SAR multi-type interference suppression and segmentation method

Abstract

The invention discloses an SAR multi-type interference suppression and segmentation method. The method comprises the following steps: obtaining an interference-containing SAR echo signal to be processed, and carrying out short-time Fourier transform and normalization to obtain a time-frequency graph to be processed; a pre-trained generative adversarial network model based on the convolutional block attention mechanism is obtained, a generator is a deep residual network integrated with the convolutional block attention mechanism, a discriminator is a PatchGAN network, and interference used in the training process is multi-type interference; inputting the to-be-processed time-frequency graph into a pre-trained generator to obtain an interference suppression result and an interference segmentation result; and carrying out inverse normalization and short-time inverse Fourier transform on the interference suppression result to obtain an SAR echo signal subjected to interference suppression, and carrying out imaging processing to obtain an SAR imaging result subjected to interference suppression. According to the method, the generative adversarial network is combined, and the problems of signal loss, single interference pattern and interference diffusion energy residue of an existing SAR interference suppression method are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular relates to a SAR multi-type interference suppression and segmentation method. Background Art

[0002] Synthetic Aperture Radar (SAR), as an advanced microwave imaging radar, can acquire high-resolution images through synthetic aperture technology. It utilizes electromagnetic waves in the microwave frequency band for imaging, unaffected by weather and lighting conditions, enabling all-day and all-weather operation. It has been widely used in military reconnaissance, disaster monitoring, and biological activities. However, due to its broadband system characteristics, SAR is vulnerable to various electromagnetic interference threats. These interferences can cause problems such as defocusing and distortion of target signals during the SAR imaging process, thereby affecting its target recognition and detection performance. SAR interference suppression, as a key step before imaging, is of great significance for improving the imaging quality of target signals. However, with the increase in interference patterns, energy diffusion, and high coupling characteristics with signals, traditional interference suppression methods are unable to meet the requirements of increasingly complex electromagnetic environments. Therefore, new technologies are needed to improve interference suppression capabilities.

[0003] The SAR interference suppression process can be viewed as the lossless separation of interference and signal in the echo signal's time-frequency diagram, or as the generation of interference-free echo time-frequency diagrams. In recent years, numerous methods have been proposed for the rapid and effective suppression of SAR interference. These methods are generally categorized as model-driven and data-driven. Traditional model-driven SAR interference suppression methods utilize mathematical models of SAR signals to model and analyze the characteristics of interference signals and design corresponding filters or algorithms to suppress interference. Alternatively, they utilize target scattering models to analyze the characteristics of target echo signals and distinguish between targets and interference. However, due to their strong reliance on prior parameters and key model assumptions, model-driven methods struggle to rapidly select interference suppression parameters and accurately suppress interference in complex and variable electromagnetic environments. Data-driven methods, on the other hand, do not require prior parameter guidance or model assumptions for target signals and interference. Data-driven SAR interference suppression methods utilize deep learning algorithms to learn and train SAR data containing interference. Deep neural networks automatically extract the spatial features of the data and establish a mapping relationship between interference-free and interference-free data, thereby enabling interference identification and suppression. However, data-driven methods have problems such as a single interference pattern, target signal loss, and poor suppression of interference diffusion energy. Faced with diverse high-energy dynamic interference in actual complex electromagnetic environments, this type of method is difficult to achieve effective migration and generalization. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a SAR multi-type interference suppression and segmentation method. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0005] The present invention provides a SAR multi-type interference suppression and segmentation method, comprising:

[0006] Obtain the interference-containing SAR echo signal to be processed, perform short-time Fourier transform and normalization, and obtain the corresponding time-frequency diagram as the time-frequency diagram to be processed;

[0007] Obtain a pre-trained generative adversarial network model based on a convolutional block attention mechanism; wherein the generative adversarial network model based on a convolutional block attention mechanism includes: a generator and a discriminator; the generator is a deep residual network integrated with a convolutional block attention mechanism, and the discriminator is a PatchGAN network;

[0008] Inputting the to-be-processed time-frequency graph into a generator of a pre-trained generative adversarial network model based on a convolutional block attention mechanism to obtain an interference suppression result and a corresponding interference segmentation result; wherein the interference used in the training process of the generative adversarial network model based on the convolutional block attention mechanism is multi-type interference;

[0009] Performing inverse normalization and inverse short-time Fourier transform on the interference suppression result to obtain an interference-suppressed SAR echo signal, wherein the interference segmentation result is used to assist in determining the effect of interference suppression;

[0010] Imaging processing is performed on the interference-suppressed SAR echo signal to obtain a SAR imaging result after interference suppression.

[0011] In one embodiment of the present invention, the training process of the generative adversarial network model based on the convolutional block attention mechanism includes:

[0012] Superimposing any one of multiple types of interference on each acquired interference-free SAR echo signal to generate a corresponding interference-containing SAR echo signal, and using short-time Fourier transform and normalization to convert a group of interference-free SAR echo signals and the corresponding interference-containing SAR echo signals into time-frequency graphs to form a time-frequency graph reference pair; in the time-frequency graph reference pair, the time-frequency graph of the interference-free SAR echo signal serves as the true label of the time-frequency graph of the interference-containing SAR echo signal; using a set threshold, pixel-level segmentation is performed on the time-frequency graph of the interference-containing SAR echo signal to obtain a corresponding segmentation label; the segmentation label is a binary image containing 0 and 1, wherein 0 indicates that the pixel at the corresponding position in the time-frequency graph of the interference-containing SAR echo signal is free of interference, and 1 indicates that the pixel at the corresponding position in the time-frequency graph of the interference-containing SAR echo signal is interfered with; using the time-frequency graph reference pair and the corresponding segmentation label as samples, constructing a data set from a number of samples, selecting a large number of samples from the data set to form a training set, and the remaining small number of samples to form a validation set;

[0013] A generative adversarial network model based on a convolutional block attention mechanism is constructed, comprising: a generator and a discriminator; the generator is used to process an input time-frequency graph containing an interfering SAR echo signal, and output an interference suppression result and an interference segmentation result, thereby completing interference suppression and segmentation; the discriminator is used to process the input interference suppression result and the true label, and output a discriminant score matrix of the interference suppression result and a discriminant score matrix of the true label, thereby completing the judgment of the interference suppression result; wherein, the interference suppression result refers to the result obtained after the generator performs interference suppression on the time-frequency graph containing the interfering SAR echo signal, and the interference segmentation result refers to the probability matrix diagram obtained by the generator performing pixel-level segmentation on the time-frequency graph containing the interfering SAR echo signal, wherein each element in the probability matrix diagram represents the probability of whether the pixel at the corresponding position in the time-frequency graph containing the interfering SAR echo signal has interference, and a larger probability value indicates a greater possibility that the pixel at the corresponding position has interference;

[0014] Design a preset loss function based on the weighted method of multiple loss functions;

[0015] Using the samples in the training set, as well as the corresponding interference suppression results, interference segmentation results, the discriminant score matrix of the interference suppression results, the discriminant score matrix of the true label and the preset loss function, the generator and the discriminator are iteratively trained and verified to obtain a trained generative adversarial network model based on the convolutional block attention mechanism.

[0016] In one embodiment of the present invention, the multiple types of interference include: sinusoidal modulation swept frequency interference, comb spectrum interference, dense false target interference and intermittent sampling direct forwarding interference.

[0017] In one embodiment of the present invention, the generator includes:

[0018] A convolutional layer, an activation layer, a first CRB module, a second CRB module, and a first RB module are connected in sequence, and the output of the first RB module is divided into two branches, wherein the first branch includes the second RB module, a CBAM module, a third CRB module, a third RB module, a fourth CRB module and a convolutional layer connected in sequence, and the second branch includes the fourth RB module, the fifth RB module, a convolutional layer, a standard normalization layer and a convolutional layer connected in sequence, and the output of the first RB module is added point by point to the input of the first CRB module to realize the residual connection; wherein, the first branch outputs an interference suppressed image, and the second branch outputs an interference segmented image; any CRB module is a residual block integrated by the convolutional block attention mechanism; any RB module is a basic residual block; the CBAM module is a convolutional block attention mechanism module.

[0019] In one embodiment of the present invention, any CRB module includes two RB modules and one CBAM module connected in sequence.

