Sea clutter suppression method, device, equipment and medium based on generative adversarial network

Through the pyramid multi-scale residual attention model based on the generative adversarial network, the problem of sea clutter suppression in radar systems in marine environments is solved, efficient sea clutter suppression and complete retention of target information are achieved, and radar detection performance is improved.

CN118897275BActive Publication Date: 2024-12-06YANTAI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411388541.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-10-08
Publication Date
2024-12-06
Estimated Expiration
2044-10-08

AI Technical Summary

Technical Problem

The prior art is difficult to effectively suppress sea clutter in radar systems in marine environments, especially in the context of strong clutter, resulting in a degradation of target detection performance, and conventional methods may lead to blurring of target information or loss of information.

Method used

Using the sea clutter suppression method based on the generative adversarial network, a generative adversarial network with multi-scale residual attention in the pyramid is constructed, and the sea clutter suppression is performed using the mutual constraints of the generator and the discriminator, and the changes of sea clutter are learned and effective. Remove clutter while retaining target information.

Benefits of technology

It significantly improves the sea clutter suppression performance of the radar system, can effectively suppress sea clutter in complex marine environments, retain real target information, and still shows excellent clutter suppression effect under a small number of training samples.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118897275B_ABST
    Figure CN118897275B_ABST
Patent Text Reader

Abstract

The present invention discloses a sea clutter suppression method, device, equipment and medium based on a generative adversarial network, which belongs to the technical field of radar clutter suppression. The method comprises the following steps: obtaining sea clutter data from radar echo signals, constructing a sea clutter data set containing simulated targets and measured targets; constructing a generative adversarial network based on pyramid multi-scale residual attention, wherein the generative adversarial network comprises a mutually constrained generator and a discriminator; training the generative adversarial network using the sea clutter data set to obtain a sea clutter suppression network; and suppressing sea clutter on measured radar images through the sea clutter suppression network. The present invention learns the changing characteristics of sea clutter through the generator in the generative adversarial network, and suppresses sea clutter by using dynamic adversarial training between the generator and the discriminator, which not only solves the interference problem caused by sea clutter on target detection, but also improves the sea clutter suppression performance of the radar.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a sea clutter suppression method, device, equipment and medium based on a generative adversarial network, and belongs to the technical field of radar clutter suppression. Background Art

[0002] Radar systems operating in marine environments are often interfered by reflected signals from the sea surface, which are usually called sea clutter. Sea clutter can mask target echoes, resulting in a decrease in radar detection performance. In order to enhance the target detection capability of radar systems in marine environments, effective methods must be adopted to suppress the influence of sea clutter.

[0003] Traditional sea clutter suppression methods mainly include statistical-based methods and adaptive filtering methods. The literature [Wang W P, Feng Y, Shan T. A sea clutter suppression method using improved time-frequency filtering method [J]. J Signal Process, 2019, 35 (2): 208-216.] proposed an improved time-frequency filtering method, which analyzes the frequency variation characteristics of sea clutter, uses the frequency energy distribution to determine the frequency range of clutter, and uses the amplitude mean of the frequency interval as the filtering threshold statistical method. However, when the spectral characteristics of clutter and target are close, the filtering effect may be affected. [Yu J, Pan B, Yu Z, et al. Sea clutter suppression based on chaotic prediction model by combining the generator and long short-term memory networks [J]. Remote Sensing, 2024, 16 (7): 1260.] proposed a dual-base sea clutter suppression method combining compressed sensing optimization and transfer learning. By optimizing the sparse basis and using the orthogonal matching pursuit algorithm to reconstruct the sea clutter, a more efficient clutter suppression effect is achieved. However, this method has limited clutter suppression effect in complex and changeable ocean environments, especially in strong clutter backgrounds, and may not be able to effectively eliminate clutter, thus affecting the overall suppression performance.

[0004] With the improvement of computing power and the enrichment of data sets, machine learning-based methods have gradually become a hot topic of research, and image processing methods have greatly promoted the development of clutter suppression. [He K, Zhang X, Ren S, et al. Deep residual learning for image recognition [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778.] proposed the residual idea, which solved the problem of the difficulty of training deep convolutional neural networks. Many people have applied it in the field of image processing. [Isola P, Zhu JY, Zhou T, et al. Image-to-image translation with conditional adversarial networks [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1125-1134.] proposed a generative adversarial network (GAN) based on Pix2pix for image denoising, and achieved good results. The CycleGAN model proposed by [Harms J, Lei Y, Wang T, et al. Paired cycle-GAN-based image correction for quantitative cone-beam computed tomography[J]. Medical physics, 2019, 46(9): 3998-4009.] achieves efficient image-to-image conversion without the need for paired training data. However, in the ocean surveillance radar echo, the target echo coexists with sea clutter. These methods may cause blurring of target details or information loss while suppressing clutter. Therefore, how to fully retain the target information while suppressing sea clutter is crucial to improving the performance of final target detection. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a sea clutter suppression method, device, equipment and medium based on a generative adversarial network, which can improve the sea clutter suppression performance of the radar and solve the interference problem caused by sea clutter on target detection.

[0006] The technical solution adopted by the present invention to solve the technical problem is:

[0007] In a first aspect, an embodiment of the present invention provides a sea clutter suppression method based on a generative adversarial network, comprising the following steps:

[0008] Step 1, obtaining sea clutter data from radar echo signals and constructing a sea clutter data set containing simulated targets and measured targets;

[0009] Step 2, constructing a generative adversarial network based on pyramid multi-scale residual attention, wherein the generative adversarial network includes a mutually constrained generator and a discriminator;

[0010] Step 3, using the sea clutter data set to train the generative adversarial network to obtain a sea clutter suppression network;

[0011] Step 4: Suppress sea clutter on the measured radar image through the sea clutter suppression network.

