Clutter suppression method and system based on complex value cyclic consistent generative adversarial network

Abstract

The invention belongs to the technical field of ocean remote sensing detection, and discloses a clutter suppression method and system based on a complex value cyclic consistent generative adversarial network, and the method comprises the steps: selecting a data analysis region, extracting two types of strong and weak sea clutter data under high and low sea conditions, and superposing target echoes, thereby constructing a model training set; preprocessing the input model data set, and unifying the magnitude order of sea clutter data; on the basis of a generative adversarial idea, performing complex valuation on a generator module and a discriminator module to realize feature extraction on radar complex value input data; training iteration is carried out based on a loss function module constraint model, and mapping fitting between strong clutters and weak clutters is realized; and taking the sea clutter improvement factor as an evaluation index of a network output result, and performing accuracy verification by using the sea clutter improvement factor based on actually measured sea clutter data. The method is wider in application range, has better model convergence and stronger robustness, can quickly realize sea clutter suppression, and further improves the target detection performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean remote sensing detection, and in particular relates to a clutter suppression method and system based on a complex-valued cyclic consistent generative adversarial network. Background Art

[0002] Sea clutter suppression is crucial for target detection in the ocean. Sea clutter is the backscattered echo generated by the interaction between radar electromagnetic waves and the sea surface. Its randomness and "three non-" characteristics (non-stationary, non-uniform, and non-Gaussian) make target detection in a sea clutter background extremely difficult. Sea clutter not only obscures target signals, but its strong amplitude can also seriously interfere with the normal operation of radar, thereby increasing the challenge of sea surface target detection. Its formation is mainly influenced by the combined influence of marine environmental factors (such as wind speed, ocean currents, temperature, humidity, and wave height) and radar equipment factors (such as frequency, illumination angle, and polarization mode), which gives it complex statistical and physical characteristics. Therefore, effective sea clutter suppression technology can not only improve the accuracy and reliability of target detection, but also has far-reaching theoretical significance and practical application value in multiple fields such as ocean monitoring, military defense, and fishery resource management.

[0003] Given the difficulty in distinguishing targets from sea clutter in radar echo signals, rapid screening of potential targets requires research on sea clutter characteristics to reduce or even eliminate them. Classic sea clutter suppression and target detection methods typically rely on energy or Doppler spectrum differences between targets and sea clutter. Commonly used methods include non-coherent integration and constant false alarm detection, as well as moving target display and detection techniques. These methods, based on statistical model assumptions, are not well adapted to the environment and are susceptible to the non-Gaussian distribution of sea clutter. They can suffer from spectral divergence and difficulty accumulating energy for non-uniformly moving targets. Furthermore, slow-moving targets, with their low Doppler frequencies, are easily absorbed into the Doppler spectrum of sea clutter and suppressed, resulting in a sharp decline in target detection performance. To address these issues with traditional sea clutter suppression and target detection methods, some researchers have proposed intelligent sea clutter suppression and target detection methods based on deep learning. However, these methods suffer from the complex time-varying and dynamic characteristics of sea clutter in practical applications, resulting in insufficient generalization and applicability. Research on sea clutter suppression for X-band marine radars has been conducted both domestically and internationally, and preliminary results have been achieved. However, the following problems still exist:

[0004] When intelligent models are applied to real ocean data, their applicability to specific scenarios and data decreases dramatically due to factors such as the complex time-varying and dynamic characteristics of sea clutter. Existing intelligent models primarily rely on real-valued radar amplitude input or extract features using real-valued neural networks from radar PPI images. This ignores the important phase characteristics of the radar complex data and fails to fully extract target motion response features. Therefore, a method with a wider range of applicability and high clutter suppression performance is urgently needed to quickly suppress sea clutter and thereby improve target detection performance.

[0005] The prior art invention patent, "A radar image sea clutter suppression assisted target detection method and system" (publication number CN117765231A, announcement date 20240326), discloses a radar image sea clutter suppression assisted target detection method and system. This patent belongs to the field of radar image processing technology and specifically relates to a radar image sea clutter suppression assisted target detection method and system based on a conditional generative adversarial network denoising algorithm. The method comprises the following steps: generating multiple image pairs through image fusion, and using each image pair as a training sample; designing a conditional generative adversarial network model and training the conditional generative adversarial network with the training samples; inputting a radar image to be detected containing sea clutter into the generator of the trained conditional generative adversarial network to obtain a sea clutter-free image, which is then input into a radar target detection module based on deep learning to obtain a target detection result. The invention improves the accuracy and reliability of target detection with minimal computational effort.

[0006] Disadvantages: This invention mainly uses a conditional generative adversarial network to achieve sea clutter suppression in radar images. It requires paired data with strong correlation (i.e., the same radar target but different backgrounds in the same pair of data) as input, resulting in poor generalization. In the prior art, the invention patent "Sea Clutter Suppression Method, Device, Equipment and Medium Based on Generative Adversarial Network" (publication number CN118897275A, announcement date 20241206) discloses a sea clutter suppression method, device, equipment and medium based on a generative adversarial network, belonging to the field of radar clutter suppression technology, comprising the steps of: constructing a sea clutter dataset containing simulated targets and measured targets; constructing a generative adversarial network based on pyramid multi-scale residual attention, including a mutually constrained generator and discriminator; training the designed generative adversarial network using the sea clutter dataset to obtain a sea clutter suppression network; and performing sea clutter suppression on measured radar images using the sea clutter suppression network. The present invention uses a generator in a generative adversarial network to learn the changing characteristics of sea clutter and uses dynamic adversarial training between the generator and the discriminator to suppress sea clutter. This not only solves the problem of interference caused by sea clutter on target detection, but also improves the sea clutter suppression performance of the radar.

[0007] Disadvantages: This invention mainly realizes sea clutter suppression in radar images based on generative adversarial networks. By using real-valued neural networks to extract features, it ignores the important phase features of the original radar complex data and cannot fully extract the target motion response features. Summary of the Invention

[0008] To overcome the problems existing in the related art, the embodiments disclosed in the present invention provide a clutter suppression method and system based on a complex-valued cyclic consistent generative adversarial network, which involves sea clutter suppression and target detection analysis of radar echoes under high sea conditions. The technical solution is as follows:

[0009] The present invention is implemented as follows: a clutter suppression method based on a complex-valued cyclic consistent generative adversarial network comprises the following steps:

[0010] S1, radar original complex echo data set construction: select the data analysis area, extract strong and weak sea clutter data of high and low sea conditions, superimpose target echoes, and construct the model training set;

[0011] S2, sea clutter suppression intelligent model construction: pre-process the input model data set to unify the order of magnitude of sea clutter data; based on the generative adversarial theory, the generator and discriminator modules are complex-valued to achieve feature extraction of radar complex-valued input data; finally, based on the loss function module, the model training iteration is constrained to achieve mapping fitting between strong and weak clutter;

[0012] S3, accuracy verification: The sea clutter improvement factor is used as an evaluation indicator for the network output results; based on the measured sea clutter data, the sea clutter improvement factor is used to verify the accuracy.

