A method and device for suppressing radar sea clutter

Abstract

A method and device for suppressing radar sea clutter of the present invention belongs to the technical field of radar clutter suppression. The method comprises the following steps: acquiring radar signal data and creating six data sets, wherein the data sets include data set A, data set B, data set C, data set D, data set E and data set F; using a DCGAN network to perform data enhancement processing on the six data sets respectively to obtain six corresponding sample data sets; constructing a TD‑AMRKNet network for suppressing sea clutter based on an AMRK module and TripletAttention; training the TD‑AMRKNet network using the six sample data sets to obtain a sea clutter suppression model, and performing sea clutter suppression processing using the sea clutter suppression model. The present invention exhibits good clutter suppression effects on both staring and scanning radar data, especially in complex sea conditions, significantly improving the accuracy and robustness of target detection.

Description

Technical Field

[0001] The invention relates to a method and a device for suppressing radar sea clutter, and belongs to the technical field of radar clutter suppression. Background Art

[0002] Sea clutter refers to the backscattered echo caused by sea waves, sea breeze and other factors when the radar beam illuminates the sea surface. It is the main interference source for sea surface radar target detection, seriously affecting the radar's ability to detect low-altitude, slow-moving and small targets. Sea clutter suppression technology has become an important topic in the field of radar signal processing.

[0003] Traditional sea clutter suppression methods include adaptive filtering, time-frequency filtering, background modeling, and motion detection. For example, Wiener filtering and Kalman filtering improve clutter suppression capabilities by dynamically adjusting parameters; time-frequency filtering reduces noise by smoothing or filtering, but may lose target details. In addition, Doppler filtering and pulse compression techniques reduce static clutter interference by processing Doppler frequency shift or echo signals, and polar coordinate transformation methods use the uniform distribution characteristics of sea clutter to enhance target recognition.

[0004] In recent years, deep learning-based methods have gradually become a new trend in sea clutter suppression. Convolutional neural networks (CNNs) significantly improve the discrimination between targets and clutter by learning the characteristic differences between sea clutter and targets [X. Chen, N. Su, Y. Huang, J. Guan. False-alarm-controllable radar detection for marine targetbased on multi features fusion via CNNs[J]. IEEE Sensors Journal, 2021, 21(7): 9099-9111.]; generative adversarial networks (GANs) enhance the suppression effect of the model by generating realistic sea scenes and target images [X. Mou, X. Chen, J. Guan, Y. Dong, N. Liu. Sea clutter suppressionfor radar PPI images based on SCS-GAN[J]. IEEE Geoscience and Remote Sensing Letters, 2020, 18(11): 1886-1890.]; the Transformer architecture uses the self-attention mechanism to capture long-range dependencies in complex sea surface images, further improving the accuracy of clutter suppression and target detection [X. Luo, G.Fu, J. Yang, Y. Cao. Multi-modal image fusion via deep laplacian pyramid hybrid network[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023, 33(12): 7354-7369.]. Although deep learning has made significant progress in sea clutter suppression, further optimization and research are still needed when dealing with extreme sea conditions and complex target scenes. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a method and device for radar sea clutter suppression, which can improve the accuracy and robustness of sea clutter suppression.

[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 method for suppressing radar sea clutter, comprising the following steps:

[0008] Step 1, acquiring radar signal data and creating six data sets, the data sets including data set A, data set B, data set C, data set D, data set E and data set F;

[0009] Step 2: Use the DCGAN (Deep Convolutional Generative Adversarial Network) network to perform data enhancement processing on the six data sets respectively to obtain the corresponding six sample data sets;

[0010] Step 3: Build the TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention;

[0011] Step 4: Use six sample data sets to train the TD-AMRKNet network to obtain a sea clutter suppression model, and use the sea clutter suppression model to perform sea clutter suppression processing.

[0012] As a possible implementation of this embodiment, the dataset A, dataset B and dataset C are time-frequency diagram datasets obtained by a staring radar, and the dataset D, dataset E and dataset F are PPI display image datasets obtained by a scanning radar.

[0013] As a possible implementation of this embodiment, in step 2, using the DCGAN network to perform data enhancement processing on the six data sets respectively to obtain six corresponding sample data sets, including:

[0014] Define the generator G and discriminator D. The generator G generates images from random noise through a convolutional neural network, and the discriminator D is used to distinguish between real samples and generated samples.

[0015] Perform adversarial training on the generator G and the discriminator D, so that the images generated by the generator G become more and more realistic, until the discriminator D cannot distinguish between real samples and generated samples;

[0016] The six datasets are augmented using the adversarially trained generator G and discriminator D.

[0017] As a possible implementation of this embodiment, the mathematical expression of the generator G is:

[0018] <1> ,

[0019] in, is the Leaky ReLU activation function, is a two-dimensional convolution operation, is the upsampling operation, is the initial feature of the input, It is the hyperbolic tangent activation function, which maps the input value to the range of -1 to 1;

[0020] The mathematical expression of the discriminator D is:

[0021] <2> ,

[0022] in, is the feature extracted by the discriminator convolution layer. is the Sigmoid function, is the transpose of the weight vector, is the bias term;

[0023] The objective function for adversarial training of the generator G and the discriminator D is:

[0024] <3> ,

[0025] in, This indicates that the parameters of the generator G are minimized. Indicates the maximization operation on the parameters of the discriminator D. represents the loss function of the discriminator D and the generator G, Represents the real data distribution expectations, is the output of the discriminator D, is the output of the generator G.

