A SAR comb spectrum interference suppression method based on a dual-channel attention residual network

Abstract

The application discloses a SAR comb spectrum interference suppression method based on a double-channel attention residual network, and comprises the following steps: acquiring a time domain echo signal with comb spectrum interference; transforming the time domain echo signal to a frequency domain to obtain a frequency domain echo signal with interference; extracting the real part and the imaginary part of the frequency domain echo signal with interference to obtain a real part signal with interference and an imaginary part signal with interference; inputting the real part signal with interference and the imaginary part signal with interference into a pre-trained double-channel attention residual network to suppress the comb spectrum interference and output a frequency domain echo signal; the pre-trained double-channel attention residual network comprises two suppression networks with the same structure, and each suppression network comprises a channel attention residual block and a global attention block; wherein one suppression network is used for reconstructing a real part signal according to an input real part signal with interference, and the other suppression network is used for reconstructing an imaginary part signal according to an input imaginary part signal with interference.

Description

Technical Field

[0001] This invention belongs to the field of interference cancellation technology, specifically relating to a SAR comb spectrum interference suppression method based on a dual-channel attention residual network. Background Technology

[0002] Synthetic Aperture Radar (SAR) boasts advantages such as all-weather, all-day operation, long range, and high-resolution imaging, playing a crucial role in both military and civilian fields, including Earth detection, resource exploration, and battlefield reconnaissance. However, interference reduces the signal-to-interference-plus-noise ratio (SNR) of SAR echoes, affecting subsequent imaging quality and the accuracy of SAR image interpretation. Researching SAR interference suppression under complex electromagnetic environments is of significant importance for improving the anti-interference capability and information acquisition ability of SAR systems.

[0003] Traditional SAR interference suppression methods can be divided into parametric estimation and non-parametric estimation methods. Parametric estimation methods require accurate modeling and parameter estimation of the interference, followed by reconstruction of the interference based on the parameter estimation results, thus achieving precise interference suppression. While this method theoretically achieves superior interference suppression performance, the high precision required for interference modeling leads to low efficiency and poor versatility. Non-parametric estimation methods typically transform the SAR echo signal to a representation domain such as the frequency domain, time-frequency domain, or wavelet domain. By analyzing the characteristic differences between the target echo signal and the interference, they utilize notch filtering, adaptive filtering, and subspace decomposition to suppress the interference. This method is simple and efficient; however, it is prone to target echo signal loss and reduced SAR imaging quality when faced with complex interference, and is only suitable for interference with sparse characteristics. For example, when the number of interference points in comb-spectrum interference is large, it loses its sparsity, and non-parametric methods can easily lead to target echo signal loss, causing SAR image defocusing. The aforementioned traditional SAR interference suppression methods all perform interference suppression before imaging processing. Some researchers have also implemented interference suppression in the image domain, which can effectively suppress sparse interference. However, these methods still lead to the loss of target image information for complex comb-like interference. Furthermore, image-domain anti-interference methods rely on the premise that motion compensation is not required during the imaging process; otherwise, the presence of interference will severely affect the accuracy of imaging parameter estimation, resulting in severe defocusing of the imaging results.

[0004] With the development of deep learning, it has achieved successful applications in image dehazing and speech denoising. Many scholars have combined deep learning with radar anti-jamming, proposing deep learning-based time-frequency domain SAR intelligent interference suppression methods. These methods can project interference and target echoes into a high-dimensional abstract feature space through deep nonlinear networks, achieving decoupling between the interference and target echoes. This not only effectively recovers the amplitude of the target echo signal but also accurately preserves the phase information of the target echo, ensuring the accuracy of target echo reconstruction. However, for comb-spectrum interference that is densely distributed in the time-frequency domain and completely obscures the target signal, the coupling between the interference and target echoes in the time-spectrum diagram is severe, making accurate target reconstruction difficult.

[0005] In other words, existing technologies are only applicable to interference with sparse characteristics. When faced with comb-spectrum interference that is heavily coupled with the target echo, it is difficult to achieve accurate decoupling between the interference and the target echo, which can easily lead to the loss of the target signal and reduce the SAR imaging quality and interpretation accuracy. Summary of the Invention

[0006] To address the aforementioned problems in related technologies, this invention provides a SAR comb spectrum interference suppression method based on a dual-channel attention residual network. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] This invention provides a SAR comb spectrum interference suppression method based on a dual-channel attention residual network, comprising:

[0008] Acquire time-domain echo signals with comb-spectral interference;

[0009] The time-domain echo signal is transformed to the frequency domain to obtain an interference-laden frequency-domain echo signal;

[0010] The real and imaginary parts of the interference frequency domain echo signal are extracted to obtain the interference real part signal and the interference imaginary part signal.

[0011] The real part signal with interference and the imaginary part signal with interference are input into a pre-trained dual-channel attention residual network to suppress comb spectrum interference and output a frequency domain echo signal. The pre-trained dual-channel attention residual network includes two suppression networks with identical structures. Each suppression network includes a channel attention residual block and a global attention block. One suppression network is used to reconstruct the real part signal based on the input real part signal with interference, and the other suppression network is used to reconstruct the imaginary part signal based on the input imaginary part signal with interference.