[0013] In step S1, the radar original complex echo data set is constructed, including:

[0014] Based on the original echo data, the target area is selected for interception to build a target database;

[0015] Construct a sea clutter database; use the coherent radar original complex echo data to construct two types of data sets, the first type is low sea state weak clutter data, the second type is high sea state strong clutter data, and the data size is determined to be a 256×256 complex matrix according to the network input:

[0016]

[0017] Where R is the complex data matrix to be judged, x rn is the real data of row r and column n, y rn is the imaginary part data of the rth row and nth column, i is the symbol of the imaginary part of the complex number;

[0018] Construct a model data set, randomly superimpose the target echo on the sea clutter database, and construct a model data set {x1,x2,x3…x n},x n =R n +T n , where x n is the nth data sample, R n is the nth echo matrix, T n is the nth target matrix; the radar original echo data under different sea conditions are processed to generate a pulse-range two-dimensional echo matrix;

[0019] The data slicing and interpolation techniques are used to reconstruct the original data into a data tensor of size 256×256. The training set and test set are constructed by adding targets with different SCRs into the sea clutter.

[0020] Furthermore, the target area is selected for interception, including buoys, fishing boats and speedboats, and the interception size is set according to the target size.

[0021] Furthermore, the expression of SCR is:

[0022]

[0023] Where, P t is the power of the target signal, P c is the power of the clutter signal, P c It is estimated by calculating the average power of 16 reference cells after removing the 6 protection cells around the target; SCR is the signal-to-clutter ratio;

[0024] According to the SCR calculation method, 200 samples are generated every 1dB interval in the range of 0dB-10dB, and divided into training set and validation set in a ratio of 8:2 for deep learning model training. Another 120 samples are generated every 2dB in the range of 0dB-10dB for testing.

[0025] In step S2, an intelligent model for sea clutter suppression is constructed, including:

[0026] Processing the real and imaginary parts of the radar complex echo separately to generate pre-processed echo data;

[0027]

[0028] Where s is a sufficiently small number, s = e -5 , e is a constant, sgn is a sign function, abs() is the sign of the absolute value, X' is the echo data generated after preprocessing, and i is the sign of the imaginary part of the complex number;

[0029] A complex-valued generator and discriminator are constructed. The complex-valued generator consists of three modules: an encoding module, a conversion module, and a decoding module. The encoding module is used to extract and learn features in the complex space of input radar data. The conversion module is used to convert features between the high-sea-state sea clutter data domain and the low-sea-state sea clutter data domain. The decoding module uses a deconvolution layer to restore the feature map. The discriminator is used to determine whether the data conforms to the measured sea clutter data domain features. The model processes radar complex echo data by jointly using a complex-valued convolutional layer (CV-Conv), a complex-valued batch normalization layer (CV-BN), and an activation function (CV-ReLU).

[0030] Construct a loss function module; the loss function consists of three parts. The first part is used to train the adversarial loss of the corresponding generator and discriminator. When learning the mapping relationship: f:X→Y corresponds to the discriminator D y , the loss is defined as:

[0031]

[0032] Where, L adv_f (f,D Y ,X,Y) is the discriminator D Y And the adversarial loss value of the generator f, To generate the expected value of the sample y and the real sample, To generate the expected value of the sample x and the real sample, D Y () is the discrimination of y and f(x), and the output value is in the range of [0,1]. x, y are the radar echo data from X before clutter suppression and Y after clutter suppression, respectively. f(x) is the clutter echo result after strong clutter x is suppressed by the generator.

[0033] For the mapping: g:Y→X, the adversarial loss is defined as:

[0034]

[0035] Where, L adv_g (g,D X ,Y,X) is the discriminator D X and the adversarial loss value of the generator G, D X () is the identification of X and g(y);

[0036] In the second part, g(y) is the mapping result from the post-suppression to the pre-suppression strictly corresponding to y, and the expected formula is:

[0037] g(f(x))≈x,f(g(y))≈y

[0038] In the formula, g() is the result of the generator g, and f() is the result of the generator f;

[0039] The cycle loss is defined as:

[0040]

[0041] where ||1 is the L1 norm, L cycle (f,g) is the cycle loss value;

[0042] In the third part, g(x) is the result generated by the generator for x, f(y) is the result generated by the generator for y, m is the final generated result after sea clutter suppression, i is the row of the matrix, and j is the column of the matrix, and the consistency loss is defined as follows:

[0043]

[0044] where λ1, λ2 are weight coefficients, λ1 is used to control the target invariance before and after sea clutter suppression, and λ1 is used to control the gradient change, reduce the residual error and suppress the influence of high frequency noise; L tar (f,g) is the consistency loss value;

[0045] The final loss function is defined as:

[0046]

[0047] The generator and discriminator are optimized by the optimization function of the sea clutter simulation intelligent model, and the optimization function of the sea clutter simulation intelligent model is represented as:

[0048]

[0049] where λ is a weight coefficient, f is the generator f, g is the generator g, D X is the discrimination of strong clutter X and generated false strong clutter g(y).

[0050] Further, a complex-valued generator and discriminator are constructed, including:

[0051] (1) The complex convolution layer extracts features from the original complex echo data, and performs complex domain convolution operation on the input complex data through complex weight parameters and complex activation functions.

[0052] Let the convolution kernel be K = K r +jK i , where K r is the real part of the convolution kernel, K i is the imaginary part of the convolution kernel, and j is the imaginary symbol.