[0026] As a possible implementation of this embodiment, the architecture of the generator G includes a fully connected layer and a multi-layer upsampling convolution module, and the output layer of the generator G uses a tanh activation function to map the pixel value range to [-1, 1]; the architecture of the discriminator D is a multi-layer convolutional network, which maps the input image to a probability value through multi-layer convolution and pooling operations.

[0027] As a possible implementation of this embodiment, step 3, constructing a TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention, includes:

[0028] Build the overall UNet framework and add the TripletAttention module to improve the UNet network;

[0029] Introduce lightweight multi-scale convolution modules and multiple small sub-networks, build the AMRK module and integrate the GSConvs module, SA module, KAN module and CRA module;

[0030] The improved AMRK module and UNet network are integrated into the diffusion model to form the TD-AMRKNet network.

[0031] As a possible implementation of this embodiment, in the encoder and decoder of UNet, the TripletAttention module performs weighted processing on the features of the skip connection through channel attention, spatial attention, and global attention mechanisms. The mathematical logic expression of the TripletAttention module is:

[0032] <4> ,

[0033] in, Represents the final tensor tensor, is a tensor of shape (C×H×W), yes A tensor rotated 90° counterclockwise along the H axis, yes The result after Z-Pool: yes A tensor rotated 90° counterclockwise along the W axis, yes The result after Z-Pool: Refers to the result after the sigmoid activation function; represents a standard 2D convolutional layer defined by kernel size k in the three branches of triplet attention: channel attention, spatial attention, and global attention.

[0034] As a possible implementation of this embodiment, the GSConvs module extracts spatial information of different scales through multi-scale convolution operations; the SA module generates a spatial attention map through a 7×7 convolution kernel to weight the spatial information of the feature map; the KAN module enhances the representation ability of the network through B-spline functions and nonlinear transformations; the CRA module weights the channel information of the feature map through a channel attention mechanism.

[0035] As a possible implementation of this embodiment, the improved AMRK module and the UNet network are integrated into the diffusion model to form a TD-AMRKNet network, including:

[0036] The improved UNet network is integrated into the diffusion model, which learns how to reverse the noise addition process through forward and reverse processes;

[0037] Embed the AMRK module in the downsampling module and upsampling module of UNet to enhance the ability to extract multi-scale features;

[0038] Insert the TripletAttention module into the middle layer of the diffusion model.

[0039] In a second aspect, an embodiment of the present invention provides a radar sea clutter suppression device, comprising:

[0040] A data acquisition module, used to acquire radar signal data and create six data sets, the data sets including data set A, data set B, data set C, data set D, data set E and data set F;

[0041] The data enhancement module is used to perform data enhancement processing on the six data sets using the DCGAN network to obtain six corresponding sample data sets;

[0042] Network construction module, used to build the TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention;

[0043] The sea clutter suppression module is used to train the TD-AMRKNet network using six sample data sets to obtain a sea clutter suppression model, and perform sea clutter suppression processing using the sea clutter suppression model.

[0044] The beneficial effects of the technical solution of the embodiment of the present invention are as follows:

[0045] The present invention proposes a TD-AMRKNet network architecture, which effectively reduces the model parameters by introducing lightweight multi-scale Multi-scale-GSConvs and a variety of small model networks, making it a small and efficient network; through the designed AMRK module, combined with the gating mechanism, it can capture the long-range dependencies of the image, and fuse multi-source knowledge through cross-resolution image information to enhance semantic understanding, while improving the comprehensive ability of details and local features; TD-AMRKNet not only performs well in the sea clutter suppression of the time-frequency diagram of staring radar data, but also has a significant clutter suppression effect on the PPI image obtained by scanning radar data, and can perform efficient sea clutter suppression on both types of data. The present invention shows good clutter suppression effects on both staring and scanning radar data, especially in complex sea conditions, significantly improving the accuracy and robustness of target detection.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] (1) This paper proposes a network architecture based on the diffusion model and Triplet Attention, TD-AMRKNet (Triplet Diffusion Attention Multi-scale Res-KAN Net). The network effectively reduces the model parameters by introducing a lightweight multi-scale convolution module (Multi-scale-GSConvs) and multiple small sub-networks. The number of parameters of TD-AMRKNet is only 3.46M, and the OA processing time is only 0.0305 seconds.

