Robust ISAR Target Deformation Recognition Method Based on Dual-Channel Fusion Network

Abstract

The present invention discloses an ISAR target deformation robust recognition method based on a dual-channel fusion network, which mainly solves the problem of low recognition rate in existing ISAR target recognition technologies. The implementation scheme is as follows: 1) Obtain the echo data of the satellite, and use the BP algorithm to perform high-resolution ISAR imaging on the echo data, and generate a training set and a test set with combined deformations; 2) Construct an ISAR target deformation robust recognition network including a deformation robust module, a feature extraction module, and a feature fusion module; 3) Design the loss function of the ISAR target deformation robust recognition network and train the network; 4) Input the test set images into the trained ISAR target deformation robust recognition network to obtain the network output results and the recognition accuracy of the test set. The present invention can enhance the robustness of the model for target feature extraction, effectively reduce the calculation cost, improve the ISAR target recognition performance, and can be used to improve the information perception of the ISAR system.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and further relates to a method for robust recognition of inverse synthetic aperture radar (ISAR) target deformation, which can be used to improve the information perception ability of existing ISAR systems. Background Art

[0002] Inverse synthetic aperture radar (ISAR) has been widely used in fields such as automatic target recognition (ATR) of non-cooperative targets in the air / space due to its advantages of all-weather, all-day, long operating range, and high resolution. With the rapid development of deep learning, a large number of methods have been successfully applied to automatic target recognition based on optical and synthetic aperture radar (SAR) images. However, most of the existing recognition methods based on two-dimensional high-resolution ISAR images use real ISAR images and ignore the recognizable features contained in the original complex images. In addition, during ISAR imaging, a high-resolution range profile (HRRP) sequence of the target can be obtained simultaneously, which contains important temporal features such as the distribution of scattering points and the change of backscattering coefficients. However, the existing ISAR target recognition methods discard the HRRP sequence features, resulting in limited recognition performance. At the same time, the unique imaging mechanism of ISAR causes unknown deformations such as stretching, compression, and rotation in the target image and HRRP sequence, and the existing technologies have not effectively adjusted them, making it difficult to extract robust deformation features. Therefore, there is an urgent need to study a method for robust fusion recognition of ISAR targets based on complex ISAR images and corresponding HRRP sequences.

[0003] The patent document with the application publication number CN110705508A discloses a "method for satellite recognition of ISAR images", which is a method for satellite recognition of ISAR images based on a deep convolutional neural network. It is completed in two parts. First, the image domain features of ISAR are extracted through a deep convolutional neural network, and then the classification layer is used to predict the category of the target. The main deficiencies of this method are mainly in two aspects: one is that since it only extracts features based on real images, the separable features contained in the original complex ISAR image are lost; the other is that since the influence of ISAR image deformation on recognition is not considered, the recognition rate is low.

[0004] The patent document with the application publication number CN113625227A discloses a "method for target recognition of radar high-resolution range profile based on an attention transformation network". It is completed in two parts. First, the local detailed features and global temporal information of the high-resolution range profile are extracted through an attention transformation network; then the classification layer is used to predict the target category. The main deficiencies of this method are mainly in two aspects: one is that the attention transformation network has a complex structure and a large number of parameters, and the attention calculation cost is relatively high; the other is that since the shape and structure features of the target are not utilized, the recognition rate is low. Summary of the Invention

[0005] The object of the present invention is to propose a robust ISAR target deformation recognition method based on a dual-channel fusion network in view of the deficiencies of the above-mentioned existing technologies, so as to improve the recognition rate, reduce network parameters, and lower the computational cost.

[0006] The technical idea for realizing the object of the present invention is as follows: by utilizing the two-dimensional structural information of the complex ISAR image and the fusion recognizable features of the ISAR image and the corresponding HRRP image, and at the same time, by making targeted adjustments to the stretching, rotation, and combined deformation of the target ISAR and HRRP images, the recognition rate is improved; by fully simplifying the network structure of the attention-based HRRP feature extraction network, the network parameters are reduced, and the computational cost is lowered.