[0053] The input complex data is: x' = x' r +jx' i , where x' r is the real part data, and x' iis the imaginary part data;

[0054] The convolution operation is expressed as:

[0055] K×x′=(K r +jK i )×(x′ r +jx′ i )=(K r ×x′ r -K i ×x′ i )+j(K r ×x′ i +K i ×x′ r )

[0056] The matrix representation is:

[0057]

[0058] In the formula, the symbol · represents the convolution operation;

[0059] The complex input of the channel is reassembled into The block of channels, the first half and the second half contain the real and imaginary parts of the input respectively; the output x' in the feature layer before the activation function l+1 Through the output o of the previous layer l and the weight matrix ω l+1 and bias Calculate, where is the real part processing sign, is the imaginary part processing symbol;

[0060]

[0061] (2) Complex-valued batch normalization layer, which uses the complex-valued BN algorithm for batch normalization;

[0062]

[0063] Where V is the covariance matrix, is the normalized output, x' is the input data, and E[] is the expected value of the input data;

[0064] By subtracting the mean and multiplying the square root of the variance, Standard normal complex distribution with mean μ = 0, covariance Γ = 1, pseudo-covariance C = 0;

[0065]

[0066] The complex-valued BN algorithm sets two trainable parameters β and γ, β is a complex value with two trainable components; the scaling parameter γ is a semi-positive definite matrix with three trainable components;

[0067]

[0068] The complex-valued BN algorithm is:

[0069]

[0070] During grid training, γ rr and γ ii Initialized to Setting up ri , the real and imaginary parts of β are 0;

[0071] (3) Using the complex-valued ReLU activation function As a complex activation function for complex-valued neural networks;

[0072]

[0073] Where, is the output of the complex-valued ReLU function, ReLU( ) is the traditional ReLU activation function, is the real part of x', is the imaginary part of x'.

[0074] Furthermore, during the network training process, the parameters of the generator are kept fixed and D is updated. X and D Y The parameters of the two discriminators are fixed, and the loss function is maximized to make the discriminator discriminate between the input strong clutter sea clutter data and the data generated by the generator. Then, the parameters of the two discriminators are fixed, the generator parameters are updated, and the loss function is minimized to make the data generated by the generator consistent with the real input sea clutter data after suppression. The optimization process is repeated until the network converges.

[0075] Furthermore, the model is trained based on the constructed dataset. The complex-valued signal generator in the network repeatedly uses the complex-valued convolution module with a convolution kernel size of 3×3 for downsampling, and the complex-valued residual modules are all trained using 3×3 convolution kernels. The upsampling part uses the complex-valued deconvolution module to restore the size of the radar echo data matrix. The complex-valued signal discriminator repeatedly uses the complex-valued convolution module with a convolution kernel size of 4×4 to extract features from the input radar data. The last layer of convolution converts the feature map into a decision output. The model is trained for 200 epochs using the Adam optimizer with an initial learning rate of 10. -3 , set to 10 after 100 epochs -4 .

[0076] In step S3, the sea clutter improvement factor σ is expressed as:

[0077] σ=SCR out -SCR in

[0078] Among them, SCR in Represents the target signal-to-clutter ratio of the input radar image, SCR out It represents the target signal-to-clutter ratio of the radar image output by the sea clutter suppression network, and the unit is dB. The target signal-to-clutter ratio SCR is expressed as:

[0079]

[0080] in, represents the average power of the target, Represents the average power of the sea clutter background.

[0081] Another object of the present invention is to provide a clutter suppression system based on a complex-valued cyclically consistent generative adversarial network, which is used to control the clutter suppression method based on the complex-valued cyclically consistent generative adversarial network. The system includes:

[0082] The radar original complex echo data set construction module is used to select the data analysis area, extract strong and weak sea clutter data in high and low sea conditions, superimpose target echoes, and construct the model training set;

[0083] The sea clutter suppression intelligent model construction module is used to preprocess the input model data set and unify the order of magnitude of the sea clutter data. Based on the generative adversarial theory, the generator and discriminator modules are complex-valued to achieve feature extraction of the radar complex-valued input data. Finally, the loss function module constrains the model training iteration to achieve mapping fitting between strong and weak clutter.

[0084] The accuracy verification module is used to use the sea clutter improvement factor as an evaluation indicator of the network output results; based on the measured sea clutter data, the sea clutter improvement factor is used to verify the accuracy

[0085] In combination with all the above technical solutions, the beneficial effects of the present invention are as follows:

[0086] First, traditional data preprocessing methods only take the logarithm of the amplitude or perform direct linear normalization, without utilizing the phase information in the echo signal. This invention proposes a new complex data processing method specifically suited for radar echo data. Through sign-preserving logarithmic operations, it effectively utilizes the phase information in the radar signal. This method generates preprocessed echo data by separately processing the real and imaginary parts of the radar complex echo without changing the sign characteristics of the original data. By constructing a complex-valued generator and discriminator to adapt to the characteristics of radar complex echo data, this invention effectively extracts and utilizes the features of radar complex echo data, thereby significantly improving the performance and application effectiveness of radar image processing.

[0087] Second, unlike traditional generative adversarial networks, to adapt to the characteristics of radar complex data, the sea clutter suppression intelligent model constructed in this paper processes radar complex echo data by jointly using a complex-valued convolutional layer (CV-Conv), a complex-valued batch normalization layer (CV-BN), and an activation function (CV-ReLU). This not only improves the effectiveness of radar sea clutter suppression, but also enhances the robustness and generalization of the model. The complex-valued convolutional layer can directly extract features from the original complex echo data, and performs a complex-domain convolution operation on the input complex data through complex weight parameters and a complex activation function, thereby fully retaining and fully utilizing the amplitude and phase information of the echo signal. Compared with the traditional real-valued convolutional layer that only processes amplitude information, the complex-valued convolution operation can simultaneously extract amplitude and phase information from the complex data, significantly improving the integrity and accuracy of feature extraction, and showing a strong advantage in suppressing sea clutter and improving image quality.

[0088] By constructing a loss function module, the present invention can more effectively extract and utilize the features of radar complex echo data, thereby significantly improving the effectiveness of sea clutter suppression. The technical benefits of the present invention include, but are not limited to, more accurate feature extraction, better model convergence, and greater robustness.

[0089] Third, the present invention has a wider range of applications and higher clutter suppression performance. Based on a complex-valued cyclic consistent generative adversarial network, it can quickly suppress sea clutter, thereby improving target detection performance. With better model convergence and stronger robustness, the present invention can be applied in various real-world marine scenarios, quickly suppressing sea clutter and thus improving target detection performance.