[0048] (2) The present invention introduces a lightweight multi-scale convolutional module (Multi-scale-GSConvs) and multiple small sub-networks, innovatively designs the AMRK module, and combines it with a gating mechanism, so that the network can effectively capture the long-range dependencies in the image; through cross-resolution image information fusion, the module not only enhances the semantic understanding ability, but also improves the comprehensive performance of details and local features, thereby improving the overall performance of the model;

[0049] (3) The present invention creates six datasets including time-frequency diagrams and PPI displays, and improves DCGAN on this basis, performs image data enhancement, and expands the size of the dataset. Through experimental verification on these datasets, TD-AMRKNet performs well in the sea clutter suppression task on both time-frequency diagrams and PPI displays, verifying that the model has good practicality and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of a method for suppressing radar sea clutter according to an exemplary embodiment;

[0051] Figure 2 is a schematic structural diagram of a radar sea clutter suppression device according to an exemplary embodiment;

[0052] Figure 3 An overall framework diagram of the present invention shown according to an exemplary embodiment;

[0053] Figure 4 is a flowchart of using a DCGAN network to perform data enhancement on a data set according to an exemplary embodiment;

[0054] Figure 5 is a structural diagram of an AMRK module according to an exemplary embodiment;

[0055] Figure 6 is a structural diagram of a TripletAttention module according to an exemplary embodiment;

[0056] Figure 7 is a structural diagram of a TD-AMRKNet network module according to an exemplary embodiment;

[0057] Figure 8 It is a schematic diagram of the comparison of radar clutter suppression in the time-frequency diagrams of the three data sets A, B, and C;

[0058] Fig. 9 This is a PSNR comparison chart of the suppression effect of each method on different data sets;

[0059] Fig.10 This is the SSIM comparison chart of the suppression effect of each method on different data sets;

[0060] Fig.11 This is a PS comparison chart of the suppression effect of each method on different data sets;

[0061] Fig.12 This is a diagram showing the effect of the present invention on suppressing actual IPIX sea clutter data;

[0062] Fig.13 This is a diagram showing the effect of the present invention on suppressing actual CSIR sea clutter data. DETAILED DESCRIPTION

[0063] 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.

[0064] like Figure 1 As shown, a radar sea clutter suppression method provided by an embodiment of the present invention includes the following steps:

[0065] Step 1, acquiring radar signal data and creating six data sets, the data sets including data set A, data set B, data set C, data set D, data set E and data set F;

[0066] Step 2: Use the DCGAN (Deep Convolutional Generative Adversarial Network) network to perform data enhancement processing on the six data sets respectively to obtain the corresponding six sample data sets;

[0067] Step 3: Build the TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention;

[0068] Step 4: Use six sample data sets to train the TD-AMRKNet network to obtain a sea clutter suppression model, and use the sea clutter suppression model to perform sea clutter suppression processing.

[0069] As a possible implementation of this embodiment, the dataset A, dataset B and dataset C are time-frequency diagram datasets obtained by a staring radar, and the dataset D, dataset E and dataset F are PPI display image datasets obtained by a scanning radar.

[0070] As a possible implementation of this embodiment, in step 2, using the DCGAN network to perform data enhancement processing on the six data sets respectively to obtain six corresponding sample data sets, including:

[0071] Define the generator G and discriminator D. The generator G generates images from random noise through a convolutional neural network, and the discriminator D is used to distinguish between real samples and generated samples.

[0072] Perform adversarial training on the generator G and the discriminator D, so that the images generated by the generator G become more and more realistic, until the discriminator D cannot distinguish between real samples and generated samples;

[0073] The six datasets are augmented using the adversarially trained generator G and discriminator D.

[0074] As a possible implementation of this embodiment, the mathematical expression of the generator G is:

[0075] <1> ,

[0076] in, is the Leaky ReLU activation function, is a two-dimensional convolution operation, is the upsampling operation, is the initial feature of the input, It is the hyperbolic tangent activation function, which maps the input value to the range of -1 to 1;

[0077] The mathematical expression of the discriminator D is:

[0078] <2> ,

[0079] in, is the feature extracted by the discriminator convolution layer. is the Sigmoid function, is the transpose of the weight vector, is the bias term;

[0080] The objective function for adversarial training of the generator G and the discriminator D is:

[0081] <3> ,

[0082] in, This indicates that the parameters of the generator G are minimized. Indicates the maximization operation on the parameters of the discriminator D. represents the loss function of the discriminator D and the generator G, Represents the real data distribution expectations, is the output of the discriminator D, is the output of the generator G.

[0083] As a possible implementation of this embodiment, the architecture of the generator G includes a fully connected layer and a multi-layer upsampling convolution module, and the output layer of the generator G uses a tanh activation function to map the pixel value range to [-1, 1]; the architecture of the discriminator D is a multi-layer convolutional network, which maps the input image to a probability value through multi-layer convolution and pooling operations.

[0084] As a possible implementation of this embodiment, step 3, constructing a TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention, includes:

[0085] Build the overall UNet framework and add the TripletAttention module to improve the UNet network;

[0086] Introduce lightweight multi-scale convolution modules and multiple small sub-networks, build the AMRK module and integrate the GSConvs module, SA module, KAN module and CRA module;

[0087] The improved AMRK module and UNet network are integrated into the diffusion model to form the TD-AMRKNet network.

[0088] As a possible implementation of this embodiment, in the encoder and decoder of UNet, the TripletAttention module performs weighted processing on the features of the skip connection through channel attention, spatial attention, and global attention mechanisms. The mathematical logic expression of the TripletAttention module is:

[0089] <4> ,

[0090] in, Represents the final tensor tensor, is a tensor of shape (C×H×W), yes A tensor rotated 90° counterclockwise along the H axis, yes The result after Z-Pool: yes A tensor rotated 90° counterclockwise along the W axis, yes The result after Z-Pool: Refers to the result after the sigmoid activation function; represents a standard 2D convolutional layer defined by kernel size k in the three branches of triplet attention: channel attention, spatial attention, and global attention.

