Intelligent Prediction Method, Device and Storage Medium for Target RCS Based on SAR Images

Abstract

The present invention discloses a method, device and storage medium for intelligent prediction of target RCS based on SAR images, which relates to the field of microwave detection and is used to improve the generalization ability of detecting target RCS based on SAR images. The present invention obtains the SAR images and total RCS of a sample model, extracts each group of scattering centers of the sample model in the SAR images, superimposes and stacks their echo signals to obtain first echo data and second echo data, obtains the RCS of the second echo data based on the RCS calibration result, then trains a PINN to obtain a prediction model, and uses the prediction model to predict the total RCS according to the SAR image of the target to be measured. The present invention has good RCS prediction accuracy and generalization performance.

Description

Technical Field

[0001] The present invention relates to the field of microwave detection, and in particular to an intelligent prediction method, device, and storage medium for the Radar Cross Section (RCS) of a target based on Synthetic Aperture Radar (SAR) images. Background Art

[0002] Detecting the RCS of a target is of great significance for target recognition (such as ship type, etc.). Correspondingly, detecting the RCS of a target also has great application value for researching target stealth technology, etc.

[0003] SAR can obtain high-resolution SAR images, and SAR images can reflect the backscattering characteristics of the target. Therefore, it can be considered to detect the RCS of the target through SAR images.

[0004] Currently, most of the research on SAR images is based on the RCS of the target at some known angles, and through means such as simulation, the RCS of the target at unknown angles is obtained, and its prediction accuracy for the RCS of the target is relatively high. However, such methods need to first obtain the RCS of the target under some angles to be measured, and are not suitable for detecting targets with unmeasured RCS, and the generalization ability is insufficient. Summary of the Invention

[0005] The object of the present invention is to provide an intelligent prediction method, device, and storage medium for the RCS of a target based on SAR images to improve the generalization ability of detecting the RCS of a target based on SAR images for all or part of the above problems.

[0006] The technical solution adopted by the present invention is as follows:

[0007] An intelligent prediction method for the RCS of a target based on SAR images, which includes:

[0008] Obtain the facet RCS and total RCS of the sample model at different azimuth angles;

[0009] Obtain SAR images based on the facet RCS;

[0010] Perform RCS calibration on the SAR images;

[0011] Extract the scattering centers of the sample model in the SAR images, and each group of scattering centers generates an echo signal respectively;

[0012] Superimpose and stack each group of the echo signals respectively to obtain first echo data and second echo data;

[0013] Calculate the RCS of the second echo data based on the RCS calibration result;

[0014] With the goal of minimizing the total loss of the physics-informed neural network, using the total RCS and the first echo data as the ground truth, and the RCS of the second echo data as the input, train the physics-informed neural network to obtain a prediction model;

[0015] Use the prediction model to predict the total RCS of the target to be measured based on the SAR image of the target to be measured.

[0016] To solve all or part of the above problems, the present invention also provides a storage medium storing a computer program, and when the computer program is run, the above-mentioned intelligent prediction method for the target RCS based on the SAR image is executed.

[0017] To solve all or part of the above problems, the present invention also provides a device including a storage medium and a processor. A computer program is stored in the storage medium, and when the processor runs the computer program, the above-mentioned intelligent prediction method for the target RCS based on the SAR image is executed.

[0018] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are:

[0019] The intelligent prediction method for the target RCS based on the SAR image in the present application uses the physics-informed neural network as the prediction model architecture, which can train a prediction model with excellent performance under the condition of limited target detection samples, and can accurately and comprehensively predict the unknown angles of the sample targets and the RCS of the unknown targets, with strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be described by way of examples with reference to the drawings, wherein:

[0021] Figure 1 is the training flowchart of the intelligent prediction method for the target RCS based on the SAR image provided by the embodiment of the present application.

[0022] Figure 2 is the schematic diagram of the RCS calibration result in the embodiment of the present application.