[0090] Fourth, it fills the gap in radar complex data processing in the domestic and international industries, applies generative adversarial networks to radar complex data, and makes up for the deficiency of being unable to fully extract target motion response characteristics due to ignoring the important phase characteristics of radar complex data. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure;

[0092] Figure 1 This is a flow chart of a sea clutter suppression method based on a complex-valued cyclic consistent generative adversarial network provided by an embodiment of the present invention;

[0093] Figure 2 This is a structural diagram of a deep learning model provided by an embodiment of the present invention;

[0094] Figure 3 : This is a radar echo data set (unit: dB) under different SCR inputs provided by an embodiment of the present invention; wherein, in Figure (a), SCR = 0 dB, in Figure (b), SCR = 2 dB, in Figure (c), SCR = 4 dB, in Figure (d), SCR = 6 dB, in Figure (e), SCR = 8 dB, and in Figure (f), SCR = 10 dB;

[0095] Figure 4 1 is a schematic diagram of sea clutter suppression results under different signal-to-clutter ratio inputs provided by an embodiment of the present invention; wherein, Figure (a) is a picture of input data before suppression when the signal-to-clutter ratio value SCR = 0 dB; Figure (b) is a picture of output data after suppression when the signal-to-clutter ratio value SCR = 0 dB; Figure (c) is a picture of input data before suppression when the signal-to-clutter ratio value SCR = 2 dB; Figure (d) is a picture of output data after suppression when the signal-to-clutter ratio value SCR = 2 dB; Figure (e) is a picture of input data before suppression when the signal-to-clutter ratio value SCR = 4 dB; Figure (f) is a picture of input data when the signal-to-clutter ratio value SCR = 4dB, the output data picture after suppression; Figure (g) is the input data picture before suppression when the signal-to-noise ratio value SCR = 6dB; Figure (h) is the output data picture after suppression when the signal-to-noise ratio value SCR = 6dB; Figure (i) is the input data picture before suppression when the signal-to-noise ratio value SCR = 8dB; Figure (j) is the output data picture after suppression when the signal-to-noise ratio value SCR = 8dB; Figure (k) is the input data picture before suppression when the signal-to-noise ratio value SCR = 10dB; Figure (bl) is the output data picture after suppression when the signal-to-noise ratio value SCR = 10dB;

[0096] Figure 5 Schematic diagram of sea clutter improvement factor results under different signal-to-clutter ratio inputs provided by an embodiment of the present invention; wherein, Figure (a) SCR = 0 dB, Figure (b) SCR = 2 dB, Figure (c) SCR = 4 dB, Figure (d) SCR = 6 dB, Figure (e) SCR = 8 dB, and Figure (f) SCR = 10 dB;

[0097] Figure 6 FIG. 4 is a schematic diagram of average sea clutter improvement factor results under different signal-to-clutter ratio inputs provided by an embodiment of the present invention.

[0098] Figure 7 The sea clutter suppression results of target slice radar echo data measured by different methods (unit: dB) DETAILED DESCRIPTION

[0099] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0100] The innovation of the present invention lies in: logarithmic processing of radar complex data, and applying a generative adversarial network to the radar complex data, which makes up for the deficiency of being unable to fully extract the target motion response characteristics due to ignoring the important phase characteristics of radar complex data. The innovative introduction of the target consistency loss function avoids the suppression of the target when the sea clutter is consistent.

[0101] Example 1, as Figure 1 As shown, the clutter suppression method based on the complex-valued cyclic consistent generative adversarial network provided by the embodiment of the present invention specifically includes the following steps:

[0102] S1: Construction of the radar raw complex echo dataset. The data analysis area is selected, and strong and weak sea clutter data for both high and low sea conditions are extracted and superimposed with target echoes to construct the model training set.

[0103] 1-1) Build a target database. Based on the original echo data, select the target area for interception. The intercepted targets include buoys, fishing boats, speedboats and other small targets on the sea surface. The interception size is set according to the target size.

[0104] 1-2) Construct a sea clutter database. Using the coherent radar's raw complex echo data, we construct two datasets: the first for low sea conditions with weak clutter, and the second for high sea conditions with strong clutter. The data size is determined to be a 256×256 complex matrix based on the network input:

[0105]

[0106] Where R is the complex data matrix to be judged, x rn is the real data of row r and column n, y rn is the imaginary part data of the rth row and nth column, i is the symbol of the imaginary part of the complex number;

[0107] 1-3) Construct a model data set, randomly superimpose the target echo on the sea clutter database, and construct a model data set {x1,x2,x3…x n},x n =R n +T n ; where x n is the nth data sample, R n is the nth echo matrix, T n is the nth target matrix; the radar raw echo data under different sea conditions are processed to generate a pulse-range two-dimensional echo matrix. Data slicing and interpolation techniques are used to reconstruct the raw data into a 256×256 data tensor, which, on the one hand, adapts to the filter morphological characteristics of the deep neural network, and on the other hand, meets the optimal receptive field configuration requirements of the multi-scale convolution kernel. The generalization ability of the deep model is limited by the scale and data quality of the training set. However, it is difficult to obtain a large number of different labeled samples in the radar area. To solve this problem, the training set and test set are constructed by adding targets with different SCRs to the sea clutter. SCR is defined as:

[0108]

[0109] Where, P t is the power of the target signal, P c is the power of the clutter signal, P c The average power of 16 reference cells is estimated by removing the six protection cells around the target. SCR is the signal-to-interference plus noise ratio (SCR).

[0110] According to the SCR calculation method, 200 samples are generated every 1dB interval in the range of 0dB-10dB, and divided into training set and validation set in a ratio of 8:2 for deep learning model training. Another 120 samples are generated every 2dB in the range of 0dB-10dB for testing.

[0111] S2, sea clutter suppression intelligent model construction. First, the S1 partial input model data set is preprocessed to unify the order of magnitude of the sea clutter data to facilitate subsequent data feature extraction. Second, based on the generative adversarial theory, the generator and discriminator modules are complex-valued to achieve feature extraction of radar complex-valued input data. This step can fully extract information from the original data and avoid the loss of radar phase information. Finally, based on the loss function module, the model training iteration is constrained to achieve mapping fitting between strong and weak clutter.

[0112] 2-1) Complex Data Preprocessing. Traditional data preprocessing methods only take the logarithm of the amplitude or directly perform linear normalization, without utilizing the phase information in the echo signal. We propose a new complex data processing method specifically suited for radar echo data. This method effectively utilizes the phase information in the radar signal through sign-preserving logarithmic operations. This method generates preprocessed echo data by processing the real and imaginary parts of the radar complex echo separately, without changing the sign characteristics of the original data:

[0113]

[0114] Where s is a sufficiently small number, s = e -5 , e is a constant, sgn is a sign function, abs() is the sign of the absolute value, X' is the echo data generated after preprocessing, and i is the sign of the imaginary part of the complex number;

[0115] Through the above logarithmic transformation processing, the problem of large differences in data distribution under high and low sea conditions is solved while the phase information of the echo signal is completely preserved.

