Multi-channel foresight SAR azimuth resolution enhancement method based on zero sample learning

Abstract

The invention discloses a zero sample learning-based multichannel forward-looking SAR azimuth resolution enhancement method, which comprises the following steps of: firstly, acquiring multichannel forward-looking SAR echo data of a to-be-imaged area, performing pulse compression processing, performing synthetic aperture imaging preprocessing to obtain a multichannel forward-looking SAR preliminary imaging result, then constructing a data set, dividing the data set into a training set and a test set, and finally performing zero sample learning on the training set and the test set; and constructing an adversarial network model, inputting the adversarial network model into the training set for training, finally inputting the test set into the trained model for network testing, and outputting a result which is a resolution enhancement result. The method provided by the invention overcomes the problem of low resolution of the azimuth angle of the adjacent area of the track during multi-channel SAR foresight imaging, does not have the problems of difficulty in parameter adjustment, large calculation amount and the like compared with an existing super-resolution algorithm, and also overcomes the problem of lack or difficulty in acquisition of high-resolution real radar image information in a deep learning method.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar imaging, and particularly relates to a method for enhancing the azimuth resolution of a multi-channel forward-looking SAR based on zero-shot learning. Background Art

[0002] Radar forward-looking imaging has important applications in fields such as aircraft autonomous landing, autonomous navigation, and forward-looking reconnaissance and guidance. However, due to the reasons of left-right ambiguity and small angle change, the existing monostatic synthetic aperture radar (SAR) or Doppler beam sharpening (DBS) technology has a forward-looking imaging blind area.

[0003] Multi-channel radar, which receives echoes through multiple channels in the azimuth direction, has the potential for forward-looking imaging. However, due to the limitation of the platform size, its azimuth resolution is usually poor. The literature "Ren Lingyun, Wu Di, Zhu Daiyin, etc. Forward-looking imaging based on iterative super-resolution estimation of airborne multi-channel radar. Journal of Radars: 1-13, 2023" can achieve forward-looking super-resolution imaging through iterative spectral estimation of a single snapshot, but there are problems such as difficult selection of iterative parameters and large computational complexity; the literature "Li Yueli, Liang Diannong, Huang Xiaotao. A method for forward-looking imaging of multi-channel deconvolution of a monopulse radar. Signal Processing, 2007, (05): 699-703" uses the method of multi-channel deconvolution for inverse imaging, and there are problems such as complex operation and noise sensitivity; the literature "W. Li, Z. Wang, R. Chen, Z. Li, J. Wu, and J. Yang, 'Traditional synthetic aperture processing assisted gan-like network for multi-channel radar forward-looking super-resolution imaging,' IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-13, 2024" realizes multi-channel radar forward-looking super-resolution imaging through a generative neural network, which can overcome problems such as difficult parameter adjustment and high computational complexity of conventional super-resolution algorithms. However, the training process of this network depends on high-resolution radar forward-looking images as references, which are often difficult to obtain in practical applications. Summary of the Invention

[0004] To solve the above problems, the present invention proposes a method for enhancing the azimuth resolution of a multi-channel forward-looking SAR based on zero-shot learning, aiming to achieve enhanced imaging of multi-channel forward-looking SAR through a learning method without high-resolution radar forward-looking images.

[0005] The technical solution adopted by the present invention is as follows: A multi-channel forward-looking SAR azimuth resolution enhancement method based on zero-shot learning, and the specific steps are as follows:

[0006] S1. Obtain multi-channel forward-looking SAR echo data of the area to be imaged based on original scene simulation;

[0007] S2. Based on step S1, perform pulse compression processing on the obtained echo data, and then perform synthetic aperture imaging preprocessing to obtain a preliminary imaging result of multi-channel forward-looking SAR;

[0008] S3. Based on step S2, construct a data set from the obtained preliminary imaging result of multi-channel forward-looking SAR, and divide it into a training set and a test set;

[0009] Based on the preliminary imaging result of multi-channel forward-looking SAR obtained in step S2, the imaging result is evenly cropped into three parts along the azimuth direction to form a left image Left, a middle image Middle, and a right image Right. The middle image Middle is used as the Y-domain image, and the left image Left and the right image Right are used as the X-domain images.

[0010] Then repeat the above cropping operation, batch process to obtain P pairs of image pairs, and set a ratio according to the actual situation, and randomly divide the data into a training set and a test set.

[0011] S4. Construct an adversarial network model and design a model loss function;

[0012] S5. Based on the network model constructed in step S4, input the training set divided in step S3 for network training;

[0013] The goal of training the model is to minimize the loss function. The training code is based on the PyTorch deep learning framework and adopts an image pool strategy, that is, when updating the parameters of the discriminator module, a new sample generated in this iteration is used, and a series of historical samples generated by the generator are input into the discriminator module together, and each accounts for 50%.

