A Synthetic Aperture Radar Undersampling Imaging Method Based on SE-Unet

Abstract

The present invention discloses a synthetic aperture radar undersampling imaging method based on the SE-Unet neural network, which is applied to the field of radar technology. Aiming at the deficiencies of the Unet reconstructed image with long running time, insufficient retention of image details, and poor performance in suppressing azimuth ambiguity of complex background targets; the present invention combines SAR signal processing with a convolutional neural network with feature compression (SE-Unet) to reduce the computational complexity of SAR undersampling imaging, reduce the neural network training time, and improve the SAR undersampling imaging quality in complex scenarios. Compared with the traditional convolutional neural network reconstruction method, this method reduces the computational complexity of the network, reduces the network training time, and more completely reconstructs the image details in complex scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar, and particularly relates to an imaging technique of a synthetic aperture radar (abbreviation: SAR) under undersampling. Background Technique

[0002] Synthetic Aperture Radar (SAR) generally operates on airborne and spaceborne platforms. It belongs to a high-resolution microwave imaging system and has the advantages of all-weather and all-time operation. It is widely used in the field of maritime security, such as ship detection and surveillance. However, the finiteness of the actual radar system's own PRF and the non-ideal antenna pattern will lead to an inevitable azimuth ambiguity problem in SAR imaging methods, and this problem will be more prominent under undersampling conditions.

[0003] In the currently published literature, a representative method among the azimuth ambiguity suppression methods for SAR undersampling imaging is the method based on the U-shaped convolutional neural network (abbreviation: Unet) (Reference 1: Liu.Z, Wu.N, X.Liao, “SAR Image Restoration From Spectrum Aliasing by Deep Learning”, IEEE Access, vol.99, pp.1-1, 2020). And the SENet module is only used in the field of SAR target recognition and has not been applied in the aspect of azimuth ambiguity suppression for SAR undersampling imaging (Reference 2: S.Mei, X.Wei, B.Jin, and J.Guo, “SAR image target recognition based on senet depth separable convolutional neural network,” in 2021 2nd China International SAR Symposium (CISS), Shanghai, China, Nov. 2021, pp.1–6.doi:10.23919 / CISS51089.2021.9652272). Although the Unet method can improve the resolution of the imaging result and suppress the virtual image caused by azimuth ambiguity in a scene without complex background, there are some problems with this method. For example, the azimuth ambiguity suppression effect is not good when dealing with a complex background scene, the details of the complex background cannot be well retained, and the operation time is long and the calculation amount is large. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a synthetic aperture radar undersampling imaging method based on the SE-Unet neural network, which can more completely and accurately retain details while effectively improving the imaging quality.

[0005] The technical solution adopted by the present invention is: a synthetic aperture radar undersampling imaging method based on SE-Unet, including:

[0006] S1. Respectively construct the radar parameter set para_set normal and the radar parameter set para_set sub ;

[0007] The parameters included in the radar parameter set para_set normal are: the pulse repetition frequency under normal sampling, denoted as PRF normal ; the azimuth modulation frequency of the radar, denoted as K a ; the pulse wavelength emitted by the radar system, denoted as λ; the pulse width of the pulse emitted by the radar system, denoted as T r ; the range sampling rate of the radar system, denoted as F r ; the synthetic aperture time of the radar, denoted as T sar ; the effective speed of the radar platform movement, denoted as V r ; the range sampling interval of the radar system, denoted as ΔT; the azimuth sampling interval of the radar system, denoted as ΔA; the speed of light is denoted as c; the center frequency is denoted as f0; the Doppler center frequency, denoted as f DOP ;