[0023] Figure 3 is the schematic diagram of constructing the second echo data in the embodiment of the present application.

[0024] Figure 4 is the architecture diagram of the physics-informed neural network provided by the embodiment of the present application.

[0025] Figure 5 is the sample example provided by the embodiment of the present application.

[0026] Figure 6 It is a comparison chart of the total RCS predicted by the embodiments of the present application for Samples 1, 2, and 3 under sampling at different azimuth angle intervals.

[0027] Figure 7 It is a comparison chart of the total RCS predicted by the embodiments of the present application for Sample 4 under different azimuth angle sampling intervals. Specific Embodiments

[0028] All features disclosed in this specification, or all steps in the disclosed methods or processes, except for mutually exclusive features and / or steps, can be combined in any manner.

[0029] Any feature disclosed in this specification (including any additional claims, abstract) can be replaced by other equivalent or similar-purpose alternative features unless specifically stated. That is, unless specifically stated, each feature is only an example of a series of equivalent or similar features.

[0030] Aiming at the problem of insufficient generalization ability of the current method for predicting the target RCS based on SAR images, the embodiments of the present application provide an intelligent prediction method, device, and storage medium for the target RCS based on SAR images to improve the generalization ability of the omnidirectional prediction of the target RCS based on SAR images.

[0031] In some embodiments, as Figure 1 shown, the intelligent prediction method for the target RCS based on SAR images includes the following processes:

[0032] S1. Obtain the element RCS and total RCS of the sample model at different azimuth angles.

[0033] As an alternative implementation, use the GO (Geometric Optics)-PO (Physical Optics) hybrid method to calculate the backscattering RCS of the sample model under the HH polarization of the radar wave, and obtain the element RCS and total RCS (Total RCS) of the sample model at different azimuth angles (in the range of 0-360°).

[0034] In addition, as an alternative embodiment of RCS calibration, at least two small targets are also set around the sample model as calibration points, and the RCS of each calibration point is known.

[0035] S2. Obtain the SAR image based on the element RCS.

[0036] Using the grid model, based on the element RCS, calculate the signal echo in the case of the radar looking directly sideward using the back projection algorithm (BPA) to obtain the SAR image.

[0037] As an alternative implementation of RCS calibration, in the obtained SAR image, there is a grayscale image including at least two set calibration points.

[0038] S3. Perform RCS calibration on the SAR image.

[0039] The so-called RCS calibration is to obtain the mapping relationship between the RCS and the amplitude value in the SAR image.

[0040] As an alternative implementation, this step S3 includes:

[0041] S3.1. Determine the amplitude value of each calibration point in the SAR image.

[0042] S3.2. Based on the amplitude value of each calibration point and the RCS of each calibration point, calibrate the mapping relationship between the RCS and the amplitude value of the SAR image.

[0043] In some feasible implementations, the method for calibrating the mapping relationship between the RCS and the amplitude value of the SAR image is to fit the mapping relationship between the RCS and the amplitude value of the SAR image based on the amplitude value of each calibration point and the RCS of each calibration point. And in some specific implementations, by using the linear regression method to fit the amplitude value of each calibration point and the RCS of each calibration point, the mapping relationship between the RCS and the amplitude value of the SAR image is fitted.

[0044] Such as Figure 2 (a) is an embodiment of the arranged calibration points. Assume that the arranged calibration points are four, namely a, b, c, and d. As an alternative implementation, all the calibration points are respectively located at: the corner positions relative to when the sample model is in the middle of the SAR image. For example Figure 2 (a), when the sample model is in the middle of the image, the four calibration points are distributed at the upper left corner and the lower right corner positions, so as to avoid the mutual interference caused by the sidelobes.