[0014] Then, input the training set into the network model for training and learning. Set to save the weight parameters once every n rounds according to the actual situation. Wait until the model loss converges, and at the same time update and save the weight parameters of the current last round. Use the PyTorch visualization tool visdom to monitor the training status in real time during training.

[0015] S6. Based on the trained network model obtained in step S5, input the test set divided in step S3 for network testing, and the output result is the resolution enhancement result;

[0016] Load the weight parameters of the last round saved in step S5 and the test set data into the network model, and perform tests on the test data set. Input the Y-domain image of the test set into the trained network model, and then combine the result output by the network with the corresponding X-domain image in the test set to output the test result, which is the final image resolution enhancement result.

[0017] Furthermore, the specific steps of step S1 are as follows:

[0018] First, set the original scenario, that is, set the channel configuration of the multi-channel forward-looking SAR to single transmit and multiple receive. Each channel is arranged on a platform with a height of h and is evenly arranged along the Y-axis direction. The platform flies in a straight line at a constant speed of v r to form a synthetic aperture with a length of L s . During the flight, the radar transmitting channel T x transmits linear frequency modulation pulses at a given pulse repetition frequency PRF, and each channel simultaneously receives the echo signal.

[0019] Suppose there is a point target P1(x0, y0, 0) in the imaging area. The coordinates of each channel can be expressed as (v r t a , y i , h). The transmit distance history and distance history expressions of each channel to the target P1 are as follows:

[0020]

[0021] where t a represents the slow time, y i represents the azimuth coordinate of the i-th receiving channel, i ∈ [1, N], and N represents the number of channels.

[0022] For the point target P1(x0, y0, 0), its distance history expression is as follows:

[0023]

[0024] Then the echo signal expressions received by multiple channels are as follows:

[0025]

[0026] where β0 represents the point target scattering coefficient, t r represents the fast time, w r represents the range envelope, K r represents the chirp rate of the transmitted signal, λ represents the wavelength of the transmitted signal, and c represents the speed of light.

[0027] Finally, Gaussian white noise is added based on the received echo signal according to actual requirements.

[0028] Further, the specific steps of step S2 are as follows:

[0029] Perform range-direction pulse compression on the echo data obtained in step S1, and set the matching function of pulse compression as Then the signal S after pulse compression c (y i ,t r ,t a ) is expressed as follows:

[0030]

[0031] Then, according to the echo model, perform synthetic aperture azimuth focusing processing. The signal S after pulse compression c (y i ,t r ,t a ) uses the back projection (BP) algorithm to perform coherent accumulation on each channel to obtain the azimuth focusing result of a single snapshot. Then, the imaging result f of the j-th snapshot bpj (t aj ,ρ,γ) is expressed as follows:

[0032]

[0033] Among them, t aj represents the sampling of the j-th slow time, and ρ and γ respectively represent the slant range and azimuth angle corresponding to the imaging grid.

[0034] Then, coherently superimpose the imaging results of different snapshots to obtain the non-ambiguous multi-channel forward-looking SAR imaging result f bp (ρ,γ), and the expression is as follows:

[0035]

[0036] Among them, M represents the number of snapshots.

[0037] Further, the specific steps of step S4 are as follows:

[0038] S41. Construct an adversarial network model;

[0039] Adopt the basic structure of a generative adversarial network to build the overall network model, including: a generator module and a discriminator module.

[0040] Among them, the overall network model structure connects the generator module and the discriminator module in series to form two groups of reverse symmetric loops, and the generator and the discriminator are shared.

[0041] One input of the network model is an X-domain image. The input X-domain image Input_X is converted into a Y-domain image Generated_Y by the generator, and then the reconstructed X-domain image Cyclic_X is obtained through the inverse generator. The difference between the two is characterized by the cyclic consistency loss. The second input of the network model is a Y-domain image. The input Y-domain image Input_Y is converted into an X-domain image Generated_X by the generator, and then the reconstructed Y-domain image Cyclic_Y is obtained through the inverse generator. The difference between the two is characterized by the cyclic consistency loss. The discriminators D X and D Y are responsible for judging whether the input image conforms to the original sample distribution of the corresponding image domain.

[0042] The generator module includes: an encoder, a converter, and a decoder. The generator module uses ReLU as the activation function, adds a dropout layer, uses instance normalization layers to normalize the features of the intermediate layers of the network, and introduces global residual connections.

[0043] Among them, the encoder extracts features from the input image using 3 convolutional layers, compresses the entire image into a feature vector, then converts the feature vector to another image domain through 9 residual blocks in the converter, and finally the decoder restores the low-level features from the feature vector to obtain the generated image. And a hybrid attention mechanism is introduced in the decoder, which is implemented through two cascaded sub-modules of spatial attention and channel attention.

[0044] The discriminator module adopts the PatchGAN method, traverses the entire image through a 70×70 sliding window, then uses 5 convolutional layers to extract features from the image patches, replaces all batch normalization layers with instance normalization layers, and uses LeakyReLU as the activation function.