[0116] 2-2) Construction of complex-valued generator and discriminator. Traditional generative adversarial networks cannot directly process raw complex-valued radar echo data. To meet the characteristics of input radar complex echo data, a new radar complex echo data processing method is now provided. This method adapts to the characteristics of radar complex echo data by constructing a complex-valued generator and a discriminator, effectively extracting and utilizing the features of radar complex echo data, thereby significantly improving the performance and application effect of radar image processing. Among them, the complex-valued generator consists of three modules: the encoding module, the conversion module, and the decoding module. The encoding module is mainly used to extract and learn the features in the complex space of the input radar data. The conversion module is used to realize the conversion between the features of the high sea state sea clutter data domain and the low sea state sea clutter data domain. The decoding module uses the deconvolution layer to restore the feature map. The discriminator is used to determine whether the data conforms to the characteristics of the measured sea clutter data domain. Unlike traditional generative adversarial networks, in order to adapt to the characteristics of radar complex data, this model processes radar complex echo data by jointly using complex-valued convolutional layers (CV-Conv), complex-valued batch normalization layers (CV-BN) and activation functions (CV-ReLU). This not only improves the effect of radar sea clutter suppression, but also enhances the robustness and generalization ability of the model.

[0117] 2-2-1) The complex-valued convolution layer can directly extract features from the original complex echo data, and perform complex domain convolution operations on the input complex data through complex weight parameters and complex activation functions, thereby completely retaining and fully utilizing the amplitude and phase information of the echo signal. Compared with the traditional real convolution layer that only processes amplitude information, the complex-valued convolution operation can simultaneously extract amplitude and phase information from the complex data, significantly improving the integrity and accuracy of feature extraction, and showing a strong advantage in suppressing sea clutter and improving image quality. Assume that the convolution kernel is: K = K r +jK i , the convolution operation can be expressed as:

[0118] K×x′=(K r +jK i )×(x′ r +jx′ i )=(K r ×x′ r -K i ×x′ i )+j(K r ×x′ i +K i ×x′ r )

[0119] The matrix representation is:

[0120]

[0121] In the formula, the symbol · represents the convolution operation;

[0122] one In this way the multiplexed inputs of the channels are recombined into a The first half and the second half of the channel block contain the real and imaginary parts of the input respectively. The output x' of the feature layer before the activation function l+1 Through the output o of the previous layer l and the weight matrix ω l+1 By performing calculations, we can obtain the specific calculation process as follows:

[0123]

[0124] 2-2-2) Complex-valued Batch Normalization Layer. Batch Normalization (BN) is a regularization technique widely used in modern networks. It can speed up training, provide better convergence, and reduce the risk of gradient vanishing or exploding. However, the conventional BN operation cannot normalize the complex-valued convolution result to obey the standard normal complex distribution with mean 0 and covariance 1. Therefore, the complex-valued BN algorithm must be used instead of the conventional BN. The normalization operation of this algorithm is:

[0125]

[0126] Where V is the covariance matrix, is the normalized output, x' is the input data, and E[ ] is the expected value of the input data;

[0127] This method can ensure has a standard normal complex distribution, by subtracting the mean and multiplying the square root of the variance, It can have a standard normal complex distribution with mean μ = 0, covariance Γ = 1, and pseudo-covariance C = 0. The corresponding calculation formula is as follows:

[0128]

[0129] Similar to the real-valued BN algorithm, the complex-valued BN algorithm also requires setting two trainable parameters β and γ. β is complex-valued and has two trainable components. The scaling parameter γ is a semi-positive definite matrix with three trainable components, which can be expressed as:

[0130]

[0131] Therefore, the complex BN algorithm can be obtained as:

[0132]

[0133] because The covariance of is 1, so during grid training, γ rr and γ ii Initialized to To ensure that the normalized square error modulus is 1, set γ ri , the real and imaginary parts of β are 0;

[0134] 2-2-3) Activation function. In this invention, complex-valued ReLU is mainly used. The activation function is a complex activation function of a complex-valued neural network. The complex-valued ReLU activation function can apply independent ReLU functions to the real and imaginary parts of the parameters respectively. When the real and imaginary parts of the complex number satisfy strictly positive or strictly negative values ​​at the same time, the complex-valued ReLU is holomorphic and can be expressed as:

[0135]

[0136] Where, is the output of the complex-valued ReLU function, ReLU() is the traditional ReLU activation function, is the real part of x', is the imaginary part of x'.

[0137] 2-3) Loss function module. The loss function consists of three parts. One of them is the adversarial loss used to train the corresponding generator and discriminator. Taking the learning mapping relationship f:X→Y as an example, the corresponding discriminator is D y , then the loss is defined as:

[0138]

[0139] Where x, y represent the radar echo data from X before clutter suppression and Y after clutter suppression, respectively. f(x) is the clutter echo result after strong clutter x is suppressed by the generator. D Y Used to identify y and f(x). Similarly, for the mapping g:Y→X, the adversarial loss is defined as:

[0140]

[0141] Where, L adv_g (g,D X ,Y,X) is the discriminator D X and the adversarial loss value of the generator G, D X () is the identification of X and g(y);

[0142] Secondly, in order to ensure that x and f(x), y and g(y) can uniquely correspond, g(y) is the mapping result from the suppression to the suppression before strict correspondence of y, and the expected formula can be obtained:

[0143] f(g(x))≈x,f(g(y))≈y

[0144] The cycle loss is thus defined as:

[0145]

[0146] Where, || ||1 is the L1 norm, L cycle (f, g) is the circulation loss value;

[0147] Third, consistency loss is the key innovation of this invention: on the one hand, when machine learning methods are applied to the computer field, they may introduce high-frequency noise points, that is, clutter residuals need to be filtered out; on the other hand, sea clutter suppression not only requires accurate removal of sea clutter in complex and changing sea environments, but more importantly, it requires effective preservation of target information. The consistency loss is defined as follows:

[0148]

[0149] Where λ1 and λ2 are weight coefficients. When λ1 is large, the network will be more inclined to retain target information. When λ2 is large, the network will make the final generated data smoother and filter out residual clutter. λ1 is used to control the sea clutter suppression to keep the target unchanged before and after, and λ2 is used to control the gradient change, reduce the residual error and suppress the influence of high-frequency noise. tar (f,g) is the consistency loss value;

[0150] In summary, the total loss function is defined as:

[0151]

[0152] λ represents the weighting coefficient. By combining the above formulas, the present invention can more effectively extract and utilize the features of radar complex echo data compared to using a single formula, thereby significantly improving sea clutter suppression. The technical benefits of the present invention include, but are not limited to, more accurate feature extraction, improved model convergence, and greater robustness.

[0153] Finally, the model generator and discriminator are optimized by the optimization function of the sea clutter simulation intelligent model. The optimization function of the sea clutter simulation intelligent model can be expressed as:

[0154]

[0155] In the formula, λ is the weight coefficient, f is the generator f, g is the generator g, D X It is the identification of strong clutter X and generated false strong clutter g(y).