[0012] As a possible implementation of this embodiment, the step 1 of acquiring sea clutter data from radar echo signals and constructing a sea clutter data set including simulated targets and measured targets includes the following steps:

[0013] Acquire a public radar data set, wherein the public radar data set includes at least an IPIX radar data set, a CSIR radar data set, and an X-band radar data set, wherein the IPIX radar data set and the CSIR radar data set both include real sea clutter and target data, and the X-band radar data set includes measured radar data of level 2 sea state, level 3 sea state, and level 4 sea state;

[0014] Acquire sea clutter data from the IPIX radar data set, and add two types of simulated targets, uniform acceleration targets and micro-motion targets, to the acquired sea clutter data as training sets;

[0015] The public radar dataset is used as a test set.

[0016] As a possible implementation of this embodiment, the step of acquiring sea clutter data from a radar echo signal is specifically as follows:

[0017] Short-time Fourier transform (STFT) is used to perform time-frequency analysis on radar signals in public radar datasets:

[0018] ,

[0019] in, Represents the radar echo signal, is the window function, is the integration variable, represents an imaginary unit;

[0020] The time-frequency diagram generated by short-time Fourier transform is used as the sea clutter dataset.

[0021] As a possible implementation of this embodiment, the step 2, constructing a generative adversarial network based on pyramid multi-scale residual attention, includes the following steps:

[0022] A generator is designed, in which a comprehensive residual attention module and a multi-scale pyramid adaptive feature extraction module are set; the comprehensive residual attention module includes a spatial attention module, a coordinate attention module and a residual connection module; the spatial attention module highlights the target area and suppresses the clutter in the background by adaptively adjusting the weights of each spatial position; the coordinate attention module locates the target area in the sea clutter image by introducing position information; the residual connection module processes the input feature map through a series of convolution layers and nonlinear activation functions to obtain an output feature map, and directly adds the input feature map to the output feature map; the multi-scale pyramid adaptive feature extraction module extracts features of different scales through multi-scale convolution kernels, and connects and fuses the extracted multi-scale features through an adaptive learning (AS) module;

[0023] Design a discriminator, which consists of a convolutional layer, a batch normalization layer, and an activation function layer;

[0024] A comprehensive loss function including adversarial loss, perceptual loss and L1 loss is set, and the generator and discriminator are optimized using the weighted sum of adversarial loss, perceptual loss and L1 loss.

[0025] As a possible implementation of this embodiment, the generator receives an input feature map containing clutter and targets, and gradually extracts features of the input feature map through multiple convolutional layers and activation functions;

[0026] The generator gradually extracts the features of the input feature map through multiple layers of convolutional layers and activation functions. The specific process is as follows:

[0027] Perform global average pooling on the input feature map to obtain two feature maps of different scales and feature map :

[0028] , ,

[0029] in, F is the input feature map, To perform global pooling on the feature map, To downsample the feature map;

[0030] For feature maps and feature map The two feature maps are respectively convolved by 1x1 to adjust the number of channels:

[0031] , ,

[0032] in To perform convolution operation on the feature map:

[0033] The feature map Upsampling to feature map Same size, and concatenate on the channel dimension:

[0034] ,

[0035] in To perform a full connection operation on the feature map;

[0036] The number of channels is adjusted through convolution, and fused with the input feature map by element-wise multiplication to obtain the final feature map:

[0037] =Conv1x1 ( ), .

[0038] As a possible implementation of this embodiment, the discriminator receives real sea clutter data and an image output by the generator, extracts image features through a series of convolutional layers, and uses a deepened network structure to distinguish clutter signals and target signals in the input image, thereby evaluating the output quality of the generator.

[0039] As a possible implementation of this embodiment, the comprehensive loss function is:

[0040] ,

[0041] ,

[0042] in, is the optimal generator, is the optimal discriminator, and They are the minimum and minimum solution functions of the generator discriminator loss functions, respectively; , and , They are the weight coefficients of adversarial loss, perceptual loss, generator and discriminator L1 loss respectively;

[0043] Represents adversarial loss:

[0044] ,

[0045] in, Represents the discriminator’s discrimination result on the real image, represents the discriminator’s judgment result on the generated image, is a real image, is the input noise of the generator;

[0046] Represents the perceptual loss:

[0047] ,

[0048] in Indicates i Feature maps of layers;

[0049] and Represents the L1 loss of the generator and the discriminator respectively:

[0050] ,

[0051] ,

[0052] in Indicates the generated image.

[0053] In a second aspect, an embodiment of the present invention provides a sea clutter suppression device based on a generative adversarial network, comprising:

[0054] A data set construction module is used to obtain sea clutter data from radar echo signals and construct a sea clutter data set containing simulated targets and measured targets;

[0055] A generative adversarial network building module, used to build a pyramid multi-scale residual attention-based generative adversarial network, wherein the generative adversarial network includes a mutually constrained generator and a discriminator;

[0056] A generative adversarial network optimization module, used to train the generative adversarial network using the sea clutter data set to obtain a sea clutter suppression network;

[0057] The sea clutter suppression module is used to suppress sea clutter on the measured radar image through the sea clutter suppression network.

[0058] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory through the bus, and the processor executes the machine-readable instructions to perform any step of the above-mentioned sea clutter suppression method based on a generative adversarial network.

[0059] In a fourth aspect, an embodiment of the present invention provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned sea clutter suppression methods based on a generative adversarial network are executed.

[0060] The beneficial effects of the technical solution of the embodiment of the present invention are as follows:

[0061] A sea clutter suppression method based on a generative adversarial network according to the technical solution of an embodiment of the present invention comprises the following steps: step 1, obtaining sea clutter data from a radar echo signal, and constructing a sea clutter data set including simulated targets and measured targets; step 2, constructing a generative adversarial network based on pyramid multi-scale residual attention, wherein the generative adversarial network comprises a mutually constrained generator and a discriminator; step 3, training the generative adversarial network using the sea clutter data set to obtain a sea clutter suppression network; step 4, suppressing sea clutter on a measured radar image through the sea clutter suppression network. The present invention learns the changing characteristics of sea clutter through a generator in a generative adversarial network, and effectively suppresses sea clutter by using dynamic adversarial training between the generator and the discriminator. The present invention effectively removes sea clutter while retaining real radar target information, and has good practicality and broad application prospects.