[0007] According to the above idea, the implementation steps of the present invention are as follows:

[0008] (1) Generate a training set and a test set:

[0009] (1a) Obtain the echo data of the satellite, and perform high-resolution ISAR imaging on the obtained echo data by using the BP algorithm;

[0010] (1b) Set two groups of different parameters, including the pitch angle, the bandwidth accumulation angle, and the azimuth angle. According to these two groups of different parameters, divide the ISAR imaging result map into a training set and a test set, and then perform inverse Fourier transform on them to obtain the corresponding high-resolution range image HRRP training set and test set;

[0011] (2) Construct a robust ISAR target deformation recognition network:

[0012] (2a) Establish a deformation adjustment module including 5 convolutional layers, 5 pooling layers, 1 fully connected layer, a grid generator, and a sampler;

[0013] (2b) Establish a feature extraction module including a complex convolutional neural network and a horizontal stripe transformer;

[0014] (2c) Establish a feature fusion module composed of weighted fusion of the output features of the complex convolutional neural network and the horizontal stripe transformer;

[0015] (2d) Cascade the deformation adjustment module, the feature extraction module, the feature fusion module, and the existing softmax classifier to form a robust ISAR target deformation recognition network;

[0016] (3) Design the loss function of the robust ISAR target deformation recognition network:

[0017] (3a) Design a distribution calibration loss function L according to the complex convolutional neural network and the horizontal stripe transformer in (2b) Proto :

[0018]

[0019]

[0020]

[0021] Among them, and respectively represent the prototypes of the i-th type of target of the complex convolutional neural network and the horizontal stripe transformer in (2b). and respectively represent the mapped features of the j-th training sample of the complex convolutional neural network and the horizontal stripe transformer in (2b), where i, i = 1,..., 4, j, j = 1,..., N, and N is the number of training samples;

[0022] (3b) Design the loss function L of the total ISAR target deformation robust recognition network according to the distribution calibration loss function L Proto and the loss functions of different data channels: total :

[0023] L total = β1L ISAR + β2L HRRP + β3L fusion + β4L Proto

[0024] where L ISAR is the cross-entropy loss function of the complex convolutional neural network in (2b), L HRRP is the cross-entropy loss function of the horizontal stripe transformer in (2b), L fusion is the cross-entropy loss function of the feature fusion module in (2c), and β1, β2, β3, and β4 are four weighted parameters with different values;

[0025] (4) Input the training sets of ISAR and HRRP into the ISAR target deformation robust recognition network, and use the forward-backward propagation method to train the network until the total loss function L total reaches convergence, and obtain the trained ISAR target deformation robust recognition network;

[0026] (5) Input the test sets of ISAR and HRRP into the trained ISAR target deformation robust recognition network for testing, and obtain the classification results and recognition accuracy rates output by the network.

[0027] The present invention has the following advantages compared with the prior art:

[0028] First, in the designed complex fusion network, the present invention realizes the extraction and fusion of the real and imaginary parts of the two-dimensional complex ISAR image by designing a complex convolutional neural network; by designing a distribution accuracy loss function to shorten the distance between the corresponding features of the ISAR image and the HRRP image of the same target, the effective fusion of features and the accurate identification of types can be achieved, thereby improving the recognition performance.

[0029] Secondly, since the deformation adjustment module is designed in the ISAR target deformation robust recognition network, the unknown scaling, rotation and combined deformation of the target ISAR image and HRRP sequence can be adjusted in a targeted manner by fixing the number and position of affine parameters; at the same time, since the image scaling adjustment hyperparameter is added to the deformation adjustment module to limit the scaling adjustment range, the loss of target boundary information is avoided, the robustness of the model for target feature extraction is enhanced, and the recognition rate is improved.

[0030] Third, the present invention designs a horizontal stripe transformer module in the ISAR target deformation robust recognition network to realize the feature extraction of the HRRP sequence, and captures the lateral attention features of the HRRP sequence by dividing the input image into narrower strips in the horizontal direction. Therefore, the problem of too many parameters caused by the existing method performing attention calculation on the entire input image is avoided, the computational cost is effectively reduced, and the recognition performance is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the implementation of the present invention;

[0032] Figure 2 It is a structural diagram of the ISAR target deformation robust recognition network constructed in the present invention. DETAILED DESCRIPTION

[0033] The implementation and effects of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0034] Reference Figure 1 , the implementation steps for this example are as follows:

[0035] Step 1: Generate training set and test set.