[0045] In the generated SAR image, read out the amplitude value of each calibration point respectively, and use p x to represent, where x ∈ {a, b, c, d}. At the same time, obtain the RCS corresponding to the four calibration points to form the data set required for RCS calibration. Use the linear regression method to fit the data points in the data set, with the linear equation as the fitting target, and fit out a first-order linear equation, expressed as:

[0046] σ x_dB = k × p x + b,

[0047] In the formula, σ x_dB represents the RCS in dB units, and k and b are the coefficients obtained by fitting. Such as Figure 2(b) shows the result of linear fitting of the amplitude values and RCS through four fixed calibration points. Among them, the overall amplitude value of the SAR image is normalized before fitting.

[0048] S4. Extract the scattering centers of each group of the sample model in the SAR image, and generate echo signals corresponding to each group of scattering centers respectively.

[0049] As an alternative implementation, the orthogonal matching pursuit (OMP) method is used to extract the scattering centers of the sample model in the SAR image. Depending on the different scattering characteristics of the sample model, several groups of scattering center parameters can be obtained. Corresponding echo signals can be generated respectively through each group of scattering center parameters.

[0050] S5. Superimpose and stack the echo signals of each group respectively to obtain the first echo data and the second echo data.

[0051] S5.1. Obtain the first echo data by superimposing the echo signals generated by each scattering center. The first echo data is a two-dimensional amplitude matrix, denoted as M ASC .

[0052] S5.2. Obtain the second echo data by stacking the echo signals generated by each scattering center on the image channel. The second echo data is a three-dimensional amplitude matrix, denoted as {M j}, j ∈ [1, n], where M j represents the echo data of the echo signal generated by the j-th scattering center, and n represents the number of scattering centers. As Figure 3 shown, where Figure 3 (a) each rectangle represents an M j , Figure 3 (b) is the stacked second echo data.

[0053] S6. Calculate the RCS of the second echo data based on the RCS calibration result.

[0054] In some feasible implementation manners, the method for calculating the RCS of the second echo data includes:

[0055] S6.1. Obtain the highest amplitude value of the echo signal corresponding to each scattering center respectively.

[0056] S6.2. Normalize the echo data of the echo signal corresponding to each scattering center based on the highest amplitude value corresponding to each scattering center.

[0057] Assume that the highest amplitude value of the echo data of the j-th scattering center is A j , for M j , the normalization process based on A j is expressed as:

[0058]

[0059] In the formula, represents the normalization result of M j .

[0060] Normalize the echo data corresponding to each scattering center, and the normalized second echo data is expressed as

[0061] S6.3. Calculate the RCS corresponding to each maximum amplitude value based on the RCS calibration result.

[0062] Obtain the RCS value corresponding to A j through the RCS calibration result, denoted as R j .

[0063] S6.4. Weight the RCS corresponding to the maximum amplitude value of each scattering center by using the normalized echo data of each scattering center, respectively, to obtain the RCS corresponding to the echo signal of each scattering center.

[0064] For the j-th scattering center, its normalized echo data is The RCS value corresponding to the maximum amplitude value is R j , then the RCS matrix S j corresponding to the echo signal of this scattering center is expressed as: S j = M j ·R j .

[0065] S6.5. Stack the RCS corresponding to the echo signals of each scattering center to form the RCS of the second echo data, denoted as {S j}.

[0066] S7. With the goal of minimizing the total loss of the physics-informed neural network, using the total RCS and the first echo data as the ground truth, and using the RCS of the second echo data as the input, train the physics-informed neural network to obtain a prediction model.