[0008] The parameters included in the radar parameter set para_set sub are: the pulse repetition frequency corresponding to the undersampled image, denoted as PRF sub ; the azimuth modulation frequency of the radar, denoted as K a ; the pulse wavelength emitted by the radar system, denoted as λ; the pulse width of the pulse emitted by the radar system, denoted as T r ; the range sampling rate of the radar system, denoted as F r ; the synthetic aperture time of the radar, denoted as T sar ; the effective speed of the radar platform movement, denoted as V r ; the range sampling interval of the radar system, denoted as ΔT; the azimuth sampling interval of the radar system, denoted as ΔA; the speed of light is denoted as c; the center frequency is denoted as f0; the Doppler center frequency, denoted as f DOP ;

[0009] S2. Based on the radar parameter set para_set constructed in step S1 normal and the radar parameter set para_set subConstruct a training data set. Each set of data in the training data set is a pair of data, specifically: based on para_set normal and para_set sub The radar echo data matrices generated after their respective simulations are then imaged by the ωKA algorithm and sheared

[0010] S3. Construct the SE-Unet network;

[0011] S4. Use the training data set generated in step S2 to train the SE-Unet network constructed in step S3;

[0012] S5. Input the undersampled azimuth ambiguous image obtained by non-uniform sampling and imaging of the echo data in the S-SAR mode into the trained SE-Unet network to obtain an image with azimuth ambiguity removed.

[0013] Advantages of the present invention: The innovation of the present invention lies in proposing a SAR undersampling imaging method based on a feature compression U-shaped convolutional neural network (abbreviation: SE-Unet) method to address the deficiencies of long running time of Unet reconstructed images, insufficient retention of image details, and poor performance in suppressing azimuth ambiguity of complex background targets. The key to the method of the present invention is to construct a suitable SE-Unet network model set for network training to achieve SAR undersampling imaging in complex scenarios.

[0014] The advantages of the present invention are that it combines SAR signal processing with a convolutional neural network with feature compression (SE-Unet) to reduce the computational complexity of SAR undersampling imaging, reduce the neural network training time, and improve the SAR undersampling imaging quality in complex scenarios. Compared with the traditional convolutional neural network reconstruction method, the method of the present invention reduces the computational complexity of the network, reduces the network training time, and more completely reconstructs the image details in complex scenarios. Brief Description of the Drawings

[0015] Figure 1 is the flowchart of the solution of the present invention;

[0016] Figure 2 is the structural diagram of the SE-Unet network;

[0017] Among them, (a) is the structural diagram of the SENet module, and (b) is the overall structural diagram of the SE-Unet network;

[0018] Figure 3 is the comparison schematic diagram of SE-Unet network and Unet undersampling imaging under complex background;

[0019] Among them, (a) is the test input image, (b) is the ideal image, (c) is the Unet output image, and (d) is the SE-Unet output image. Detailed implementation mode

[0020] To facilitate the description of the content of the present invention, the following term definitions are made first:

[0021] Definition 1, SE-Unet network

[0022] The overall structure of the SE-Unet network is as Figure 2 (b) shown.

[0023] The overall mathematical model of the SE-Unet network is

[0024]

[0025]

[0026] where T(·) is the network equivalent conversion function, is the result of the network training for t + 1 times, x true,k is the k-th non-azimuth-blurred image data in the training set, y k is the k-th azimuth-blurred image data in the training set, is the result of y k after being input into the SE-Unet network and trained for t times. N is the total number of the training set, ζ is a parameter, and L is the L1-norm cost function of the network. As the number of iteration rounds of the network increases, the value of the cost function becomes smaller and smaller.

[0027] The SENet module in the SE-Unet network is as Figure 2 (a) shown.

[0028] The mathematical model of the SENet structure is divided into a fully connected operation F tr , a compression operation F sq , an expansion operation F ex , and a weight superposition operation F scale :

[0029]

[0030]

[0031]

[0032]

[0033] where ": " is a colon, used to indicate that the expression after the colon is the specific calculation formula of each operation before the colon. U and X are feature maps, is the fully connected function, u c(i, j) represents the element in the i-th row and j-th column of the c-th two-dimensional matrix in U, and z is the feature obtained by compressing the feature map U. z c is the c-th element of z, W'×H' and W×H are the dimensions of the feature map, C' and C are the number of channels, S is the weight of the obtained feature map, σ(·) is the sigmoid function, δ(·) is the relu function, W1, W2 are parameters, is the feature map with weights superimposed, represents the set of complex numbers, W1, W2 are dimensions, r is the scaling parameter.