[0012] The present invention has the following beneficial technical effects:

[0013] This invention utilizes a SAR comb spectrum interference suppression network based on a dual-channel attention residual network to solve the problem of severe target signal loss in existing technologies when suppressing comb spectrum interference. Furthermore, this invention projects the heavily coupled interference and target echo signals to a high-dimensional separable feature space through this network, and introduces channel attention and global attention to improve the accuracy of feature extraction and characterization. This achieves intelligent decoupling of interference and target echo, greatly reducing the loss of target echo signals, thereby improving SAR imaging quality and interpretation accuracy.

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 (a) An exemplary time-domain comparison diagram of interference-free and interference-containing signals provided in an embodiment of the present invention;

[0016] Figure 1 (b) A schematic diagram of the comparison between interference-free and interference signals in the time-frequency domain, provided by an embodiment of the present invention;

[0017] Figure 1 (c) A schematic diagram of the comparison of interference-free and interference signals in the frequency domain provided by an embodiment of the present invention;

[0018] Figure 2 A flowchart of a SAR comb spectrum interference suppression method based on a dual-channel attention residual network provided in an embodiment of the present invention;

[0019] Figure 3 A schematic diagram of the structure of an exemplary attention residual block for each channel provided in an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of the structure of an exemplary attention block for each channel provided in an embodiment of the present invention;

[0021] Figure 5 A schematic diagram of an exemplary intra-frequency self-attention network provided in an embodiment of the present invention;

[0022] Figure 6 A schematic diagram illustrating the feature processing principle of an exemplary intra-frequency self-attention network provided in an embodiment of the present invention;

[0023] Figure 7 A schematic diagram illustrating the processing principle of an exemplary dual-channel attention residual network for input signals, provided in an embodiment of the present invention.

[0024] Figure 8 (a) An exemplary simulation data imaging result without interference suppression provided in an embodiment of the present invention;

[0025] Figure 8 (b) An exemplary simulation data imaging result obtained using the frequency domain notch method provided in an embodiment of the present invention;

[0026] Figure 8 (c) An exemplary simulation data imaging result obtained using the U-Net method according to an embodiment of the present invention;

[0027] Figure 8 (d) is an exemplary simulation data imaging result obtained using the DPARNet method proposed in this invention, provided in an embodiment of the invention;

[0028] Figure 9 (a) Exemplary original admission data provided for embodiments of the present invention;

[0029] Figure 9 (b) An imaging result obtained by frequency domain notch filtering of exemplary raw data provided in an embodiment of the present invention;

[0030] Figure 9 (c) Imaging results of exemplary raw data provided in an embodiment of the present invention after passing through a U-Net network;

[0031] Figure 9 (d) Imaging results of exemplary raw data provided in an embodiment of the present invention processed by the DPARNet proposed in this invention. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0033] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0034] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0035] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, disclosure, and appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0036] Comb Spectrum Modulation Jamming (CSMJ) is the superposition of a series of narrowband interference signals within a specific bandwidth. The time-domain expression for CSMJ is: Among them, A k (t), f k and These represent the amplitude, carrier frequency, and phase of the k-th narrowband interference signal, respectively.

[0037] Figure 1 (a), 1(b), and 1(c) show the comparison results of the measured comb-spectrum interference characteristics in the time domain, time-frequency domain, and frequency domain, respectively. Figure 1 (a) Comparison of interference-free and interference-affected signals in the time domain. Figure 1 (b) Comparison of interference-free and interference-containing signals in the time-frequency domain. Figure 1 (c) Comparison of interference-free and interference-interfered signals in the frequency domain. It can be seen that the interference and target echo signals are heavily coupled in both the time and time-frequency domains, making it difficult to separate them using traditional methods. However, the interference spectrum is distributed in the frequency domain as equally spaced instantaneous peak pulses, with a maximum amplitude much higher than the target echo. Therefore, this significant difference between the interference and target signals makes them separable, providing a basis for further interference suppression.

[0038] Figure 2 This is a flowchart of a SAR comb spectrum interference suppression method based on a dual-channel attention residual network provided in an embodiment of the present invention, as follows: Figure 2 As shown, the method includes the following steps:

[0039] S101. Obtain the time-domain echo signal with comb spectrum interference.

[0040] Here, the acquired time-domain echo signal with comb-like spectrum interference can be one or multiple.

[0041] S102. Transform the time-domain echo signal to the frequency domain to obtain the frequency-domain echo signal with interference.

[0042] Here, Fourier transform can be used to transform from the time domain to the frequency domain.

[0043] S103. Extract the real and imaginary parts of the frequency domain echo signal with interference to obtain the real part signal with interference and the imaginary part signal with interference.

[0044] S104. Input the real part signal with interference and the imaginary part signal with interference into the pre-trained dual-channel attention residual network to suppress comb spectrum interference and output the frequency domain echo signal. The pre-trained dual-channel attention residual network includes two suppression networks with the same structure. Each suppression network includes a channel attention residual block and a global attention block. One suppression network is used to reconstruct the real part signal from the input real part signal with interference, and the other suppression network is used to reconstruct the imaginary part signal from the input imaginary part signal with interference.