[0045] S42. Based on step S41, design the model loss function;

[0046] The loss function of the model includes: adversarial loss, cyclic consistency loss, identity mapping loss, and frequency domain loss, which are specifically as follows:

[0047] (1) Adversarial loss L GAN (G A , D Y , X, Y);

[0048] The adversarial loss consists of two parts. For the mapping G A (x): X→Y, the expression of the adversarial loss is as follows:

[0049]

[0050] Among them, X and Y respectively represent two image domains, GA Denote the mapping G A (x): The generator of X→Y, where x and y represent the images belonging to the X domain and the Y domain respectively; D Y Denote the discriminator of the Y domain; p data (x) and p data (y) represent the sample distributions of the X domain and the Y domain respectively; Denote the mathematical expectation.

[0051] Similarly, for the mapping G B (y): Y→X, the adversarial loss is denoted as L GAN (G B , D X , Y, X).

[0052] Among them, G B Denote the inverse mapping G A of G B (y): The generator of Y→X, D X Denote the discriminator of the X domain.

[0053] (2) Cycle consistency loss L cycle (G A , G B );

[0054] For each image in the X domain, it will go through the cycle: x→G A (x)→G B (G A (x)).

[0055] Among them, the difference between x and G B (G A (x)) is the forward cycle consistency loss. Similarly, if the input is an image in the Y domain, it is called the backward consistency loss, and the expression of this cycle consistency loss is as follows:

[0056]

[0057] Among them, ||·||1 represents the L1 norm.

[0058] (3) Identity mapping loss L idt (G A , G B );

[0059] The expression of the identity mapping loss is as follows:

[0060]

[0061] (4) Frequency domain loss L fre (G A , G B );

[0062] The expression of the frequency-domain loss function is as follows:

[0063]

[0064] Among them, K and L represent the width and length of the reconstructed image and the input image, u and v represent the frequency-domain coordinates, corresponding to the frequency components of the image in the horizontal and vertical directions respectively, and F o (u, v), F r (u, v) represent the frequency-domain functions of the input image and the reconstructed image respectively, and w(u, v) represents the introduced frequency-domain weight matrix, and its defining expression is as follows:

[0065] w(u, v) = |F o (u, v) - F r (u, v)| γ (11)

[0066] Among them, γ ≥ 0 represents a scaling factor.

[0067] Finally, the expression of the overall loss function of the network model is as follows:

[0068]

[0069] Among them, λ1, λ2, and λ3 represent weight parameters, which are used to control the relative importance of the three loss functions of cycle consistency loss, identity mapping loss, and frequency loss.

[0070] Advantages of the present invention: The method of the present invention first obtains multi-channel forward-looking SAR echo data of the area to be imaged for pulse compression processing, then performs synthetic aperture imaging preprocessing to obtain a multi-channel forward-looking SAR preliminary imaging result, then constructs a data set, divides it into a training set and a test set, constructs an adversarial network model and inputs the training set for training, and finally inputs the test set into the trained model for network testing, and the output result is the resolution enhancement result. The method of the present invention utilizes the characteristic that the azimuth resolution of multi-channel forward-looking SAR imaging is spatially variant, and realizes azimuth resolution enhancement under the condition of no high-resolution real radar image based on internal learning of the pre-imaging result. Since the high-resolution real radar image information is not used in the whole training process, it can be called zero-sample learning. It overcomes the problem of low azimuth resolution in the track adjacent area during multi-channel SAR forward-looking imaging. Compared with the existing super-resolution algorithms, there are no problems such as difficult parameter adjustment and large computational amount, and it also overcomes the problem of lack or difficult acquisition of high-resolution real radar image information in deep learning methods. Description of the Drawings

[0071] Figure 1Flow chart of a multi-channel forward-looking SAR azimuth resolution enhancement method based on zero-shot learning according to the present invention.

[0072] Figure 2 Schematic diagram of the dataset construction process in an embodiment of the present invention.

[0073] Figure 3 Geometric model diagram of multi-channel SAR forward-looking imaging in an embodiment of the present invention.

[0074] Figure 4 Schematic diagram of the adversarial network model structure in an embodiment of the present invention.

[0075] Figure 5 Generator structure diagram in an embodiment of the present invention.

[0076] Figure 6 Schematic diagram of the hybrid attention module structure in an embodiment of the present invention.

[0077] Figure 7 Discriminator structure diagram in an embodiment of the present invention.

[0078] Figure 8 Schematic diagram of the learning process in an embodiment of the present invention.

[0079] Figure 9 Schematic diagram of the data processing process in an embodiment of the present invention.

[0080] Figure 10 Schematic diagram of the original observation scene in an embodiment of the present invention.

[0081] Figure 11 Forward-looking imaging result diagram of the original observation scene in an embodiment of the present invention.