[0062] The present invention introduces a comprehensive residual attention mechanism in the generator, and significantly improves the prominence and positioning accuracy of the target area by adaptively adjusting the spatial position weight and introducing position information encoding; the present invention uses convolution kernels of different sizes for multi-scale feature extraction, and performs feature fusion through an adaptive learning module, which can generate high-quality radar images after sea clutter suppression; the method proposed in the present invention can still show excellent clutter suppression performance even when there are only a small number of training samples. The present invention not only solves the interference problem caused by sea clutter on target detection, but also improves the sea clutter suppression performance of the radar. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flow chart of a sea clutter suppression method based on a generative adversarial network according to an exemplary embodiment;

[0064] Figure 2 is a schematic structural diagram of a sea clutter suppression device based on a generative adversarial network according to an exemplary embodiment;

[0065] Figure 3 It is a specific flow chart of sea clutter suppression using the sea clutter suppression device based on the generative adversarial network of the present invention;

[0066] Figure 4 is a framework diagram of a generator according to an exemplary embodiment;

[0067] Figure 5 is a network diagram of channel attention and coordinate attention according to an exemplary embodiment;

[0068] Figure 6 is a structural diagram of a multi-scale pyramid adaptive feature extraction module according to an exemplary embodiment;

[0069] Figure 7 is a structural diagram of a discriminator according to an exemplary embodiment;

[0070] Figure 8 is a time-frequency diagram comparison diagram of an IPIX uniform acceleration target under different sea clutter suppression methods according to an exemplary embodiment;

[0071] Fig. 9 is a time-frequency diagram comparison diagram of an IPIX micro-motion target under different sea clutter suppression methods according to an exemplary embodiment;

[0072] Fig.10 is a time-frequency diagram comparison diagram of a CSIR measured data set under different sea clutter suppression methods according to an exemplary embodiment;

[0073] Fig.11 is a diagram showing the clutter suppression effect of an X-band radar of the present invention under different sea conditions according to an exemplary embodiment;

[0074] Fig.12 is a comparison diagram of clutter suppression effects of different loss functions according to an exemplary embodiment. DETAILED DESCRIPTION

[0075] In order to more clearly illustrate the technical features of the solution of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0076] like Figure 1 As shown, a sea clutter suppression method based on a generative adversarial network provided by an embodiment of the present invention comprises the following steps:

[0077] Step 1, obtaining sea clutter data from radar echo signals and constructing a sea clutter data set containing simulated targets and measured targets;

[0078] Step 2, constructing a generative adversarial network based on pyramid multi-scale residual attention, wherein the generative adversarial network includes a mutually constrained generator and a discriminator;

[0079] Step 3, using the sea clutter data set to train the generative adversarial network to obtain a sea clutter suppression network;

[0080] Step 4: Suppress sea clutter on the measured radar image through the sea clutter suppression network.

[0081] As a possible implementation of this embodiment, the step 1 of acquiring sea clutter data from radar echo signals and constructing a sea clutter data set including simulated targets and measured targets includes the following steps:

[0082] Acquire a public radar data set, wherein the public radar data set includes at least an IPIX radar data set, a CSIR radar data set, and an X-band radar data set, wherein the IPIX radar data set and the CSIR radar data set both include real sea clutter and target data, and the X-band radar data set includes measured radar data of level 2 sea state, level 3 sea state, and level 4 sea state;

[0083] Acquire sea clutter data from the IPIX radar data set, and add two types of simulated targets, uniform acceleration targets and micro-motion targets, to the acquired sea clutter data as training sets;

[0084] The public radar dataset is used as a test set.

[0085] As a possible implementation of this embodiment, the step of acquiring sea clutter data from a radar echo signal is specifically as follows:

[0086] Short-time Fourier transform (STFT) is used to perform time-frequency analysis on radar signals in public radar datasets:

[0087] ,

[0088] in, Represents the radar echo signal, is the window function, is the integration variable, represents an imaginary unit;

[0089] The time-frequency diagram generated by short-time Fourier transform is used as the sea clutter dataset.

[0090] As a possible implementation of this embodiment, the step 2, constructing a generative adversarial network based on pyramid multi-scale residual attention, includes the following steps:

[0091] A generator is designed, in which a comprehensive residual attention module and a multi-scale pyramid adaptive feature extraction module are set; the comprehensive residual attention module includes a spatial attention module, a coordinate attention module and a residual connection module; the spatial attention module highlights the target area and suppresses the clutter in the background by adaptively adjusting the weights of each spatial position; the coordinate attention module locates the target area in the sea clutter image by introducing position information; the residual connection module processes the input feature map through a series of convolution layers and nonlinear activation functions to obtain an output feature map, and directly adds the input feature map to the output feature map; the multi-scale pyramid adaptive feature extraction module extracts features of different scales through multi-scale convolution kernels, and connects and fuses the extracted multi-scale features through an adaptive learning (AS) module;

[0092] Design a discriminator, which consists of a convolutional layer, a batch normalization layer, and an activation function layer;

[0093] A comprehensive loss function including adversarial loss, perceptual loss and L1 loss is set, and the generator and discriminator are optimized using the weighted sum of adversarial loss, perceptual loss and L1 loss.

