Interference suppression method and device based on multi-scale implicit neural representation network

Abstract

The invention discloses an interference suppression method and device based on a multi-scale implicit neural representation network, and the method comprises the steps: inputting an original time-frequency graph corresponding to a radar echo signal into a pre-trained target network model, carrying out the down-sampling of the original time-frequency graph through the target network model, and obtaining at least one sub-time-frequency graph, and performing feature extraction on the at least one sub-time-frequency graph to obtain a corresponding implicit neural representation feature, then performing interference separation based on the original time-frequency graph and the implicit neural representation feature, and outputting a target signal. According to the method, the end-to-end neural network model is utilized to perform joint processing on the original time-frequency graphs and the sub-time-frequency graphs of different scales, and the implicit neural representation features corresponding to the sub-time-frequency graphs are fused, so that the potential features of the target signal can be fully mined, the decoupling performance of the target signal and the interference signal is improved, and the loss of useful signals is avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and in particular relates to an interference suppression method and device based on a multi-scale implicit neural representation network. Background Art

[0002] Synthetic Aperture Radar (SAR), as an advanced remote sensing technology, plays an irreplaceable role in a variety of fields, including geological disaster relief, crop surveys, and military reconnaissance. However, with the advent of the information age and the vigorous development of military electronic countermeasures equipment by various countries, the widespread use of various electronic devices and jammers has led to an increasingly complex electromagnetic environment. This also means that SAR is increasingly subject to co-frequency active interference. The presence of active interference poses a serious threat to the survival environment of SAR, not only reducing military reconnaissance and precision strike capabilities but also severely impacting civilian remote sensing applications such as geological monitoring and disaster assessment.

[0003] In the prior art, interference suppression algorithms for SAR have poor decoupling when faced with complex interference types and strong coupling between interference and signals, and are prone to causing loss of useful signals. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides an interference suppression method and device based on a multi-scale implicit neural representation network.

[0005] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] In a first aspect, the present invention provides an interference suppression method based on a multi-scale implicit neural representation network, comprising:

[0007] Acquiring a radar echo signal for detecting a target object, where the radar echo signal includes an interference signal and a target signal;

[0008] The original time-frequency graph corresponding to the radar echo signal is input into a pre-trained target network model. The original time-frequency graph is downsampled by the target network model to obtain at least one sub-time-frequency graph. Feature extraction is performed on the at least one sub-time-frequency graph to obtain the corresponding implicit neural representation feature. Interference separation is performed based on the original time-frequency graph and the implicit neural representation feature, and the target signal is output.

[0009] In a second aspect, the present invention provides an interference suppression device based on a multi-scale implicit neural representation network, comprising:

[0010] An acquisition module is used to acquire a radar echo signal for detecting a target object, where the radar echo signal includes an interference signal and a target signal;

[0011] The prediction module is used to input the original time-frequency map corresponding to the radar echo signal into a pre-trained target network model, downsample the original time-frequency map through the target network model to obtain at least one sub-time-frequency map, extract features from the at least one sub-time-frequency map to obtain the corresponding implicit neural representation features, perform interference separation based on the original time-frequency map and the implicit neural representation features, and output the target signal.

[0012] The present invention provides an interference suppression method and device based on a multi-scale implicit neural representation network. First, the original time-frequency graph corresponding to the radar echo signal is input into a pre-trained target network model. The original time-frequency graph is downsampled by the target network model to obtain at least one sub-time-frequency graph. Feature extraction is then performed on the at least one sub-time-frequency graph to obtain the corresponding implicit neural representation feature. After that, interference separation is performed based on the original time-frequency graph and the implicit neural representation feature, and the target signal is output. The present invention utilizes an end-to-end neural network model to jointly process the original time-frequency graph and the sub-time-frequency graphs of different scales, and fuses the implicit neural representation features corresponding to the sub-time-frequency graphs. This can fully tap the potential features of the target signal, thereby improving the decoupling of the target signal and the interference signal and avoiding the loss of useful signals.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 1 is a flow chart of an interference suppression method based on a multi-scale implicit neural representation network provided by an embodiment of the present invention;

[0015] Figure 2A Schematic diagram of the prediction process of a target network model in an interference suppression method based on a multi-scale implicit neural representation network provided by an embodiment of the present invention;

[0016] Figure 2B 1 is a schematic diagram of a processing process of an implicit neural representation network in an interference suppression method based on a multi-scale implicit neural representation network provided by an embodiment of the present invention;

[0017] Figure 3 Schematic diagram of SAR interference suppression comparison results of an ablation experiment according to an embodiment of the present invention;

[0018] Figure 4 A comparison diagram of imaging results before and after anti-interference for different JNR suppression interferences using different interference suppression algorithms in an embodiment of the present invention;

[0019] Figure 5 FIG1 is a schematic diagram showing a comparison of suppression results of different interference suppression algorithms under interference suppression conditions according to an exemplary embodiment of the present invention;

[0020] Figure 6FIG1 is another schematic diagram showing a comparison of suppression results of different interference suppression algorithms under interference suppression conditions according to another exemplary embodiment of the present invention;

[0021] Figure 7 FIG1 is another schematic diagram showing a comparison of suppression results of different interference suppression algorithms under interference suppression conditions according to another exemplary embodiment of the present invention;