[0082] Figure 12 Resolution enhancement result diagram in an embodiment of the present invention. Detailed implementation manners

[0083] The method of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0084] As Figure 1 shown, the flow chart of a multi-channel forward-looking SAR azimuth resolution enhancement method based on zero-shot learning according to the present invention is as follows:

[0085] As Figure 1 shown, the specific steps of the present invention are as follows:

[0086] S1. Obtain multi-channel forward-looking SAR echo data of the area to be imaged based on the simulation of the original scene;

[0087] In this embodiment, the high-resolution SAR image is first randomly cropped into the size of 128×128 to obtain the original observed scene image, and then the echo is obtained according to the multi-channel forward-looking SAR working mode.

[0088] S2. Based on step S1, the obtained echo data is subjected to pulse compression processing, and then synthetic aperture imaging preprocessing is performed to obtain the preliminary imaging result of the multi-channel forward-looking SAR;

[0089] The preliminary imaging result image is saved as the size of 128×128, which is the same as the randomly cropped image size of the imaging scene image in step S1.

[0090] S3. Based on step S2, a data set is constructed from the obtained preliminary imaging result of the multi-channel forward-looking SAR and divided into a training set and a test set;

[0091] Based on the preliminary imaging result of the multi-channel forward-looking SAR obtained in step S2, the imaging result is evenly cropped into three parts along the azimuth direction to form the left image Left, the middle image Middle, and the right image Right. The middle image Middle is used as the Y-domain image, and the left image Left and the right image Right are used as the X-domain images. The data set construction process is as Figure 2 shown.

[0092] Then, the above cropping operation is repeated. In this embodiment, 2400 pairs of image pairs are obtained through batch processing, and the ratio is set according to the actual situation. The data is randomly divided into a training set and a test set. In this embodiment, the numbers of the training set and the test set are 2000 pairs and 400 pairs respectively.

[0093] S4. Construct an adversarial network model and design the model loss function;

[0094] S5. Based on the network model constructed in step S4, the training set divided in step S3 is input for network training;

[0095] The goal of training the model is to minimize the loss function. The training code is based on the PyTorch deep learning framework and adopts the image pool strategy, that is, when updating the discriminator module parameters, a series of historical samples generated by the generator are input into the discriminator module together with the newly generated samples in this iteration, and each accounts for 50%, so as to reduce model oscillation and improve the stability of the training process. In this embodiment, the size of the buffer pool is set to 50. The inactivation ratio of the generator random inactivation layer is set to 0.5 to increase network randomness and prevent overfitting.

[0096] In this embodiment, the optimization algorithm uses the Adam optimization algorithm. The initial learning rate of 0.0002 is maintained for the first 100 epochs and linearly decays to 0 starting from the 100th epoch. The Batch Size can be modified according to the actual hardware situation. In this embodiment, it is set to 1, and the number of iterations is set to 200 epochs. The cycle consistency loss L cycle The weight λ1 is set to 10, and the identity mapping loss L idt The weight λ2 is set to 5, and the frequency loss L fre The weight λ3 is set to 5. The specific training parameters in this embodiment are shown in Table 2.

[0097] Table 2

[0098]

[0099] Then, the training set is input into the network model for training and learning. In this embodiment, a total of 200 rounds of training are performed. The weight parameters are saved every 5 iterations. When the model loss converges, the weight parameters of the current last round are updated and saved. During training, the PyTorch visualization tool visdom is used to monitor the training status in real time to adjust various parameters in a timely manner and avoid time consumption caused by incorrect operations. During the training process, the adversarial loss, cycle consistency loss, identity mapping loss, and frequency loss of each round can be obtained from equations (7), (8), (9), and (10) respectively. After 200 iterations, the loss value decreases significantly, and the model gradually tends to be stable.

[0100] S6. Based on the trained network model obtained in step S5, input the test set divided in step S3 for network testing, and the output result is the resolution enhancement result;

[0101] Load the weight parameters of the last round saved in step S5 and the test set data into the network model, and perform tests on the test data set. Input the Y-domain image of the test set into the trained network model, and then combine the output result of the network with the corresponding X-domain image in the test set to output the test result, which is the final image resolution enhancement result.

[0102] In this embodiment, step S1 is specifically as follows:

[0103] First, set the original scenario, that is, set the channel configuration of the multi-channel forward-looking SAR to single transmit and multiple receive. Each channel is arranged on a platform with a height of h and evenly arranged along the Y-axis direction. The platform flies at a constant speed of v r along the X-axis direction, forming a synthetic aperture with a length of L s . During the flight, the radar transmit channel T x transmits chirp pulses at a given pulse repetition frequency PRF, and each channel simultaneously receives the echo signals.

[0104] In this embodiment, the geometric configuration of the multi-channel forward-looking SAR is as Figure 3 shown, and the parameters of the multi-channel forward-looking SAR are shown in Table 1.