[0045] In some embodiments, each suppression network may include: five channel attention residual blocks with identical structures and a global attention block; the five channel attention residual blocks are connected in series, and the first channel attention residual block has the same number of channels as the fifth channel attention residual block, the second channel attention residual block has the same number of channels as the fourth channel attention residual block, and the third channel attention residual block has the largest number of channels; the global attention block includes: two self-attention networks with identical structures connected in series.

[0046] Specifically, a global attention block can include: an intra-band self-attention network and an inter-band self-attention network; the intra-band self-attention network and the inter-band self-attention network are connected in series and have the same structure.

[0047] For example, the number of channels in the five channel attention residual blocks included in each suppression network are 32, 64, 128, 64, and 32 respectively; thus, by first increasing and then decreasing the number of channels, it is beneficial to refine the separation of echo signals and interference signals.

[0048] In some embodiments, each channel attention residual block sequentially includes: two first convolutional blocks, a second convolutional block, a channel attention block, and a feature combination block.

[0049] For example, Figure 3 A schematic diagram of the structure of the attention residual block for each channel, as shown below. Figure 3As shown, the first convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, and a Leaky ReLU activation function; the second convolutional block consists of a one-dimensional convolutional layer and a batch normalization layer. The feature combining block consists of an element-wise addition layer and a ReLU activation function, and the input of the feature combining block is the output of the channel attention block and the input of the first convolutional block. The number of convolutional kernels in the convolutional layers of the first and second convolutional blocks is the same as the number of output channels of the corresponding channel attention residual block, with a kernel size of 3, a stride of 1, and padding of 1. Assume the input of the channel attention residual block is F∈R B×C×L The output is F. 1 ∈R B×C×L Where B is the number of input features (i.e., the number of real or imaginary signals with interference input to the dual-channel attention residual network), and C and L represent the number of channels and length of each input feature (i.e., each sequence), respectively. The channel attention residual block can then be represented as: F 1 =CA(W c (F))+F; where W c The weights represent the weights of the convolutional block, and CA represents the weights of the channel attention block.

[0050] For example, Figure 4 This is a schematic diagram of the channel attention block structure, such as... Figure 4 As shown, the channel attention block includes an average pooling layer, a one-dimensional convolutional layer, a sigmoid activation function, and a dot product layer. The input to the dot product layer is the feature processed by the sigmoid activation function and the input to the average pooling layer. With negligible additional learning parameters and computational cost, the channel attention block can adaptively generate high-resolution channel features, thereby further improving network performance. Specifically, it first aggregates features using adaptive average pooling, and then uses a one-dimensional convolutional layer with a kernel size of k to extract the dependencies between k channels, achieving the effect of adaptively selecting important channel features. Assume the feature input to the channel attention block is F. (in) ∈R B×C×L The output is F (out) ∈R B×C×L The output of the channel attention block is: In the formula, σ(·) is the Sigmoid activation function, GAP(·) represents adaptive average pooling, and W C ∈R C×k represents the weights of a one-dimensional convolutional layer. k is an important adaptive parameter that varies depending on the number of channels C in the channel attention residual block. Larger k values ​​are better at capturing long-range dependencies, while smaller k values ​​are better at capturing short-range dependencies. The value of k is controlled by the number of channels C, and its calculation expression can be: C represents the number of channels, and b and γ are preset values ​​(for example, b and γ can be set to 1 and 2 respectively). The stride of the convolution in the channel attention block can be 1, and the padding can be (k-1) / 2 rounded down.

[0051] In some embodiments, each suppression network further includes a feature slicing layer, which is concatenated after the 5-channel attention residual block and before the global attention block. The feature slicing layer segments (partitions) the input along its length dimension L, and then concatenates the resulting sub-outputs to a first preset dimension to obtain the updated input. Assume the input is F∈R. B×C×L After segmenting along dimension L, the segments are concatenated to the new dimensions N and P, resulting in the expression: F∈R B×C×L →F′∈R B×C×P×N In the formula, N represents the number of frequency bands, and P represents the length of each frequency band. Here, considering that the computational complexity of subsequent self-attention calculations is proportional to the square of the sequence length, N can be approximated as... In this way, the computational complexity can be reduced from O(L) 2 ) downgraded to This minimizes the cost of attention calculations.

[0052] In some embodiments, each self-attention network (intra-frequency self-attention network or inter-frequency self-attention network) sequentially includes: a dimension reorganization layer, a position encoding layer, a global attention layer, and an addition operation layer.