[0094] As a possible implementation of this embodiment, the generator receives an input feature map F containing clutter and targets, and gradually extracts features of the input feature map F through multiple convolutional layers and activation functions;

[0095] The generator gradually extracts the features of the input feature map F through multiple layers of convolutional layers and activation functions. The specific process is as follows:

[0096] Perform global average pooling on the input feature map F to obtain two feature maps of different scales and feature map :

[0097] , ,

[0098] in To perform global pooling on the feature map, To downsample the feature map;

[0099] For feature maps and feature map The two feature maps are respectively convolved by 1x1 to adjust the number of channels:

[0100] , ,

[0101] in To perform convolution operation on the feature map:

[0102] The feature map Upsampling to feature map Same size, and concatenate on the channel dimension:

[0103] ,

[0104] in To perform a full connection operation on the feature map;

[0105] The number of channels is adjusted through convolution, and fused with the input feature map by element-wise multiplication to obtain the final feature map:

[0106] =Conv1x1 ( ), .

[0107] As a possible implementation of this embodiment, the discriminator receives real sea clutter data and an image output by the generator, extracts image features through a series of convolutional layers, and uses a deepened network structure to distinguish clutter signals and target signals in the input image, thereby evaluating the output quality of the generator.

[0108] As a possible implementation of this embodiment, the comprehensive loss function is:

[0109] ,

[0110] ,

[0111] in, is the optimal generator, is the optimal discriminator, and They are the minimum and minimum solution functions of the generator discriminator loss functions, respectively; , and , They are the weight coefficients of adversarial loss, perceptual loss, generator and discriminator L1 loss respectively;

[0112] Represents adversarial loss:

[0113] ,

[0114] in, Represents the discriminator’s discrimination result on the real image, represents the discriminator’s judgment result on the generated image, is a real image, is the input noise of the generator;

[0115] Represents the perceptual loss:

[0116] ,

[0117] in Indicates i Feature maps of layers;

[0118] and Represents the L1 loss of the generator and the discriminator respectively:

[0119] ,

[0120] ,

[0121] in Indicates the generated image.

[0122] like Figure 2 As shown, an embodiment of the present invention provides a sea clutter suppression device based on a generative adversarial network, comprising:

[0123] A data set construction module is used to obtain sea clutter data from radar echo signals and construct a sea clutter data set containing simulated targets and measured targets;

[0124] A generative adversarial network building module, used to build a pyramid multi-scale residual attention-based generative adversarial network, wherein the generative adversarial network includes a mutually constrained generator and a discriminator;

[0125] A generative adversarial network optimization module, used to train the generative adversarial network using the sea clutter data set to obtain a sea clutter suppression network;

[0126] The sea clutter suppression module is used to suppress sea clutter on the measured radar image through the sea clutter suppression network.

[0127] like Figure 3 As shown, the specific process of performing sea clutter suppression using the sea clutter suppression device based on the generative adversarial network of the present invention is as follows.

[0128] First, obtain sea clutter data and construct a data set containing simulated targets and measured targets under the real sea clutter background.

[0129] The present invention uses three different data sets, namely, the IPIX radar data set, the CSIR radar data set and the X-band radar maritime detection data set published by the Naval Aviation University.

[0130] The IPIX radar dataset is a classic sea clutter dataset. The present invention adds two types of simulated targets to the dataset, namely uniformly accelerated targets and micro-motion targets, to simulate the complex situations in different radar detection scenarios. This enables the dataset to not only test the network's ability to suppress actual sea clutter, but also to test its performance in complex moving target detection.

[0131] The CSIR radar dataset provides real sea clutter and target data, which is mainly used to evaluate the suppression effects of different methods under real data. In this dataset, the focus is on the comparison of the visual effects of each method after suppression, so that we can observe the suppression effect from an intuitive perspective.

[0132] The X-band radar dataset used focuses more on the suppression effect under different sea conditions. This dataset is made public by the Naval Aviation University. The present invention uses measured radar data of level 2, level 3, and level 4 sea conditions. These data simulate sea conditions ranging from relatively calm to more complex, allowing us to evaluate the robustness of the proposed method under different environmental conditions.

[0133] The specific information of each dataset and its usage are shown in Table 1.

[0134] Table 1 Dataset settings

[0135]

[0136] In order to evaluate the performance of the network more comprehensively, the present invention adds two simulated targets to the IPIX band data set to simulate the complex scenes in radar detection. The two simulated targets have different motion parameters, as shown in Table 2.

[0137] Table 2 Target motion parameters

[0138]

[0139] In order to use these data more effectively, the present invention uses short-time Fourier transform (STFT) to perform time-frequency analysis on radar signals. STFT is suitable for real-time processing and online analysis due to its fast calculation speed and good time-frequency resolution. Its formula is as follows:

[0140] <1> ,

[0141] in, Represents the radar echo signal, is the window function, is the integration variable, represents the imaginary unit. The time-frequency diagram generated by this time-frequency transformation is used as the data set for the model.

[0142] 2. Construct a pyramid-based multi-scale residual attention generative adversarial network, in which the generator and discriminator constrain each other.

[0143] 2.1 Design of the generator.

[0144] The generator proposed by the present invention is as follows Figure 4 As shown in the figure. The generator receives an image containing clutter and targets, and gradually extracts the features of the image through multiple layers of convolutional layers and activation functions. In this process, the generator continuously learns and models the changing characteristics of sea clutter to generate images similar to the input but with lower clutter interference. To improve the performance of the generator, residual connection modules are introduced into the generator. These modules improve the training stability and convergence speed of the network by alleviating the gradient vanishing problem, enabling the generator to better suppress sea clutter and highlight the target signal.

[0145] The pyramid multi-scale residual attention module is a deep convolutional network, and its attention module consists of three parts: spatial attention module, coordinate attention module and multi-scale pyramid adaptive feature extraction module.

[0146] The spatial attention and coordinate attention modules used in the present invention are as follows: Figure 5 As shown in the figure, when processing sea clutter images, the spatial attention module can effectively highlight the target area and suppress the clutter in the background by adaptively adjusting the weights of each spatial position. The input feature map is subjected to global average pooling and maximum pooling respectively to obtain two different feature maps. The average feature map and the maximum feature map are concatenated in the channel dimension to form a comprehensive feature map. The concatenated feature map is processed by a convolution layer, mapped to a single channel, and a spatial attention map is generated. The output of the convolution layer is normalized using an activation function to obtain a spatial attention map, which can adaptively highlight the features of the target area. The spatial attention map is element-wise multiplied with the input sea clutter image to enhance the features of the target area while suppressing the clutter in the background.