[0034] The present invention mainly uses the method of simulation experiments to verify the feasibility of the scheme, and all steps and conclusions are verified correctly on MATLAB R2018b. The specific implementation steps are as follows:

[0035] Step 1: Initialize parameters

[0036] Initialize the radar parameters required for the subsequent steps, including: the pulse repetition frequency (PRF) corresponding to the undersampled image, denoted as PRF sub ; the pulse repetition frequency under normal sampling, denoted as PRF normal ; the azimuth chirp rate of the radar, denoted as K a ; the pulse wavelength emitted by the radar system, denoted as λ; the pulse width of the pulse emitted by the radar system, denoted as T r ; the range sampling rate of the radar system, denoted as F r ; the synthetic aperture time of the radar, denoted as T sar ; the effective speed of the radar platform movement, denoted as V r ; the range sampling interval of the radar system, denoted as ΔT; the azimuth sampling interval of the radar system, denoted as ΔA; the speed of light is denoted as c; the center frequency is denoted as f0; the Doppler center frequency is denoted as f DOP ; the parameters are shown in Table 1:

[0037] Table 1 Radar parameter list

[0038] Radar Parameters Symbol Value Unit Transmission Pulse Width <![CDATA[T r > 41.75 μs Range Sampling Rate <![CDATA[F r > 32.317 MHz Effective Radar Velocity <![CDATA[V r > 7062 m / s Radar Operating Wavelength λ 0.057 m Azimuth Chirp Rate <![CDATA[K a > 1733 Hz / s Synthetic Aperture Time <![CDATA[T sar > 6.5956 ms Normal Sampling Azimuth Sampling Rate <![CDATA[PRF normal > 1256.98 Hz Under-Sampling Azimuth Sampling Rate <![CDATA[PRF sub > 75 Hz Range Sampling Time Interval ΔT 0.031 μs Azimuth Sampling Time Interval ΔA 0.080 μs Center Frequency <![CDATA[f0]]> 5300 MHz Doppler Center Frequency <![CDATA[f DOP > -6900 Hz

[0039] Step 2: Generate training data and test data

[0040] Construct training and test data sets according to the radar parameters provided in Step 1. The settings of multiple training data sets are as follows:

[0041] Using PRF normal and other required parameters to form the radar parameter set para_set normal Using PRF sub and other required parameters to form the radar parameter set para_set subThe two sets of radar parameters para_set obtained normal and para_set sub only differ in PRF. The parameters used for the test data to be processed in the present invention are para_set sub Generate 500 sets of training data with a resolution of 1024*128. Each training pair includes the undersampled image Y0 obtained by non-uniformly sampling and imaging the echo data in the S-SAR mode, and the desired image obtained by uniformly sampling and imaging the echo data in the SAR mode

[0042] The test data set is set as follows:

[0043] The test data set consists of undersampled images obtained by non-uniformly sampling and imaging the echo data of a scene with a resolution of 512*256 in the S-SAR mode

[0044] Step 3: Train the SE-Unet network

[0045] The structure of the SE-Unet network is as shown in the appendix Figure 2 It has a left-right symmetric structure. The left side is the compression path, the right side is the expansion path, and the middle is the SENet module. The compression path includes two parts: convolution and max pooling, which are used to extract features from the data. The expansion path includes three parts: upsampling, concatenation, and convolution, which are used to collect the features extracted by the compression path. The feature map before max pooling in the compression path will pass through the SENet module, and the SENet outputs a superimposed feature map with weights of the same size, which is concatenated with the feature map after upsampling in the corresponding expansion path to improve the feature extraction ability. In addition, there is a copy and add operation at the bottom of the SE-Unet. The feature map after the last max pooling operation is copied using a 1×1 convolution, and the copied feature map is superimposed on the feature map before the first upsampling to enhance image details