[0105] Table 1

[0106]

[0107] As Figure 3 shown, in an xyz space coordinate system, O represents the origin of the coordinate system. The forward-looking multi-channel radar flies at a constant speed v r = 400 m / s and altitude h = 5000 m along the X-axis direction in a uniform straight line, forming a synthetic aperture with a length of L s = 52.5 m. X1 represents the starting point of the synthetic aperture, and X2 represents the ending point of the synthetic aperture.

[0108] The central position of the target scene is X c = 5000 m, Y c = 0 m, Z c = 0 m, and the angle with the scene center is 45°. On the platform, each channel is evenly arranged along the Y-axis direction. The transmitting channel T x transmits signals at a pulse repetition frequency PRF = 400 Hz, and each receiving channel receives signals simultaneously.

[0109] During the simulation process of this embodiment, it is set that the wavelength of the radar transmitting signal is λ = 0.0315 m, the pulse width T r = 1 μs, and the linear frequency modulation pulse signal with a bandwidth B = 60 MHz. The number of sampling points N r in the range direction is 256, the number of channels N a is 16, and the number of sampling points N sa in the synthetic aperture dimension is 128.

[0110] First, for the working mode of the multi-channel forward-looking SAR, the echo data of the area to be imaged is obtained. Suppose there is a point target P1(x0, y0, 0) in the imaging area, and the coordinates of each channel can be expressed as (v r t a , y i , h). The transmitted range history and range history from each channel to the target P1 are expressed as follows:

[0111]

[0112] Among them, t a represents the slow time, y i represents the azimuth coordinate of the i-th receiving channel, i ∈ [1, N], and N represents the number of channels.

[0113] For the point target P1(x0, y0, 0), its range history expression is as follows:

[0114]

[0115] Then the echo signal expressions received by multiple channels are as follows:

[0116]

[0117] Among them, β0 represents the point target scattering coefficient, t r represents the fast time, w r represents the range envelope, K r represents the chirp rate of the transmitted signal, λ represents the wavelength of the transmitted signal, and c represents the speed of light.

[0118] Finally, Gaussian white noise is added based on the received echo signal according to actual requirements. In this embodiment, Gaussian white noise with a signal-to-noise ratio of 30 dB is added to the echo.

[0119] In this embodiment, the specific steps of step S2 are as follows:

[0120] Perform range-direction pulse compression on the echo data obtained in step S1, and set the matching function of the pulse compression as Then the signal S after pulse compression c (y i , t r , t a ) has the following expression:

[0121]

[0122] Then, according to the echo model, perform synthetic aperture azimuth focusing processing. The signal S after pulse compression c (y i , t r , t a ) uses the Back Projection (BP) algorithm to perform coherent accumulation on each channel to obtain the azimuth focusing result of a single snapshot. Then the imaging result f of the j-th snapshot bpj (t aj , ρ, γ) has the following expression:

[0123]

[0124] Among them, t aj represents the sampling of the j-th slow time, and ρ and γ represent the slant range and azimuth angle corresponding to the imaging grid, respectively.

[0125] Then, the different snapshot imaging results are coherently superposed to obtain the non-blurred multi-channel forward-looking SAR imaging result f bp (ρ, γ), and the expression is as follows:

[0126]

[0127] Among them, M represents the number of snapshots.

[0128] In this embodiment, the step S4 is specifically as follows:

[0129] S41. Construct an adversarial network model;

[0130] Adopt the basic structure of the generative adversarial network, and utilize its end-to-end mapping characteristic to achieve super-resolution. Build the overall network model, including: a generator module and a discriminator module.

[0131] Among them, the overall network structure is as Figure 4 shown. The overall network model structure connects the generator module and the discriminator module in series, forms two groups of reverse-symmetric loops, and shares the generator and the discriminator.

[0132] One input of the network model is the X-domain image. The input X-domain image Input_X is converted into the Y-domain image Generated_Y by the generator, and then the reconstructed X-domain image Cyclic_X is obtained through the reverse generator. The difference between the two is characterized by the cycle consistency loss. The second input of the network model is the Y-domain image. The input Y-domain image Input_Y is converted into the X-domain image Generated_X by the generator, and then the reconstructed Y-domain image Cyclic_Y is obtained through the reverse generator. The difference between the two is characterized by the cycle consistency loss. The discriminators D X and D Y are responsible for judging whether the input images conform to the original sample distribution of the corresponding image domain.

[0133] The generator network structure diagram is as Figure 5 shown. The generator module includes: an encoder, a converter, and a decoder. The generator module uses ReLU as the activation function, adds a random inactivation layer (Dropout layer) to increase the randomness of the network and prevent overfitting. It uses the instance normalization layer (Instance Normalization) to normalize the features of the intermediate layer of the network, replaces the commonly used batch normalization layer in the image classification task to maintain the independence between each image instance, accelerates the convergence of the model, and introduces a global residual connection to make the network converge faster, enhance the generalization ability of the model, and prevent overfitting.