[0020] In one embodiment of the present invention, any RB module includes one convolution layer, one standard normalization layer, one convolution layer, one standard normalization layer and one activation layer connected in sequence, and the output of the activation layer in any RB module is added point by point to the input of the RB module to realize residual connection.

[0021] In one embodiment of the present invention, the discriminator includes one convolutional layer, one activation layer, seven CBL modules and one convolutional layer connected in sequence, wherein any CBL module includes one convolutional layer, one standard normalization layer and one activation layer connected in sequence.

[0022] In one embodiment of the present invention, the preset loss function is composed of a combination of the generator's loss function and the discriminator's loss function, wherein the generator's loss function is a weighted combination of an improved mean square error loss function, a cross entropy loss function, and an adversarial loss function between the generator and the discriminator; the expression of the preset loss function is:

[0023] Loss=minmax(L D +L G )

[0024] L G =λL Mo_MSE +λ1L CE +λ2L A

[0025] Among them, Loss is the preset loss function; L G is the loss function of the generator; LD is the loss function of the discriminator; L Mo_MSE is the improved mean square error loss function; L CE is the cross entropy loss function; L A is the adversarial loss function between the generator and the discriminator; λ, λ1, and λ2 are three hyperparameters.

[0026] In one embodiment of the present invention, the generator and the discriminator are iteratively trained and verified using the samples in the training set, the corresponding interference suppression results, the interference segmentation results, the discriminant score matrix of the interference suppression results, the discriminant score matrix of the true label, and the preset loss function to obtain a trained generative adversarial network model based on the convolutional block attention mechanism, including:

[0027] Step a1, selecting a preset number of samples from the training set as a batch;

[0028] Step a2: inputting the time-frequency diagram of the interfering SAR echo signal in the batch of samples into the generator to obtain corresponding interference suppression results and interference segmentation results;

[0029] Step a3: inputting the interference suppression result and the true label into the discriminator to obtain a discriminant score matrix of the interference suppression result and a discriminant score matrix of the true label;

[0030] Step a4, subtracting the discriminant score matrix of the interference suppression result from the discriminant score matrix of the true label and taking the global average to obtain the final discriminant score;

[0031] Step a5: Based on the final discriminant score and the loss function L of the discriminator D The penalty term in the discriminator is used to calculate the loss function L D The numerical value of

[0032] Step a6, according to the loss function L of the discriminator D The value of is used to update the discriminator parameters using the back propagation algorithm;

[0033] Step a7, calculating the loss function L of the generator based on the discrimination score matrix of the interference suppression result, the interference suppression result, the true label, the interference segmentation result and the segmentation label G The numerical value of

[0034] Step a8, according to the loss function L of the generator G The value of the back propagation algorithm is used to update the generator parameters; wherein the improved mean square error loss function L Mo_MSE Utilize the back-propagation algorithm to include a segmentation guidance mechanism;

[0035] Step a9: Repeat steps a1-a8 for a preset number of iterations and then perform a verification.

[0036] Step a10: Repeat step a9 until the preset epoch is reached to obtain a trained generative adversarial network based on the convolutional block attention mechanism, where a single epoch is the process of the model performing a complete traversal of the training set.

[0037] In one embodiment of the present invention, the implementation process of the segmentation guidance mechanism includes:

[0038] Resetting the values ​​in the segmentation tag, including resetting the positions with values ​​of 0 in the segmentation tag to 1, and resetting the positions with values ​​of 1 to integers greater than 1;

[0039] The expression of the segmentation guidance mechanism is:

[0040] dL Mo_MSE =M×(SG(J))

[0041] Where d represents differentiation, M represents the reset segmentation label, S represents the time-frequency diagram of the SAR echo signal without interference, J represents the time-frequency diagram of the SAR echo signal with interference, and G(J) represents the interference suppression result output by the generator.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] In a SAR multi-type interference suppression and segmentation method provided by an embodiment of the present invention, the interference-containing SAR echo signal to be processed is first obtained, and short-time Fourier transform and normalization are performed to obtain the corresponding time-frequency map as the time-frequency map to be processed; then a pre-trained generative adversarial network model based on the convolutional block attention mechanism is obtained; wherein the generative adversarial network model based on the convolutional block attention mechanism includes: a generator and a discriminator; the generator is a deep residual network integrating the convolutional block attention mechanism, and the discriminator is a PatchGAN network; then the time-frequency map to be processed is input into An interference suppression result and a corresponding interference segmentation result are obtained from a pre-trained generator of a generative adversarial network model based on a convolutional block attention mechanism; wherein the interference used in the training process of the generative adversarial network model based on the convolutional block attention mechanism is multi-type interference; then, the interference suppression result is inversely normalized and inversely short-time Fourier transformed to obtain an interference-suppressed SAR echo signal, wherein the interference segmentation result is used to assist in judging the effect of interference suppression; finally, imaging processing is performed on the interference-suppressed SAR echo signal to obtain a SAR imaging result after interference suppression.

[0044] This paper introduces the Convolutional Block Attention Module (CBAM) and Generative Adversarial Network (GAN) technology to the problem of generating interference-free target data for SAR, transforming the interference suppression process into an interference-free data generation process. Simultaneously, by introducing the Residual Neural Network (ResNet) and PatchGAN network, a SAR multi-type interference suppression and segmentation method based on the Generative Adversarial Network Based on Convolutional Block Attention Module (CBAM-GAN) is proposed. This method can accurately segment the statistical characteristics of different types of interference in the time-frequency domain and optimize the interference suppression effect, thereby improving the ability to efficiently recover and reconstruct the target echo signal. This paper solves the problems of existing SAR interference suppression methods in terms of signal loss, single interference pattern, and residual interference diffusion energy. It can achieve effective transfer and generalization when facing diverse, high-energy, dynamic interference in actual complex electromagnetic environments. It also proposes new ideas for further promoting the innovation and development of SAR interference suppression technology, and has important theoretical significance and application value.

[0045] Furthermore, the present invention introduces an interference segmentation guidance mechanism in the training process of the network model. Through the interference segmentation guidance mechanism, preset weights are added to the interference area during the reverse propagation process of the network, thereby increasing the focus on the interference area and promoting the improvement of the interference suppression effect.

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 1 is a flow chart of a SAR multi-type interference suppression and segmentation method provided by an embodiment of the present invention;

[0048] Figure 2These are time domain, frequency domain, and time-frequency domain representation diagrams of echoes containing sinusoidal modulated swept frequency interference and dense false target interference, respectively, at JNR = 25 dB, provided by an embodiment of the present invention. (a) is a time domain representation diagram of an echo containing sinusoidal modulated swept frequency interference, (b) is a frequency domain representation diagram of an echo containing sinusoidal modulated swept frequency interference, (c) is a time-frequency domain representation diagram of an echo containing sinusoidal modulated swept frequency interference, (d) is a time domain representation diagram of an echo containing dense false target interference, (e) is a frequency domain representation diagram of an echo containing dense false target interference, and (f) is a time-frequency domain representation diagram of an echo containing dense false target interference.

[0049] Figure 3 This is a structural diagram of the generative adversarial network based on the convolutional block attention mechanism of the present invention;

[0050] Figure 4 This is a network structure diagram of the convolutional block attention mechanism of the present invention;

[0051] Figure 5 : The following are the imaging results of adding simulated interference to the actual interference-free SAR echo signal, where (a) is the imaging result of the actual interference-free SAR echo signal, (b) is the imaging result after adding sinusoidal modulation swept frequency interference to the actual interference-free SAR echo signal, (c) is the imaging result after adding comb spectrum interference to the actual interference-free SAR echo signal, (d) is the imaging result after adding dense false target interference to the actual interference-free SAR echo signal, and (e) is the imaging result after adding intermittent sampling direct forwarding interference to the actual interference-free SAR echo signal.