[0147] The coordinate attention module can introduce position information to help the network locate the target area in the sea clutter image more accurately. The input sea clutter image is globally averaged pooled in the height and width dimensions to obtain two feature maps, which represent the global height and width information respectively. The above two feature maps are spliced ​​in the channel dimension and mapped through a convolution layer to obtain a low-dimensional feature map. The low-dimensional feature map is processed using a batch normalization layer and an activation function to improve the feature expression ability. The low-dimensional feature map is divided into two parts, height and width, and the corresponding coordinate attention map is generated through two convolution layers respectively. The convolution layer output is normalized using an activation function to obtain the final coordinate attention map. The generated coordinate attention map is element-wise multiplied with the input sea clutter image to guide the network to focus on the target area and further suppress the clutter in the background.

[0148] Residual connection, the input feature map is processed through a series of convolutional layers and nonlinear activation functions to obtain the output feature map. The input feature map and the output feature map are directly added to form a residual connection. Through the residual connection, the network can retain more information when transmitting features, improving the training efficiency and effect of the network.

[0149] In deep feature processing, regularization-multi-scale connection is used to effectively adjust the features of sea clutter images through standardization and scaling operations to enhance the robustness and adaptability of the network. The mean and standard deviation of the feature matrix of the nth layer are calculated as:

[0150] <2> ,

[0151] <3> ,

[0152] <4> ,

[0153] in is the output of the new feature tensor, It is Layer input neurons, is the number of neurons in this layer, Add a small constant to the input feature tensor To prevent the standard deviation from being 0, set , , for , A collection of , are the standard deviation and mean extracted from the regularization process.

[0154] In this way, the network can learn the optimal scale and distribution of feature maps, further improving the suppression effect of sea clutter.

[0155] like Figure 6 As shown in the figure, the output features after the comprehensive residual attention module are input to the multi-scale pyramid adaptive feature extraction module. Multi-scale features are extracted by using 3x3, 5x5 and 7x7 convolution kernels to capture targets and clutter information of different sizes. The features extracted by 3x3 and 5x5 convolution kernels are processed by adaptive learning module 1. The feature map output by adaptive learning module 1 and the features extracted by 7x7 convolution kernel are processed in adaptive learning module 2.

[0156] The specific process is:

[0157] Input feature map F and perform global average pooling to obtain two feature maps of different scales and :

[0158] , <5> ,

[0159] in To perform global pooling on the feature map, To downsample the feature map.

[0160] Through 1x1 convolution on the two feature maps, adjust the number of channels:

[0161] , <6> ,

[0162] in To perform convolution operation on the feature map.

[0163] Upsample the second feature map to the same size as the first feature map and concatenate along the channel dimension:

[0164] <7> ,

[0165] in To perform a full connection operation on the feature map.

[0166] The number of channels is adjusted again through convolution, and fused with the input feature map by element-wise multiplication to obtain the final feature map:

[0167] =Conv1x1 ( ), <8> .

[0168] 2.2 Design of the discriminator.

[0169] The discriminator proposed in the present invention is as follows Figure 7 As shown in the figure, the discriminator receives the real sea clutter data and the image output by the generator, extracts image features through a series of convolutional layers, and uses a deepened network structure to enhance its discriminative ability. The role of the discriminator is to distinguish the clutter signal and the target signal in the input image, so as to evaluate the output quality of the generator. By continuously optimizing the discriminator, the clutter and target in the real data can be separated more accurately, providing a reliable basis for further sea clutter suppression.

[0170] 2.3 Training process

[0171] The generator and the discriminator are optimized by alternating training, so that the generator continuously generates more realistic images, while the discriminator continuously improves its ability to distinguish between real and generated data. This paper proposes a new loss function, which uses the weighted sum of adversarial loss, perceptual loss and L1 loss to optimize the generator and the discriminator, optimizes the performance of the generation model from different levels and angles, and ensures that the generated images meet high standards in terms of authenticity, perceptual quality and pixel-level consistency. The comprehensive loss function is as follows:

[0172] <9> ,

[0173] <10> ,

[0174] in, is the optimal generator, is the optimal discriminator, and They are the minimum and minimum solution functions of the generator discriminator loss functions, respectively; , and , They are the weight coefficients of adversarial loss, perceptual loss, generator and discriminator L1 loss respectively.

[0175] Specifically, Stands for adversarial loss: Through the adversarial relationship between the generator and the discriminator, the generator is forced to generate more realistic images, thereby improving the authenticity of the generated images.

[0176] <11> ,

[0177] in, Represents the discriminator’s discrimination result on the real image, represents the discriminator’s judgment result on the generated image, is a real image, is the input noise of the generator.

[0178] It stands for perceptual loss. It uses the pre-trained VGG19 to extract features and compare the generated image with the real image in the high-level feature space to ensure the consistency of the perceptual quality of the generated image and improve the realism of details and textures.

[0179] <12> ,

[0180] in Indicates i The feature map of the layer.

[0181] and They represent the L1 loss of the generator and the discriminator respectively, and directly calculate the difference between the generated image and the real image in the pixel space to ensure the consistency of the overall structure between the generated image and the real image and reduce large-scale deviations and blurs.

[0182] <13> ,

[0183] <14> ,

[0184] The training goal of the generator is to maximize the fidelity of the generated data and at the same time maximize the probability of it being correctly classified. The training goal of the discriminator is to enhance the ability to distinguish between real data and fake data as much as possible. Through the back-propagation algorithm, the gradient of the loss function is propagated from the output layer to the input layer, the gradient of each weight layer is calculated layer by layer, and these gradients are used to modify the parameters of the network. The above process is repeated until the model reaches a good equilibrium state. Finally, all the parameters of the discriminator are saved.