[0091] As a possible implementation of this embodiment, the GSConvs module extracts spatial information of different scales through multi-scale convolution operations; the SA module generates a spatial attention map through a 7×7 convolution kernel to weight the spatial information of the feature map; the KAN module enhances the representation ability of the network through B-spline functions and nonlinear transformations; the CRA module weights the channel information of the feature map through a channel attention mechanism.

[0092] As a possible implementation of this embodiment, the improved AMRK module and the UNet network are integrated into the diffusion model to form a TD-AMRKNet network, including:

[0093] The improved UNet network is integrated into the diffusion model, which learns how to reverse the noise addition process through forward and reverse processes;

[0094] Embed the AMRK module in the downsampling module and upsampling module of UNet to enhance the ability to extract multi-scale features;

[0095] Insert the TripletAttention module into the middle layer of the diffusion model.

[0096] like Figure 2 As shown, an embodiment of the present invention provides a radar sea clutter suppression device, comprising:

[0097] A data acquisition module, used to acquire radar signal data and create six data sets, the data sets including data set A, data set B, data set C, data set D, data set E and data set F;

[0098] The data enhancement module is used to perform data enhancement processing on the six data sets using the DCGAN network to obtain six corresponding sample data sets;

[0099] Network construction module, used to build the TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention;

[0100] The sea clutter suppression module is used to train the TD-AMRKNet network using six sample data sets to obtain a sea clutter suppression model, and perform sea clutter suppression processing using the sea clutter suppression model.

[0101] Figure 3 It is the overall framework diagram of the present invention; the implementation of the present invention mainly includes four parts: constructing a data set, data enhancement, training a TD-AMRKNet network model, and testing a TD-AMRKNet network model.

[0102] Step 1: Create six data sets, including different data sets, time-frequency diagrams, and PPI displays.

[0103] 1.1 Create six data sets: A, B, C, D, E, and F.

[0104] Datasets A, B, and C are obtained by staring radar, and datasets D, E, and F are obtained by scanning radar. Dataset A is the IPIX (McMaster IPIX Radar Data Set) dataset, which is a classic radar signal processing research dataset provided by McMaster University in Canada and an important public resource for marine radar target echo data. The training dataset consists of actual sea clutter and simulated targets, which is also the training set of datasets B and C. The test set is composed of actual sea clutter. Among them, those with bright colors and certain shapes are targets, and those without certain shapes and that affect the targets are clutter. Dataset B is a CSIR dataset, which is researched and developed by the Indian CSIR organization. The generation of the dataset is based on synthetic aperture radar (SAR) or other radar systems, including different terrains and complex weather conditions. It captures the target's echo signal, reflection characteristics and other information to provide high-resolution radar images or time-frequency domain data. Its dataset composition is the same as IPIX. Dataset C is the data selected from the Journal of Radars, 2022, No. 1 [G. Jian, L. Ningbo, W. Guoqing, D. Hao, D. Yunlong. Sea-detecting radar experiment and target feature data acquisition for dual polarization multistate scattering dataset of marine targets [J]. Journal of Radars, 2023, 12(2): 456–469. doi: 10.12000 / JR23029.], where the target is a buoy. This data is also a staring radar, so it is displayed using a time-frequency diagram. Its dataset structure is the same as IPIX. Therefore, the structures of the three datasets A, B, and C are the same. Datasets D and E are the actual data obtained by cropping the datasets collected on the Yantai coastline. Here we call this dataset R_S (Rain_Sun) dataset. We regard the dataset collected on rainy days as clutter, and the dataset collected on sunny days as the dataset with only targets. However, the number of ships moored on the shore is different in the case of multiple lands. So here we transfer the clutter part in the rainy day dataset to the sunny day dataset to obtain a paired training set. The test set is the dataset collected on rainy days.Dataset D (More_lands) contains most of the land and ships at the test location. Because the radar used to collect data is on the coast, it contains most of the land. The large yellow area in the picture is the land, the black area is the ocean, and the dots in the black ocean are sea clutter. Dataset E (Less_lands) removes most of the land and contains only 2-3 ships and island targets. The larger yellow area with the same shape in each picture is the target, and the dots in the black ocean are clutter. The composition of the dataset is the same as More_lands. Dataset F is a dataset from the Journal of Radars, 2021, 10(1), doi: 10.12000 / JR21011. This dataset is a scanning radar dataset with the land part removed. It is a fan-shaped dataset. The training set intercepts the datasets under different clutters obtained at different speeds in the same place. The datasets with almost no clutter are regarded as containing only targets, and the other cases are regarded as containing clutter and targets. The brighter ones with unchanged shapes in each picture are targets, the large blue area is the ocean, and the rest that affects these two parts are clutter. The test set is randomly intercepted, and the dataset composition is the same as More_lands. In datasets D, E, and F, clean targets are first selected as targets in the training set, and then real clutter in the dataset is randomly added to the target image to form a clutter dataset that contains both clutter and targets and is paired with the target. Therefore, the construction methods of these three datasets are the same.