[0052] Figure 6 3. The figures are comparison diagrams of imaging results after using different interference suppression methods to suppress sinusoidal modulation swept frequency interference, wherein (a) is the imaging result diagram after adding sinusoidal modulation swept frequency interference to the actual interference-free SAR echo signal, (b) is the imaging result diagram obtained by using the interference suppression method based on residual network (IMN) for (a), (c) is the imaging result diagram obtained by using the wideband RF interference suppression method based on generative adversarial network (WBIM-GAN) for (a), and (d) is the imaging result diagram obtained by using the method of the present invention for (a);

[0053] Figure 7: These are comparisons of imaging results after comb spectrum interference suppression using different interference suppression methods, where (a) is the imaging result after adding comb spectrum interference to the actual interference-free SAR echo signal, (b) is the imaging result obtained by using the IMN method for (a), (c) is the imaging result obtained by using the WBIM-GAN method for (a), and (d) is the imaging result obtained by using the method of the present invention for (a);

[0054] Figure 8 : These are comparisons of imaging results after suppressing dense false target interference using different interference suppression methods, where (a) is the imaging result after adding dense false target interference to the actual interference-free SAR echo signal, (b) is the imaging result obtained by using the IMN method for (a), (c) is the imaging result obtained by using the WBIM-GAN method for (a), and (d) is the imaging result obtained by using the method of the present invention for (a);

[0055] Figure 9 : These are comparisons of imaging results after intermittent sampling direct forwarding interference suppression using different interference suppression methods, where (a) is the imaging result after adding intermittent sampling direct forwarding interference to the actual interference-free SAR echo signal, (b) is the imaging result obtained by using the IMN method for (a), (c) is the imaging result obtained by using the WBIM-GAN method for (a), and (d) is the imaging result obtained by using the method of the present invention for (a);

[0056] Figure 10 : This is the segmentation mask map of the actual interference SAR echo signal generated by the generative adversarial network based on the convolutional block attention mechanism of the present invention, where (a) is the segmentation mask map of the network for the 422nd actual interference echo, and (b) is the segmentation mask map of the network for the 566th actual interference echo;

[0057] Figure 11 3. The figures are comparison diagrams of the imaging results after suppressing the actual interference-containing SAR echo signal using different interference suppression methods, wherein (a) is the imaging result diagram of the actual interference-containing SAR echo signal recorded by the Sentinel-1 satellite in February 2025, (b) is the imaging result diagram obtained by using the IMN method for (a), (c) is the imaging result diagram obtained by using the WBIM-GAN method for (a), (d) is the imaging result diagram obtained by using the method of the present invention for (a), and (e) is the imaging result diagram of the revisited recorded data of the Sentinel-1 satellite at the same location in February 2025. DETAILED DESCRIPTION

[0058] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0059] The embodiment of the present invention provides a SAR multi-type interference suppression and segmentation method, see Figure 1 , which may include the following steps:

[0060] S1, obtaining the interference-containing SAR echo signal to be processed, performing short-time Fourier transform and normalization, and obtaining the corresponding time-frequency diagram as the time-frequency diagram to be processed;

[0061] S2, obtaining a pre-trained generative adversarial network model based on a convolutional block attention mechanism; wherein the generative adversarial network model based on a convolutional block attention mechanism includes: a generator and a discriminator; the generator is a deep residual network integrated with a convolutional block attention mechanism, and the discriminator is a PatchGAN network;

[0062] S3, inputting the to-be-processed time-frequency graph into a generator of a pre-trained generative adversarial network model based on a convolutional block attention mechanism to obtain an interference suppression result and a corresponding interference segmentation result; wherein the interference used in the training process of the generative adversarial network model based on the convolutional block attention mechanism is multi-type interference;

[0063] S4, performing inverse normalization and inverse short-time Fourier transform on the interference suppression result to obtain an interference-suppressed SAR echo signal, wherein the interference segmentation result is used to assist in determining the effect of interference suppression;

[0064] S5, performing imaging processing on the interference suppressed SAR echo signal to obtain an interference suppressed SAR imaging result.

[0065] To facilitate understanding of the method of the embodiment of the present invention, the training process of the generative adversarial network model based on the convolutional block attention mechanism is first described, including:

[0066] Each acquired interference-free SAR echo signal is superimposed with any one of multiple types of interference to generate a corresponding interference-containing SAR echo signal. Using short-time Fourier transform and normalization, a set of interference-free SAR echo signals and the corresponding interference-containing SAR echo signal are converted into time-frequency maps, forming a reference time-frequency map pair. In the reference time-frequency map pair, the time-frequency map of the interference-free SAR echo signal serves as the true label of the time-frequency map of the interference-containing SAR echo signal. The time-frequency map of the interference-containing SAR echo signal is pixel-wise segmented using a set threshold to obtain a corresponding segmentation label. The segmentation label is a binary image containing 0 and 1, where 0 indicates that the corresponding pixel in the time-frequency map of the interference-containing SAR echo signal is free of interference, and 1 indicates that the corresponding pixel in the time-frequency map of the interference-containing SAR echo signal is interference-containing. The time-frequency map reference pairs and their corresponding segmentation labels are used as samples to construct a dataset. A large number of samples are selected from the dataset to form a training set, and a small number of remaining samples form a validation set.

[0067] Typically, a SAR system's echo signal consists of three components: the target echo, noise, and interference. Interference is generally categorized as unintentional or intentional. This paper focuses on suppressing intentional interference. Intentional interference is typically categorized into suppressive jamming and deceptive jamming. The former suppresses the target signal by emitting high-intensity noise or continuous waves, for example, in frequency or power. The latter simulates or falsifies the target signal's characteristics, emitting false signals to cover the real signal. This is closely related to the key parameters of the target signal.

[0068] The present invention uses four types of interference as network training data, including two types of suppression interference: sinusoidal modulation sweep interference and comb spectrum interference, and two types of deceptive interference: dense false target interference and intermittent sampling direct forwarding interference. That is to say, the multi-type interference of the present invention includes the above four types. In a specific embodiment, the jamming-noise-ratio (JNR) is set to 25dB, and the characteristic representation of the SAR echo signal containing the above interference in the time domain, frequency domain and time-frequency domain is analyzed, wherein the time-frequency domain representation uses the short-time Fourier transform to transform the SAR signal from the time domain to the time-frequency domain. Figure 2As shown, it can be seen that in the time domain, the interference varies with time and the amplitude of the interference is much higher than that of the signal. In the frequency domain, the interference varies with frequency and the interference component occupies the majority of the proportion. Since the time-frequency domain represents the three-dimensional distribution of time-frequency-energy, it is obvious that the time and frequency characteristics of the interference are captured at the same time. In addition, the instantaneous change law, energy distribution and statistical characteristics of the interference are also accurately characterized, and effective interference characterization can improve the effect of interference identification, thereby promoting the separation of interference and target signals and thus achieving interference suppression. In addition, normalizing the time-frequency diagram facilitates network training. Therefore, the present invention uses the normalized time-frequency diagram of the SAR signal as the input of the network. In addition, since the SAR signal is characterized by the real part and the imaginary part to represent the integrity of the signal, the present invention further uses the two-dimensional normalized time-frequency diagram spliced ​​by the real part and the imaginary part as the input of the network.

[0069] In one embodiment, the present invention selects the data recorded by the Sentinel-1A satellite in February 2019 as the original interference-free echo, and constructs a simulated interference model based on the echo parameters, thereby adding simulated interference to the basic data to generate corresponding interference-containing echo data. The interference-free and interference-containing echoes are converted into time-frequency graph reference pairs through short-time Fourier transform and normalization. To facilitate network training, the size of the time-frequency graph in the present invention is uniformly set to 128×128×2, where 128×128 represents the height × width of the time-frequency graph, and 2 represents the channel dimension of the time-frequency graph. The present invention uses two dimensions, real and imaginary, to characterize the integrity of the SAR signal characteristics, so the dimension value is 2.

[0070] Constructing a training set is a necessary prerequisite for network training and is crucial for building an accurate network model. Similar to conventional model training, the present invention can divide the dataset into a training set and a validation set according to a preset ratio. The training set is used for model training, and the validation set is used to verify the model during training.

[0071] A generative adversarial network model based on a convolutional block attention mechanism is constructed, comprising: a generator and a discriminator. The generator is used to process the input time-frequency graph of the interfering SAR echo signal, outputting an interference suppression result and an interference segmentation result, thereby completing interference suppression and segmentation. The discriminator is used to process the input interference suppression result and the true label, outputting a discriminant score matrix of the interference suppression result and a discriminant score matrix of the true label, thereby completing the judgment of the interference suppression result. The interference suppression result refers to the result obtained by the generator after performing interference suppression on the time-frequency graph of the interfering SAR echo signal, and the interference segmentation result refers to the probability matrix obtained by the generator after performing pixel-level segmentation on the time-frequency graph of the interfering SAR echo signal. Each element in the probability matrix represents the probability of whether the pixel at the corresponding position in the time-frequency graph of the interfering SAR echo signal has interference, and a larger probability value indicates a greater likelihood that the pixel at the corresponding position has interference.