[0022] Figure 8 Schematic diagram of how each interference suppression evaluation index changes with the interference suppression JNR in an embodiment of the present invention;

[0023] Figure 9 This is a comparison diagram of imaging results before and after anti-interference of different interference suppression algorithms for different JNR deception interferences in an embodiment of the present invention;

[0024] Figure 10 FIG1 is a schematic diagram showing a comparison of suppression results of different interference suppression algorithms under deceptive interference conditions according to an exemplary embodiment of the present invention;

[0025] Figure 11 FIG1 is another schematic diagram comparing suppression results of different interference suppression algorithms under deceptive interference conditions according to the present invention;

[0026] Figure 12 FIG1 is another schematic diagram comparing suppression results of different interference suppression algorithms under deceptive interference conditions according to the present invention;

[0027] Figure 13 Schematic diagram of how each interference suppression evaluation index changes with deceptive interference JNR in an embodiment of the present invention;

[0028] Figure 14 It is a structural schematic diagram of an interference suppression device based on a multi-scale implicit neural representation network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] 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.

[0030] The embodiment of the present invention provides an interference suppression method based on a multi-scale implicit neural representation network. Figure 1 , the method comprises the following steps:

[0031] S10: Acquire a radar echo signal for detecting a target object.

[0032] Among them, the radar echo signal includes interference signal and target signal.

[0033] Specifically, the radar echo signal may be a SAR echo signal, which is a time domain signal.

[0034] S20. Input the original time-frequency graph corresponding to the radar echo signal into a pre-trained target network model, downsample the original time-frequency graph through the target network model to obtain at least one sub-time-frequency graph, extract features from the at least one sub-time-frequency graph to obtain corresponding implicit neural representation features, perform interference separation based on the original time-frequency graph and the implicit neural representation features, and output the target signal.

[0035] Specifically, a short-time Fourier transform (STFT) may be used to perform time-frequency domain conversion on the radar echo signal to obtain an original time-frequency graph corresponding to the radar echo signal. The size of the original time-frequency graph may be 128×128.

[0036] Here, data normalization is required to eliminate the impact of the different dimensions of different evaluation indicators on the data analysis results. Because the amplitude of the interference signal is much higher than that of the target signal, the commonly used maximum and minimum value normalization method easily causes the normalized interference signal result to approach 1, while the target signal normalized result approaches 0. This leads to a large gap between the normalized interference signal and target signal results. This data distribution method is not conducive to the convergence of the target network model.

[0037] To this end, this embodiment may process the original time-frequency graph using a logarithmic normalization method, and its mathematical expression is as follows:

[0038]

[0039] Where x represents the amplitude in the original time-frequency graph, and x' represents the amplitude after normalization of the original time-frequency graph.

[0040] Afterwards, the normalized result of the original time-frequency graph (denoted as ) Input the target network model, the target network model downsamples the original time-frequency graph by bilinear interpolation. For example, the downsampling scale can be 1 / 2 and 1 / 4, and two sub-time-frequency graphs can be obtained, namely the first sub-time-frequency graph (1 / 2 size) ) and the second sub-spectrogram (1 / 4 size ).

[0041] Reference Figure 2A Taking the first sub-frequency graph and the second sub-frequency graph obtained by downsampling as an example, the target network model can include a first implicit neural representation network, a second implicit neural representation network, a self-attention network corresponding to the original frequency graph, a first self-attention network, a second self-attention network and a convolution module.

[0042] First, the first implicit neural representation network extracts features from the first sub-spectrogram to obtain the corresponding implicit neural representation features, and the second implicit neural representation network extracts features from the second sub-spectrogram to obtain the corresponding implicit neural representation features. Specifically, the first implicit neural representation network and the second implicit neural representation network are applied to the 1 / 4-scale sub-spectrogram and the 1 / 2-scale sub-spectrogram, respectively, creating a coarse image feature representation and a fine image feature representation. This sequential prediction from coarse to fine effectively transfers radar echo signal information and enables information sharing across different scales.

[0043] Optionally, in step S20, feature extraction is performed on at least one sub-spectrogram to obtain a plurality of implicit neural representation features, which may specifically include:

[0044] S201 : Extract coordinates from each sub-spectrogram and encode the extracted coordinate data to obtain coordinate encoding features corresponding to each sub-spectrogram.

[0045] Specifically, the processing procedures of the first implicit neural representation network and the second implicit neural representation network are the same. Here, the processing procedure of the first implicit neural representation network is taken as an example for explanation.

[0046] Reference Figure 2B , the first implicit neural representation network extracts the coordinates of the first sub-time-frequency graph and encodes the extracted coordinate data to obtain the coordinate encoding features corresponding to the first sub-time-frequency graph, which is recorded as Where H and W are the sizes of the first sub-spectrogram, and the number of channels 2 represents the horizontal and vertical coordinates.

[0047] S202: Encode each sub-spectrogram to obtain image coding features of each sub-spectrogram.

[0048] Specifically, refer to Figure 2B , the first sub-spectrogram is encoded by the encoder to obtain the image coding feature of the first sub-spectrogram Where C represents the number of channels.