[0036] 1.1) Electromagnetic calculations based on coupled physical optics in FEKO software were used to obtain echo data of four types of satellites: OCO-2, Cloud-sat, CALIPSO and Janson-3. The parameters were set as follows:

[0037] Radar elevation angle They are 50° and 55° respectively, and the corresponding azimuth angle θ observation range is 0° to 360°, the angle interval is 0.05°, and the radar observation bandwidth is 1GHz, 1.5GHz and 2GHz respectively;

[0038] 1.2) After obtaining the electromagnetic scattering echoes of the four types of targets, set the imaging accumulation angles Δθ to 3°, 4°, 5°, and 6° respectively. At the same time, make the azimuth interval between adjacent images be 1°. Use the BP algorithm to perform high-resolution ISAR imaging on the echo data, and adjust the image size to 120×120 pixels by symmetric zero-padding in the image domain;

[0039] 1.3) Generate the ISAR training set and HRRP training set according to the ISAR imaging result map:

[0040] Take the imaging results when B = 2GHz and Δθ = 6° and the imaging results when B = 1.5GHz and Δθ = 5° and divide them into the ISAR training set, and rotate the equivalent azimuth angle of each ISAR sample in this ISAR training set to 0°;

[0041] Set the training set angle range θ = -5° to 5°, and randomly select 5 angles from the given angle range to rotate each sample in the ISAR training set in the image domain to obtain the final ISAR training set;

[0042] Finally, perform inverse Fourier transform on the ISAR training set to obtain the corresponding high-resolution range profile HRRP training set;

[0043] 1.4) Generate the ISAR test set and HRRP test set according to the ISAR imaging result map:

[0044] Take the imaging results when B = 1GHz and Δθ = 6° and the imaging results when B = 2GHz and Δθ = 3° and divide them into the ISAR test set, and rotate the equivalent azimuth angle of each ISAR sample in this ISAR test set to 0°;

[0045] Set the test set angle range θ = -40° to 40°, and randomly select 5 angles from the given angle range to rotate each sample in the ISAR test set in the image domain to obtain the final ISAR test set;

[0046] Finally, perform inverse Fourier transform on the ISAR test set to obtain the corresponding high-resolution range profile HRRP test set.

[0047] Step 2, construct a robust recognition network for ISAR target deformation.

[0048] Refer to Figure 2 , the implementation of this step is as follows:

[0049] 2.1) Construct a deformation adjustment module, and the specific steps are as follows:

[0050] 2.1.1) Set 5 convolutional layers with the convolutional kernel sizes being 9×9, 7×7, 6×6, 3×3, and 3×3 in sequence, corresponding to 4, 8, 16, 32, and 64 feature maps respectively;

[0051] 2.1.2) Set 5 pooling layers with each pooling size being 2×2 and the stride being 2 for each;

[0052] 2.1.3) Cross - connect each convolutional layer with each pooling layer, and cascade the output of the last pooling layer with the fully - connected layer to form a convolutional module;

[0053] 2.1.4) Cascade the convolutional module with the existing grid generator and the existing sampler to form a deformation adjustment module.

[0054] 2.2) Construct a complex convolutional neural network including complex convolutional layers and complex pooling layers:

[0055] 2.2.1) Set 5 complex convolutional layers with the convolutional kernel sizes being 9×9, 7×7, 6×6, 3×3, and 3×3 in sequence, corresponding to 16, 32, 64, 128, and 256 feature maps respectively;

[0056] 2.2.2) Set 5 complex pooling layers with each pooling size being 2×2 and the stride being 2 for each;

[0057] 2.2.3) Cross - connect each complex convolutional layer with each complex pooling layer to form a complex convolutional neural network;

[0058] 2.3) Construct a horizontal stripe transformer module including a convolutional module, a feature extraction module, and a fully - connected layer:

