is (196×64).

[0118] iv. The feature vector of size (196×64) is transposed and expanded, and the output feature map size is (64×13×13).

[0119] v. The feature vector of size (64×196) is passed through a 1×1 convolution kernel, the number of channels is expanded to the original number of channels, and the output feature map size is (128×13×13).

[0120] 4) Input the feature map of size (128×13×13) to the third residual module, and output the feature map of size (256×7×7). In a more specific embodiment, firstly, two 3×3 deconvolution layers are passed, each of which is followed by a batch normalization function and a ReLU function, and the feature map extracted by the last convolution layer is added to the original input feature map to obtain an intermediate feature map, and then the intermediate feature map is input into the enhanced channel attention module.

[0121] 5) Input the feature map of size (256×7×7) to the fourth residual module, and output the feature map of size (512×4×4). In a more specific embodiment, firstly, two 3×3 deconvolution layers are passed, each convolution layer is followed by a batch normalization function and a ReLU function, and the feature map extracted by the last convolution layer is added to the original input feature map to obtain an intermediate feature map, and then the intermediate feature map is input to the enhanced channel attention module.

[0122] 6) In each residual module, combine Figure 8As shown in the figure, the enhanced channel attention module is used to assign different weights to different channels of the input. Specifically, for the feature with an input size of (c, h, w), the feature map is compressed into a feature vector (c, 1, 1) through global average pooling, and then the channel dimension of the feature map vector is reduced to 1 / 16 of the original through the fully connected layer. At this time, the feature vector size is (c / 16, 1, 1), and the channel dimension of the feature map vector is restored to the previous size (c, 1, 1) through the fully connected layer, and then converted into a normalized vector of [0, 1] through the sigmoid function. The normalized weight is multiplied by the original input feature map channel channel by channel, and the output image size is a weighted feature map of (c, h, w).

[0123] Step 8: Send the output of the fourth layer of the image classification network to the global average pooling module to obtain the final feature output vector, and calculate the cross entropy loss in combination with the SAR image label. The weighted sum of the mean absolute error loss and the cross entropy loss is used as the total loss function, combined with Fig. 9 As shown, the specific process is:

[0124] 1) Send the image size (512×4×4) output from the fourth layer of the image classification network to the global average pooling layer. The calculation process is as follows: Where m and n represent the length and width of each instance in the output feature of the fourth layer, respectively, and the value is 4.

[0125] 2) The final output feature vector Input to the fully connected layer, and then get the prediction score through softmax transformation The cross entropy loss is expressed as N is the total number of categories in the training set, y i is the SAR target label of the sample;

[0126] 3) The weighted sum of the mean absolute error loss and the cross entropy loss is used as the total loss function, expressed as: Ltotal = +L rec +L cls

[0127] Step 9: Perform backpropagation according to the overall loss function Ltotal, update the network weights, and continue the next training.

[0128] Step 10: Repeat steps 6 to 9 until the loss approaches zero and the network finally converges. Save the model parameters of the last iteration. The learning rate is initialized to 0.01; a total of 200 epochs are trained, and the learning rate is decayed every 20 epochs with a decay coefficient of 0.1. The specific process is as follows:

[0129]

[0130] Example 4 Target Identification

[0131] The process of using the SAR image classification model fused with electromagnetic scattering features for SAR image target recognition specifically includes the following steps:

[0132] Step 11: Adjust the size of the original SAR image for testing to 100×100 and normalize it.

[0133] Step 12: Input the test image into the trained model to obtain the final output feature vector.

[0134] Step 13: Obtain the index of the maximum value of the final feature output vector as the recognition result. Although the present invention has been described in detail above with general descriptions, specific implementations and experiments, it is obvious to those skilled in the art that some modifications or improvements can be made to it based on the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention belong to the scope of protection claimed by the present invention. The implementation methods and functional operations of the subject matter described in this specification can be implemented in the following: digital electronic circuits, tangibly implemented computer software or firmware, computer hardware, including the structures disclosed in this specification and their structural equivalents, or a combination of more than one of the above. The implementation methods of the subject matter described in this specification can be implemented as one or more computer programs, that is, one or more modules of computer program instructions encoded on one or more tangible non-temporary program carriers, for being executed by a data processing device or controlling the operation of a data processing device. As an alternative or in addition, the program instructions can be encoded on an artificially generated propagation signal, for example, a machine-generated electrical signal, optical signal or electromagnetic signal, which is generated as encoded information to be transmitted to an appropriate receiver device executed by a data processing device. The computer storage medium can be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of the above devices.