[0053] For example, Figure 5 and 6 These are schematic diagrams of the intra-frequency self-attention network structure and feature processing principle, respectively. Figure 5 and 6 As shown, when the input of the intra-frequency self-attention network is F′∈R B×C×P×N First, F′∈R is reorganized through a dimension restructuring layer. B ×C×P×N The first dimension B and the fourth dimension N are merged to obtain the recombined input F″∈R. (B×N)×C×P The input for the recombination is B frequency band sequences (B is an integer greater than or equal to 1), each containing multiple frequency points. A positional encoding layer encodes the positional information of each frequency point in each frequency band sequence, resulting in a positional encoding vector for each frequency point. Positional encoding injects the positional information of each frequency point within the frequency band sequence, thereby enhancing the input to the attention model. For a frequency band sequence of length P, its positional encoding vector has a length of C. Let t represent the position of the frequency point in the sequence. Let represent the value of the encoding vector at frequency t at position i. The encoding rule is defined as follows: in q represents odd or even, and i is 0, 1, 2, ..., C-1. After obtaining the reconstructed input position encoding vector, the addition layer adds the position encoding vector to the reconstructed input to obtain the superimposed features, which are then used by the global attention layer for subsequent self-attention calculations.

[0054] The global attention layer consists of two normalization layers, two random deactivation layers, a self-attention computation layer, and a feedforward network layer. First, the stacked features are normalized by the first normalization layer to obtain normalized features. Then, the normalized features are input into the self-attention computation layer to calculate multiple independent attention masks (the number of attention heads can be up to 8). Specifically, the query vector, key vector, and value vector of each attention head are generated by a linear mapping of the input, and can be represented as follows: Among them, d k Let represent the dimensions of the query vector, key vector, and value vector, and h be the number of attention heads, where j = 1, 2, ..., h. Each attention head performs matrix multiplication on the query vector and key vector to obtain an attention mask, which is then multiplied by the value vector to obtain the feature output of the hybrid attention. Its calculation formula can be expressed as: Thus, the multi-head attention feature output is obtained. This output is then fed into the first random deactivation layer. The output of the first random deactivation layer is added to the superimposed features. The resulting feature is then fed into the second normalization layer. The output of the second normalization layer is then fed into the feedforward network layer. The feedforward network layer consists of, in order: a linear connection layer, a ReLU activation layer, a random deactivation layer, and another linear connection layer. The first linear connection layer magnifies the dimension of the input features by a factor of four, and the second linear connection layer restores the dimension to the input dimension. After obtaining the output of the feedforward network layer, it is fed into the second random deactivation layer. The output of the second random deactivation layer is then added element-wise to the input of the feedforward network layer. The resulting feature is then combined with the reconstructed input F″∈R. (B ×N)×C×P Perform the addition operation to restore the dimension to R. B×C×P×N The final output F″′ of the self-attention network is obtained.

[0055] Here, unlike intra-frequency self-attention networks, inter-frequency self-attention networks perform dimensional recombination by merging the first dimension B and the third dimension P. For example, to obtain the sequence input K∈R for calculating self-attention. (B×P)×C×NThe other processing principles are the same: first, positional information is added to the frequency band using positional encoding 'e'; then, layer normalization is used to normalize samples from the same batch; next, a multi-head attention mask is calculated and multiplied with the input feature map to obtain the attention output; finally, the output of the feedforward network layer is skipped and linked with the input before positional encoding to obtain the final output result. For example, the calculation processes for intra-band self-attention and inter-band self-attention can both be represented as follows:

[0056] K′=K+e;

[0057] K″=MultiHeadAttention(LayerNorm(K′))+K′;

[0058] K″′=FeedForward(LayerNorm(K″)+K″);

[0059] Y = K + K″′;

[0060] Where Y is the output of the self-attention network, and e is the position code.

[0061] In some embodiments, the above-described S104 can be implemented through the following steps:

[0062] S1041. Input the real part signal with interference into the first suppression network, and simultaneously input the imaginary part signal with interference into the second suppression network.

[0063] S1042. Each suppression network processes the input signal through five sequentially connected channel attention residual blocks with identical structures to obtain the first output.

[0064] Specifically, after the input signal enters the first channel attention residual block, it sequentially undergoes convolution, batch normalization, and first activation processing in two first convolution blocks, followed by convolution and batch normalization processing in the second convolution block to obtain the output of the second convolution block. The output of the second convolution block undergoes average pooling, convolution, and second activation processing in the channel attention block to obtain an intermediate output. The intermediate output and the output of the second convolution block are multiplied by the channel attention block to obtain the output of the channel attention block. The input and output of the first channel attention residual block are added by the feature combination block and then processed by the third activation processing to obtain the output of the first channel attention residual block. The output of the m-th channel attention residual block is processed by the (m+1)-th channel attention residual block to obtain the output of the (m+1)-th channel attention residual block, until the output of the 5th channel attention residual block is obtained. The output of the 5th channel attention residual block is used as the first output mentioned above. m is an integer from 1 to 4.

[0065] S1043. After segmenting the first output input along the length dimension, the sub-outputs obtained after segmentation are concatenated to the first preset dimension to obtain the updated first output.

[0066] For example, when the first output is F∈R B×C×L When the first output is updated, it is F′∈R B×C×P×N The first output includes B frequency band sequences, where B is the number of real or imaginary signals with interference input to the pre-trained dual-channel attention residual network; C is the number of channels in each frequency band sequence; L is the length dimension of each frequency band sequence; N is the number of frequency bands; and P is the length of each frequency band. N and P are the first preset dimensions. That is, the three dimensions of each frequency band sequence are C, P, and N.