[0049] Here, steps S201 and S202 may be performed simultaneously.

[0050] S203 , performing high-dimensional space encoding on the coordinate coding features and image coding features corresponding to each sub-spectrogram to obtain a feature sequence corresponding to each sub-spectrogram.

[0051] Specifically, refer to Figure 2B , the coordinate coding features and image coding features corresponding to the first sub-time-frequency graph can be respectively encoded in high-dimensional space using periodic space coding, and the coding results can be spliced ​​to obtain the feature sequence corresponding to the first sub-time-frequency graph.

[0052] For example, the process of encoding the coordinate encoding features corresponding to the first sub-frequency spectrum is as follows:

[0053] M′=[sin(y),cos(y),…sin(2 L-1 y),cos(2 L-1 y)];

[0054] in, The coordinates in the coordinate encoding feature are normalized in the interval [-1, 1]. L is a pre-set hyperparameter for determining the dimension value, and L can be set to 4.

[0055] S204: Perform multi-layer perceptron network decoding on the feature sequence corresponding to each sub-frequency graph to obtain the implicit neural representation feature corresponding to each sub-frequency graph.

[0056] Specifically, refer to Figure 2B A multilayer perceptron can be used to decode the feature sequence corresponding to the first sub-spectrogram to obtain an implicit neural representation feature corresponding to the first sub-spectrogram. The implicit neural representation feature can be an implicit neural representation feature image. The multilayer perceptron includes three layers, each layer including 256 hidden dimensions.

[0057] The process of multi-layer perceptron outputting feature sequences into implicit neural representation feature images can be regarded as an implicit neural interpolation process, which can be expressed as:

[0058]

[0059] Among them, s is the gray value of the implicit neural representation feature image, are four adjacent coordinate sets around y, k represents the coordinate set index, ω k is the weight of the coordinate set, and satisfies Represents an implicit function.

[0060] It can be expressed as:

[0061]

[0062] in, is the spatial coordinate point of each pixel corresponding to the input sub-frequency graph, is the grayscale value corresponding to the coordinate point.

[0063] Optionally, if the at least one sub-spectrogram includes a first sub-spectrogram, the target network model may include a self-attention network corresponding to the original spectrogram and a first self-attention network corresponding to the first sub-spectrogram.

[0064] In this regard, in step S20, interference separation is performed based on the original time-frequency graph and the implicit neural representation features, and the target signal is output, which may specifically include:

[0065] S205 , extracting original features of the original time-frequency graph and fusing them with implicit neural representation features corresponding to the first sub-time-frequency graph to obtain a first fused feature.

[0066] Specifically, refer to Figure 2A , extracting original features from the input original time-frequency graph and fusing them with the implicit neural representation features corresponding to the first sub-time-frequency graph to obtain a first fused feature. The first fused feature can then be convolved using a convolution module and input into the self-attention network corresponding to the original time-frequency graph.

[0067] S206. Use the self-attention network corresponding to the original time-frequency graph to process the first fusion feature to obtain a first output feature.

[0068] Specifically, refer to Figure 2A The self-attention network corresponding to the original time-frequency graph can include multiple sub-network layers, for example, 3, each sub-network layer includes a decoder, a bottleneck layer and a decoder, wherein the encoder is responsible for extracting the high-dimensional features of the interference signal and the target signal in the first fusion feature, the bottleneck layer performs dimensionality compression on the high-dimensional features, and the decoder performs feature recovery on the interference signal and the target signal, thereby separating the coupled interference signal and the target signal, and obtaining the first output feature corresponding to the original time-frequency graph after interference suppression.

[0069] S208. Use the first self-attention network to process the first sub-time-frequency graph to obtain a second output feature.

[0070] Specifically, refer to Figure 2A The first self-attention network corresponding to the first time-frequency graph can include multiple sub-network layers, for example, two. The number of sub-network layers in the first self-attention network is less than the number of sub-network layers in the self-attention network corresponding to the original time-frequency graph. By processing the first sub-time-frequency graph using the first self-attention network, a second output feature corresponding to the first sub-time-frequency graph after interference suppression can be obtained.

[0071] S211 , fusing the first output feature and the second output feature to obtain a second fused feature.

[0072] S212: Convolve the second fusion feature and output the target signal.

[0073] Specifically, refer to Figure 2AThe convolution module is used to convolve the second fusion feature and output the target signal. Here, the time-frequency map corresponding to the target signal can be output, and then the time-frequency map corresponding to the target signal is denormalized to restore the target signal time-frequency map. The target signal time-frequency map is then converted into the time domain to obtain the target signal.

[0074] In this embodiment, the implicit neural representation features corresponding to the first sub-spectrogram are integrated into the original features of the original spectrogram, and the information correlation between the pixels of the spectrogram can be used to repair the features of the target signal that were damaged during the feature separation process.

[0075] Optionally, if the at least one sub-spectrogram further includes a second sub-spectrogram, the target network model further includes a second self-attention network corresponding to the second sub-spectrogram. Here, the size of the second sub-spectrogram is smaller than the size of the first sub-spectrogram.

[0076] To this end, before step S208, the method of this embodiment further includes:

[0077] S207 , extracting a first feature of the first sub-spectrogram and fusing it with an implicit neural representation feature corresponding to the second sub-spectrogram to obtain a third fused feature.