[0046] During the training process, the cost function Cost Function is selected as the L1-norm. The training parameters of the entire SE-Unet network are set as shown in Table 2. The training data set obtained in Step 2 is input into the network for training to obtain a trained network

[0047] Table 2 List of training parameter settings for the SE-UNE network

[0048] Network Parameters Value Learning Rate 0.001 Batch Size 4 Number of Epochs 100 Scaling Parameter r 16

[0049] The trained SE-Unet network has the ability to suppress azimuth ambiguity. The undersampled azimuth ambiguous image obtained by non-uniformly sampling and imaging the echo data in the S-SAR mode is input into the trained SE-Unet network to obtain an image with azimuth ambiguity removed

[0050] Step 4: Input the test data into the trained generation network to obtain the final result

[0051] Input the test data into the generation network obtained in Step 3 to obtain the final test result.

[0052] Through the above steps, the work of removing azimuth ambiguity in SAR imaging based on the SE-Unet network is completed. The effectiveness and universality of the above method are verified by three groups of test data (including land scenes, semi-ocean semi-land scenes, and ocean scenes).

[0053] Table 3 shows the comparison of the defocusing performance and reconstruction performance between the SE-Unet network and the Unet network under complex backgrounds. Table 4 shows the comparison of the running time and computational complexity between the SE-Unet network and the Unet network. Figure 3 For scenes with complex backgrounds, the SAR undersampled imaging results of the SE-Unet network and the Unet network are shown.

[0054] Table 3 Comparison of the performance between the SE-Unet network and the Unet

[0055] Method rMSE (dB) TAR Original -2.1882 4.7725 U-net -1.2246 5.559 SE-Unet -9.0217 21.6357

[0056] Table 4 Comparison of the running time and computational complexity between the SE-Unet network and the Unet network

[0057] Method Running Time (ms) Computational Load UNet <![CDATA[9.2×10 6 > <![CDATA[6.82×10 5 ×MNL]]> SE-Unet <![CDATA[2.3×10 6 > <![CDATA[1.02×10 5 ×MNL]]>

[0058] From Table 3, Table 4, and Figure 3 it can be concluded that the running time of this model is about one-fourth of that of the Unet, and the computational complexity is one-sixth of that of the Unet; the rMSE (relative mean square error) of this network is much smaller than that of the Unet, indicating that the image reconstruction performance of this network is better than that of the Unet; the TAR (true-to-blur target ratio) of this network is greater than that of the Unet, indicating that the false target suppression effect of this network is better than that of the Unet. The test results show that compared with the Unet network, the performance of the SE-Unet network has been greatly improved in terms of rMSE and TAR. Therefore, the method proposed in the present invention can effectively remove azimuth ambiguity in SAR imaging, while retaining the complex background details in the original scene more completely and accurately, and has a shorter running time and a smaller computational complexity.