[0067] S1044. Input the updated first output into the global attention block to obtain the second output, and use the second output as the signal for the suppression network output; wherein, the first suppression network output real part prediction signal and the second suppression network output imaginary part prediction signal.

[0068] Here, the global attention block includes: an intra-band self-attention network and an inter-band self-attention network with the same structure. After the updated first output is input into the global attention block, it is processed by the intra-band self-attention network to obtain the output of the intra-band self-attention network. The output of the intra-band self-attention network is processed by the inter-band self-attention network to obtain the second output.

[0069] Specifically, after updating the first output and inputting it into the intra-band self-attention network, a dimension recombination layer merges the second preset dimension (e.g., the first dimension B and the fourth dimension N) of the updated first output to obtain a recombined feature. The recombined feature is a frequency band sequence including multiple frequency points. A position encoding layer encodes each frequency point according to its position information in the frequency band sequence to obtain an encoded feature. An addition layer adds the recombined feature and the encoded feature to obtain a superimposed feature. After processing by a global attention layer, the superimposed feature yields a linear output. The addition layer adds the recombined feature and the linear output to obtain the output of the intra-band self-attention network. The inter-band self-attention network uses the same processing principle. After processing the output of the intra-band self-attention network, the output of the inter-band self-attention network, i.e., the output of the suppression network, is obtained. However, unlike the intra-band self-attention network, the inter-band self-attention network merges the first dimension B and the third dimension P during dimension recombination.

[0070] S1045. The real part prediction signal and the imaginary part prediction signal are fused to obtain the frequency domain echo signal.

[0071] For example, Figure 7This is a schematic diagram illustrating the processing principle of an input signal using a dual-channel attention residual network. Figure 7 As shown, after extracting the real and imaginary parts of the interference-laden frequency domain echo signal, the interference-laden real and imaginary signals are simultaneously input into two channels (two suppression networks) of the network. For example, taking the interference-laden real signal as an example, it is processed sequentially through five channel attention residual blocks, then through a feature slice layer, and finally into an intra-band self-attention network. The output of the intra-band self-attention network is then processed by an inter-band self-attention network, and the predicted real part signal is output. Similarly, for the other channel that inputs the interference-laden imaginary signal, the predicted imaginary part signal is output. Finally, the predicted real and imaginary signals are fused to obtain the predicted frequency domain echo signal, thus eliminating comb spectrum interference.

[0072] In some embodiments, prior to S104, the method includes:

[0073] S001. Acquire multiple training samples; each training sample is an interference-enabled frequency domain echo sample signal; the interference-enabled frequency domain echo sample signal corresponds to an interference-free frequency domain echo sample signal.

[0074] Specifically, a comb-shaped spectrum interference time-domain signal can be simulated and superimposed with the measured target echo signal. The interference spectrum is then obtained through Fourier transform and normalized, serving as a benchmark dataset for interference suppression.

[0075] S002. Select multiple training samples each time, and determine the imaginary part and real part of each selected training sample to obtain the imaginary part sample signal and the real part sample signal of each training sample.

[0076] S003. Input the imaginary part and real part of each training sample selected this time into the dual-channel attention residual network trained last time to obtain the predicted real part and predicted imaginary part of each training sample selected this time.

[0077] S004. Based on the predicted real and predicted imaginary signals corresponding to each training sample selected this time, and the real and imaginary signals of the interference-free frequency domain echo sample signals corresponding to the multiple training samples selected this time, determine the loss value for this time.

[0078] Specifically, when selecting B training samples, for each training sample, calculate the first loss between the real part of the predicted signal corresponding to the training sample and the real part of the frequency domain echo sample signal without interference, and calculate the second loss between the imaginary part of the predicted signal corresponding to the training sample and the imaginary part of the frequency domain echo sample signal without interference. The sum of the first loss and the second loss is taken as the loss corresponding to the training sample, resulting in B losses for the B training samples. The sum of the B losses is taken as the loss value for this test.

[0079] S005. Based on the current loss value, adjust the network parameters of the previously trained dual-channel attention residual network to obtain the current trained dual-channel attention residual network. This process is iterated until the preset conditions are met, at which point training stops, resulting in a pre-trained dual-channel attention residual network.

[0080] For example, the formula for calculating the loss for each training sample is as follows:

[0081] L z =(f(x) z1 )-y z1 ) 2 +(f(x z2 )-y z2 ) 2 ;

[0082] Where z is 1, 2, ..., B, f(x) z1 Let y be the predicted real part of the signal corresponding to the z-th training sample. z1 Let f(x) be the real part of the interference-free frequency domain echo sample signal corresponding to the z-th training sample. z2 Let y be the predicted imaginary part signal corresponding to the z-th training sample. z2 Let be the imaginary part of the frequency domain echo sample signal without interference corresponding to the z-th training sample.