[0078] Then, in step S208, the first self-attention network is used to process the first sub-spectrogram to obtain a second output feature, which may specifically include:

[0079] The third fusion feature is processed using the first self-attention network to obtain the second output feature.

[0080] Specifically, refer to Figure 2A In the network structure of this embodiment, the corresponding self-attention network and implicit neural representation feature network can be added according to the downsampling scale setting. In the above embodiment, taking the first sub-time-frequency graph obtained by downsampling as an example, the target network model may include the self-attention network, the first self-attention network and the first implicit neural representation feature network of the original time-frequency graph. If the sampling also obtains the second sub-time-frequency graph, the target network model also includes a second self-attention network and a second implicit neural representation feature network. Among them, the second implicit neural representation feature network has the same structure as the first implicit neural representation feature network, and the second self-attention network may include multiple sub-network layers, for example, 1, and the number of sub-network layers in the second self-attention network is less than the number of sub-network layers in the first self-attention network.

[0081] Accordingly, refer to Figure 2ASimilar to the process of processing the original spectrogram, during the processing of the first sub-spectrogram, the first feature of the first sub-spectrogram can be extracted and fused with the implicit neural representation feature corresponding to the second sub-spectrogram output by the second implicit neural representation network to obtain a third fused feature. Subsequently, the third fused feature is processed using the first self-attention network to obtain the second output feature corresponding to the first sub-spectrogram after interference suppression.

[0082] Optionally, in step S208, the first sub-spectrogram is processed using the first self-attention network to obtain a second output feature, which may specifically include:

[0083] S2081. Obtain the first bottleneck layer feature obtained in the process of processing the first fusion feature by the self-attention network corresponding to the original time-frequency graph.

[0084] S2082. Use the first self-attention network to process the first fusion feature and the first bottleneck layer feature to obtain a second output feature.

[0085] Specifically, refer to Figure 2A To avoid the problem where inaccurate coarse-scale feature estimation affects subsequent-scale feature estimation, this embodiment introduces a feedback branch in the bottleneck layer of each sub-network layer during the multi-scale transformation process to achieve backpropagation of features from fine to coarse. Specifically, the first bottleneck layer features obtained by processing the first fused features in the self-attention network corresponding to the original time-frequency graph are backpropagated to the bottleneck layer of the first self-attention network.

[0086] Feature backpropagation can use information at a finer scale to help recover features at the current scale and enhance the robustness of feature recovery.

[0087] In the process of back propagation, each feedback branch fuses the two finer-scale bottleneck layer features to obtain the fused bottleneck layer features, and then propagates the fused bottleneck layer features to the bottleneck layer in the coarser-scale self-attention network. For example, the two first bottleneck layer features F a and F b Fusion is performed to obtain the first bottleneck layer feature F after fusion out , and then the fused first bottleneck layer feature F out The feature is propagated to the bottleneck layer in the coarser-sized first self-attention network, and the first fusion feature and the fused first bottleneck layer feature are processed by the first self-attention network to obtain the second output feature.

[0088] Fusion bottleneck layer features F out It can be expressed as:

[0089]

[0090] Among them, σ(·) is the Sigmod function, is matrix multiplication, [·] represents channel concatenation, and Conv(·) represents convolution operation.

[0091] Optionally, before step S211, the method of this embodiment further includes:

[0092] S209. Obtain a second bottleneck layer feature obtained in the process of the first self-attention network processing the third fusion feature.

[0093] S210: Use a second self-attention network to process the second sub-time-frequency graph and the second bottleneck layer features to obtain a third output feature.

[0094] Then, in step S211, the first output feature and the second output feature are fused to obtain a second fused feature, which may specifically include:

[0095] The first output feature, the second output feature, and the third output feature are fused to obtain a second fused feature.

[0096] Specifically, refer to Figure 2A Similar to the process of processing the first sub-spectrogram, when processing the second sub-spectrogram, the two second bottleneck layer features obtained from two adjacent bottleneck layers are fused during the first self-attention network's processing of the third fused feature. This fusion is then back-propagated to the bottleneck layer of the second self-attention network. The second self-attention network then processes the second sub-spectrogram and the fused second bottleneck layer features to obtain the third output feature. Accordingly, the first, second, and third output features are subsequently fused to obtain the second fused feature.

[0097] Optionally, the target loss function used when training the target network model can be expressed as:

[0098] Loss = loss MSE (F1)+λloss MSE (F2)+ηloss MSE (F3);

[0099] Among them, Loss represents the target loss function, F1 represents the prediction result corresponding to the original time-frequency graph of the sample, and loss MSE (F1) represents the loss function corresponding to F1, F2 represents the first prediction result output by the first self-attention network for the first sub-frequency graph of the sample, loss MSE (F2) represents the loss function corresponding to F2, and F3 represents the second prediction result output by the second self-attention network for the second sub-frequency graph of the sample. MSE(F3) represents the loss function corresponding to F3, and λ and η represent different loss weight parameters respectively.