[0072] The essence of interference suppression in the time-frequency domain is to separate the interference and target by exploiting the statistical differences between them. This process can be viewed as the generation of interference-free target signals and the generation of interference segmentation labels. Because generative adversarial networks (GANs) have a well-established application and significant value in image generation, and because these methods can generate highly realistic data through adversarial game play between the generator and the discriminator, GANs can be considered as the underlying network structure for this method.

[0073] The generative adversarial network consists of a generator and a discriminator. The former is mainly used to generate interference suppression results and interference segmentation results, and the latter is mainly used to judge the interference suppression results generated by the generator, and to encourage the generator to generate more realistic interference suppression results through the penalty function. It can be seen that the generator and the discriminator are in a process of continuous game. The two confront and promote each other until the network converges and obtains a parameter set with optimal performance. Considering that the main functions of the generator and the discriminator are quite different, the basic network structures of the two are also very different. The basic structure adopted by the generator in the present invention is a deep residual network structure with an integrated convolutional block attention mechanism, and the discriminator adopts a PatchGAN network structure.

[0074] Considering that the generator needs to establish a mapping relationship between interference data and non-interference data, the input of the generator is the time-frequency diagram of the interference-containing SAR echo signal, and the output is the interference suppression result and the interference segmentation result. The discriminator needs to judge the interference suppression result, evaluate whether it is the generated interference suppression result or the time-frequency diagram of the original non-interference SAR echo signal, and obtain an evaluation result to punish the generator. Therefore, the input of the discriminator is the interference suppression result and the true label, and the output is the discriminant score matrix of the interference suppression result and the discriminant score matrix of the true label. The discriminant score matrix is ​​a scalar matrix, and each scalar in the matrix represents the discrimination result of a fixed area. In the present invention, the fixed area is a 16×16 matrix area.

[0075] In the generator of the generative adversarial network (GAN), interference suppression and interference segmentation processes share the same network, using a deep residual network (DRN) as the underlying architecture. Through residual learning and skip connections, DRNs fuse shallow texture detail features with deep, abstract semantic information, preserving information integrity and preventing network degradation, addressing vanishing and exploding gradients.

[0076] For details, see Figure 3 As shown in the dotted box in the upper middle section, the generator includes: a sequentially connected convolutional layer, an activation layer, the first CRB (Convolutional Residual Block, a residual block integrated with the convolutional block attention mechanism) module, the second CRB module, and the first RB (Residual Block, basic residual block) module. The output of the first RB module is divided into two branches, wherein the first branch contains the sequentially connected second RB module, a CBAM module, the third CRB module, the third RB module, the fourth CRB module, and a convolutional layer; the second branch contains the sequentially connected fourth RB module, the fifth RB module, a convolutional layer, a standard normalization layer, and a convolutional layer, and the output of the first RB module is added point by point to the input of the first CRB module to realize the residual connection. The first branch outputs an interference suppressed image, and the second branch outputs an interference segmented image. Any CRB module is a residual block integrated with the convolutional block attention mechanism, including two sequentially connected RB modules and a CBAM module. Each RB module is a basic residual block, consisting of a convolutional layer, a standard normalization layer, a convolutional layer, a standard normalization layer, and an activation layer connected in sequence. The output of the activation layer in any RB module is added point by point to the input of the RB module to achieve a residual connection. The CBAM module is a convolutional block attention mechanism module.

[0077] In recent years, the convolutional block attention mechanism (CBAM) has been widely used in fields such as target detection and semantic segmentation. This mechanism focuses on important features of interest by adaptively adjusting the feature responses in the channel and spatial dimensions of the feature map, thereby improving the representation ability of the model. As a lightweight attention mechanism, the convolutional block attention mechanism includes channel attention and spatial attention, which can adaptively adjust the feature responses in the channel and spatial dimensions of the feature map. The channel attention module pays attention to which channels in the feature map contain more important information, thereby enhancing the feature responses of these channels; the spatial attention focuses on which spatial positions in the feature map contain key information, thereby enhancing the features of these positions. This adaptive feature optimization mechanism enables the model to pay more attention to important features and suppress irrelevant features. The present invention integrates the convolutional block attention mechanism into the deep residual network as a generator of the generative adversarial network to improve the representation ability of the model.

[0078] See Figure 4 The channel attention and spatial attention modules are combined in series, with the channel attention layer preceding the spatial attention layer. The channel attention module consists of one global max pooling layer, one global average pooling layer, one multilayer perceptron with 3 layers, and one broadcast layer. The multilayer perceptron has a dimensionality of 16. First, global max pooling and global average pooling are used in parallel to focus on specific features. Then, multilayer perceptrons are used to increase the nonlinear representation capability of the network. The two features are then fused and broadcasted. Finally, the input and output features are multiplied point by point to produce the channel attention feature map output. The spatial attention module consists of one global max pooling layer, one global average pooling layer, one full convolutional layer with a 1×1 kernel, and one broadcast layer. Similarly, global max pooling and global average pooling are used in parallel, but the processing is performed in the spatial dimension to focus on specific spatial features. Then, features from both dimensions are fused using convolution. Finally, the output of the channel attention feature map is fused with the output of the convolutional block attention mechanism to produce the final output. The expression of the above process is:

[0079]

[0080] Formula (1) reflects the serial characteristics of the channel and spatial attention modules in the convolutional block attention mechanism, where F is the input feature image of the convolutional block attention module, F′ is the output of the channel attention module, and F″ is the final output of the convolutional block attention module. Indicates element-by-element multiplication, M C (F) represents the channel attention module processing, M S (F′) represents the spatial attention module processing.

[0081] Channel attention module processes MC The expression of (F) is:

[0082]

[0083] Among them, σ represents the sigmoid activation operation, MLP represents a shared multi-layer perceptron with 1 hidden layer, gmp represents global maximum pooling, gap represents global average pooling, and Represents the weight of the multi-layer perceptron, C represents the dimension, r represents the channel dimensionality reduction and dimension increase size in the multi-layer perceptron, which is set to 16 in the present invention. represents the output of F after global maximum pooling, represents the output of F after global average pooling.

[0084] The spatial attention module processes M S The expression of (F′) is:

[0085] M S (F′)=σ(Conv 7×7 (cat[gmp(F′);gap(F′)])) (3)

[0086] Among them, Conv 7×7 represents a convolution operation with a convolution kernel of 7×7, and cat[·] represents a feature map concatenation operation.

[0087] Through the description of the above principles and formulas, it can be seen that the convolutional block attention mechanism itself has good interpretability, and the network structure has visualization characteristics, such as Figure 4 As shown, it can be seen intuitively that the convolutional block attention mechanism improves the deep residual network, providing strong support for the purpose of SAR multi-type interference suppression of the present invention.

[0088] The activation layer of the generator uses the ReLU function, the convolution kernel size and step size are set to 3×3 and 1 respectively, and the Adam (Adaptive Moment Estimation) optimizer is used. The first-order decay rate and the second-order decay rate are set to 0.5 and 0.9 respectively. The learning rate is initialized to 10e-3 and the learning rate is dynamically adjusted according to the cosine descent strategy.

[0089] The basic structure used by the discriminator of the generative adversarial network in the present invention is the PatchGAN network. Unlike the traditional generative adversarial network discriminator, which only outputs a single scalar as the discrimination score, PatchGAN judges the local area of ​​the image and outputs a two-dimensional matrix. Each element in the matrix corresponds to the authenticity judgment of a small area of ​​the image, thereby improving the discriminator's attention to local texture and details rather than the global structure of the entire image. Because each discrimination result only depends on the local area, a balance is achieved in the image generation task between efficient computing and detail optimization. It is often used in high-resolution image processing and is suitable for the SAR image interference suppression and segmentation tasks in the present invention. Therefore, the present invention uses PatchGAN technology to improve the image generation quality of the network.

[0090] For details, see Figure 3 As shown in the dotted box in the lower middle section, the discriminator includes one convolutional layer, one activation layer, seven CBL modules, and one convolutional layer connected in sequence. The CBL module is a Conv-BatchNormalization-LeakyReLU integrated module, and any CBL module includes one convolutional layer, one standard normalization layer, and one activation layer connected in sequence.