[0105] Step 2: Use the DCGAN network to enhance the data set to expand the samples and increase the diversity of samples. The structure is as follows: Figure 4 shown.

[0106] 2.1 For the DCGAN network, it is an improved version of the Generative Adversarial Network (GAN), which learns to generate realistic images through the generator and discriminator of the convolutional neural network. The goal of the generator G is to generate Generate images in , making it impossible for the discriminator D to distinguish and real samples x; the discriminator aims to distinguish real samples from generated samples as much as possible. The objective function of this adversarial training can be expressed as the following formula:

[0107] <5> ,

[0108] In the code, the generator and discriminator are defined in PyTorch respectively. The function of the generator is to project the low-dimensional latent vector z into the high-dimensional image space. Its architecture includes a fully connected layer and a multi-layer upsampling convolution module to gradually enlarge the resolution of the feature map. Specifically, the random noise z is transformed into an initial feature map F with a dimension of (128, H / 4, W / 4) through a fully connected layer, and then gradually restored to the target resolution H×W through two upsampling and convolution layer operations. The output layer of the generator uses the tanh activation function to map the pixel value range to [−1, 1]. Therefore, the mathematical expression of the generator can be expressed as:

[0109] <6> ,

[0110] in, is the Leaky ReLU activation function. The architecture of the discriminator is a multi-layer convolutional network that aims to map the input image to a probability value, indicating whether the image is a real sample. It gradually reduces the resolution of the feature map through multi-layer convolution and pooling operations, and finally flattens the features and outputs a scalar through the fully connected layer. In order to improve stability, the discriminator uses batch normalization (BatchNorm) and Dropout mechanisms to alleviate the overfitting problem, and adopts Leaky ReLU as the activation function. The mathematical logic of the discriminator can be expressed as:

[0111] <7> ,

[0112] in, is the feature extracted by the discriminator convolution layer. is the Sigmoid function. During the training process, the discriminator aims to maximize the output confidence of the real sample and minimize the confidence of the generated sample. In the implementation, in each training cycle, the generator is first updated to generate fake samples through random noise and minimize Then update the discriminator to maximize by optimizing the discriminative ability of real samples and generated samples at the same time. The optimizer uses the Adam method, combined with the momentum parameter and Further accelerate convergence. For the code, the network structure is changed, and new convolution and sampling layers are added to generate better PPI images. The original data set of 440 images has clutter and target data sets, and 336 data sets are generated to match the 777 data sets containing only target data. During the training process, the code uses PyTorch's DataLoader to load image data and applies a series of standardization and resizing preprocessing steps. In addition, by using torch.nn.DataParallel, multi-GPU parallel computing is supported to accelerate training. The images generated by the generator are saved every several iterations, and the model weights are stored regularly. The present invention implements a complete DCGAN framework, from the random noise sampling of the generator to the true and false classification confrontation of the discriminator, and finally enables the generator to learn to generate realistic images through the game between the two. Mathematically, this reflects the classic minimum-maximum optimization problem in GAN, which gradually improves the quality of generated samples through mutual confrontation.

[0113] Step 3: Construct the TD-AMRKNet diffusion model network structure based on the AMRK module and TripletAttention to suppress sea clutter. The overall structure is as follows: Figure 5 shown.

[0114] 3.1 proposes the GSConvs module. GSConvs is a multi-scale sub-module that contains a variety of convolution and activation operations. The cv1 convolution layer uses a 1×1 convolution kernel to compress the input feature map of 32 channels into 16 channels.

[0115] The GSConvs submodule contains a variety of convolution and activation operations, among which the cv1 convolution layer uses a 1×1 convolution kernel to compress the input feature map of 32 channels into 16 channels. The purpose of this convolution operation is to reduce the number of channels while maintaining the spatial information of the feature map. Next, the convolution result is standardized by batch normalization (BatchNorm2d), and then the Mish activation function is applied. The Mish activation function helps the network learn more complex features through smooth nonlinear transformations. The cv2 part is a multi-scale convolution module, which contains three convolution layers with different dilation rates, namely standard 3×3 convolution, dilated convolution, and stronger dilated convolution. These convolution operations have different receptive fields, which help the model extract richer spatial information at different scales. The ReLU activation function is applied after each convolution layer to enhance the nonlinear feature representation. Suffer is a simple 1×1 convolution layer, which maps the input 16-channel feature map back to 32 channels to prepare suitable input features for subsequent operations. Multi-scale convolution operations are added to the original GSConv, making the model more robust in processing sea clutter images under various models.

[0116] 3.2 The CRA and KAN modules are skip-connected, which greatly reduces the network parameters of the diffusion model.