[0067] The physics-informed neural network is one of the focuses for improving the prediction method of this application. In some alternative embodiments, as Figure 4 shown, the physics-informed neural network includes a convolutional neural network CNN and a deep neural network DNN;

[0068] The CNN includes seven convolutional layers, namely the first convolutional layer Conv0 to the seventh convolutional layer Conv6. Among them, the first convolutional layer Conv0, the second convolutional layer Conv1, the third convolutional layer Conv2, and the sixth convolutional layer Conv5 are all connected with a batch normalization layer BatchNormalization; the first convolutional layer Conv0, the third convolutional layer Conv2, the fourth convolutional layer Conv3, the fifth convolutional layer Conv4, and the sixth convolutional layer Conv5 are all connected with a dropout layer Dropout; after the first convolutional layer Conv0, the third convolutional layer Conv2, and the fourth convolutional layer Conv3, residual blocks Res (Res0, Res1, and Res2 in sequence) are connected; after the third convolutional layer Conv2, the fifth convolutional layer Conv4, and the sixth convolutional layer Conv5, squeeze-and-excitation layers Se (Se0, Se1, and Se2 in sequence) are connected; after the seventh convolutional layer Conv6, a PINN layer is connected;

[0069] The DNN includes four DNN subnets, namely the first DNN subnet Dnn0 to the fourth DNN subnet Dnn3. Among them, after the first DNN subnet Dnn0, the second DNN subnet Dnn1, and the fourth DNN subnet Dnn3, a batch normalization layer BatchNormalization is connected; after the second DNN subnet Dnn1 and the fourth DNN subnet Dnn3, a dropout layer Dropout is connected;

[0070] The CNN is connected to the DNN through a flattening layer Flatten. The output layer Output of the DNN is the output layer of the physics-informed neural network, and after training, it is the output layer of the prediction model.

[0071] The total loss of the physics-informed neural network includes the physics information loss PINN Loss and the data information loss Data Loss. As Figure 1 shown, the total RCS is used as the ground truth to optimize the data information loss Data Loss, and the first echo data is used as the ground truth to optimize the physics information loss PINN Loss.

[0072] During the training process, the matrix I output by the seventh convolutional layer Conv6 is obtained pred , and the number of filters in this layer is 1 (filters = 1), so I pred is single-channel data. I pred is regarded as the scattering center image reconstructed by the neural network. I pred is close to the target shape in the first echo data M ASC , but its amplitude is inverted by the neural network to predict the target RCS information. Based on this, the physics information loss PINN Loss is designed by comparing I pred with MASC To obtain the PINN Loss, the structural similarity loss function (SSIM) can be used according to the structural similarity of the target in the middle. The DataLoss can be obtained using the Huber function.

[0073] When calculating the PINN Loss, normalization processing is required. In some alternative embodiments, for I pred The normalization process is as follows:

[0074]

[0075] In the formula, I norm represents the result after normalization, represents the average value of I pred , σ I represents the standard deviation of the image, and ε1 is a small constant term used to prevent division by zero. The normalization process for M ASC is the same.

[0076] S8. Using the prediction model, predict the total RCS of the target to be measured based on the SAR image of the target to be measured.

[0077] For the SAR image of the target to be measured, refer to Figure 1 , obtain the corresponding second echo data through the orthogonal matching pursuit method, and then further obtain the RCS of the second echo data through the RCS calibration result, and input it into the prediction model to obtain the RCS of all directions of the target to be measured.

[0078] The embodiments of the present application also verify the performance of the above prediction method. As Figure 5 shown, the embodiments of the present application provide four different types of ship targets as experimental samples, which are sequentially labeled as samples 1-4 (denoted as shiptarget1, ship target2, ship target3, and ship target4) in sequence. Sample 4 does not participate in the training and is used for verifying the generalization ability.

[0079] 7879 groups of data ([[]] M ASC and the total RCS) from different azimuth angles (randomly sampled at a minimum interval of 0.1°, and the theoretical maximum number of samples is 10800) of samples 1, 2, and 3 are prepared by the methods mentioned above. Among them, 4076 groups of data are randomly assigned to participate in the training. Compared with the maximum number of samples, the proportion of data participating in the training is about 37.74%.