[0091] Similar to the generator, the activation layer of the discriminator uses the LeakyReLU function, the convolution kernel size is 3×3, the step size is 2, the padding is 1, the Adam optimizer is used, the first-order decay rate and the second-order decay rate are set to 0.5 and 0.9 respectively, the learning rate is initialized to 10e-3, and the learning rate strategy is dynamically adjusted according to the cosine descent.

[0092] Then, a preset loss function is designed based on the weighted method of multiple loss functions.

[0093] In deep learning, the loss function is a key component that measures the model's prediction effect and guides the optimization direction. It is an indispensable part of the neural network design process. It converts the deviation between the model's prediction result and the true label into a quantifiable scalar value through a mathematical expression. In the process of backpropagation to update the parameters, the gradient of the loss function directly determines the direction and amplitude of the parameter adjustment, and different tasks require the design of different loss functions to guide the model to learn the relevant features of different tasks. According to the task division, the task of the present invention includes the suppression and segmentation of multiple types of SAR interference, so the loss function must design an interference suppression loss function and an interference segmentation loss function. According to the network structure division, the present invention uses a generative adversarial network, which includes a generator and a discriminator. Therefore, the design of the loss function needs to consider the loss function of the generator, the loss function of the discriminator, and the adversarial loss function of the two. According to the relationship between tasks and structures, the generator needs to complete the suppression and segmentation of interference, and the discriminator needs to complete the judgment of the interference suppression results. Therefore, no matter from which dimension, the loss function of the generative adversarial network (CBAM-GAN) based on the convolutional attention mechanism of the present invention is composed of the loss function of the generator and the loss function of the discriminator. Among them, the loss function of the generator is a weighted combination of three parts: the improved mean square error loss function, the cross entropy loss function, and the adversarial loss function between the generator and the discriminator, thereby achieving balanced optimization of multiple objectives. Specifically, the mathematical expression of the network loss function is:

[0094] Loss=minmax(L D +L G )

[0095] L G =λL Mo_MSE +λ1L CE +λ2L A (4)

[0096] Among them, Loss is the preset loss function; L G is the loss function of the generator; L D is the loss function of the generative adversarial network discriminator; L Mo_MSE is the improved mean square error loss function; L CE is the cross entropy loss function; L A is the adversarial loss function between the generator and the discriminator; λ, λ1, and λ2 are three hyperparameters with values ​​of 1, 10e-5, and 10e-5, respectively, which are used to balance the optimization of multiple objectives such as the suppression and segmentation of SAR multi-type interference in the present invention.

[0097] The loss function L of the generative adversarial network discriminator DBy introducing a gradient penalty term to constrain the gradient of the discriminator, the training is made more stable. This loss function is used in the present invention to optimize the judgment task of the discriminator, and its expression is:

[0098]

[0099] in, Indicates the expectation, S represents the time-frequency diagram of the interference-free SAR echo signal, that is, the true label, J represents the time-frequency diagram of the interference-containing SAR echo signal, D(S) represents the discriminant score matrix of the true label output by the discriminator, G(J) represents the interference suppression result output by the generator, and D(G(J)) represents the discriminant score matrix of the interference suppression result output by the discriminator. represents the WGAN-GP (Wasserstein GAN with Gradient Penalty, gradient penalty WGAN) penalty term, where The value of is related to the time-frequency diagram S of the interference-free SAR echo signal and the interference suppression result G(J) output by the generator. It is the linear interpolation result of S and G(J). represents the gradient, ||·|| represents the 2-norm, (·) 2 represents the square operation, and λ3 is a hyperparameter with a value of 10.

[0100] Generally, the mean square error loss function is used to measure the average square distance between the predicted value and the true value, and the predicted value is approached to the true value by adjusting the distance. The present invention uses an improved mean square error loss function L Mo_MSE To optimize the task of interference suppression in the generator, its expression is:

[0101]

[0102] At the same time, the segmentation guidance mechanism is used to optimize the network parameter update during back propagation, and the interference segmentation results are used to strengthen the focus on the interference location and improve the ability to suppress and guide interference. The expression of the segmentation guidance mechanism is:

[0103] dL Mo_MSE =M×(SG(J)) (7)

[0104] Where d represents differentiation, M represents the reset segmentation label, S represents the time-frequency diagram of the SAR echo signal without interference, J represents the time-frequency diagram of the SAR echo signal with interference, and G(J) represents the interference suppression result output by the generator.

[0105] The implementation process of the segmentation guidance mechanism includes resetting the values ​​in the segmentation label, including resetting the positions with a value of 0 in the segmentation label to 1, and resetting the positions with a value of 1 to an integer greater than 1, which is set to 10 in the present invention. At this time, the calculation of the non-interference position is not affected, while the weight of the interference position is increased, so that the product result pays more attention to the interference area.

[0106] Cross entropy loss function L CE , which is used to measure the difference between the predicted probability distribution and the true distribution. In this paper, the cross entropy loss function is used to optimize the image segmentation task, and its expression is:

[0107]

[0108] Among them, y ic Indicates the true situation that the i-th interference segmentation result belongs to category c, and takes a value of 0 or 1, where 1 means it belongs to the category and 0 means it does not belong to the category. represents the probability that the interference segmentation result belongs to category c, N represents the total number of interference segmentation results, C represents the total number of categories, which is set to 2 in the present invention, indicating two categories: interference and no interference, log represents the logarithm with base 10, and ∑· represents the summation operation.

[0109] L A The adversarial loss function between the generator and the discriminator is used in this paper to force the generator to deceive the discriminator and promote the model to learn the complex structure of data distribution. Its expression is:

[0110]

[0111] Wherein, eps is a constant, and the constant value is set to 10e-12 in the present invention.

[0112] In general, the loss function is not only a tool for measuring model error, but also the core driving force for model training and learning. By quantifying errors, guiding parameter updates, and strengthening task objectives, the training of the generative adversarial network in this invention is transformed into an optimizable mathematical problem. Its importance runs through the entire process from model design to training and tuning.

[0113] Using the samples in the training set, as well as the corresponding interference suppression results, interference segmentation results, the discriminant score matrix of the interference suppression results, the discriminant score matrix of the true label and the preset loss function, the generator and the discriminator are iteratively trained and verified to obtain a trained generative adversarial network model based on the convolutional block attention mechanism, specifically including:

[0114] Step a1, selecting a preset number of samples from the training set as a batch;

[0115] The batch size used in the present invention is 4;

[0116] Step a2: inputting the time-frequency diagram of the interfering SAR echo signal in the batch of samples into the generator to obtain corresponding interference suppression results and interference segmentation results;

[0117] The size of the interference suppression result and the interference segmentation result are both 128×128;

[0118] Step a3: inputting the interference suppression result and the true label into a discriminator to obtain a discriminant score matrix of the interference suppression image and a discriminant score matrix of the true label;

[0119] The above discriminant score matrices are all scalar matrices, and each element in the matrix corresponds to the authenticity judgment of a small area of ​​the image, thereby improving the discriminator's attention to local textures and details rather than the global structure of the entire image.

[0120] Step a4, subtracting the discriminant score matrix of the interference suppression result from the discriminant score matrix of the true label and taking the global average to obtain the final discriminant score;

[0121] The matrix subtraction operation mentioned above refers to subtracting the elements of the corresponding positions of the two matrices, and the global average operation is to add the results of the subtraction of the two matrices element by element and then calculate the average. The global average represents the final discrimination score, which is used to calculate the discriminator loss function L D , and then complete the update of the discriminator parameters.

[0122] Step a5: Based on the final discriminant score and the loss function L of the discriminator D The penalty term in the discriminator loss function L is calculated D The value of

[0123] See formula (5), That is, the operation of obtaining the final discriminant score in step a4 is equivalent to first subtracting and then averaging. In addition, the discriminator loss function L D The gradient penalty term is introduced to constrain the gradient of the discriminator, making the training more stable.

[0124] Step a6, according to the loss function L of the discriminator D The value of is used to update the discriminator parameters using the back propagation algorithm;

[0125] The automatic differentiation method (Autograde) provided by Pytorch is used to process the back-propagation process, and the Adam optimizer is used to accelerate network convergence and complete the update of the discriminator parameters.