[0156] During the network training process, each round of training can be regarded as a "max-min" process. First, keep the parameters of the generator fixed and update D X and D Y The parameters of the two discriminators are fixed, and the loss function is maximized to enable the discriminator to accurately distinguish the input strong clutter sea clutter data from the data generated by the generator. Then, the parameters of the two discriminators are fixed, the generator parameters are updated, and the loss function is minimized to make the data generated by the generator as consistent as possible with the real input sea clutter data after suppression. The above optimization process is repeated until the network converges.

[0157] The model is trained based on the dataset constructed above. On the one hand, the complex-valued signal generator in the network repeatedly uses the complex-valued convolution module with a convolution kernel size of 3×3 for downsampling, and the complex-valued residual modules are all trained using 3×3 convolution kernels. Finally, the upsampling part uses the complex-valued deconvolution module to restore the size of the radar echo data matrix; on the other hand, the complex-valued signal discriminator repeatedly uses the complex-valued convolution module with a convolution kernel size of 4×4 to extract features from the input radar data. The last layer of convolution converts the feature map into a decision output. The model is trained for 200 epochs using the Adam optimizer, with an initial learning rate of 10 -3 , set to 10 after 100 epochs -4 The structure diagram of the deep learning model is as follows: Figure 2 shown.

[0158] S3, accuracy verification, is based on using the sea clutter improvement factor σ as the evaluation indicator of the network output results.

[0159] The expression of sea clutter improvement factor σ is:

[0160] σ=SCR out -SCR in

[0161] Among them, SCR in Represents the target signal-to-clutter ratio of the input radar image, SCR out It represents the target signal-to-clutter ratio of the radar image output by the sea clutter suppression network, and the unit is dB. The target signal-to-clutter ratio SCR is expressed as:

[0162]

[0163] in, represents the average power of the target, Represents the average power of the sea clutter background.

[0164] In embodiment 2, the clutter suppression system based on a complex-valued cyclic consistent generative adversarial network provided by the embodiment of the present invention specifically includes:

[0165] The radar original complex echo data set construction module is used to select the data analysis area, extract strong and weak sea clutter data in high and low sea conditions, superimpose target echoes, and construct the model training set;

[0166] The sea clutter suppression intelligent model construction module is used to preprocess the input model data set and unify the order of magnitude of the sea clutter data. Based on the generative adversarial theory, the generator and discriminator modules are complex-valued to achieve feature extraction of the radar complex-valued input data. Finally, the loss function module constrains the model training iteration to achieve mapping fitting between strong and weak clutter.

[0167] The accuracy verification module is used to use the sea clutter improvement factor as an evaluation indicator of the network output result; based on the measured sea clutter data, the sea clutter improvement factor is used to perform accuracy verification.

[0168] In order to further demonstrate the positive effects of the above embodiment, the present invention conducts the following experiments based on the above technical solution.

[0169] The clutter suppression method based on complex-valued cyclic consistent generative adversarial networks provided in this experiment specifically includes:

[0170] Step 1: construct the radar original complex echo dataset.

[0171] The present invention is based on a radar sea detection test conducted by a university in Harbin in September 2023 at a test point in the southeast waters of Lingshan Island, Qingdao (120°11.237′E-120°11.473′E; 35°45.123′N-35°45.394′N). The test was based on an X-band solid-state fully coherent radar to detect the sea surface. The sea surface radar echo complex data was obtained by orthogonal sampling of the sea clutter data in each range gate. The specific radar system parameters and marine environment parameters are shown in Table 1.

[0172] Table 1 Parameters of the coherent radar system of a university in Harbin

[0173] parameter Parameter value Operating frequency (GHz) 9.4 Antenna height (m) 30m Antenna pattern Circular Sweep Antenna gain (dB) ≥30 Antenna rotation speed / (r / min) 24-36 Peak power / (W) ≥100 Distance resolution / (m) 6 Pulse repetition frequency (Hz) 3000 Bandwidth / (MHz) 80 Beam width / (.) 0.95±0.1 Polarization Horizontal polarization

[0174] The test uses equipment such as wave riders, anemometers, and AIS signal receivers to synchronously collect sea surface environment and target information, with sea condition levels covering 1-5.

[0175] The raw radar echo data under different sea conditions is processed to generate a two-dimensional pulse-range echo matrix. Through operations such as data cropping and interpolation, the data is output to a 256×256 size, creating target and background samples suitable for the deep learning network size and convolution operator morphology, facilitating optimal network performance.

[0176] The performance of deep learning models depends largely on the number and quality of samples in the training set. However, it is difficult to obtain a large number of different labeled samples in the radar area. Usually, targets with different SCRs are added to the real clutter to construct the training and test sets. The model dataset is constructed with strong clutter data of level 3 and 4 sea conditions and weak clutter data of level 1 sea conditions, and random buoy and ship echo data are added, such as Figure 3 As shown in Figure 2, 200 sets of data were selected for each of the two datasets, and the training set and test set were divided into 8:2 for training.

[0177] Step 2: Build an intelligent model for sea clutter suppression.

[0178] The model is trained based on the dataset constructed above. On the one hand, the complex-valued signal generator in the network repeatedly uses the complex-valued convolution module with a convolution kernel size of 3×3 for downsampling, and the complex-valued residual modules are all trained using 3×3 convolution kernels. Finally, the upsampling part uses the complex-valued deconvolution module to restore the size of the radar echo data matrix; on the other hand, the complex-valued signal discriminator repeatedly uses the complex-valued convolution module with a convolution kernel size of 4×4 to extract features from the input radar data. The last layer of convolution converts the feature map into a decision output. The model is trained for 200 epochs using the Adam optimizer, with an initial learning rate of 10 -3 , set to 10 after 100 epochs -4 .

[0179] Step 3: Accuracy verification.

[0180] After the training is completed, check the test results as follows Figure 4-Figure 7 shown.

[0181] To demonstrate the superiority of the proposed algorithm, ablation experiments were conducted to evaluate the effectiveness and impact of model complexification and consistency loss. Method 1 is the original CycleGAN network without complexification, and all internal modules used are real-valued processing modules. Method 2 is a complex-valued CycleGAN network. Compared with Method 1, its internal modules use complex-valued convolutional layers, complex-valued batch normalization layers, and complex-valued activation functions. The input data is the aforementioned test set, and 120 additional samples are generated every 2dB in the range of 0dB-10dB for testing. The results are shown in Table 2.