[0080] In the training data, 80% is used as the training set and 20% is used as the validation set. As the input layer data of the physics-informed neural network; M ASCParticipate in the calculation of PINN Loss; the total RCS is used as the true value data of the Output layer of the physics-informed neural network and participates in the calculation of Data Loss. The loss value of Data Loss is obtained using the Huber function. The total loss TotalLoss can be expressed as:

[0081] Total Loss = W1 × Data Loss + W2 × PINN Loss,

[0082] As an example, in the formula, W1 = W2 = 0.5 indicates that the loss values calculated by DataLoss and PINNLoss have the same weight during training.

[0083] After training, 3803 groups of data of samples No. 1, 2, and 3 that were not involved in training were used as comparison samples, and the total RCS of samples No. 1, 2, and 3 was predicted. The results are as Figure 6 (a) shown. In addition, to further evaluate the prediction performance, comparison samples with azimuth intervals of approximately 1°, 5°, and 10° were selected for further result display. The results are respectively as Figure 6 (b), Figure 6 (c) and Figure 6 (d) shown. Among them, the total RCS calculated using the GO-PO hybrid method is used as the true value and is represented by a blue broken line; in Figure 6 (a) and Figure 6 (b), the predicted values of the prediction model are represented by scatter points. To keep them clearly visible, in Figure 6 (c) and Figure 6 (d), the predicted values of the prediction model are connected by red lines to scatter the points. Since the test set is obtained by random sampling of the total samples, it is difficult to find all comparison samples at strict angular intervals. Since the RCS data used in the training process of the physics-informed neural network is represented in decibels (dB) units, Figure 6 the mean value shown in

[0084] is the average result of the dB unit of the directly calculated RCS value. This method can intuitively reflect the prediction performance of the model. From the comparison results, it can be seen that the prediction model shows excellent prediction performance at different sampling intervals. Figure 6 As can be seen, the prediction model has a better prediction for samples with a stronger median value of the total RCS, and the predicted points with better performance under continuous azimuth angles are in line with the trend of the true values. From the perspective of the mean value, the prediction error of all samples is less than 1 dBsm, and there is no obvious difference in the prediction performance under different azimuth sampling intervals, which verifies that the prediction model has good prediction performance under the training conditions of limited target detection samples.

[0085] The generalization performance of the prediction model was tested using sample No. 4 that had never participated in training. The prediction results are asFigure 7 As shown. It can be seen that when the sampling intervals are approximately 5° and 10° (5° corresponds to Figure 7 (a), 10° corresponds to Figure 7 (b)), the mean error predicted by the prediction model is relatively high compared to Samples 1, 2, and 3 that participated in the training. However, it can still basically match the change trend of the true value at continuous azimuth angles, especially for samples with a relatively strong median total RCS, the prediction effect is better. This indicates that the prediction model has good generalization ability.

[0086] For the same purpose, an embodiment of the present application also provides a storage medium that stores a computer program. When the computer program is run, it executes the method for intelligent prediction of the target RCS based on SAR images in the above embodiment.

[0087] In addition, an embodiment of the present application also provides a device, including a storage medium and a processor. The storage medium stores a computer program. When the processor runs the computer program, it executes the method for intelligent prediction of the target RCS based on SAR images in the above embodiment.

[0088] The present invention is not limited to the foregoing specific embodiments. The present invention extends to any new feature or any new combination disclosed in this specification, as well as any new combination of steps of any new method or process disclosed.

Claims

1. An intelligent prediction method for target RCS based on SAR images, characterized in that, Including: Obtain the element radar cross section (RCS) and the total RCS of the sample model at different azimuth angles; Obtain a synthetic aperture radar (SAR) image based on the element RCS; Perform RCS calibration on the SAR image; Extract each group of scattering centers of the sample model in the SAR image, and each group of scattering centers generates an echo signal respectively; Superimpose and stack each group of the echo signals respectively. By superimposing the echo signals generated by each scattering center, first echo data is obtained. By stacking the echo signals generated by each scattering center on the image channel, second echo data is obtained; Calculate the RCS of the second echo data based on the RCS calibration result; With the goal of minimizing the total loss of the physics-informed neural network, using the total RCS and the first echo data as the ground truth, and using the RCS of the second echo data as the input, train the physics-informed neural network to obtain a prediction model; Use the prediction model to predict the total RCS of the target to be measured based on the SAR image of the target to be measured.