[0100] Specifically, λ can be set to 0.5, and η can be set to 0.25. MSE (F1), loss MSE (F2), loss MSE (F3) uses Mean Square Error Loss (MSE), and the calculation formula is:

[0101]

[0102] Among them, N represents the total number of training samples, Represents the prediction result, x i Represents the training label, and i represents the training sample index.

[0103] Here, during the model training process, the second output feature of the first sub-frequency graph of the sample can be convolved to obtain the first prediction result output by the first self-attention network for the first sub-frequency graph of the sample; similarly, the third output feature of the second sub-frequency graph of the sample can be convolved to obtain the second prediction result output by the second self-attention network for the second sub-frequency graph of the sample.

[0104] This embodiment uses mean squared error loss as the loss function. The gradient of the mean squared error loss decreases as the error decreases, making it less likely to miss the global minimum. Furthermore, the mean squared error loss converges quickly, which can improve training efficiency.

[0105] The interference suppression method based on the multi-scale implicit neural representation network of this embodiment first inputs the original time-frequency graph corresponding to the radar echo signal into a pre-trained target network model, downsamples the original time-frequency graph through the target network model to obtain at least one sub-time-frequency graph, then extracts features from the at least one sub-time-frequency graph to obtain the corresponding implicit neural representation features, and then performs interference separation based on the original time-frequency graph and the implicit neural representation features to output the target signal. The present invention utilizes an end-to-end neural network model to jointly process the original time-frequency graph and sub-time-frequency graphs of different scales, and fuses the implicit neural representation features corresponding to the sub-time-frequency graphs, which can fully tap the potential features of the target signal, thereby improving the decoupling of the target signal and the interference signal and avoiding the loss of useful signals.

[0106] The interference suppression method based on the multi-scale implicit neural representation network provided by the present invention is further illustrated by simulation experiments below.

[0107] In order to compare the interference suppression effect and target signal recovery effect of the method of the present invention and the comparative method, this embodiment, on the one hand, qualitatively evaluates the interference suppression effect by observing the SAR imaging profile and the clarity of scene details, and on the other hand, uses evaluation indicators such as Signal Distortion Ratio (SDR), Multiplicative Noise Ratio (MNR), Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) to quantitatively evaluate the interference signal suppression effect and target signal recovery effect.

[0108] This embodiment selects three different interference suppression methods as comparison methods, specifically including a time-frequency domain notch method, a GoDec algorithm, and an SRGAN algorithm.

[0109] First, the effectiveness and necessity of the method of the present invention for multi-scale joint processing, implicit neural representation and segmentation results are verified through ablation experiments. Table 1 shows the ablation experiment results under the interference suppression condition of JNR=20dB. JNR refers to the Jammer-to-Noise Ratio, and the best indicators are highlighted in black bold format. It can be seen from the results in the table that the application of multi-scale joint processing improves SDR by 6.29dB and SSIM by 0.32. The application of INR improves SDR and SSIM by 1.6dB and 0.06 respectively, indicating that the two can effectively improve the reconstruction accuracy and imaging quality of the target echo signal. The improvement of various indicators verifies that multi-scale joint processing, implicit neural representation and segmentation results can enhance the target network's ability to decouple complex interference signals from target signals. Figure 3 This is the comparison result of interference suppression imaging in the ablation experiment, where the red and blue boxes represent the strong and weak scattering areas used in the MNR calculation, respectively.

[0110] Table 1 Ablation experiment results

[0111]

[0112]

[0113] To verify the robustness of the proposed algorithm, and considering that interference power can significantly impact target signals, this example discusses the algorithm's performance based on different JNRs for suppression and deception jamming. Suppression jamming requires a high-power signal to mask or drown out the desired signal, while deception jamming, due to its high similarity to the SAR signal, can achieve pulse compression processing gain, allowing for slightly lower signal power compared to suppression jamming. Therefore, this example sets the JNR for suppression jamming to a range of 20dB to 40dB, and the JNR for deception jamming to a range of 10dB to 20dB. Specific experimental results are as follows.

[0114] Figure 4 This paper presents a comparison of SAR imaging results using different interference suppression algorithms before and after interference suppression under varying JNRs. The SAR echo signal receives three types of suppressive interference, sequentially in azimuth: sinusoidal swept-frequency interference, sawtooth swept-frequency interference, and comb spectrum interference. The interference results show that the widespread suppressive interference in SAR echo data appears as high-energy bands in the imaging results, completely covering the detection scene. As the JNR increases, the brightness of the bands increases, and the scene information is more completely submerged, severely hindering SAR target observation and image interpretation. The interference suppression results of each algorithm show that the time-frequency domain notching method has a certain suppression effect on all types of interference, highlighting the detection scene under low JNR conditions. However, as the JNR increases, the suppression effect of the notch decreases significantly, resulting in a large number of artifacts and a more blurred scene in the imaging results. The GoDec algorithm's suppression results show that the imaging results contain both a large number of bands and scene information. Combined with the interference release order, it is clear that this algorithm has no significant suppression effect on swept-frequency interference, but has a significant suppression effect on comb spectrum interference. This is because the GoDec algorithm requires the interference to exhibit low-rank characteristics in the time-frequency domain, which swept-frequency interference does not meet. However, the multiple similar bright lines presented by comb spectrum interference in the time-frequency spectrum meet the low-rank condition, resulting in partial scene recovery and partial scene interference. At the same time, the GoDec algorithm also suffers from the problem of algorithm failure under high JNR conditions. SRGAN can completely and clearly recover scene information in the face of various interferences under low JNR conditions, but some interference strips remain under high JNR conditions, resulting in incomplete interference suppression. This indicates that SRGAN is unable to separate interference and useful signals in the face of high-intensity, strongly coupled interference. MSINR-Net can clearly restore the detected target and scene under various interference and JNR conditions, and the algorithm performance does not decrease significantly when the JNR increases, demonstrating that the proposed algorithm has good suppression effect and robustness against swept-frequency interference and comb spectrum interference.