[0059] 2.3.1) Set 2 convolutional layers with the convolutional kernel sizes being 7×7 and 3×3 in sequence, corresponding to 8 and 16 feature maps respectively;

[0060] 2.3.2) Set 1 max - pooling layer with the pooling size being [pooling size value not given in the original] and the stride being 2;

[0061] 2.3.3) Cascade the two convolutional layers and the max - pooling layer to form the convolutional module of the horizontal stripe transformer;

[0062] 2.3.4) Construct a horizontal stripe attention module that evenly divides the input features into non - overlapping horizontal stripes of equal width;

[0063] 2.3.5) Construct a first feature extraction module formed by cascading the existing position encoding and the horizontal stripe attention module;

[0064] 2.3.6) Construct a second feature extraction module formed by cascading the existing position encoding and the existing multi - layer perceptron;

[0065] 2.3.7) Cascade the first feature extraction module, the second feature extraction module, the first feature extraction module, and the second feature extraction module in sequence to form the feature extraction module of the horizontal stripe transformer;

[0066] 2.3.8) Cascade the convolution module, the feature extraction module, and the fully connected layer to form the horizontal stripe transformer;

[0067] 2.4) Parallelly connect the complex convolutional neural network and the horizontal stripe transformer to form the feature extraction module of the ISAR target deformation robust recognition network;

[0068] 2.5) Construct a feature fusion module composed of weighted fusion of the output features of the complex convolutional neural network and the horizontal stripe transformer, and its formula is expressed as follows:

[0069]

[0070] where I ISAR is the output feature of the complex convolutional neural network, I HRRP is the output feature of the horizontal stripe transformer, I fusion is the fusion recognition feature obtained after weighted fusion, α is the weight coefficient, represents the vector concatenation operation.

[0071] 2.6) Cascade the deformation adjustment module, the feature extraction module, the feature fusion module, and the existing softmax classifier to form the ISAR target deformation robust recognition network.

[0072] Step 3, design the loss function of the ISAR target deformation robust recognition network.

[0073] 3.1) Design the distribution calibration loss function L Proto :

[0074]

[0075]

[0076]

[0077] where, and respectively represent the prototypes of the i-th class of targets of the complex convolutional neural network and the horizontal stripe transformer, and respectively represent the mapped features of the j-th training sample of the complex convolutional neural network and the horizontal stripe transformer, i, i = 1,..., 4, j, j = 1,..., N, and N is the number of training samples;

[0078] 3.2) According to the distribution calibration loss function LProto Design the loss function \(L\) of the overall ISAR target deformation robust recognition network for different data channels total :

[0079] \(L\) total =\(\beta_1L_{ce}^{cnn}+\beta_2L_{ce}^{hst}+\beta_3L_{ce}^{ffm}+\beta_4L_{ce}^{srm}\) ISAR +\(\beta_2L_{ce}^{hst}\) HRRP +\(\beta_3L_{ce}^{ffm}\) fusion +\(\beta_4L_{ce}^{srm}\) Proto

[0080] where \(L_{ce}^{cnn}\) is the cross - entropy loss function of the complex convolutional neural network, \(L_{ce}^{hst}\) is the cross - entropy loss function of the horizontal stripe transformer, \(L_{ce}^{ffm}\) is the cross - entropy loss function of the feature fusion module, and \(\beta_1,\beta_2,\beta_3,\beta_4\) are four weighted parameters with different values ISAR is the cross - entropy loss function of the complex convolutional neural network, \(L_{ce}^{hst}\) HRRP is the cross - entropy loss function of the horizontal stripe transformer, \(L_{ce}^{ffm}\) fusion is the cross - entropy loss function of the feature fusion module, and \(\beta_1,\beta_2,\beta_3,\beta_4\) are four weighted parameters with different values

[0081] Step 4: Train the ISAR target deformation robust recognition network

[0082] Input the training set into the ISAR target deformation robust recognition network and train the recognition network using the loss function designed in Step 3

[0083] 4.1) Randomly initialize the weight values

[0084] 4.2) Calculate the value of the loss function \(L\) through forward propagation total ;