[0083] This invention first extracts the real and imaginary parts of the one-dimensional frequency domain signal input to the network. Two identical suppression networks are then used to recover the real and imaginary parts respectively. This effectively separates the characteristics of the target echo signal and the interference signal, achieving suppression, while preserving the phase information of the target echo during suppression. Furthermore, the use of channel attention residual blocks for feature processing enriches semantic information through inter-channel information interaction, enhancing the model's feature extraction capabilities. Subsequently, global attention blocks are used for further feature processing, allowing for the correction of the input feature map using global information and recovering information lost in the previous stage, thereby enhancing the model's representation and learning abilities.

[0084] This invention utilizes a SAR comb spectrum interference suppression framework based on a dual-channel attention residual network to solve the problem of severe target signal loss in existing technologies when suppressing comb spectrum interference. Furthermore, this invention projects the heavily coupled interference and target echo signals to a high-dimensional separable feature space through a deep nonlinear network, and introduces channel attention and global attention to improve the accuracy of feature extraction and representation, thereby achieving intelligent decoupling between interference and target echo. The interference suppression effect is significantly better than existing interference suppression methods.

[0085] The following experiments further illustrate the effects achieved by the embodiments of the present invention.

[0086] I. Dataset

[0087] To evaluate the performance of the interference suppression method proposed in this invention, comb spectrum interference was first simulated using the MATLAB software platform. The number of interference frequency points was set between 100 and 300, and the number of echo points was 16384. Then, the simulated interference echo was generated by superimposing it with the measured SAR echo signal. The interference spectrum was obtained by Fourier transform, and a total of 10000 samples were generated.

[0088] The simulated test set uses real-world SAR echo signals, different from the training set, superimposed with simulated comb-spectrum interference. The data is then generated through Fourier transform and normalization. The number of interference frequencies is set to approximately 600, resulting in 1410 samples, each with a size of 1×16384. The real-world test set, on the other hand, is derived from actual measured SAR interference data.

[0089] II. Implementation Details

[0090] (1) Experimental conditions

[0091] The hardware platform for the simulation experiment of this invention is as follows: CPU is AMD Ryzen 9 5900HX Radeon Graphics, 16 cores, with a main frequency of 3.30GHz, and memory size is 32GB; video memory size is 16GB.

[0092] The target echo signals of the simulation experiment training sample set and test sample set of this invention are obtained from the measured interference-free echo data under different times and scenarios in the TopSAR mode recorded by the Sentinel-1 satellite; the measured test sample set is the measured interference echo data in the TopSAR mode recorded by the Sentinel-1 satellite.

[0093] (2) This invention is implemented in PyTorch, using the AdamW optimizer and StepLR learning rate update strategy for 200 training rounds. The learning rate is set to 0.001, the update step size and update rate are 50 and 0.1 respectively, and the batch size is 8. The data is constructed as training pairs consisting of target echo signals and echo signals containing interference. After each training round, the training is monitored and parameters are tuned using the validation set, and the optimal model with the minimum mean square loss on the validation set is saved. Finally, the interference suppression effect and target echo reconstruction accuracy of the model are evaluated on the test set using SAR imaging quality.

[0094] (3) Evaluation indicators

[0095] To verify the effectiveness of the proposed method, quantitative and qualitative analyses were performed on the interference suppression effects of different WBI suppression methods. For qualitative evaluation, we can directly compare the recovered target echo signal SAR imaging results. Furthermore, we used three Image Quality Assessment (IQA) metrics to quantitatively evaluate the SAR imaging results. The three IQA metrics include Multiplicative Noise Ratio (MNR), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity (SSIM).

[0096] 1) Multiplicative Noise Ratio (MNR)

[0097] Multiplicative Noise Ratio (MNR) is defined as the ratio of the average image intensity of weakly scattering regions to that of strongly scattering regions. Weakly scattering regions are areas with virtually no echo, such as very smooth areas like a calm river, while strongly scattering regions are such as cities. The MNR expression is defined as:

[0098] Where M and N represent the number of pixels in the weak scattering region and the strong scattering region, respectively, and I n and I m These represent the pixel values ​​in the weakly scattering region and the strongly scattering region, respectively. The smaller the MNR value, the better the contrast of the recovered image.

[0099] 2) Peak Signal-to-Noise Ratio (PSNR)

[0100] Peak Signal-to-Noise Ratio (PSNR) is defined as the ratio of the sum of the squares of the maximum gray values ​​in an image to the mean square error between the reconstructed image and the original image. A higher PSNR indicates less image distortion and higher quality reconstructed image. Its expression is defined as:

[0101]

[0102] 3) Structural Similarity (SSIM)

[0103] Structural similarity is a metric for measuring the similarity between two images. It detects changes in structural information to perceive approximate information about image distortion. Its expression is defined as:

[0104] Where, μ x μ y , and σ xy c1 and c2 represent the mean, variance, and covariance of the restored and target images, respectively. c1 and c2 are two very small constants to avoid zero denominators. A larger SSIM value indicates less image distortion and greater similarity between the two images.

[0105] III. Experimental Results

[0106] To highlight the performance of the proposed method, the interference suppression effects of the proposed method (DPARNet), frequency domain-notched filtering (Fnotch), and U-Net network are compared, and evaluated from both qualitative and quantitative perspectives. Imaging results are compared as follows: Figure 8 As shown in (a) to 8(d), the results of the interference suppression effectiveness evaluation are shown in Table 1.