[0117] The CRA (Channel Reduction Attention) module is a more complex attention mechanism, which is an attention mechanism used to enhance the channel expression ability of convolutional neural networks (CNNs). Its core idea is to selectively focus on important feature channels by reducing the number of channels and suppress unimportant channels, thereby improving the network's representation ability and computational efficiency. CRA combines the self-attention mechanism and channel reduction, and weights and filters the channels of the feature map by learning a channel importance weight to highlight key information. Mathematically, the operation of CRA can be divided into several steps. First, for the input feature map , where C is the number of channels, H and W are the height and width of the feature map, CRA calculates the weight of each channel through a channel attention module. In order to obtain the correlation of channels, CRA captures the global information of the channel by performing global pooling (e.g., average pooling) on ​​the input feature map. Specifically, first, the input feature map undergoes a global average pooling (GAP) operation to compress the information of each channel into a single scalar. This process transforms X into represents the overall response of each channel. Next, this global information is further processed by a set of fully connected layers (usually two layers of MLP, using activation functions such as ReLU or GELU) to obtain the attention weights of each channel. Assume is the pooled feature vector, which CRA feeds into two linear layers and Transformed and activated by the function Generating channel attention weights , this process can be expressed as:

[0118] <8> ,

[0119] in The activation function uses GELU. and is the weight matrix, and is the bias term. The channel attention weight obtained It will be used to weight the channels of the original feature map. Specifically, the output feature map It is generated by multiplying each channel of the input feature map with the corresponding channel attention weight:

[0120] <9> ,

[0121] in Represents an element-by-element multiplication operation, which weights the features of each channel and suppresses unimportant channels. In this process, the channel attention mechanism of CRA dynamically adjusts the importance of each channel in an adaptive manner. By reducing unimportant channels, CRA can improve the network's expressiveness while reducing the amount of computation. Its goal is to enable the network to automatically focus on the feature channels that are most useful for the task and reduce unnecessary redundant information. In general, the mathematical logic of CRA is mainly based on global pooling, fully connected transformation, and element-by-element weighted operations of channel information. Through these steps, the attention weights of the channels are calculated, thereby effectively improving the performance of the network.

[0122] By weighting the convolution kernels and channels, KAN provides a flexible and efficient way to enhance the network's representation capabilities, allowing it to better handle complex sea clutter images and tasks. The core idea behind it is Kolmogorov–Arnold, that is, for any multivariate continuous function, it can be represented as a combination of a finite number of single-variable functions and addition. KART (Knowledge Attention Network) theorem, given a continuous function , which states that there exists a set of one-dimensional nonlinear basis functions So that: ,in is a nonlinear combination of each basis function.

[0123] KAN uses B-Splines, which is a linear combination of multiple local basis functions to form a spline. Therefore, KAN is decomposed into multiple KANLinear layers, each of which has an independent nonlinear transformation and uses a combination method to generate output. Its mathematical logic can be expressed by the following formula. Assume that the input is Then the output of the KAN module can be expressed as:

[0124] <10> ,

[0125] in, is the weight matrix of each nonlinear transformation, is the activation function, is bias.

[0126] 3.3 Finally, the TripletAttention module is used to greatly improve the sea clutter suppression performance of the network.

[0127] Define the TripletAttention module in the code, such as Figure 6As shown. The TripletAttention module usually consists of three main parts: channel attention, spatial attention, and global attention. The channel attention module focuses on the weights of different channels, while the spatial attention module focuses on the spatial position in the image. These attention mechanisms are used to filter important features. Then, the module needs to be used after the convolutional layers of the encoder and decoder of UNet to weight the extracted features. The mathematical logic is as follows:

[0128] <11> ,

[0129] in Represents the final tensor tensor, is a tensor of shape (C×H×W), yes A tensor rotated 90° counterclockwise along the H axis, yes The result after Z-Pool: yes A tensor rotated 90° counterclockwise along the W axis, yes The result after Z-Pool: Refers to the result after the sigmoid activation function; Represents a standard 2D convolutional layer defined by kernel size k in the three branches of triplet attention.

[0130] In UNet, the TripletAttention module needs to work with the skip connection. In specific implementation, when the decoder receives the skip connection from the encoder, the TripletAttention module will first pay attention to these skip connections before the decoder performs upsampling. This operation ensures that the decoder receives weighted high-quality features, which helps to improve the accuracy of the segmentation results.

[0131] Finally, when compiling and training the model, the TripletAttention module will help UNet focus on more critical features during training and improve segmentation results. During model training, the introduction of TripletAttention will not change the optimization process, but it will improve the training effect by enhancing the representation of important features and reducing the impact of irrelevant features.

[0132] Figure 7 It is an overall diffusion model based on the TD-AMRKNet network. The next step is to train this model.

[0133] Step 4: Train TD-AMRKNet on the six datasets proposed in step 1 to obtain a sea clutter suppression model, and verify the sea clutter suppression performance of the sea clutter suppression model on actual datasets.

[0134] The time-frequency diagrams of datasets A, B, and C, and the PPI display diagrams of datasets D, E, and F are shown in Figure 1. The time-frequency diagrams of datasets A, B, and C are shown in Figure 2. Different network models are tested and compared with TD-AMRKNet. Figure 8 The specific indicator results are shown in Figure 9-11 As shown, Fig. 9 This is a PSNR comparison chart of the suppression effect of each method on different data sets. Fig.10 This is the SSIM comparison chart of the suppression effect of each method on different data sets. Fig.11 This is a PS comparison chart of the inhibition effect of each method on different data sets.

[0135] The effect of the present invention is further illustrated by the following simulation comparison test:

[0136] 1. Simulation conditions.

[0137] 1.1 Dataset composition

[0138] The specific composition of the dataset is shown in Table 1.