[0126] Step a7, calculating the loss function L of the generator based on the discrimination score matrix of the interference suppression result, the interference suppression result, the true label, the interference segmentation result and the segmentation label G The value of

[0127] The loss function L of the generator G It includes an improved mean square error loss function, a cross entropy loss function, and a network adversarial loss function. The improved mean square error loss function makes the interference suppression result closer to the true label. The cross loss function measures the distribution difference between the generated segmentation result and the true segmentation result. The adversarial loss function is used to force the generator to deceive the discriminator. The generator's loss function L G The expression is:

[0128] L G =λL Mo_MSE +λ1L CE +λ2L A (10)

[0129] Step a8, according to the loss function L of the generator G The value of the back propagation algorithm is used to update the generator parameters; wherein the improved mean square error loss function L Mo_MSE Utilize the back-propagation algorithm to include a segmentation guidance mechanism;

[0130] The back propagation process is processed by using Pytorch's own automatic differentiation method (Autograde), and the Adam optimizer is used to accelerate network convergence and complete the update of the generator parameters; in particular, the improved mean square error loss function L in this patent is Mo_MSE The back propagation algorithm of the mean square error loss function is redefined, that is, a segmentation guidance mechanism is used in the algorithm to increase the attention to the interference area.

[0131] The implementation process of the segmentation guidance mechanism includes:

[0132] Resetting the values ​​in the segmentation tag, including resetting the positions with values ​​of 0 in the segmentation tag to 1, and resetting the positions with values ​​of 1 to integers greater than 1;

[0133] For details, please refer to the previous description and will not be described in detail here.

[0134] Step a9: Repeat steps a1-a8 for a preset number of iterations and then perform a verification.

[0135] Step a10: Repeat step a9 until the preset epoch is reached to obtain a trained generative adversarial network based on the convolutional block attention mechanism, where a single epoch is the process of the model performing a complete traversal of the training set.

[0136] The present invention sets the epoch to 40.

[0137] The present invention sets the network to iterate 50 times for one verification, and logs the losses during the training process. At the same time, the interference suppression results and interference segmentation results during the training process are saved in the form of images, and the interference suppression results and interference segmentation results during the verification process are saved in the form of images, so as to monitor the training process and facilitate our judgment and determination of the optimization direction. If the loss does not decrease, the interference suppression is overfitting or underfitting, or the segmentation is not executed, the training can be stopped in time and the relevant parameters can be adjusted or the data can be enhanced before continuing the training, thereby gradually achieving troubleshooting and ensuring the normal learning of the model.

[0138] After training and verification are completed, the generative adversarial network based on the convolutional block attention mechanism is the parameter optimization setting network of the generator (inhibition network + segmentation network) and the discriminator (judgment network).

[0139] During the testing phase, the SAR echo signal is similarly converted from the time domain to the time-frequency domain and the normalized time-frequency graph is input into a trained generative adversarial network based on the convolutional block attention mechanism to generate interference suppression and interference segmentation results. The interference suppression result is then first inversely normalized and then inversely short-time Fourier transformed to the time domain for SAR imaging. Finally, the interference suppression effect is analyzed. The interference segmentation result is a probability matrix obtained by the generator performing pixel-level segmentation on the time-frequency graph containing the interfering SAR echo signal, recording the pixel locations where interference may exist in the time-frequency graph containing the interfering SAR echo signal. Since the interference in the interference suppression result has been removed, that is, it can be considered that there is no interference in the interference suppression result, the interference segmentation result can be used to assist in observing the corresponding position of the interference in the interference suppression result, thereby judging the effect of interference suppression and providing support for the suppression decision.

[0140] The testing phase of the present invention is divided into simulated interference testing and actual interference testing. The simulated interference testing involves adding simulated interference to actual interference-free SAR echo signal data to test the effectiveness and generalizability of the present invention. The actual interference testing involves performing interference suppression on actual interference-containing SAR echo signals to further verify the portability and superiority of the present invention. The specific steps are as follows:

[0141] (1) Simulation interference test

[0142] Step 1: Select a new interference-free SAR echo scene (i.e., the data of the interference-free SAR echo signal in this scene is different from the data of the interference-free SAR echo signal used in the training set and the validation set), obtain the interference-free SAR echo signal in this scene, add the interference component to the interference-free SAR echo signal according to the echo parameters, and perform short-time Fourier transform and normalization to obtain a time-frequency diagram of the interference-containing SAR echo signal with a size of 128×128×2 (width×height×number of channels), and then construct a test set. The simulated interference of the test set of the present invention includes sinusoidal modulation swept frequency interference, comb spectrum interference, dense false target interference, and intermittent sampling direct forwarding interference, and the interference-to-noise ratio (JNR) is set to 25dB.

[0143] Step 2: Input the time-frequency graph generated in step 1 into the trained generative adversarial network based on the convolutional block attention mechanism for suppression and segmentation, and record and save the interference suppression results and interference segmentation results.

[0144] Step 3: first perform inverse normalization on the interference suppression result recorded in step 2 and then perform inverse short-time Fourier transform, and save the inverse normalized time-frequency diagram.

[0145] Step 4: Image the data of the interference-free SAR echo signal in step 1 to obtain interference-free SAR imaging results. Select two fixed areas as the non-reflection area and the reflection area to calculate the quantitative analysis indicator MNR (Multiplicative Noise Ratio). The non-reflection area is the darker area in the imaging result (see the attached figure). Figure 5 (a) in yellow box), the reflection area is the strong reflection area in the imaging result (see attached Figure 5 blue box in (a).

[0146] In step 5, the data after the short-time Fourier inverse transform processing in step 3 is imaged to obtain the final SAR imaging result after interference suppression. By analyzing the image, the interference suppression effect of the method of the present invention can be intuitively seen. At the same time, combined with the non-reflection area and the reflection area described in step 4, the corresponding indicators are calculated for quantitative analysis.

[0147] (2) Actual interference test

[0148] Step 1: Select the actual interfering SAR echo signal, perform short-time Fourier transform and normalization, and obtain a time-frequency diagram of the interfering SAR echo signal with a size of 128×128×2 (width×height×number of channels). Then, construct a test set. At the same time, select the data of the non-interfering SAR echo signal corresponding to the same scene and SAR system for comparative analysis with the experimental results.

[0149] Step 2: Input the time-frequency graph generated in step 1 into the trained generative adversarial network based on the convolutional block attention mechanism for suppression and segmentation, and record and save the interference suppression results and interference segmentation results.

[0150] In step 3, the interference suppression result recorded in step 2 is first inverse normalized and then subjected to inverse short-time Fourier transform, and the time-frequency diagram of the unsuppressed interference SAR echo signal is imported. The time-frequency diagram of the interference suppression result during the inverse normalization process and the time-frequency diagram of the unsuppressed interference SAR echo signal are recorded simultaneously to observe the suppression effect in the time-frequency domain.

[0151] Step 4: Image the data of the interference-free SAR echo signal described in step 1 to obtain interference-free SAR imaging results. Select two fixed areas as the non-reflection area and the reflection area to calculate the quantitative analysis indicator MNR. The non-reflection area is the darker area in the imaging result (see the attached figure). Figure 11 (e) in yellow box), the reflection area is the strong reflection area in the imaging result (see attached Figure 11 blue box in (e).

[0152] In step 5, the data after the short-time Fourier inverse transform in step 3 is subjected to SAR imaging to obtain the final SAR imaging result after interference suppression. By analyzing the image, the interference suppression effect of the method of the present invention can be intuitively seen. At the same time, combined with the non-reflection area and the reflection area described in step 4, the corresponding indicators are calculated and quantitative analysis is performed.