[0182] Table 2 Ablation experiment results

[0183]

[0184] As can be seen from Table 2, under the same signal-to-clutter ratio radar echo input, the complex-valued neural network has a higher sea clutter suppression effect than the real-valued neural network due to its ability to extract the phase characteristic information of the input data, which is more obvious under low signal-to-clutter ratio input. In addition, the consistency loss proposed in this invention effectively reduces the residual clutter residual, further improving the sea clutter suppression effect. Taking SCR = 2dB and SCR = 8dB input as examples, under 2dB input, the method proposed in this invention improves the signal-to-clutter ratio by 15.05dB and 5.184dB respectively compared with Method 1 and Method 2; under 8dB input, the method proposed in this invention improves the signal-to-clutter ratio by 16.021dB and 3.594dB respectively compared with Method 1 and Method 2, showing stronger clutter suppression performance.

[0185] Furthermore, to verify the performance of the proposed method in practical applications, the generalization analysis and verification of the intelligent sea clutter suppression method proposed in this invention is conducted based on the SDRDSP database released by a naval university in 2019. The SDRDSP database originates from the "Radar Sea Detection Data Sharing Project," led by a naval university's marine target perception research team. This project systematically conducts multimodal observation experiments of the marine environment using an X-band all-digital phased array radar system. Since 2019, six batches of datasets have been released in phases on the Journal of Radars platform. The observation base is located at the Yantai Offshore Comprehensive Test Field in Shandong Province and is deployed at three observation stations with different altitude gradients, ranging from 30 to 150 meters, forming a multi-perspective collaborative observation network. The dataset was constructed using radar echo data in scanning mode. The datasets that meet this condition primarily include the first phase of 2019 data and the second phase of 2020 data. The first phase of 2019 data is used as an example for further discussion below.

[0186] The first phase of data collection is located in the coastal area of ​​Yangma Island, Yantai. The station is only 50 meters away from the coastline and 30 meters above sea level. The radar system has a 180° omnidirectional scanning capability, with effective detection coverage of 0.3° to 15° pitch angles. The monitoring targets are mainly small and medium-sized ships, covering a variety of marine mobile platforms. The first phase of data in 2019 is divided into 3 groups, including 45 available radar scan data and the corresponding 6 meteorological and hydrological data and 4 AIS data files. The sea state level covers level 3-4 and there are abundant targets. The first group of data is the data of 20 continuous circular scans of the radar, the sea state level is level 3-4, there are four channel buoys in the radar field of view, and there are non-cooperative small fishing boat targets.

[0187] In order to verify the performance of the proposed method in practical applications, a generalization analysis is conducted. In order to address the clutter suppression problem in complex ocean environments, the study focuses on the target detection performance under sea conditions of level 3-4. The results show the clutter suppression results of target slices in measurement scenarios under different methods. Although the data of the training set and validation set are obtained from different ocean environments and different radar equipment, the scene backgrounds are different, and the types and numbers of targets are different, it can be seen from the results that the clutter suppression network can still effectively suppress sea clutter. Compared with the Singular Value Decomposition-Fractional Fourier Transform (SVD-FRFT) and the traditional CycleGAN network, the proposed complex-valued neural network CV-CGAN can better protect the target information. Therefore, the good performance and generalization ability of the proposed method are verified by the clutter suppression results in the measured scene.

[0188] The above description is only a preferred specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A clutter suppression method based on a complex-valued cyclic consistent generative adversarial network, characterized in that: The method comprises the following steps: S1, radar original complex echo data set construction: select the data analysis area, extract strong and weak sea clutter data of high and low sea conditions, superimpose target echoes, and construct the model training set; S2, sea clutter suppression intelligent model construction: pre-process the input model data set to unify the order of magnitude of sea clutter data; based on the generative adversarial theory, the generator and discriminator modules are complex-valued to achieve feature extraction of radar complex-valued input data; finally, based on the loss function module, the model training iteration is constrained to achieve mapping fitting between strong and weak clutter; S3, accuracy verification: The sea clutter improvement factor is used as an evaluation indicator for the network output results; based on the measured sea clutter data, the sea clutter improvement factor is used to verify the accuracy.

2. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 1 is characterized in that: In step S1, the radar original complex echo data set is constructed, including: Based on the original echo data, the target area is selected for interception to build a target database; Construct a sea clutter database; use the coherent radar original complex echo data to construct two types of data sets, the first type is low sea state weak clutter data, the second type is high sea state strong clutter data, and the data size is determined to be a 256×256 complex matrix according to the network input: Where R is the complex data matrix to be judged, x rn is the real data of row r and column n, y rn is the imaginary part data of the rth row and nth column, i is the symbol of the imaginary part of the complex number; Construct a model data set, randomly superimpose the target echo on the sea clutter database, and construct a model data set {x1,x2,x3…x n },x n =R n +T n , where x n is the nth data sample, R n is the nth echo matrix, T n is the nth target matrix; the radar original echo data under different sea conditions are processed to generate a pulse-range two-dimensional echo matrix; The data slicing and interpolation techniques are used to reconstruct the original data into a data tensor of size 256×256. The training set and test set are constructed by adding targets with different SCRs into the sea clutter.

3. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 2 is characterized in that: Select the target area for interception, including buoys, fishing boats and speedboats. The interception size is set according to the target size.

4. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 2 is characterized in that: The expression of SCR is: Where, P t is the power of the target signal, P c is the power of the clutter signal, P c It is estimated by calculating the average power of 16 reference cells after removing the 6 protection cells around the target; SCR is the signal-to-clutter ratio; According to the SCR calculation method, 200 samples are generated every 1dB interval in the range of 0dB-10dB, and divided into training set and validation set in a ratio of 8:2 for deep learning model training. Another 120 samples are generated every 2dB in the range of 0dB-10dB for testing.

5. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 1 is characterized in that: In step S2, an intelligent model for sea clutter suppression is constructed, including: Processing the real and imaginary parts of the radar complex echo separately to generate pre-processed echo data; Where s is a sufficiently small number, s = e -5 , e is a constant, sgn is a sign function, abs() is the sign of the absolute value, X' is the echo data generated after preprocessing, and i is the sign of the imaginary part of the complex number; A complex-valued generator and discriminator are constructed. The complex-valued generator consists of three modules: an encoding module, a conversion module, and a decoding module. The encoding module is used to extract and learn features in the complex space of input radar data. The conversion module is used to convert features between the high-sea-state sea clutter data domain and the low-sea-state sea clutter data domain. The decoding module uses a deconvolution layer to restore the feature map. The discriminator is used to determine whether the data conforms to the measured sea clutter data domain features. The model processes radar complex echo data by jointly using a complex-valued convolutional layer (CV-Conv), a complex-valued batch normalization layer (CV-BN), and an activation function (CV-ReLU). Construct a loss function module; the loss function consists of three parts. The first part is used to train the adversarial loss of the corresponding generator and discriminator. When learning the mapping relationship: f:X→Y corresponds to the discriminator D Y , the loss is defined as: Where, L adv_f (f,D Y ,X,Y) is the discriminator D Y And the adversarial loss value of the generator f, is the expected value of the generated sample y and the real sample, where To generate the expected value of the sample x and the real sample, D Y () is the discrimination of y and f(x), and the output value is in the range of [0,1]. x, y are the radar echo data from X before clutter suppression and Y after clutter suppression, respectively. f(x) is the clutter echo result after strong clutter x is suppressed by the generator. For the mapping: g:Y→X, the adversarial loss is defined as: Where, L adv_g (g,D X ,Y,X) is the discriminator D X and the adversarial loss value of the generator G, D X () is the identification of X and g(y); In the second part, g(y) is the mapping result from the post-suppression to the pre-suppression strictly corresponding to y, and the expected formula is: g(f(x))≈x,f(g(y))≈y In the formula, g() is the result of the generator g, and f() is the result of the generator f; The cycle loss is defined as: In the formula, ||||1 is the L1 norm, L cycle (f, g) is the circulation loss value; In the third part, g(x) is the result of x generated in the generator, f(y) is the result of y generated in the generator, m is the final result after sea clutter suppression, i is the row of the matrix, j is the column of the matrix, and the consistency loss is defined as follows: Where, λ1 and λ2 are weight coefficients. λ1 is used to control the sea clutter suppression so that the target remains unchanged before and after the suppression. λ2 is used to control the gradient change, reduce the residual error and suppress the influence of high-frequency noise. tar (f,g) is the consistency loss value; The final loss function is defined as: The model generator and discriminator are optimized by the optimization function of the sea clutter simulation intelligent model. The optimization function of the sea clutter simulation intelligent model is expressed as: In the formula, λ is the weight coefficient, f is the generator f, g is the generator g, D X It is the identification of strong clutter X and generated false strong clutter g(y).

6. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 5 is characterized in that: Build complex value generators and discriminators, including: (1) The complex-valued convolution layer extracts features from the original complex echo data and performs complex domain convolution operations on the input complex data through complex weight parameters and complex activation functions; Let the convolution kernel be K=K r +jK i , where K r is the real convolution kernel, K i is the imaginary convolution kernel, j is the imaginary part symbol; The complex data input is: x'=x' r +jx' i , where x' r is the real part data, x' i is the imaginary part data; The convolution operation is expressed as: K×x′=(K r +jK i )×(x′ r +jx′ i )=(K r ×x′ r -K i ×x′ i )+j(K r ×x i +K i ×x′ r ) The matrix representation is: In the formula, the symbol · represents the convolution operation; The complex input of the channel is reassembled into The block of channels, the first half and the second half contain the real and imaginary parts of the input respectively; the output x' in the feature layer before the activation function l+1 Through the output o of the previous layer l and the weight matrix ω l+1 and bias Calculate, where is the real part processing sign, is the imaginary part processing symbol; (2) Complex-valued batch normalization layer, which uses the complex-valued BN algorithm for batch normalization; Where V is the covariance matrix, is the normalized output, x' is the input data, and E[] is the expected value of the input data; By subtracting the mean and multiplying the square root of the variance, Standard normal complex distribution with mean μ = 0, covariance Γ = 1, pseudo-covariance C = 0; The complex-valued BN algorithm sets two trainable parameters β and γ, β is a complex value with two trainable components; the scaling parameter γ is a semi-positive definite matrix with three trainable components; The complex-valued BN algorithm is: During grid training, γ rr and γ ii Initialized to Setting up ri , the real and imaginary parts of β are 0; (3) Using the complex-valued ReLU activation function As a complex activation function for complex-valued neural networks; Where, is the output of the complex-valued ReLU function, ReLU() is the traditional ReLU activation function, is the real part of x', is the imaginary part of x'.

7. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 5 is characterized in that: During network training, keep the parameters of the generator fixed and update D X and D Y The parameters of the two discriminators are fixed, and the loss function is maximized to make the discriminator discriminate between the input strong clutter sea clutter data and the data generated by the generator. Then, the parameters of the two discriminators are fixed, the generator parameters are updated, and the loss function is minimized to make the data generated by the generator consistent with the real input sea clutter data after suppression. The optimization process is repeated until the network converges.

8. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 7 is characterized in that: The model is trained based on the constructed dataset. The complex-valued signal generator in the network repeatedly uses the complex-valued convolution module with a convolution kernel size of 3×3 for downsampling. The complex-valued residual modules are all trained with a 3×3 convolution kernel. The upsampling part uses a complex-valued deconvolution module to restore the size of the radar echo data matrix. The complex-valued signal discriminator repeatedly uses the complex-valued convolution module with a convolution kernel size of 4×4 to extract features from the input radar data. The last layer of convolution converts the feature map into a decision output. The model is trained for 200 epochs using the Adam optimizer with an initial learning rate of 10. -3 , set to 10 after 100 epochs -4 .

9. The clutter suppression method based on complex-valued cyclic consistent generative adversarial network according to claim 1 is characterized in that: In step S3, the sea clutter improvement factor σ is expressed as: σ=SCR out -SCR in Among them, SCR in Represents the target signal-to-clutter ratio of the input radar image, SCR out It represents the target signal-to-clutter ratio of the radar image output by the sea clutter suppression network, and the unit is dB. The target signal-to-clutter ratio SCR is expressed as: in, represents the average power of the target, Represents the average power of the sea clutter background.

10. A clutter suppression system based on a complex-valued cyclic consistent generative adversarial network, characterized in that: The system is used to control the clutter suppression method based on the complex-valued cyclic consistent generative adversarial network described in any one of claims 1 to 9, and the system includes: The radar original complex echo data set construction module is used to select the data analysis area, extract strong and weak sea clutter data in high and low sea conditions, superimpose target echoes, and construct the model training set; The sea clutter suppression intelligent model construction module is used to preprocess the input model data set and unify the order of magnitude of the sea clutter data. Based on the generative adversarial theory, the generator and discriminator modules are complex-valued to achieve feature extraction of the radar complex-valued input data. Finally, the loss function module constrains the model training iteration to achieve mapping fitting between strong and weak clutter. The accuracy verification module is used to use the sea clutter improvement factor as an evaluation indicator of the network output result; based on the measured sea clutter data, the sea clutter improvement factor is used to perform accuracy verification.

Images