[0139] Table 1 Dataset information table

[0140]

[0141] 1.2 Experimental setup

[0142] All experiments were performed on a high-performance computer with the following configuration: Intel(R) Xeon(R) CPU E5-2686 v4 CPU @ 3.70 GHz processor, 64 GB memory, 18 cores, and graphics card driver version 530.30.02 dual 3090 graphics cards, 24 GB video memory. Network uplink bandwidth: 70 Mbps / s, downlink bandwidth: 1000 Mbps / s. The operating system is Linux, the framework is Pytorch, the framework version is 2.0.0, cuda is 11.8, and Python is 3.8.

[0143] The parameter comparison of the TD-AMRKNet network model proposed in the present invention and other models is shown in Table 2, where the parameter scale unit is million (M). (1) Comparing TD-AMRKNet with the existing WeatherDiffusion-main and RADARDIFF diffusion model networks, the AMRK module proposed in the present invention replaces the ResNet module in the traditional network, and significantly reduces the model complexity by adopting a network structure with small parameters. (2) Comparing TD-AMRKNet with lightweight diffusion models including LWTDM and LightGrad-master, the experimental results show that TD-AMRKNet still has an advantage in parameter scale, but the design idea of ​​the above-mentioned lightweight diffusion model still has important reference value. (3) Comparing with other network structures. For example, the currently popular U-KAN-main framework achieves lightweight by embedding the KAN module in UNet. In contrast, the TD-AMRKNet of the present application further reduces the model complexity with a parameter scale of 3.46M. The parameter scale of the multi-head self-attention mechanism (MHA) in MHA-DNet is larger than the TripletAttention module used in the AMRK module of the present invention. TripletAttention achieves a smaller parameter scale while maintaining the high efficiency of multi-head self-attention.

[0144] Table 2 Network parameter comparison table

[0145]

[0146] 2 Experimental content.

[0147] 2.1 Test the network’s sea clutter suppression performance under actual data sets.

[0148] A. Select the 7th column of the VH_19931118_162155_stareC0000.mat data in the IPIX dataset. The resulting time-frequency diagram contains both the target and the sea clutter x. After clutter suppression, x~ is obtained. It can be seen that the actual sea clutter suppression effect is obvious, such as Fig.12 shown.

[0149] B. Select the 9th column of the TFC17_006.03.mat data in the CSIR dataset. The resulting time-frequency diagram contains both the target and the sea clutter x. After clutter suppression, x~ is obtained. It can be seen that the actual sea clutter suppression effect is obvious, such as Fig.13 shown.

[0150] C. More_lands dataset, the image x containing actual sea clutter has sea clutter in the whole image. The first PPI dataset is this type. There are more land targets, and it has not had a big impact on the land targets after completion. The second row has a few more islands than the first row, and it is also fully preserved. The actual dataset was collected in rainy weather on the coast of Yantai. It can be seen that the actual sea clutter suppression effect is obvious.

[0151] D. Less_lands dataset, the actual dataset, is the dataset after reducing the influence of land targets. x is the PPI display image containing sea clutter and rain clutter, and x~ is the effect after suppression.

[0152] 2.2 Evaluation indicators.

[0153] The above train training set is divided into a training set and a test set in a ratio of 8:2, and then training is performed. The number of training times is set to 1200 times, and the clutter suppression performance is quantitatively compared.

[0154] The training results are compared in terms of peak signal-to-noise ratio (PSNR) (i.e., clutter suppression ratio) and structural similarity index (SSIM). PSNR (peak signal-to-noise ratio): The larger the value, the better, indicating a higher image quality. SSIM (structural similarity index): The larger the value, the better, indicating a higher similarity in structure and visual quality of the images. The PS index is defined as:

[0155] <12>

[0156] <13>

[0157] By maximizing the objective function And the objective function It tends to 0 to comprehensively evaluate PSNR and SSIM.

[0158] The larger the values ​​of these three evaluation indicators, the better the effect. According to the training results as shown in Table 3, it can be seen that the method proposed in the present invention is superior to the other seven algorithms in these three aspects. The algorithms compared are ADN, SCS-GAN, RADARDIFF, MHA-DNet, CycleGAN, pix2pix, and ResNet.

[0159] Table 3 Index results of the suppression effect of each method on different data sets

[0160]

[0161] Figure 9-11 It can be clearly concluded from the perspective of visualization that the clutter suppression method proposed in the present invention can achieve the most ideal effect in the above indicators.

[0162] 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 method for suppressing radar sea clutter, characterized in that: The steps include: Step 1, obtain radar signal data and create six data sets, the data sets include data set A, data set B, data set C, data set D, data set E and data set F, the data set A, data set B and data set C are time-frequency diagram data sets obtained by staring radar, the data set D, data set E and data set F are PPI display image data sets obtained by scanning radar; Step 2: Use the DCGAN network to perform data enhancement processing on the six data sets respectively to obtain six corresponding sample data sets; Step 3: Build the TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention; Step 4, using six sample data sets to train the TD-AMRKNet network, obtain a sea clutter suppression model, and use the sea clutter suppression model to perform sea clutter suppression processing; The step 3, constructing a TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention, includes: Build the overall UNet framework and add the TripletAttention module to improve the UNet network; Introduce lightweight multi-scale convolution modules and multiple small sub-networks, build the AMRK module and integrate the GSConvs module, SA module, KAN module and CRA module; The improved AMRK module and UNet network are integrated into the diffusion model to form the TD-AMRKNet network; The GSConvs module extracts spatial information of different scales through multi-scale convolution operations; the SA module generates a spatial attention map through a 7×7 convolution kernel to weight the spatial information of the feature map; the KAN module enhances the representation ability of the network through B-spline functions and nonlinear transformations; the CRA module weights the channel information of the feature map through a channel attention mechanism; The improved AMRK module and UNet network are integrated into the diffusion model to form the TD-AMRKNet network, including: The improved UNet network is integrated into the diffusion model, which learns how to reverse the noise addition process through forward and reverse processes; Embed the AMRK module in the downsampling module and upsampling module of UNet to enhance the ability to extract multi-scale features; Insert the TripletAttention module into the middle layer of the diffusion model.