[0153] The effectiveness of the generative adversarial network based on the convolutional block attention mechanism proposed in the present invention is verified by simulating interference data and actual interference data. At the same time, the evaluation indicators MNR (Multiplicative Noise Ratio, multiplicative noise ratio), SSIM (Structural Similarity, structural similarity), PSNR (Peak signal-to-noise ratio, peak signal-to-noise ratio), and RMSE (Root Mean Squared Error) commonly used in SAR image processing are used to further verify the superiority of the method proposed in the present invention. It should be noted that for these four evaluation indicators, the smaller the MNR and RMSE values, the better the suppression effect. Conversely, the larger the SSIM and PSNR values, the better the suppression effect. The simulation experiments of the present invention are as follows:

[0154] (1) Simulation interference data suppression experiment

[0155] In the process of simulating interference data suppression, the actual interference-free SAR echo signal data selected by the present invention is the interference-free echo recorded by the Sentinel-1 satellite in February 2019, with an echo number of 1548, and a simulated interference component is added to this interference-free echo data. The simulated interference used in the present invention is sinusoidal modulation sweep interference, comb spectrum interference, dense false target interference, and intermittent sampling direct forwarding interference, and the interference-to-noise ratio JNR is set to 25dB. The imaging result diagram of the actual interference-free SAR echo signal data and the imaging result diagram of the echo signal data after adding the above four types of interference are shown in Figure 2. Figure 5 As shown in the figure, it can be seen that different interferences appear in different styles in the scene, but they all have different degrees of submergence and obstruction on the target signal. For example, suppression interference appears as a bright broadband in the SAR imaging result image, completely submerging the target signal. Deception interference appears as a narrow bright band or a bright block in the SAR imaging result image, covering part of the target scene, making it impossible to intuitively identify and detect the scene target.

[0156] The time-frequency graph containing the interfering SAR echo signal is input into the network for interference suppression. The present invention uses an interference suppression method based on a residual network (Interference Mitigation Network, IMN), a broadband RF interference suppression method based on a generative adversarial network (Generative Adversarial Network Based Wideband Interference Mitigation Model, WBIM-GAN), and a proposed suppression method based on a convolutional block attention mechanism (CBAM-GAN). Figure 6 The figure shows a comparison of the imaging results after using different interference suppression methods to suppress the sinusoidal modulation swept frequency interference. It can be seen that the sinusoidal swept frequency interference presents a bright broadband that completely covers the background scene information. The above three interference suppression methods can effectively restore the background information, and texture features such as mountains and lakes can be effectively observed. However, there are serious narrow-band bright lines in the imaging results using the IMN suppression method, which cover part of the ground object information and cannot be observed. The WBIM-GAN method also has non-negligible interference blur residues. In contrast, the CBAM-GAN method of the present invention can completely restore the ground object features, and texture details such as mountains and lakes can be clearly and intuitively observed. Figure 7-Figure 9The following figures show the imaging results after using different interference suppression methods to suppress comb spectrum interference, dense false target interference, and intermittent sampling direct forwarding interference. Similarly, the IMN and WBIM-GAN methods cannot completely suppress the interference, and there are non-negligible interference residues in the imaging result images. However, the CBAM-GAN method of the present invention can restore the scene target information with higher quality. This shows that in the comparison of the imaging result images, the method proposed in the present invention has the best interference suppression effect.

[0157] Table 1 shows the comparison results of evaluation indicators after using different interference suppression methods to suppress sinusoidal modulation swept frequency interference. It can be seen that the CBAM-GAN of the present invention has the best interference suppression effect.

[0158] Table 1 Comparison of evaluation indicators after sinusoidal modulation swept frequency interference suppression

[0159] method MNR SSIM PSNR RMSE IMN -5.3835 0.8448 37.5475 0.1784 WBIM-GAN -9.5198 0.9073 39.7838 0.0729 CBAM-GAN -13.4977 0.9868 49.4283 0.0415

[0160] Similarly, Tables 2, 3, and 4 show the comparative results of different evaluation metrics after using different interference suppression methods to suppress comb spectrum interference, dense false target interference, and intermittent sampling direct forwarding interference. It can be seen that the interference suppression method using the CBAM-GAN of the present invention has the smallest MNR and RMSE, and the largest SSIM and PSNR, fully verifying the effectiveness of the proposed method.

[0161] Table 2 Comb spectrum interference suppression performance comparison

[0162] method MNR SSIM PSNR RMSE IMN -12.7821 0.8644 33.0060 0.1282 WBIM-GAN -14.5417 0.9023 39.6533 0.1204 CBAM-GAN -15.7849 0.9890 50.2607 0.0416

[0163] Table 3 Comparison of dense false target interference suppression performance

[0164] method MNR SSIM PSNR RMSE IMN -7.1183 0.8650 38.0457 0.1936 WBIM-GAN -12.0536 0.9486 42.4930 0.0701 CBAM-GAN -16.6087 0.9873 49.1167 0.0285

[0165] Table 4 Intermittent sampling direct forwarding interference suppression performance comparison

[0166] method MNR SSIM PSNR RMSE IMN -6.5085 0.8481 37.5097 0.1651 WBIM-GAN -11.5218 0.8778 38.2765 0.0901 CBAM-GAN -16.4789 0.9355 41.2198 0.0630

[0167] (2) Actual interference data suppression experiment

[0168] The effectiveness and generalizability of the proposed method were further verified using measured data. Interference suppression was performed on data containing interfering SAR echo signals recorded by the Sentinel-1 satellite in February 2025, with an echo count of 1409. Similar to the simulated interference data suppression experiment, the data containing interfering SAR echo signals was converted into time-frequency plots and normalized before being input into different interference suppression networks: the IMN, WBIM-GAN, and the CBAM-GAN of the present invention. The actual interference suppression results were compared and analyzed. Figure 10 (a) and Figure 10 (b) shows the network's segmentation masks for the 422nd and 566th actual interference-containing SAR echo signals, respectively. To better demonstrate the interference segmentation effect, a binary classification mask generation operation is performed on the interference segmentation results output by the network. That is, the channel dimension maximum index of the interference segmentation result is taken to obtain the above segmentation mask. It can be seen that the CBAM-GAN method of the present invention can accurately capture the spatial location and statistical features of the interference. It can also be seen that the interference is typical swept frequency interference. Figure 11 The figure shows the comparison of imaging results after suppressing the actual interference-containing SAR echo signal using different interference suppression methods. It can be seen that before interference suppression, the SAR image is covered by an interference bright band, which makes it difficult to observe the details of landform features such as farmland and mountains. After IMN and WBIM-GAN suppression, the scene information in the SAR imaging is largely restored, and the interference bright band is significantly weakened. However, there are still significant interference residues in the image that cover the scene targets, and details such as farmland and mountains cannot be fully displayed. After using the CBAM-GAN of the present invention for suppression, the interference bright band is almost completely removed, and the ground object information is effectively restored. Overall, the CBAM-GAN proposed in the present invention has the best suppression effect, followed by the WBIM-GAN suppression method, and the IMN method has the worst suppression effect. Figure 6-Figure 9 The displayed effects remain consistent, indicating that the method of the present invention can be effectively transferred to the actual SAR interference suppression scenario.

[0169] Similarly, the four evaluation indicators of MNR, SSIM, PSNR, and RMSE are used to further verify the method of the present invention. Table 5 shows the comparative results of suppressing the actual data containing interference SAR echo signals using different interference suppression methods. It can be seen that the interference suppression method using the CBAM-GAN of the present invention has the smallest MNR, RMSE and the largest SSIM, PSNR, which is consistent with the results analysis of Tables 1 to 4, and is consistent with Figure 11 The results of the analysis are consistent with those of the previous one, which further verifies the generalization and effectiveness of the method proposed in this paper.

[0170] Table 5 Comparison of different methods for suppressing actual interference-containing SAR echo signal data

[0171] method MNR SSIM PSNR RMSE IMN -0.4568 0.7771 36.0547 0.6015 WBIM-GAN -1.2693 0.7866 36.3989 0.5410 CBAM-GAN -2.3012 0.8718 38.0258 0.2827

[0172] As can be seen, the embodiments of the present invention provide a SAR multi-type interference suppression and segmentation method. This invention introduces the convolutional block attention mechanism and generative adversarial network technology to the problem of SAR interference-free target data generation, transforming the interference suppression process into an interference-free data generation process. At the same time, by introducing a residual neural network and a PatchGAN network, a SAR multi-type interference suppression and segmentation method based on a generative adversarial network with a convolutional block attention mechanism is proposed. This method can accurately segment the statistical characteristics of different types of interference in the time and frequency domains and optimize the interference suppression effect, thereby improving the ability to efficiently recover and reconstruct the target echo signal. The present invention solves the problems of existing SAR interference suppression methods in terms of signal loss, single interference pattern, and residual interference diffusion energy. When faced with diverse, high-energy, dynamic interference in actual complex electromagnetic environments, it can achieve effective transfer and generalization. At the same time, it proposes new ideas for further promoting the innovation and development of SAR interference suppression technology, and has important theoretical significance and application value.