[0115] Based on the imaging comparison results of various algorithms, the time-frequency domain filtering method has broad requirements for interference types and strong applicability, but the interference residue is serious and the suppression effect is limited; the GoDec algorithm has a good suppression effect on comb spectrum interference, but fails to suppress swept frequency interference, that is, it has high requirements for interference types and poor applicability; SRGAN has a significant suppression effect on various interference types, but its decoupling ability is limited, and the suppression effect on interference with high intensity and strong coupling is insufficient. The method of the present invention has a good suppression effect on various types of suppression interference with JNR changes, showing excellent effectiveness and robustness.

[0116] Figures 5 to 7 The time-frequency plots of the 300th, 800th, and 1300th azimuth echo signals recovered by the interference suppression algorithm under the interference suppression condition of JNR = 20dB are compared. The corresponding interferences are sinusoidal swept frequency interference, sawtooth swept frequency interference, and comb spectrum interference, respectively. As shown in Figure (c) of each figure, the time-frequency domain notch filter method has a significant suppression effect on all types of interference. However, due to the limitation of the threshold, this method can only remove interference above the threshold, leaving a significant low-energy interference residue, and its ability to suppress interference and recover the target echo signal is insufficient. Figure 5 and Figure 6 (d) shows that since the swept frequency interference does not have low rank in the time-frequency domain, although the interference energy is greatly reduced after the GoDec algorithm is suppressed, there is still a large amount of interference remaining, which in turn affects the imaging results; Figure 7 Figure (d) shows that the GoDec algorithm can effectively suppress low-rank interference based on the similarity of comb spectrum interference, but its ability to recover useful signals is insufficient, and the recovered signal exhibits obvious discontinuities at the location of the interference. Figures (e) and (f) in each figure show that the interference suppression algorithm using deep learning is more effective than the traditional algorithm in suppressing interference, and the suppressed signal is more similar to the original signal's time-frequency spectrum. However, compared with the two methods, SRGAN has a higher degree of distortion, while the method of the present invention has a higher degree of signal restoration and a closer time-frequency spectrum, thus achieving better interference suppression.

[0117] Table 2 shows in detail the interference suppression evaluation index results of different algorithms under different JNR suppression interference. In order to more intuitively show the changes in the indicators of different interference suppression algorithms, Figure 8Based on the data in the table, a line graph is plotted showing how each interference suppression evaluation metric changes with the JNR. It can be seen that the time-frequency domain notch filter method and GoDec yield poor evaluation metric results due to residual interference and damaged target signals, but the metric results are consistent with the imaging results. Compared to traditional methods, SRGAN and the proposed method achieve better overall results. Furthermore, the proposed method achieves lower SDR and MNR values ​​than other algorithms, indicating that the proposed algorithm reconstructs target signals with lower loss and restores better image contrast. Furthermore, the proposed method achieves higher and more stable PSNR and SSIM values ​​than other algorithms, indicating that the proposed algorithm achieves higher target signal reconstruction accuracy and better restored image quality.

[0118] Table 2 Interference suppression effects of different interference suppression algorithms under different JNR suppression interference

[0119]

[0120] Figure 9 This paper presents SAR imaging results using different interference suppression algorithms before and after interference suppression under different JNR deceptive jamming conditions. The SAR echo signal receives three types of deceptive jamming, sequentially in azimuth: intermittent sampling and forwarding jamming, multiple false targets jamming, and dense false targets jamming. The jamming results show that the deceptive jamming present in the SAR echo signal appears as bright lines representing false targets in the imaging results. As the JNR increases, the brightness of the bright lines increases. The presence of false targets seriously affects the SAR's ability to detect the target and acquire scene information. The interference suppression results of each algorithm show that due to the large coverage area of ​​dense false targets in the time-frequency domain, the notching method, which sets the frequency bins to zero, results in significant loss of useful signals, resulting in poor interference suppression. The imaging results also demonstrate that while the notching method in the time-frequency domain reduces the intensity of the jamming, a large number of defocused false targets still remain in the image. While the intensity of the false targets is significantly reduced in the GoDec algorithm imaging results, the presence of low-energy bands indicates the presence of residual low-energy jamming in the jamming suppression results. SRGAN achieves good interference suppression results when JNR = 10dB. However, as JNR increases, false targets of varying strengths and locations are present in the suppression results, reflecting SRGAN's inability to separate high-intensity, strongly coupled interference from signals. The proposed method can clearly restore the detection scene and target under various interference and JNR conditions, and its performance does not significantly decrease as JNR increases, demonstrating the proposed algorithm's robustness and effectiveness in suppressing deceptive interference.