[0185] Third, the constructed dataset is used to train the constructed generative adversarial network, and finally the sea clutter suppression network is obtained.

[0186] Since the generative adversarial network in the training phase requires paired data, but it is very difficult to obtain measured sea clutter data that meets the requirements in practice, the present invention uses the IPIX measured sea clutter data of 1993 as the pure sea clutter background in the training phase. Based on this background, two types of simulated targets, uniform acceleration and micro motion, are added, and the combined image containing the simulated target and the measured sea clutter is used as the input of the generator, while the output of the generator and the corresponding image containing only the simulated target are used as the input of the discriminator, and finally a network model with the ability to suppress sea clutter is obtained.

[0187] In addition, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) are used as performance test indicators.

[0188] Among them, PSNR is used to measure the difference between the image after clutter removal and the real radar image at the pixel level. A higher PSNR value indicates better overall clutter removal performance and higher image fidelity. Its calculation formula is as follows:

[0189] <15> ,

[0190] in, is the maximum value of the image pixel value, It represents the mean square error between the original image and the image after denoising.

[0191] On the other hand, SSIM measures the similarity between the denoised image and the real image in terms of target retention and detail presentation by analyzing the brightness, contrast and structural information of the image. The value range of this indicator is [0,1]. The closer the value is to 1, the less distortion there is in the structure and detail level of the image. The calculation formula of SSIM is:

[0192] <16> ,

[0193] in represents the original radar image including sea clutter and targets, represents the image after sea clutter suppression, and Respectively represent images and The mean of and Respectively represent images and The standard deviation of It means the image and The covariance between and is a constant used to stabilize the calculation.

[0194] The effect of the present invention is further illustrated by the following comparative experiments:

[0195] 1. Experimental environment configuration.

[0196] All experiments were conducted on a high-performance computer with the following configuration: Intel(R) Xeon(R) W-2255 CPU @ 3.70GHz, 64GB memory and dual NVIDIA RTX A4000 graphics cards (16GB each). The operating system was Windows 10, and the experimental framework was Pytorch. The experiment used the Adam optimizer with an initial learning rate of 0.0002. The entire training process was performed for 100 iterations.

[0197] 2. Analysis and evaluation of experimental results.

[0198] In order to verify the effectiveness of the proposed algorithm, the present invention is compared with several sea clutter removal methods, including the classic ResNet method as well as Pix2pix and CycleGAN, to comprehensively analyze the relative performance of the proposed algorithm.

[0199] 2.1 IPIX radar dataset

[0200] The sea clutter suppression results under uniform acceleration target are shown in Figure 8 As shown, Figure 8(a) is the original time-frequency image, including the target and sea clutter. Figure 8 (b) is the time-frequency diagram after suppressing sea clutter using the CycleGAN method. It can be seen that this method can suppress most of the clutter, but there is still some clutter interference. Figure 8 (c) Figure 8 (d) are the time-frequency diagrams after suppressing sea clutter using the Pix2pix method and the ResNet method. Compared with CycleGAN, these two methods can effectively suppress sea clutter, but the target signal is also affected to a certain extent. Figure 8 (e) is the time-frequency diagram after suppressing sea clutter using the PMS-RA GAN method proposed in this invention. This method can highlight the target signal most clearly while suppressing sea clutter, and performs best among all the methods.

[0201] Fig. 9 is the result of sea clutter suppression under micro-motion target, Fig. 9 (a) is the original time-frequency image, Fig. 9 (b) Fig. 9 (c) Fig. 9 (d) The time-frequency diagram after suppressing sea clutter using the CycleGAN method, Pix2pix method, and ResNet method, respectively. Fig. 9 (e) is the time-frequency diagram after suppressing sea clutter using the PMS-RA GAN method proposed in this invention. Fig. 9 (a) It can be seen that the target is submerged in the sea clutter. The Cycle GAN method performs well in suppressing sea clutter, and the target signal is clearer, but there is still a small amount of residual clutter. The Pix2pix method suppresses part of the target as clutter while suppressing the clutter, and the ResNet method cannot completely retain the target signal. The PMS-RA GAN method proposed in this invention shows good suppression performance even when the target is difficult to distinguish from the clutter.

[0202] The PSNR and SSIM of various methods for the data are shown in Table 3. From the results in Table 3, it can be seen that PMS-RA GAN performs best on the IPIX radar dataset. Its PSNR value is 33.6761dB, which is significantly higher than other methods, indicating that it is the most superior in restoring the clarity of the target image. CycleGAN follows closely with a PSNR of 31.2027dB, which is slightly inferior to PMS-RA GAN, but still significantly better than Pix2pix's 29.4869 dB and ResNet's 27.8387 dB. In terms of structural similarity index (SSIM), PMS-RA GAN also performs best, with an SSIM of 0.9861, which is much higher than other methods. CycleGAN's SSIM is 0.9624, ranking second, while Pix2pix and ResNet are 0.9385 and 0.9240 respectively. Overall, the high PSNR and SSIM values ​​of PMS-RA GAN indicate its significant advantages in suppressing sea clutter and preserving target information, demonstrating its superior performance in practical applications.

[0203] Table 3 PSNR and SSIM (dB) of different methods

[0204]

[0205] 2.2 CSIR measured data

[0206] Fig.10 The time-frequency diagram comparison of sea clutter suppression using different methods under the CSIR measured data set is shown. Fig. 9 (a) is the original time-frequency image, Fig. 9 (b) Fig. 9 (c) Fig. 9 (d) The time-frequency diagram after suppressing sea clutter using the CycleGAN method, Pix2pix method, and ResNet method, respectively. Fig. 9 (e) is a time-frequency diagram after suppressing sea clutter using the PMS-RA GAN method proposed in the present invention. The target and background sea clutter are both measured. It can be seen that the CycleGAN method is not ideal in suppressing sea clutter, and there is still a lot of clutter. The Pix2pix and ResNet methods also weaken the target signal when suppressing sea clutter. The target is almost completely suppressed and difficult to identify. The PMS-RA GAN proposed in the present invention performs best among all methods, can effectively suppress sea clutter, and clearly highlight the target signal, showing excellent sea clutter suppression performance.