[0134] Among them, the encoder uses 3 convolutional layers (Conv layers) to extract features from the input image, compresses the entire image into a feature vector, then converts the feature vector to another image domain through 9 residual blocks in the transformer, and finally the decoder recovers low-level features from the feature vector through a deconvolution layer (DeConv layer) to obtain the generated image. And since the azimuth resolution of the preliminary imaging result is variable, this means that the processing difficulty in different regions is different. Therefore, an attention mechanism is adopted in the decoder. Considering that separate spatial attention may not be able to effectively distinguish important and irrelevant channels, while channel attention can improve the feature expression ability, a hybrid attention module is adopted here, and it is implemented through two cascaded sub-modules of spatial attention and channel attention, as Figure 6 shown.

[0135] The discriminator network structure is as Figure 7 shown. The discriminator module adopts the PatchGAN method, traverses the entire image through a 70×70 sliding window pane to focus on local information in the image. Then 5 convolutional layers are used to extract features from the image patches to capture local texture and spatial information. And all batch normalization layers (Batch Norm layers) are replaced with instance normalization layers (Instance Norm layers), and LeakyReLU is used as the activation function.

[0136] S42. Based on step S41, design the model loss function;

[0137] The loss function of the model includes: adversarial loss, cycle consistency loss, identity mapping loss, and frequency domain loss, which are specifically as follows:

[0138] Among them, the adversarial loss aims to make the generated image as close as possible to the real data distribution; in order to make the converted image retain the original image information and prevent mode collapse, the cycle consistency loss is introduced; the identity mapping loss is a supplement to the cycle consistency loss, and its purpose is to further constrain the generated image to be consistent with the original input in content. The frequency domain loss function is used to narrow the difference between the input image and the reconstructed image in the frequency domain and improve the enhancement performance.

[0139] (1) Adversarial loss L GAN (G A ,D Y ,X,Y);

[0140] The purpose of the adversarial loss is to let the generator generate a realistic image so that the discriminator cannot distinguish the generated image from the real image. The adversarial loss consists of two parts. For the mapping G A (x): X→Y, the expression of the adversarial loss is as follows:

[0141]

[0142] Among them, X and Y respectively represent two image domains, and G A represents the generator of the mapping G A (x): X → Y, where x and y respectively represent the images belonging to the X domain and the Y domain; D Y represents the discriminator in the Y domain; p data (x) and p data (y) respectively represent the sample distributions in the X domain and the Y domain; represents the mathematical expectation.

[0143] Similarly, for the mapping G B (y): Y → X, the adversarial loss is expressed as L GAN (G B , D X , Y, X).

[0144] Among them, G B represents the generator of the inverse mapping G A (y): Y → X of G B , and D X represents the discriminator in the X domain.

[0145] (2) Cyclic consistency loss L cycle (G A , G B );

[0146] The cyclic consistency loss ensures that the original information is not lost when the transformed image is reconstructed. This loss constrains the generator of the model, so that when the generator performs image transformation between the two domains, it still retains the key structure and content features of the original image, rather than just performing random style transformation.

[0147] For each image in the X domain, it will go through the cycle: x → G A (x) → G B (G A (x)).

[0148] Among them, the difference between x and G B (G A (x)) is the forward cyclic consistency loss. Similarly, if the input is an image in the Y domain, it is called the backward consistency loss, and the expression of this cyclic consistency loss is as follows:

[0149]

[0150] Among them, ||·||1 represents the L1 norm.

[0151] (3) Identity mapping loss L idt (G A , GB )

[0152] The same mapping loss is complementary to the cycle consistency loss, which can measure the difference between the predicted value and the true value and is used to enhance the stability of the model. The expression is as follows:

[0153]

[0154] (4) Frequency domain loss L fre (G A , G B )

[0155] The frequency domain loss function is used to reduce the difference between the input image and the reconstructed image in the frequency domain, which can improve the enhancement performance. The expression is as follows:

[0156]

[0157] Among them, K and L represent the width and length of the reconstructed image and the input image, u and v represent the frequency domain coordinates, corresponding to the frequency components of the image in the horizontal and vertical directions respectively, F o (u, v), F r (u, v) represent the frequency domain functions of the input image and the reconstructed image respectively, and w(u, v) represents the introduced frequency domain weight matrix. Its definition expression is as follows:

[0158] w(u, v) = |F o (u, v) - F r (u, v)| γ (11)

[0159] Among them, γ ≥ 0 represents a scaling factor used to flexibly adjust the attention to high-frequency and low-frequency details.

[0160] Finally, the expression of the overall loss function of the network model is as follows:

[0161]

[0162] Among them, λ1, λ2, and λ3 represent weight parameters used to control the relative importance of the three loss functions of cycle consistency loss, same mapping loss, and frequency loss.