[0121] Based on the suppression results of various algorithms for deceptive interference, although the time-frequency domain notch method can reduce the intensity of false targets, it will cause obvious defocusing problems and the overall suppression effect is poor; the GoDec algorithm has a significant suppression effect on false targets, but the suppression is incomplete and there is still residual interference; SRGAN's insufficient decoupling ability for interference leads to the inability to completely suppress false targets in certain locations; the method of the present invention has a good suppression effect under the conditions of changing JNR and interference type, showing superior algorithm performance.

[0122] Figures 10 to 12 The time-frequency plots of the 300th, 800th, and 1300th azimuth echo signals recovered by the interference suppression algorithm under the deceptive jamming condition of JNR = 20dB are compared. The corresponding jamming is intermittent sampling forwarding jamming, multiple false target jamming, and dense false target jamming, respectively. As shown in Figure (c), the time-frequency domain notch method does not completely suppress the jamming due to the threshold limitation, and the recovered signal is discontinuous in the time-frequency domain. Therefore, the signal imaging results after suppression show multiple defocused false targets. Figure (d) shows that compared with the time-frequency domain notch method, the GoDec algorithm can suppress the jamming to a certain extent based on the false targets with similar information to the deceptive jamming. However, the problems of insufficient recovery of the target signal and some residual jamming still lead to the appearance of weak false targets in the imaging results. Figures (e) and (f) show that both SRGAN and the proposed method can restore the basic structure of the target signal. However, a comparison of the suppression results of various types of interference shows that SRGAN has lower restoration accuracy in terms of the true value and details of the signal. The signal restored by the proposed method is closer to the true value of the original signal and has more similar details, demonstrating the superior performance of the proposed algorithm in interference suppression and signal restoration.

[0123] Table 3 shows in detail the results of the interference suppression evaluation indexes of different algorithms under different JNR deception interference. The figure is a line graph of the interference suppression evaluation indexes as the JNR changes based on the data in the table. Figure 13 It can be seen that the various algorithms have significantly improved their indicators when facing deception interference. The method of the present invention maintains a low level in SDR and MNR calculation by virtue of its high decoupling ability of radar echo signals and accurate recovery ability of target signals. The slow decrease of MNR when JNR continues to increase also shows that both strong and weak targets in the image of the proposed method can be well recovered; various indicators of the restored image prove the good suppression ability of the proposed algorithm against SAR interference and the target information recovery ability.

[0124] Table 3 Interference suppression effects of different interference suppression algorithms under different JNR deception interference

[0125]

[0126]

[0127] Summary: In view of the problems that existing interference suppression algorithms are prone to causing useful signal loss and imaging quality degradation when facing complex interferences of multiple types, high dynamics, and strong coupling, the method of the present invention constructs a target network model based on the significant feature differences between the interference signal and the target signal in the time and frequency domain. The target network model uses multi-scale joint processing and implicit neural representation to extract features of the coupled interference signal and target signal, and then performs feature separation on the target signal to achieve intelligent decoupling of the interference signal and the target signal. Finally, in the test results of data on various interference suppression and deception interference, the effectiveness and robustness of the method of the present invention are verified using the imaging results after various interference suppression and quantitative evaluation indicators such as SDR, MNR, PSNR and SSIM, proving that the method of the present invention has strong interference suppression capabilities.

[0128] Corresponding to the above-mentioned interference suppression method based on a multi-scale implicit neural representation network, an embodiment of the present invention further provides an interference suppression device based on a multi-scale implicit neural representation network; Figure 14 As shown, the device may include:

[0129] An acquisition module 1401 is configured to acquire a radar echo signal for target object detection, where the radar echo signal includes an interference signal and a target signal;

[0130] The prediction module 1402 is used to input the original time-frequency diagram corresponding to the radar echo signal into a pre-trained target network model, downsample the original time-frequency diagram through the target network model to obtain at least one sub-time-frequency diagram, extract features from the at least one sub-time-frequency diagram to obtain corresponding implicit neural representation features, perform interference separation based on the original time-frequency diagram and the implicit neural representation features, and output the target signal.

[0131] For the specific content of the device, please refer to the steps of the interference suppression method based on the multi-scale implicit neural representation network provided in the first aspect, which will not be described in detail here.

[0132] The present invention provides an interference suppression device based on a multi-scale implicit neural representation network. The device inputs the original time-frequency graph corresponding to the radar echo signal into a pre-trained target network model, downsamples the original time-frequency graph through the target network model to obtain at least one sub-time-frequency graph, and then extracts features from the at least one sub-time-frequency graph to obtain the corresponding implicit neural representation features. Subsequently, interference separation is performed based on the original time-frequency graph and the implicit neural representation features, and the target signal is output. The present invention utilizes an end-to-end neural network model to jointly process the original time-frequency graph and the sub-time-frequency graphs of different scales, and fuses the implicit neural representation features corresponding to the sub-time-frequency graphs. This can fully tap the potential features of the target signal, thereby improving the decoupling of the target signal and the interference signal and avoiding the loss of useful signals.

[0133] It should be noted that, as for the device, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0134] It should be noted that the terms "first," "second," and the like are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in sequences other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention.