[0085] 4.3) Use the backpropagation algorithm to calculate the derivative of the loss function \(L\) with respect to the weights and update the weights using the Adam algorithm total ;

[0086] 4.4) Loop 4.2) and 4.3) until the loss function \(L\) converges to obtain the trained ISAR target deformation robust recognition network total ;

[0087] Step 5: Test the ISAR target deformation robust recognition network

[0088] Input the test sets of ISAR and HRRP into the trained ISAR target deformation robust recognition network for testing to obtain the classification results output by the network and the recognition accuracy rate \(c\) of the test set

[0089]

[0090] where \(\hat{c}=\frac{1}{M}\sum_{i = 1}^{M}I(h(\mathbf{x}_i)=t_i)\) represents the recognition accuracy rate of the test set, \(M\) represents the number of samples in the test set, \(h(\cdot)\) represents the classification discrimination function, \(t_i\) i represents the true category of the \(i\) - th test sample in the test set, and \(y_i\) idenotes the network output result corresponding to the i-th test sample in the test set. When t i and y i are equal, h(t i , y i ) is equal to 1; otherwise, h(t i , y i ) is equal to 0.

[0091] The effects of the present invention will be further described below in conjunction with simulation experiments.

[0092] 1. Simulation experiment conditions:

[0093] The data used in the simulation experiment is a simulation data set. The physical optics method based on coupling in FEKO software is used to perform electromagnetic calculations to obtain the echo data of four types of satellites, namely OCO-2, Cloud-sat, CALIPSO, and Janson-3. The radar elevation angles are 50° and 55° respectively, and the corresponding azimuth angle θ observation ranges are both 0° to 360°, with an angular interval of 0.05°. The radar observation bandwidths are 1 GHz, 1.5 GHz, and 2 GHz respectively.

[0094] After obtaining the electromagnetic scattering echoes of the four types of targets, the BP algorithm is used for high-resolution ISAR imaging. Let the imaging accumulation angles Δθ be 3°, 4°, 5°, and 6° respectively for high-resolution ISAR imaging, and at the same time, let the azimuth angle interval between adjacent images be 1°. Finally, the size is adjusted to 120×120 by symmetric zero-padding in the image domain. Finally, the obtained complex ISAR image is subjected to inverse Fourier transform to obtain its corresponding HRRP sample. Finally, a combined deformation data set is constructed, that is, different combinations of elevation angles, azimuth angles, bandwidths, and accumulation angles are selected for the training set and the test set to verify the deformation robustness of the model.

[0095] The simulation experiment hardware platform is an Intel Xeon E5-2683@2.00GHz CPU, 64GB RAM, and an NVIDIA GeForce GTX1080 Ti GPU. The simulation experiment software platform is Python 3.7 and pytorch 1.3.

[0096] 2. Simulation experiment content and result analysis:

[0097] Under the above experimental conditions, using the same data set, the method of the present invention and the existing ISAR and HRRP recognition methods are applied to classify the targets, and the recognition accuracies of the three methods are calculated respectively. The results are shown in Table 1.

[0098] Table 1 Comparison of recognition rates between the present invention and the prior art

[0099] Recognition rate Prior art of ISAR recognition 79.40％ Prior art of HRRP recognition 70.88％ The present invention 92.50％

[0100] As can be seen from Table 1, the recognition rate of the ISAR target deformation robust recognition method based on the dual-channel fusion network proposed in the present invention reaches 92.50%, which is 13.1% higher than the existing ISAR image recognition technology and 21.62% higher than the existing HRRP recognition technology. It shows that the present invention significantly improves the correct recognition rate of ISAR targets by effectively fusing the features of these two channels of ISAR and HRRP.