[0107]

[0108]

[0109] Table 1

[0110] Figure 8 (a) Shows the simulation data imaging results without interference suppression, where the white shadows completely cover the scene information due to the influence of dense comb-like spectral interference. Figure 8 (b) Shows the results obtained using the frequency domain notch method, where signal loss leads to severe degradation of image quality. Figure 8 (c) and Figure 8(d) Images obtained using U-Net and the proposed DPARNet are shown respectively. These two methods exhibit similar performance, producing high-quality images, effectively distinguishing features such as cities, rivers, and farmland. However, U-Net's suppression results still retain white artifacts. Comparison of the white rectangular frames reveals that the SAR imaging results obtained by the proposed DPARNet show better contrast and clearer edges. Table 1 quantitatively presents the comparison results of the proposed DPARNet with traditional frequency-domain notch filtering and U-Net frequency-domain anti-interference effects. It can be seen that DPARNet outperforms the comparison algorithms in MNR, PSNR, and SSIM, and compared to the traditional frequency-domain notch filtering method, all indicators are significantly improved: MNR is increased by 7.37 dB, PSNR by 18.41 dB, and SSIM by 0.52 dB.

[0111] This section uses measured data to verify the portability of the DPARNet proposed in this invention. The data comes from interference echo data in TopSAR mode collected by the Sentinel-1 satellite in some regions. Figure 9 (a) shows the original admission data. The scene in the figure has very little visible scene information, and most areas are covered by bright lines, resulting in the loss of a lot of useful information. Figure 9 (b) is the image obtained by frequency domain notch filtering of the original data. Due to severe signal distortion, the image is out of focus and almost no information can be deciphered from the SAR image. Figure 9 (c) shows the imaging results of the original data after passing through the U-Net network. Most of the interference is suppressed, but there are still obvious artifacts caused by residual interference. Figure 9 (d) shows the imaging results of the DPARNet proposed in this invention. Compared with the results of other algorithms, it effectively recovers the target echo signal from the interference, and the generated SAR imaging results have higher signal-to-noise ratio and contrast. Further quantitative analysis of the results is shown in Table 2.

[0112] Models MNR↓(dB) PSNR↑(dB) SSIM↑(dB) Fnotch -1.80 26.23 0.86 UNet -3.16 27.96 0.84 DPARNet -7.30 30.25 0.91 Improvement (%) 305 / 131 15.32 / 8.19 5.81 / 8.33

[0113] Table 2

[0114] Table 2 shows that the proposed DPARNet significantly improves the MNR, PSNR, and SSIM metrics compared to other comparative algorithms. The MNR is improved by 5.5 dB and 4.14 dB compared to the frequency domain notch filter and U-Net network, respectively; the PSNR is improved by 4.02 dB and 2.39 dB compared to the frequency domain notch filter and U-Net network, respectively; and the SSIM is improved by 0.05 and 0.07. The results demonstrate that the proposed method can be extended to measured interference data, exhibiting good interference suppression and transfer performance.

[0115] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A SAR comb-SPR interference suppression method based on a dual-channel attention residual network, characterized in that, include: Acquire time-domain echo signals with comb-spectral interference; The time-domain echo signal is transformed to the frequency domain to obtain an interference-laden frequency-domain echo signal; The real and imaginary parts of the interference frequency domain echo signal are extracted to obtain the interference real part signal and the interference imaginary part signal. The real part signal with interference and the imaginary part signal with interference are input into a pre-trained dual-channel attention residual network to suppress comb spectrum interference and output frequency domain echo signal. The pre-trained dual-channel attention residual network includes two identical suppression networks, each of which includes a channel attention residual block and a global attention block; wherein, one suppression network is used to reconstruct the real part signal from the input signal with interference, and the other suppression network is used to reconstruct the imaginary part signal from the input signal with interference. Each suppression network includes: five channel attention residual blocks with identical structures and one global attention block; the five channel attention residual blocks are connected in series, and the first channel attention residual block and the fifth channel attention residual block have the same number of channels, the second channel attention residual block and the fourth channel attention residual block have the same number of channels, and the third channel attention residual block has the largest number of channels; the global attention block includes: an intra-band self-attention network and an inter-band self-attention network; the intra-band self-attention network and the inter-band self-attention network are connected in series and have identical structures.

2. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 1, characterized in that, The step of inputting the real part signal with interference and the imaginary part signal with interference into a pre-trained dual-channel attention residual network to suppress comb spectrum interference and output a frequency domain echo signal includes: The signal with the real part of the interference is input into the first suppression network, and the signal with the imaginary part of the interference is input into the second suppression network. Each suppression network processes the input signal through five sequentially connected, identically structured channel attention residual blocks to obtain the first output; After segmenting the first output input along the length dimension, the sub-outputs obtained after segmentation are concatenated to the first preset dimension to obtain the updated first output; The updated first output is input into the global attention block to obtain a second output, which is then used as the signal output by the suppression network; wherein the first suppression network outputs a real part prediction signal, and the second suppression network outputs an imaginary part prediction signal. The real part prediction signal and the imaginary part prediction signal are fused to obtain the frequency domain echo signal.

3. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 2, characterized in that, Each attention residual block sequentially includes: two first convolutional blocks, a second convolutional block, a channel attention block, and a feature binding block; each suppression network processes the input signal through five sequentially connected and structurally identical channel attention residual blocks to obtain a first output, including: After the input signal enters the first channel attention residual block, it goes through the convolution, batch normalization and first activation processing of the two first convolution blocks in sequence, and then goes through the convolution and batch normalization processing of the second convolution block to obtain the output of the second convolution block. The output of the second convolutional block is processed by average pooling, convolution, and second activation of the channel attention block to obtain an intermediate output. The intermediate output and the output of the second convolutional block are multiplied by the channel attention block to obtain the output of the channel attention block. The input and output of the first channel attention residual block are added by the feature combining block and then processed by the third activation process to obtain the output of the first channel attention residual block. The output of the m-th channel attention residual block is processed by the (m+1)-th channel attention residual block to obtain the output of the (m+1)-th channel attention residual block, until the output of the 5th channel attention residual block is obtained. The output of the 5th channel attention residual block is used as the first output; m is an integer from 1 to 4.

4. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 2, characterized in that, The first output is The updated first output is The first output includes Frequency band sequence, The number of the real part signal with interference or the imaginary part signal with interference input to the pre-trained dual-channel attention residual network; The number of channels for each frequency band sequence. The length dimension of each frequency band sequence, For the number of frequency bands, Frequency band length; and This is the first preset dimension.

5. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 2, characterized in that, The step of inputting the updated first output into the global attention block to obtain the second output includes: After the updated first output is input into the global attention block, it is processed by the self-attention network within the frequency band to obtain the output of the self-attention network within the frequency band. The output of the intra-band self-attention network is processed by the inter-band self-attention network to obtain the second output.

6. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 5, characterized in that, Each self-attention network sequentially includes: a dimension reorganization layer, a position encoding layer, a global attention layer, and an addition operation layer; after the updated first output is input into the global attention block, it is processed by the self-attention network within the frequency band to obtain the output of the self-attention network within the frequency band, including: After the updated first output is input into the self-attention network within the frequency band, the second preset dimension of the updated first output is merged through a dimension recombination layer to obtain a recombined feature; the recombined feature is a frequency band sequence including multiple frequency points; The position coding layer performs position coding on each frequency point according to its position information in the frequency band sequence to obtain coding features; The addition layer adds the recombined feature to the encoded feature to obtain the superimposed feature; The superimposed features are processed by the global attention layer to obtain a linear output; The addition layer adds the recombined features to the linear output to obtain the output of the self-attention network within the frequency band.

7. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 1, characterized in that, Before inputting the interfering real part signal and the interfering imaginary part signal into a pre-trained dual-channel attention residual network to suppress comb-spectrum interference and output a frequency domain echo signal, the method includes: Multiple training samples are acquired; each training sample is an interference-interfered frequency domain echo sample signal; the interference-interfered frequency domain echo sample signal corresponds to an interference-free frequency domain echo sample signal. Multiple training samples are selected each time, and the imaginary and real parts of each selected training sample are determined to obtain the imaginary part sample signal and the real part sample signal of each training sample. The imaginary part and real part of each training sample selected this time are input into the dual-channel attention residual network trained last time to obtain the predicted real part and predicted imaginary part of each training sample selected this time. The loss value is determined based on the predicted real and imaginary part signals corresponding to each selected training sample, as well as the real and imaginary part signals of the interference-free frequency domain echo sample signals corresponding to multiple selected training samples. Based on the current loss value, adjust the network parameters of the previously trained dual-channel attention residual network to obtain the current trained dual-channel attention residual network. This process is iterated until a preset condition is met, at which point training stops, resulting in the pre-trained dual-channel attention residual network.

8. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 7, characterized in that, The step of determining the current loss value based on the predicted real and imaginary part signals corresponding to each selected training sample, and the real and imaginary part signals of the interference-free frequency domain echo sample signals corresponding to the multiple selected training samples, includes: When selecting B training samples, for each training sample, calculate the first loss between the real part of the predicted signal corresponding to the training sample and the real part of the frequency domain echo sample signal without interference corresponding to the training sample, and calculate the second loss between the imaginary part of the predicted signal corresponding to the training sample and the imaginary part of the frequency domain echo sample signal without interference corresponding to the training sample; B is an integer greater than 1. The sum of the first loss and the second loss is used as the loss corresponding to the training sample, resulting in B losses corresponding to B training samples. The sum of the B losses is taken as the current loss value.

9. The SAR comb spectrum interference suppression method based on a dual-channel attention residual network according to claim 8, characterized in that, The formula for calculating the loss for each training sample is as follows: ； in, for , For the first The predicted real part signal corresponding to each training sample. For the first The real part of the frequency domain echo sample signal corresponding to each training sample without interference. For the first The predicted imaginary part signal corresponding to each training sample. For the first The imaginary part of the frequency domain echo sample signal without interference corresponding to each training sample.

Images