2. The method for suppressing radar sea clutter according to claim 1, characterized in that: In step 2, the DCGAN network is used to perform data enhancement processing on the six data sets respectively to obtain six corresponding sample data sets, including: Define the generator G and discriminator D. The generator G generates images from random noise through a convolutional neural network, and the discriminator D is used to distinguish between real samples and generated samples. Perform adversarial training on the generator G and the discriminator D, so that the images generated by the generator G become more and more realistic, until the discriminator D cannot distinguish between real samples and generated samples; The six datasets are augmented using the adversarially trained generator G and discriminator D.

3. The method for suppressing radar sea clutter according to claim 2, characterized in that: The mathematical expression of the generator G is: <1>， in, is the Leaky ReLU activation function, is a two-dimensional convolution operation, is the upsampling operation, is the initial feature of the input, It is the hyperbolic tangent activation function, which maps the input value to the range of -1 to 1; The mathematical expression of the discriminator D is: <2>， in, is the feature extracted by the discriminator convolution layer, is the Sigmoid function, is the transpose of the weight vector, is the bias term; The objective function for adversarial training of the generator G and the discriminator D is: <3>， in, This indicates that the parameters of the generator G are minimized. Indicates the maximization operation on the parameters of the discriminator D. represents the loss function of the discriminator D and the generator G, Represents the real data distribution expectations, is the output of the discriminator D, is the output of the generator G.

4. The method for suppressing radar sea clutter according to claim 2, characterized in that: The architecture of the generator G includes a fully connected layer and a multi-layer upsampling convolution module. The output layer of the generator G uses a tanh activation function to map the pixel value range to [-1, 1]. The architecture of the discriminator D is a multi-layer convolutional network that maps the input image to a probability value through multi-layer convolution and pooling operations.

5. The method for suppressing radar sea clutter according to claim 1, characterized in that: In the encoder and decoder of UNet, the TripletAttention module performs weighted processing on the features of the skip connection through channel attention, spatial attention, and global attention mechanisms. The mathematical logic expression of the TripletAttention module is: <4>， in, Represents the final tensor tensor, is a tensor of shape C×H×W, where C is the number of channels, H and W are the height and width of the feature map, yes A tensor rotated 90° counterclockwise along the H axis, yes The result after Z-Pool: yes A tensor rotated 90° counterclockwise along the W axis, yes The result after Z-Pool: Refers to the result after the sigmoid activation function; represents a standard 2D convolutional layer defined by kernel size k in the three branches of triplet attention: channel attention, spatial attention, and global attention.

6. A radar sea clutter suppression device, characterized in that: include: A data acquisition module is used to acquire radar signal data and create six data sets, wherein the data sets include data set A, data set B, data set C, data set D, data set E and data set F. The data sets A, B and C are time-frequency diagram data sets obtained by the staring radar, and the data sets D, E and F are PPI display image data sets obtained by the scanning radar. The data enhancement module is used to perform data enhancement processing on the six data sets using the DCGAN network to obtain the corresponding six sample data sets; Network construction module, used to build the TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention; The sea clutter suppression module is used to train the TD-AMRKNet network using six sample data sets to obtain a sea clutter suppression model, and perform sea clutter suppression processing using the sea clutter suppression model; The specific process of the network construction module to construct the TD-AMRKNet network for sea clutter suppression based on the AMRK module and TripletAttention includes: Build the overall UNet framework and add the TripletAttention module to improve the UNet network; Introduce lightweight multi-scale convolution modules and multiple small sub-networks, build the AMRK module and integrate the GSConvs module, SA module, KAN module and CRA module; The improved AMRK module and UNet network are integrated into the diffusion model to form the TD-AMRKNet network; The GSConvs module extracts spatial information of different scales through multi-scale convolution operations; the SA module generates a spatial attention map through a 7×7 convolution kernel to weight the spatial information of the feature map; the KAN module enhances the representation ability of the network through B-spline functions and nonlinear transformations; the CRA module weights the channel information of the feature map through a channel attention mechanism; The improved AMRK module and UNet network are integrated into the diffusion model to form the TD-AMRKNet network, including: The improved UNet network is integrated into the diffusion model, which learns how to reverse the noise addition process through forward and reverse processes; Embed the AMRK module in the downsampling module and upsampling module of UNet to enhance the ability to extract multi-scale features; Insert the TripletAttention module into the middle layer of the diffusion model.

Images