[0059] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, various changes and modifications can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A synthetic aperture radar undersampling imaging method based on SE-Unet, characterized in that Including: S1. Construct a radar parameter set para_set respectively normal and the radar parameter set para_set sub ; The radar parameter set para_set normal The parameters included are: the pulse repetition frequency under normal sampling, denoted as PRF normal ; the azimuth modulation frequency of the radar, denoted as K a ; the pulse wavelength emitted by the radar system, denoted as λ; the pulse width of the pulse emitted by the radar system, denoted as T r ; the range sampling rate of the radar system, denoted as F r ; the synthetic aperture time of the radar, denoted as T sar ; the effective velocity of the radar platform movement, denoted as V r ; the range sampling interval of the radar system, denoted as ΔT; the azimuth sampling interval of the radar system, denoted as ΔA; the speed of light is denoted as c; the center frequency is denoted as f0; the Doppler center frequency is denoted as f DOP ; The radar parameter set para_set sub includes the following parameters: the pulse repetition frequency corresponding to the undersampled image, denoted as PRF sub ; the azimuth modulation frequency of the radar, denoted as K a ; the pulse wavelength emitted by the radar system, denoted as λ; the pulse width of the radar system, denoted as T r ; the range sampling rate of the radar system, denoted as F r ; the synthetic aperture time of the radar, denoted as T sar ; the effective velocity of the radar platform movement, denoted as V r ; the range sampling interval of the radar system, denoted as ΔT; the azimuth sampling interval of the radar system, denoted as ΔA; the speed of light is denoted as c; the center frequency is denoted as f0; the Doppler center frequency is denoted as f DOP ; S2. Based on the radar parameter set para_set constructed in step S1 normal and the radar parameter set para_set sub Construct a training data set. Each set of data in the training data set is a pair of data, and this pair of data is specifically: based on para_set normal and para_set sub The data matrix obtained by imaging and shearing the radar echoes generated after their respective simulations through the ωKA algorithm S3. Construct an SE-Unet network; the SE-Unet network described in step S3 has a left-right symmetric structure, with the left side being the compression path and the right side being the expansion path; The compression path includes two parts: convolution and max pooling, which are used to extract features from the data; The expansion path includes three parts: upsampling, concatenation, and convolution, which are used to collect the features extracted by the compression path; The feature map before max pooling in the compression path passes through the SENet module to output a weighted feature map of the same size, which is concatenated with the feature map after upsampling in the corresponding expansion path; At the bottom of the SE-Unet, there is also an add operation. The feature map after the last max pooling operation is replicated by a 1×1 convolution, and the replicated feature map is superimposed on the feature map before the first upsampling; The SENet module includes: a fully connected operation unit, a compression operation unit, an expansion operation unit, and a weight superimposition operation unit; the input of the fully connected operation unit is the feature map before max pooling in the current compression path, the output of the fully connected operation unit is used as the input of the compression operation, the output of the compression operation unit is used as the input of the expansion operation unit, the output of the expansion operation unit and the output of the fully connected operation unit are used as the input of the weight superimposition operation unit, and the output of the weight superimposition operation unit is a weighted feature map; The calculation process of the fully connected operation unit is: Among them, F tr is a fully connected operation, is the fully connected function, X represents the feature map before the maximum pooling of the current compression path, U represents the feature map obtained after the fully connected operation, represents the set of complex numbers, W×H is the dimension corresponding to X, C is the number of channels corresponding to X, W′×H′ is the dimension corresponding to U, and C′ is the number of channels corresponding to U; The calculation process of the compression operation unit is: Among them, F sq is the compression operation, z is the feature obtained after U undergoes the compression operation, z c is the c-th element of z, u c (i, j) represents the element in the i-th row and j-th column of the c-th two-dimensional matrix in U; The calculation process of the expansion operation unit is: Among them, F ex is an expansion operation, S is the obtained feature map weight, σ(·) is the sigmoid function, δ(·) is the relu function, W1 and W2 are dimensions, r is the scaling parameter; The calculation process of the weight superimposition operation unit is: Among them, F scale is the weight superposition operation, is the feature map with weights superimposed; S4. Use the training dataset generated in step S2 to train the SE-Unet network constructed in step S3; S5. Input the undersampled azimuth ambiguous image obtained by non-uniform sampling imaging of the echo data in the S-SAR mode into the trained SE-Unet network to obtain an image with azimuth ambiguity removed.

2. The synthetic aperture radar undersampling imaging method based on SE-Unet according to claim 1, wherein, During the training process of the SE-Unet network, the cost function is selected as the L 1- norm.

Images