Claims

1. A robust ISAR target deformation recognition method based on a dual-channel fusion network, characterized in that, The following are included: (1) Generate a training set and a test set: (1a) Obtain the echo data of the satellite, and perform high-resolution ISAR imaging on the obtained echo data using the BP algorithm; (1b) Set two groups of different parameters, including pitch angle, bandwidth, accumulation angle, and azimuth angle. Divide the ISAR imaging result map into a training set and a test set according to these two groups of different parameters, and then perform inverse Fourier transform on them to obtain the corresponding high-resolution range profile (HRRP) training set and test set; (2) Construct a robust ISAR target deformation recognition network: (2a) Establish a deformation adjustment module consisting of 5 convolutional layers, 5 pooling layers, 1 fully connected layer, a grid generator, and a sampler; (2b) Establish a feature extraction module including a complex convolutional neural network and a horizontal stripe transformer; (2c) Establish a feature fusion module composed of weighted fusion of the output features of the complex convolutional neural network and the horizontal stripe transformer; (2d) Cascade the deformation adjustment module, the feature extraction module, the feature fusion module, and the existing softmax classifier to construct a robust ISAR target deformation recognition network; (3) Design the loss function of the robust ISAR target deformation recognition network; (3a) Design the distribution calibration loss function L according to the complex convolutional neural network and the horizontal stripe transformer in (2b). Proto : Among them, and respectively represent the prototypes of the $i$-th class of objects of the complex convolutional neural network and the horizontal stripe transformer in (2b), and respectively represent the mapped features of the $j$-th training sample of the complex convolutional neural network and the horizontal stripe transformer in (2b), where $i, i = 1, \ldots, 4$, $j, j = 1, \ldots, N$, and $N$ is the number of training samples; (3b) Calibrate the loss function \(L\) according to the distribution Proto and design the loss function \(L\) of the total ISAR target deformation robust recognition network for the loss functions of different data channels total : L total = β1L ISAR + β2L HRRP + β3L fusion + β4L Proto where L ISAR is the cross-entropy loss function of the complex convolutional neural network in (2b), and L HRRP is the cross-entropy loss function of the horizontal stripe transformer in (2b), and L fusion is the cross-entropy loss function of the feature fusion module in (2c), and β1, β2, β3, and β4 are four weighted parameters with different values respectively; (4) Input the training sets of ISAR and HRRP into the ISAR target deformation robust recognition network, and use the forward-backward propagation method to train the network until the total loss function L total reaches convergence to obtain the trained ISAR target deformation robust recognition network; (5) Input the test sets of ISAR and HRRP into the trained robust ISAR target deformation recognition network for testing to obtain the classification results and recognition accuracy rates output by the network.

2. The ISAR target deformation robust recognition method according to claim 1, characterized in that: In step (1a), obtaining the echo data and high-resolution ISAR imaging are realized as follows: (1a1) The echo data of four types of satellites, namely OCO-2, Cloud-sat, CALIPSO, and Janson-3, are obtained through electromagnetic calculation using the coupled physical optics method in FEKO software. The radar elevation angles are 50° and 55° respectively. The corresponding azimuth angle θ observation ranges are both 0° to 360°, the angular interval is 0.05°, and the radar observation bandwidths are 1 GHz, 1.5 GHz, and 2 GHz respectively; (1a2) After obtaining the electromagnetic scattering echoes of four types of targets, set the imaging accumulation angles Δθ to 3°, 4°, 5°, and 6° respectively, and at the same time set the azimuth angle interval between adjacent images to 1°. Perform high-resolution ISAR imaging on the echo data using the BP algorithm, and adjust the image size to 120×120 pixels by symmetric zero-padding in the image domain.

3. The method according to claim 1, characterized in that: In step (1b), generating the training set and the test set is realized as follows: (1b1) Generate the ISAR and HRRP training sets: First, Set B = 2 GHz, Δθ = 6°, and divide the imaging results at B = 1.5 GHz and Δθ = 5° into an ISAR training set, and rotate the equivalent azimuth angle of each ISAR sample in the ISAR training set to 0°; Then, set the training set angle range θ = -5° to 5°, and randomly select 5 angles from the given angle range to rotate each sample in the ISAR training set in the image domain to obtain the final ISAR training set; Finally, perform inverse Fourier transform on the ISAR training set to obtain the corresponding high-resolution range profile (HRRP) training set; (1b2) Generate the ISAR and HRRP test sets: First, Set B = 1 GHz, Δθ = 6°, and When B = 2 GHz and Δθ = 3°, divide the imaging results into an ISAR test set, and rotate the equivalent azimuth angle of each ISAR sample in this ISAR test set to 0°; Then, set the test set angle range θ = -40° to 40°, and randomly select 5 angles from the given angle range to rotate each sample in the ISAR test set in the image domain to obtain the final ISAR test set; Finally, perform inverse Fourier transform on the ISAR test set to obtain the corresponding high-resolution range profile (HRRP) test set.