[0163] In this embodiment, the step S6 is specifically as follows:

[0164] The zero-shot learning process of this embodiment is as Figure 8 shown. Load the weight parameters of the last round into the network model, process the data according to the Figure 9 shown process, and the output result is the enhanced imaging result. As Figure 10, as shown in FIGS. 11 and 12, are the original observed scene after 30 dB Gaussian white noise in forward-looking imaging, the forward-looking pre-imaging result, and the resolution enhancement result, respectively. It can be seen that the overall structure of the resolution enhancement result map is basically the same as that of the original scene, the scene contour is clear, especially the azimuth resolution in the area adjacent to the track is greatly improved.

[0165] In summary, the method of the present invention utilizes the characteristic of the azimuth resolution variation with space in multi-channel forward-looking SAR imaging, and realizes the azimuth resolution enhancement without high-resolution real radar images based on internal learning of the pre-imaging result. Since no high-resolution real radar image information is used in the whole training process, it can be called zero-shot learning. It overcomes the problem of low azimuth resolution in the area adjacent to the track during multi-channel SAR forward-looking imaging. Compared with the existing super-resolution algorithms, there are no problems such as difficult parameter adjustment and large computational amount, and it also overcomes the problem of lack or difficult acquisition of high-resolution real radar image information in deep learning methods.

[0166] The above is the specific implementation manner of the present invention. However, the protection scope of the present invention is not limited thereto. Any person skilled in the art, without departing from the essence of the present invention, makes equivalent deformations or substitutions, which shall fall within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A method for enhancing the azimuth resolution of multi-channel forward-looking SAR based on zero-shot learning, the specific steps are as follows: S1. Obtain the multi-channel forward-looking SAR echo data of the area to be imaged based on the original scene simulation; S2. Based on step S1, perform pulse compression processing on the obtained echo data, and then perform synthetic aperture imaging preprocessing to obtain the preliminary imaging result of the multi-channel forward-looking SAR; S3. Based on step S2, construct a data set through the obtained preliminary imaging result of the multi-channel forward-looking SAR, and divide it into a training set and a test set; Based on the preliminary imaging result of the multi-channel forward-looking SAR obtained in step S2, the imaging result is evenly cropped into three parts along the azimuth direction to form the left image Left, the middle image Middle, and the right image Right; the middle image Middle is used as the Y-domain image, and the left image Left and the right image Right are used as the X-domain images; Then repeat the above cropping operation, batch process to obtain P pairs of image pairs, and set the ratio according to the actual situation, and randomly divide the data into a training set and a test set; S4. Construct an adversarial network model and design the model loss function; S5. Based on the network model constructed in step S4, input the training set divided in step S3 for network training; The goal of training the model is to minimize the loss function. The training code is based on the PyTorch deep learning framework and adopts an image pool strategy, that is, when updating the parameters of the discriminator module, a new sample generated in this iteration is used while a series of historical samples generated by the generator are input into the discriminator module together, and each accounts for 50%; Then, input the training set into the network model for training and learning; set to save the weight parameters once every n rounds according to the actual situation. Wait until the model loss converges, and at the same time update and save the weight parameters of the current last round. Use the PyTorch visualization tool visdom to monitor the training status in real time during training; S6. Based on the trained network model obtained in step S5, input the test set divided in step S3 for network testing, and the output result is the resolution enhancement result; Load the weight parameters of the last round saved in step S5 and the test set data into the network model, and perform testing on the test data set. Input the test set Y-domain image into the trained network model, and then combine the result output by the network with the corresponding X-domain image in the test set to output the test result, which is the final image resolution enhancement result.

2. A method for enhancing the azimuth resolution of a multi-channel forward-looking SAR based on zero-shot learning according to claim 1, characterized in that The specific content of step S1 is as follows: First, set the original scenario, that is, set the channel configuration of the multi-channel forward-looking SAR to single-transmission and multi-reception. Each channel is arranged on a platform at a height of h and evenly arranged along the Y-axis direction. The platform flies in a straight line at a uniform speed of v r to form a synthetic aperture with a length of L s . During the flight, the radar transmitting channel T x emits a chirp pulse at a given pulse repetition frequency PRF, and each channel simultaneously receives the echo signal; Suppose there is a point target P1(x0, y0, 0) in the imaging area, and the coordinates of each channel can be expressed as (v r t a , y i , h). The emission distance history and the distance history from each channel to the target P1 are expressed as follows: where t a represents the slow time, and y i represents the azimuth coordinate of the i-th receiving channel, where i ∈ [1, N] and N represents the number of channels; For the point target P1(x0, y0, 0), its distance history expression is as follows: Then the echo signal expressions received by multiple channels are as follows: Among them, β0 represents the point target scattering coefficient, t r represents the fast time, w r represents the range envelope, K r represents the chirp rate of the transmitted signal, λ represents the wavelength of the transmitted signal, and c represents the speed of light; Finally, add Gaussian white noise to the received echo signal according to actual requirements.