Claims

1. A SAR image target recognition method integrating electromagnetic scattering features, characterized in that: The SAR image target recognition method comprises the following steps: generating a first intermediate feature map based on a deep convolutional neural network for a SAR test image, and inputting the SAR test image and the first intermediate feature map into an image classification network to obtain a recognition result; The method for obtaining the model parameters of the deep convolutional neural network comprises the following steps: generating an attribute scattering center image based on an attribute scattering center model from a SAR training image, generating a reconstructed attribute scattering center image based on a deep convolutional neural network from the SAR training image, and calculating the mean absolute error loss between the attribute scattering center image and the reconstructed attribute scattering center image to obtain the model parameters of the deep convolutional neural network; The image classification network includes several residual modules, feature fusion modules, and global average pooling modules, where: The residual module is used to input the SAR test image or the first-size SAR image, and output the second-size SAR image after being processed by the deconvolution layer, the normalization function, the activation function, and the enhanced channel attention module; The feature fusion module is used to input the second-size SAR image and the feature map output by the reconstruction module, and output the fused SAR feature map after convolution kernel, dimension expansion, normalization, transposition and expansion rearrangement processing; The global average pooling module is used to input the second-size SAR image and output a final feature output vector.

2. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 1, characterized in that: The deep convolutional neural network also includes a deep convolutional neural network baseline model and several layers of reconstruction modules, wherein: The reconstruction module is configured to input the deep convolutional neural network baseline model or the feature map output by the reconstruction module of the previous layer, and after being processed by the deconvolution layer, normalization function and activation function, output or output the feature map of the reconstruction module of the next layer or the first intermediate feature map.

3. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 2, characterized in that: The reconstruction module consists of three layers.

4. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 3 is characterized in that: The reconstruction module includes 1 to 3 3×3 deconvolution layers, 1 batch normalization function and 1 ReLU function.

5. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 4 is characterized in that: The deep convolutional neural network baseline model includes but is not limited to the ResNet18 network.

6. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 1, characterized in that: The second-size SAR image, the feature map output by the reconstruction module, and the fused SAR feature map have the same size.

7. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 1, characterized in that: The enhanced channel attention module is configured as follows: The second intermediate feature map is input, and is processed through global average pooling, a fully connected layer, and a sigmoid function to output a weighted feature map, where the weighted feature map is the second-size SAR image.

8. The SAR image target recognition method integrating electromagnetic scattering features according to any one of claims 1 to 7, characterized in that: The method of generating attribute scattering center images based on the attribute scattering center model from SAR training images is as follows: Resizing the SAR training image to obtain a preprocessed SAR training image; The preprocessed SAR training image is subjected to image segmentation, model structure selection, parameter initialization, parameter optimization and image generation to obtain an attribute scattering center parameter estimation set; The attribute scattering center image is obtained by reconstructing the scattering center parameter estimation set.

9. The SAR image target recognition method integrating electromagnetic scattering features according to any one of claims 1 to 7, characterized in that: The deep convolutional neural network and the image classification network constitute a joint model, and the method for obtaining the model parameters of the joint model comprises the following steps: Calculating the mean absolute error loss between the attribute scatter center image and the reconstructed attribute scatter center image; Calculating a cross entropy loss based on the final feature output vector and the SAR image label; The weighted sum of the mean absolute error loss and the cross entropy loss is used to obtain a total loss function; The model parameters of the joint model are obtained according to the total loss function.

10. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 9, characterized in that: The step of inputting the SAR test image and the first intermediate feature map into an image classification network to obtain a recognition result comprises: Inputting the SAR test image and the first intermediate feature map into the joint model to obtain the final feature output vector; The index of the maximum value of the final feature output vector is obtained as the recognition result.

11. The SAR image target recognition method integrating electromagnetic scattering features as claimed in claim 8, characterized in that: The SAR image target recognition method comprises one or more of the following steps: the image segmentation is a watershed threshold method; The model structure selection is to determine the local scattering center and the distributed scattering center; The parameter optimization is to obtain the optimal value of the parameter based on the initial value of the parameter; The image generation is to obtain a simulated SAR image based on the optimal value of the parameter; Repeating the model structure selection, parameter initialization, parameter optimization and image generation steps to obtain the attribute scattering center parameter estimation set; Based on the attribute scattering center parameter estimation set, obtaining the reconstructed attribute scattering center image through inverse Fourier transform; The reconstructed attribute scattering center image is normalized.

12. A SAR image target recognition system integrating electromagnetic scattering features, characterized in that: The system comprises at least one processor; and a memory storing instructions, which, when executed by the at least one processor, implement the steps of the method according to any one of claims 1 to 11.

Images