[0207] 2.3 X-band radar dataset.

[0208] In the experiment of X-band radar dataset, the performance of PMS-RA GAN was tested in level 2, level 3 and level 4 sea conditions to verify the performance of the method in different sea conditions. Fig.11 shown. Fig.11 (a) Fig.11 (b) and Fig.11 (c) are the clutter suppression effects in sea conditions of level 2, level 3, and level 4, respectively. For sea conditions of level 2, the sea surface is relatively calm, and PMS-RA GAN can easily suppress clutter and retain the target clearly and completely; in sea conditions of level 3, the sea surface waves increase, and the proposed method still effectively reduces clutter interference, and the target information is well preserved; even in sea conditions of level 4, in the face of enhanced clutter, PMS-RA GAN still shows excellent robustness and successfully retains the main target structure. These results show that regardless of the sea conditions, PMS-RA GAN can maintain good suppression effects in various environments, proving its reliability and adaptability in practical applications.

[0209] 2.4 Loss Function Performance.

[0210] Fig.12 The comparison of the clutter suppression effect of the proposed algorithm on the time-frequency graph based on different loss functions in the IPIX dataset is shown: Fig.12 (a) is the original image, showing the time-frequency image without sea clutter suppression. Obvious sea clutter and target features can be seen. Fig.12 (b) shows the clutter suppression effect of using MAE loss alone. It can be seen that clutter can be suppressed well. However, since MAE loss mainly focuses on minimizing the global error, the detailed information of the target is partially lost, especially in areas with low signal-to-clutter ratio, where part of the target information is lost. Fig.12 (c) Using only perceptual loss, the image retains more detail information. However, the sea clutter suppression effect is slightly worse than that of MAE loss, and there is still a certain degree of clutter interference in the image. The results of using only adversarial loss are shown in Figure 2. Fig.12 (d) The sea clutter suppression image generated by the model reduces clutter interference overall. However, the adversarial loss causes the loss of some detail information, and the local information of the target is not as clear as when the perceptual loss is used. Fig.12 (e) Using the comprehensive loss function, it can be seen that the proposed method effectively retains the target information in the time-frequency diagram and suppresses most of the sea clutter interference. The target is well presented in the image, and the overall image is clearer, showing excellent performance in a complex sea clutter environment.

[0211] An electronic device provided by an embodiment of the present invention includes a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory through the bus, and the processor executes the machine-readable instructions to perform any step of the sea clutter suppression method based on a generative adversarial network as described above.

[0212] Specifically, the above-mentioned memory and processor can be general-purpose memory and processor, which are not specifically limited here. When the processor runs the computer program stored in the memory, the above-mentioned sea clutter suppression method based on the generative adversarial network can be executed.

[0213] Those skilled in the art will appreciate that the structure of the computer device does not constitute a limitation on the computer device, and may include more or fewer components than shown in the figure, or combine certain components, or split certain components, or arrange the components differently.

[0214] In some embodiments, the computer device may also include a touch screen that can be used to display a graphical user interface (e.g., a startup interface of an application) and receive user operations on the graphical user interface (e.g., startup operations on an application). The specific touch screen may include a display panel and a touch panel. The display panel may be configured in the form of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), etc. The touch panel may collect the user's contact or non-contact operations on or near it, and generate pre-set operation instructions, for example, the user uses any suitable object such as a finger, a stylus, or an accessory on or near the touch panel. In addition, the touch panel may include two parts: a touch detection device and a touch controller. The touch detection device detects the user's touch position and posture, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into information that the processor can process, and then sends it to the processor, and can receive and execute commands from the processor. In addition, the touch panel can be implemented by various types such as resistive, capacitive, infrared and surface acoustic wave, and any technology developed in the future can also be used to implement the touch panel. Further, the touch panel can cover the display panel, and the user can operate on or near the touch panel covered on the display panel according to the graphical user interface displayed on the display panel. After the touch panel detects the operation on or near it, it is transmitted to the processor to determine the user input, and then the processor provides corresponding visual output on the display panel in response to the user input. In addition, the touch panel and the display panel can be implemented as two independent components or integrated.

[0215] Corresponding to the method for starting the above-mentioned application, an embodiment of the present invention further provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned sea clutter suppression methods based on a generative adversarial network are executed.

[0216] The application startup device provided in the embodiment of the present application can be specific hardware on the device or software or firmware installed on the device. The device provided in the embodiment of the present application, its implementation principle and the technical effect produced are the same as those in the aforementioned method embodiment. For the sake of brief description, the parts not mentioned in the device embodiment can refer to the corresponding contents in the aforementioned method embodiment. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can all refer to the corresponding processes in the aforementioned method embodiment, and will not be repeated here.

[0217] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0218] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0219] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0220] In addition, each functional module in the embodiments provided in the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0221] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0222] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0223] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0224] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A sea clutter suppression method based on generative adversarial network, It is characterized in that The steps include: Step 1, obtaining sea clutter data from radar echo signals and constructing a sea clutter data set containing simulated targets and measured targets; Step 2, constructing a generative adversarial network based on pyramid multi-scale residual attention, wherein the generative adversarial network includes a mutually constrained generator and a discriminator; Step 3, using the sea clutter data set to train the generative adversarial network to obtain a sea clutter suppression network; Step 4, performing sea clutter suppression on the measured radar image through the sea clutter suppression network; The step 2, constructing a generative adversarial network based on pyramid multi-scale residual attention, includes the following steps: A generator is designed, in which a comprehensive residual attention module and a multi-scale pyramid adaptive feature extraction module are set; the comprehensive residual attention module includes a spatial attention module, a coordinate attention module and a residual connection module; the spatial attention module highlights the target area and suppresses the clutter in the background by adaptively adjusting the weight of each spatial position; the coordinate attention module locates the target area in the sea clutter image by introducing position information; The residual connection module processes the input feature map through a series of convolution layers and nonlinear activation functions to obtain an output feature map, and directly adds the input feature map to the output feature map; the multi-scale pyramid adaptive feature extraction module extracts features of different scales through a multi-scale convolution kernel, and connects and fuses the extracted multi-scale features through an adaptive learning module; Design a discriminator, which consists of a convolutional layer, a batch normalization layer, and an activation function layer; A comprehensive loss function including adversarial loss, perceptual loss and L1 loss is set, and the generator and discriminator are optimized using the weighted sum of adversarial loss, perceptual loss and L1 loss.