3. A method for enhancing the azimuth resolution of a multi-channel forward-looking SAR based on zero-shot learning according to claim 1, characterized in that, The specific content of step S2 is as follows: Perform range pulse compression on the echo data obtained in step S1, and set the matching function of the pulse compression to Then the signal S after pulse compression is c (y i ,t r ,t a ) The expression is as follows: Then, according to the echo model, synthetic aperture azimuth focusing processing is performed on the signal S after pulse compression. c (y i ,t r ,t a ) The back projection (BP) algorithm is used to perform coherent accumulation on each channel to obtain the azimuth focusing result of a single snapshot. Then, the imaging result f bpj (t aj , ρ, γ) is expressed as follows: where t aj represents the sampling of the j-th slow time, and ρ and γ represent the slant range and azimuth angle corresponding to the imaging grid, respectively; Then, the different snapshot imaging results are coherently superposed to obtain the non-blurred multi-channel forward-looking SAR imaging result f bp (ρ,γ), and the expression is as follows: Where M represents the number of snapshots.

4. A multi-channel forward-looking SAR azimuth resolution enhancement method based on zero-shot learning according to claim 1, characterized in that The specific content of step S4 is as follows: S41. Construct an adversarial network model; Adopt the basic structure of the generative adversarial network to build the overall network model, including: a generator module and a discriminator module; Among them, the overall network model structure connects the generator module and the discriminator module in series to form two sets of reverse symmetric loops, and shares the generator and the discriminator; One of the inputs of the network model is an X-domain image. The input X-domain image Input_X is converted into a Y-domain image Generated_Y by the generator, and then the reconstructed X-domain image Cyclic_X is obtained through the inverse generator. The difference between the two is characterized by the cycle consistency loss. The other input of the network model is a Y-domain image. The input Y-domain image Input_Y is converted into an X-domain image Generated_X by the generator, and then the reconstructed Y-domain image Cyclic_Y is obtained through the inverse generator. The difference between the two is characterized by the cycle consistency loss. The discriminators D X and D Y are responsible for judging whether the images input into them conform to the original sample distribution of the corresponding image domain; The generator module includes: an encoder, a converter, and a decoder; the generator module uses ReLU as the activation function, adds a dropout layer, uses instance normalization layers to normalize the features of the intermediate layers of the network, and introduces global residual connections; Among them, the encoder extracts features from the input image using 3 convolutional layers, compresses the entire image into a feature vector, then converts the feature vector to another image domain through 9 residual blocks in the converter, and finally the decoder recovers the low-level features from the feature vector to obtain the generated image; and a hybrid attention mechanism is introduced in the decoder, which is implemented through two cascaded sub-modules of spatial attention and channel attention; The discriminator module adopts the PatchGAN method, traverses the entire image through a 70×70 sliding window pane, then uses 5 convolutional layers to extract features from the image patches, replaces all batch normalization layers with instance normalization layers, and uses LeakyReLU as the activation function; S42. Based on step S41, design the model loss function; The loss function of the model includes: adversarial loss, cycle consistency loss, identity mapping loss, and frequency domain loss, which are specifically as follows: (1) Adversarial loss L GAN (G A , D Y , X, Y); The adversarial loss consists of two parts. For the mapping G A (x): X → Y, the adversarial loss expression is as follows: Among them, X and Y respectively represent two image domains, G A represents the generator of the mapping G A (x): X → Y, where x and y respectively represent the images belonging to the X domain and the Y domain; D Y represents the discriminator in the Y domain; p data (x) and p data (y) respectively represent the sample distributions in the X domain and the Y domain; represents the mathematical expectation; Similarly, for the mapping G B (y): Y → X, the adversarial loss is expressed as L GAN (G B , D X , Y, X); Among them, G B represents the inverse mapping of G A G B (y): a generator from Y to X, D X represents the discriminator in the X domain; (2) Cyclic consistency loss L cycle (G A , G B ); For each image in the X domain, it will go through the loop: x → G A (x) → G B (G A (x)); Among them, x and G B (G A (x)) The difference is the forward cycle consistency loss; similarly, if the input is an image in the Y domain, it is called the backward consistency loss, and the expression of this cycle consistency loss is as follows: Among them, ||·||1 represents the L1 norm; (3) The same mapping loss L idt (G A , G B ); The expression of the identity mapping loss is as follows: (4) Frequency-domain loss L fre (G A , G B ); The expression of the frequency domain loss function is as follows: where K and L represent the width and length of the reconstructed image and the input image, u and v represent the frequency domain coordinates, corresponding to the frequency components of the image in the horizontal and vertical directions respectively, and F o (u, v) and F r (u, v) represent the frequency domain functions of the input image and the reconstructed image respectively, and w(u, v) represents the introduced frequency domain weight matrix, and its defining expression is as follows: w(u,v) = |F o (u,v) - F r (u,v)| γ (11) Among them, γ≥0 represents a scaling factor; Finally, the expression of the overall loss function of the network model is as follows: Among them, λ1, λ2, and λ3 represent weight parameters, which are used to control the relative importance of the three loss functions of cycle consistency loss, identity mapping loss, and frequency loss.

Images