[0173] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A SAR multi-type interference suppression and segmentation method, characterized in that: include: Obtain the interference-containing SAR echo signal to be processed, perform short-time Fourier transform and normalization, and obtain the corresponding time-frequency diagram as the time-frequency diagram to be processed; Obtain a pre-trained generative adversarial network model based on a convolutional block attention mechanism; wherein the generative adversarial network model based on a convolutional block attention mechanism includes: a generator and a discriminator; the generator is a deep residual network integrated with a convolutional block attention mechanism, and the discriminator is a PatchGAN network; Inputting the to-be-processed time-frequency graph into a generator of a pre-trained generative adversarial network model based on a convolutional block attention mechanism to obtain an interference suppression result and a corresponding interference segmentation result; wherein the interference used in the training process of the generative adversarial network model based on the convolutional block attention mechanism is multi-type interference; Performing inverse normalization and inverse short-time Fourier transform on the interference suppression result to obtain an interference-suppressed SAR echo signal, wherein the interference segmentation result is used to assist in determining the effect of interference suppression; Imaging processing is performed on the interference-suppressed SAR echo signal to obtain a SAR imaging result after interference suppression.

2. The method according to claim 1, characterized in that The training process of the generative adversarial network model based on the convolutional block attention mechanism includes: Superimposing any one of multiple types of interference on each acquired interference-free SAR echo signal to generate a corresponding interference-containing SAR echo signal, and using short-time Fourier transform and normalization to convert a group of interference-free SAR echo signals and the corresponding interference-containing SAR echo signals into time-frequency graphs to form a time-frequency graph reference pair; in the time-frequency graph reference pair, the time-frequency graph of the interference-free SAR echo signal serves as the true label of the time-frequency graph of the interference-containing SAR echo signal; using a set threshold, pixel-level segmentation is performed on the time-frequency graph of the interference-containing SAR echo signal to obtain a corresponding segmentation label; the segmentation label is a binary image containing 0 and 1, wherein 0 indicates that the pixel at the corresponding position in the time-frequency graph of the interference-containing SAR echo signal is free of interference, and 1 indicates that the pixel at the corresponding position in the time-frequency graph of the interference-containing SAR echo signal is interfered with; using the time-frequency graph reference pair and the corresponding segmentation label as samples, constructing a data set from a number of samples, selecting a large number of samples from the data set to form a training set, and the remaining small number of samples to form a validation set; A generative adversarial network model based on a convolutional block attention mechanism is constructed, comprising: a generator and a discriminator; the generator is used to process an input time-frequency graph containing an interfering SAR echo signal, and output an interference suppression result and an interference segmentation result, thereby completing interference suppression and segmentation; the discriminator is used to process the input interference suppression result and the true label, and output a discriminant score matrix of the interference suppression result and a discriminant score matrix of the true label, thereby completing the judgment of the interference suppression result; wherein, the interference suppression result refers to the result obtained after the generator performs interference suppression on the time-frequency graph containing the interfering SAR echo signal, and the interference segmentation result refers to the probability matrix diagram obtained by the generator performing pixel-level segmentation on the time-frequency graph containing the interfering SAR echo signal, wherein each element in the probability matrix diagram represents the probability of whether the pixel at the corresponding position in the time-frequency graph containing the interfering SAR echo signal has interference, and a larger probability value indicates a greater possibility that the pixel at the corresponding position has interference; Design a preset loss function based on the weighted method of multiple loss functions; Using the samples in the training set, as well as the corresponding interference suppression results, interference segmentation results, the discriminant score matrix of the interference suppression results, the discriminant score matrix of the true label and the preset loss function, the generator and the discriminator are iteratively trained and verified to obtain a trained generative adversarial network model based on the convolutional block attention mechanism.

3. The method according to claim 1, characterized in that The multiple types of interference include: sinusoidal modulation sweep frequency interference, comb spectrum interference, dense false target interference and intermittent sampling direct forwarding interference.

4. The method according to claim 2, characterized in that The generator includes: A convolutional layer, an activation layer, a first CRB module, a second CRB module, and a first RB module are connected in sequence, and the output of the first RB module is divided into two branches, wherein the first branch includes the second RB module, a CBAM module, a third CRB module, a third RB module, a fourth CRB module and a convolutional layer connected in sequence, and the second branch includes the fourth RB module, the fifth RB module, a convolutional layer, a standard normalization layer and a convolutional layer connected in sequence, and the output of the first RB module is added point by point to the input of the first CRB module to realize the residual connection; wherein, the first branch outputs an interference suppressed image, and the second branch outputs an interference segmented image; any CRB module is a residual block integrated by the convolutional block attention mechanism; any RB module is a basic residual block; the CBAM module is a convolutional block attention mechanism module.

5. The method according to claim 4, characterized in that Any CRB module consists of two RB modules and one CBAM module connected in sequence.

6. The method according to claim 5, characterized in that Any RB module includes a convolution layer, a standard normalization layer, a convolution layer, a standard normalization layer and an activation layer connected in sequence, and the output of the activation layer in any RB module is added point by point to the input of the RB module to realize the residual connection.

7. The method according to claim 2, characterized in that The discriminator includes one convolutional layer, one activation layer, seven CBL modules and one convolutional layer connected in sequence, wherein any CBL module includes one convolutional layer, one standard normalization layer and one activation layer connected in sequence.

8. The method according to claim 2, characterized in that The preset loss function is composed of the loss function of the generator and the loss function of the discriminator, wherein the loss function of the generator is a weighted combination of the improved mean square error loss function, the cross entropy loss function, and the adversarial loss function between the generator and the discriminator; the expression of the preset loss function is: Loss=minmax(L D +L G ) L G =λL Mo_MSE +λ1L CE +λ2L A Among them, Loss is the preset loss function; L G is the loss function of the generator; L D is the loss function of the discriminator; L Mo_MSE is the improved mean square error loss function; L CE is the cross entropy loss function; L A is the adversarial loss function between the generator and the discriminator; λ, λ1, and λ2 are three hyperparameters.

9. The method according to claim 8, characterized in that Using the samples in the training set, as well as the corresponding interference suppression results, interference segmentation results, the discriminant score matrix of the interference suppression results, the discriminant score matrix of the true label and the preset loss function, the generator and the discriminator are iteratively trained and verified to obtain a trained generative adversarial network model based on the convolutional block attention mechanism, including: Step a1, selecting a preset number of samples from the training set as a batch; Step a2: inputting the time-frequency diagram of the interfering SAR echo signal in the batch of samples into the generator to obtain corresponding interference suppression results and interference segmentation results; Step a3: inputting the interference suppression result and the true label into the discriminator to obtain a discriminant score matrix of the interference suppression result and a discriminant score matrix of the true label; Step a4, subtracting the discriminant score matrix of the interference suppression result from the discriminant score matrix of the true label and taking the global average to obtain the final discriminant score; Step a5: Based on the final discriminant score and the loss function L of the discriminator D The penalty term in the discriminator is used to calculate the loss function L D The value of Step a6, according to the loss function L of the discriminator D The value of is used to update the discriminator parameters using the back propagation algorithm; Step a7, calculating the loss function L of the generator based on the discrimination score matrix of the interference suppression result, the interference suppression result, the true label, the interference segmentation result and the segmentation label G The value of Step a8, according to the loss function L of the generator G The value of the back propagation algorithm is used to update the generator parameters; wherein the improved mean square error loss function L Mo_MSE Utilize the back-propagation algorithm to include a segmentation guidance mechanism; Step a9: Repeat steps a1-a8 for a preset number of iterations and then perform a verification. Step a10: Repeat step a9 until the preset epoch is reached to obtain a trained generative adversarial network based on the convolutional block attention mechanism, where a single epoch is the process of the model performing a complete traversal of the training set.

10. The method according to claim 9, characterized in that The implementation process of the segmentation guidance mechanism includes: Resetting the values ​​in the segmentation tag, including resetting the positions with values ​​of 0 in the segmentation tag to 1, and resetting the positions with values ​​of 1 to integers greater than 1; The expression of the segmentation guidance mechanism is: dL Mo_MSE =M×(S-G(J)) Where d represents differentiation, M represents the reset segmentation label, S represents the time-frequency diagram of the SAR echo signal without interference, J represents the time-frequency diagram of the SAR echo signal with interference, and G(J) represents the interference suppression result output by the generator.

Images