2. The intelligent prediction method for target RCS based on SAR images according to claim 1, wherein The SAR image contains at least two calibration points; Performing RCS calibration based on the SAR image includes: Determine the amplitude values of each of the calibration points in the SAR image; Based on the amplitude values of each of the calibration points and the RCS of each of the calibration points, calibrate the mapping relationship between the RCS and the amplitude value of the SAR image.

3. The intelligent prediction method for target RCS based on SAR images according to claim 2, characterized in that, The calibrating the mapping relationship between the RCS and the amplitude value of the SAR image based on the amplitude values of each of the calibration points and the RCS of each of the calibration points includes: Based on the amplitude values of each of the calibration points and the RCS of each of the calibration points, fit the mapping relationship between the RCS and the amplitude value of the SAR image.

4. The intelligent prediction method for target RCS based on SAR images according to claim 3, wherein By using the linear regression method to fit the amplitude values of each of the calibration points and the RCS of each of the calibration points, fit the mapping relationship between the RCS and the amplitude value of the SAR image.

5. The intelligent prediction method for the target RCS based on SAR images according to any one of claims 2-4, characterized in that, Each of the calibration points is respectively located at: the corner positions relative to the sample model when it is in the middle of the SAR image.

6. The intelligent prediction method for target RCS based on SAR images according to claim 1, characterized in that The total loss of the physics-informed neural network includes a physics information loss and a data information loss. The total RCS is used to optimize the data information loss, and the first echo data is used to optimize the physics information loss.

7. The intelligent prediction method for target RCS based on SAR images according to claim 1, wherein Calculating the RCS of the second echo data based on the RCS calibration result includes: Respectively obtain the highest amplitude values of the echo signals corresponding to each scattering center; Normalize the echo data of the echo signal corresponding to each scattering center based on the highest amplitude value corresponding to the scattering center; Calculate the RCS corresponding to each highest amplitude value respectively based on the RCS calibration result; Respectively weight the RCS corresponding to the highest amplitude value of each scattering center by using the normalized echo data of each scattering center to obtain the RCS corresponding to the echo signal of each scattering center; The RCS corresponding to the echo signals of each scattering center is stacked to form the RCS of the second echo data.

8. The intelligent prediction method for the target RCS based on SAR images according to claim 1, wherein The physics-informed neural network includes a convolutional neural network (CNN) and a deep neural network (DNN); The CNN includes seven convolutional layers. Among them, batch normalization layers are connected to the first convolutional layer, the second convolutional layer, the third convolutional layer, and the sixth convolutional layer; dropout layers are connected to the first convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer; residual blocks are connected after the first convolutional layer, the third convolutional layer, and the fourth convolutional layer; attention layers are connected after the third convolutional layer, the fifth convolutional layer, and the sixth convolutional layer; a PINN layer is connected after the seventh convolutional layer. The DNN includes four DNN subnets. Among them, batch normalization layers are connected after the first DNN subnet, the second DNN subnet, and the fourth DNN subnet; dropout layers are connected after the second DNN subnet and the fourth DNN subnet. The CNN is connected to the DNN through a flattening layer.

9. A storage medium storing a computer program, characterized in that, When the computer program is run, the intelligent prediction method for the target RCS based on SAR images according to any one of claims 1-8 is executed.

10. An intelligent RCS prediction device for targets based on SAR images, comprising a storage medium and a processor, wherein a computer program is stored in the storage medium, and is characterized in that, When the processor runs the computer program, the intelligent prediction method for the target RCS based on SAR images according to any one of claims 1-8 is executed.

Images