4. The method according to claim 1, wherein: In step (2a), constructing the deformation adjustment module is realized as follows: (2a1) Construct a convolutional module: First, set 5 convolutional layers, and the sizes of their convolutional kernels are 9×9, 7×7, 6×6, 3×3, and 3×3 in sequence, corresponding to 4, 8, 16, 32, and 64 feature maps respectively; Second, set 5 pooling layers, and the size of each pooling is 2×2, and the stride is 2; Finally, cross-connect each convolutional layer with each pooling layer, and cascade the output of the last pooling layer with the fully connected layer to form a convolutional module; (2a2) Cascade the convolutional module with the existing mesh generator and the existing sampler to form a deformation adjustment module.

5. The method according to claim 1, wherein: In step (2b), construct a complex convolutional neural network as follows: (2b1) Set 5 complex convolutional layers with convolutional kernel sizes of 9×9, 7×7, 6×6, 3×3, and 3×3 respectively, corresponding to 16, 32, 64, 128, and 256 feature maps; (2b2) Set 5 complex pooling layers with each pooling size of 2×2 and a stride of 2; (2b3) Cross-connect each complex convolutional layer with each complex pooling layer to form a complex convolutional neural network.

6. The method according to claim 1, characterized in that: In step (2b), construct a horizontal stripe transformer as follows: (2b4) Construct the convolutional module of the horizontal stripe transformer: Set 2 convolutional layers with convolutional kernel sizes of 7×7 and 3×3 respectively, corresponding to 8 and 16 feature maps; Set 1 max pooling layer with a pooling size of 2×2 and a stride of 2; Cascade the two convolutional layers with the max pooling layer to form the convolutional module of the horizontal stripe transformer; (2b5) Construct the feature extraction module of the horizontal stripe transformer: Construct a horizontal stripe attention module that evenly divides the input features into non-overlapping horizontal stripes of equal width; Construct a first feature extraction module formed by cascading the existing position encoding and the horizontal stripe attention module; Construct a second feature extraction module formed by cascading the existing position encoding and the existing multi-layer perceptron; Cascade the first feature extraction module, the second feature extraction module, the first feature extraction module, and the second feature extraction module in sequence to form the feature extraction module of the horizontal stripe transformer; (2b6) Cascade the convolutional module, the feature extraction module, and the fully connected layer to form a horizontal stripe transformer.

7. The method according to claim 1, characterized in that: In step (2c), the output features of the complex convolutional neural network and the horizontal stripe transformer in the feature fusion module are weighted and fused, which is achieved by the following formula: Where I ISAR is the output feature of the complex convolutional neural network, and I HRRP is the output feature of the horizontal stripe transformer, and I fusion is the fusion recognition feature obtained after weighted fusion, and α is the weight coefficient, represents the vector concatenation operation.

8. The method according to claim 1, wherein: In step (4), use the forward-backward propagation method to train the ISAR target deformation robust recognition network as follows: (4a) Randomly initialize the weight values; (4b) Calculate the value of the loss function L total by forward propagation; (4c) Calculate the loss function L using the backpropagation algorithm total Derivative with respect to the weights, and use the Adam algorithm to update the weights, iterating until the loss function L total Converges to obtain a trained ISAR target deformation robust recognition network.

9. The method according to claim 1, characterized in that: The recognition accuracy rate output in step (5) is expressed as follows: Among them, c represents the recognition accuracy rate of the test set, M represents the number of samples in the test set, h(·) represents the classification discrimination function, and t i represents the true category of the i-th test sample in the test set, and y i represents the network output result corresponding to the i-th test sample in the test set. When t i and y i are equal, h(t i , y i ) is equal to 1; otherwise, h(t i , y i ) is equal to 0.

Images