2. The sea clutter suppression method based on generative adversarial network according to claim 1, It is characterized in that The step 1, obtaining sea clutter data from radar echo signals, and constructing a sea clutter data set including simulated targets and measured targets, comprises the following steps: Acquire a public radar data set, wherein the public radar data set includes at least an IPIX radar data set, a CSIR radar data set, and an X-band radar data set, wherein the IPIX radar data set and the CSIR radar data set both include real sea clutter and target data, and the X-band radar data set includes measured radar data of level 2 sea state, level 3 sea state, and level 4 sea state; Acquire sea clutter data from the IPIX radar data set, and add two types of simulated targets, uniform acceleration targets and micro-motion targets, to the acquired sea clutter data as training sets; The public radar dataset is used as a test set.

3. The sea clutter suppression method based on generative adversarial network according to claim 2, It is characterized in that The step of obtaining sea clutter data from radar echo signals is specifically as follows: Short-time Fourier transform is used to perform time-frequency analysis on radar signals in public radar datasets: , in, Represents the radar echo signal, is the window function, is the integration variable, represents an imaginary unit; The time-frequency diagram generated by short-time Fourier transform is used as the sea clutter dataset.

4. The sea clutter suppression method based on generative adversarial network according to claim 1, It is characterized in that The generator receives an input feature map containing clutter and targets, and gradually extracts features of the input feature map through multiple convolutional layers and activation functions; The generator gradually extracts the features of the input feature map through multiple layers of convolutional layers and activation functions. The specific process is as follows: Perform global average pooling on the input feature map to obtain two feature maps of different scales and feature map : , , in, F is the input feature map, To perform global pooling on the feature map, To downsample the feature map; For feature maps and feature map The two feature maps are respectively convolved by 1x1 to adjust the number of channels: , , in To perform convolution operation on the feature map: The feature map Upsampling to feature map Same size, and concatenate on the channel dimension: , in To perform a full connection operation on the feature map; The number of channels is adjusted through convolution, and fused with the input feature map by element-wise multiplication to obtain the final feature map: =Conv1x1 ( ), 。 5. The sea clutter suppression method based on generative adversarial network according to claim 1, It is characterized in that The discriminator receives real sea clutter data and the image output by the generator, extracts image features through a series of convolutional layers, and uses a deepened network structure to distinguish clutter signals and target signals in the input image, thereby evaluating the output quality of the generator.

6. The sea clutter suppression method based on generative adversarial network according to claim 1, It is characterized in that The comprehensive loss function is: , , in, is the optimal generator, is the optimal discriminator, and They are the minimum and minimum solution functions of the generator discriminator loss functions, respectively; , and , They are the weight coefficients of adversarial loss, perceptual loss, generator and discriminator L1 loss respectively; Represents adversarial loss: , in, Represents the discriminator’s discrimination result on the real image, represents the discriminator’s judgment result on the generated image, is a real image, is the input noise of the generator; Represents the perceptual loss: , in Indicates i Feature maps of layers; and Represents the L1 loss of the generator and the discriminator respectively: , , in Indicates the generated image.

7. A sea clutter suppression device based on generative adversarial network, It is characterized in that include: A data set construction module is used to obtain sea clutter data from radar echo signals and construct a sea clutter data set containing simulated targets and measured targets; A generative adversarial network building module, used to build a pyramid multi-scale residual attention-based generative adversarial network, wherein the generative adversarial network includes a mutually constrained generator and a discriminator; A generative adversarial network optimization module, used to train the generative adversarial network using the sea clutter data set to obtain a sea clutter suppression network; A sea clutter suppression module, used for suppressing sea clutter on the measured radar image through the sea clutter suppression network; The specific process of the generative adversarial network construction module constructing a generative adversarial network based on pyramid multi-scale residual attention includes the following steps: A generator is designed, in which a comprehensive residual attention module and a multi-scale pyramid adaptive feature extraction module are set; the comprehensive residual attention module includes a spatial attention module, a coordinate attention module and a residual connection module; the spatial attention module highlights the target area and suppresses the clutter in the background by adaptively adjusting the weight of each spatial position; the coordinate attention module locates the target area in the sea clutter image by introducing position information; The residual connection module processes the input feature map through a series of convolution layers and nonlinear activation functions to obtain an output feature map, and directly adds the input feature map to the output feature map; the multi-scale pyramid adaptive feature extraction module extracts features of different scales through a multi-scale convolution kernel, and connects and fuses the extracted multi-scale features through an adaptive learning module; Design a discriminator, which consists of a convolutional layer, a batch normalization layer, and an activation function layer; A comprehensive loss function including adversarial loss, perceptual loss and L1 loss is set, and the generator and discriminator are optimized using the weighted sum of adversarial loss, perceptual loss and L1 loss.

8. An electronic device, It is characterized in that The electronic device comprises a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory through the bus, and the processor executes the machine-readable instructions to perform the steps of the sea clutter suppression method based on a generative adversarial network as described in any one of claims 1 to 6.

9. A storage medium, It is characterized in that The storage medium stores a computer program, which, when executed by a processor, executes the steps of the sea clutter suppression method based on a generative adversarial network as claimed in any one of claims 1 to 6.