[0135] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0136] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other variations of the disclosed embodiments by viewing the drawings and the disclosed content. In the description of the present invention, the word "comprising" does not exclude other components or steps, "one" or "a" does not exclude multiple situations, and "multiple" means two or more, unless otherwise clearly and specifically defined. In addition, certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0137] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. An interference suppression method based on a multi-scale implicit neural representation network, characterized in that: include: Acquiring a radar echo signal for detecting a target object, wherein the radar echo signal includes an interference signal and a target signal; The original time-frequency graph corresponding to the radar echo signal is input into a pre-trained target network model, the original time-frequency graph is downsampled by the target network model to obtain at least one sub-time-frequency graph, features are extracted from at least one of the sub-time-frequency graphs to obtain corresponding implicit neural representation features, interference separation is performed based on the original time-frequency graph and the implicit neural representation features, and the target signal is output.

2. The interference suppression method based on a multi-scale implicit neural representation network according to claim 1, characterized in that: The extracting features of at least one of the sub-spectrograms to obtain a plurality of implicit neural representation features includes: Extract the coordinates of each sub-frequency graph and encode the extracted coordinate data to obtain the coordinate encoding features corresponding to each sub-frequency graph; Encode each sub-spectrogram to obtain the image coding features of each sub-spectrogram; Performing high-dimensional spatial encoding on the coordinate coding features and the image coding features corresponding to each sub-spectrogram to obtain a feature sequence corresponding to each sub-spectrogram; The feature sequence corresponding to each sub-time-frequency graph is decoded using a multi-layer perceptron network to obtain the implicit neural representation features corresponding to each sub-time-frequency graph.

3. The interference suppression method based on a multi-scale implicit neural representation network according to claim 2, characterized in that: If the at least one sub-spectrogram includes a first sub-spectrogram, the target network model includes a self-attention network corresponding to the original spectrogram and a first self-attention network corresponding to the first sub-spectrogram; The performing interference separation based on the original time-frequency graph and the implicit neural representation feature and outputting the target signal includes: Extracting original features of the original time-frequency graph and fusing them with implicit neural representation features corresponding to the first sub-time-frequency graph to obtain a first fused feature; Processing the first fusion feature using a self-attention network corresponding to the original time-frequency graph to obtain a first output feature; Processing the first sub-spectrogram using the first self-attention network to obtain a second output feature; fusing the first output feature and the second output feature to obtain a second fused feature; Convolution is performed on the second fusion feature to output the target signal.

4. The interference suppression method based on a multi-scale implicit neural representation network according to claim 3, characterized in that: The using the first self-attention network to process the first sub-time-frequency graph to obtain a second output feature includes: Obtaining a first bottleneck layer feature obtained by processing the first fusion feature by the self-attention network corresponding to the original time-frequency graph; The first self-attention network is used to process the first fusion feature and the first bottleneck layer feature to obtain a second output feature.

5. The interference suppression method based on a multi-scale implicit neural representation network according to claim 4, characterized in that: If the at least one sub-spectrogram further includes a second sub-spectrogram, the target network model further includes a second self-attention network corresponding to the second sub-spectrogram, and the size of the second sub-spectrogram is smaller than that of the first sub-spectrogram; The method further comprises: Extracting a first feature of the first sub-spectrogram and fusing it with an implicit neural representation feature corresponding to the second sub-spectrogram to obtain a third fused feature; The using the first self-attention network to process the first sub-time-frequency graph to obtain a second output feature includes: The third fusion feature is processed using the first self-attention network to obtain a second output feature.

6. The interference suppression method based on a multi-scale implicit neural representation network according to claim 5, characterized in that: The method further comprises: Obtaining a second bottleneck layer feature obtained in the process of processing the third fusion feature by the first self-attention network; Processing the second sub-spectrogram and the second bottleneck layer features using the second self-attention network to obtain a third output feature; The fusing the first output feature and the second output feature to obtain a second fused feature includes: The first output feature, the second output feature, and the third output feature are fused to obtain a second fused feature.

7. The interference suppression method based on a multi-scale implicit neural representation network according to claim 6, characterized in that: The target loss function used when training the target network model is expressed as: Loss=loss MSE (F1)+λloss MSE (F2)+ηloss MSE (F3); Among them, Loss represents the target loss function, F1 represents the prediction result corresponding to the original time-frequency graph of the sample, and loss MSE (F1) represents the loss function corresponding to F1, F2 represents the first prediction result output by the first self-attention network for the first sub-frequency graph of the sample, loss MSE (F2) represents the loss function corresponding to F2, F3 represents the second prediction result output by the second self-attention network for the second sub-frequency graph of the sample, loss MSE (F3) represents the loss function corresponding to F3, and λ and η represent different loss weight parameters respectively.

8. An interference suppression device based on a multi-scale implicit neural representation network, characterized in that: include: An acquisition module, configured to acquire a radar echo signal for detecting a target object, wherein the radar echo signal includes an interference signal and a target signal; The prediction module is used to input the original time-frequency diagram corresponding to the radar echo signal into a pre-trained target network model, downsample the original time-frequency diagram through the target network model to obtain at least one sub-time-frequency diagram, extract features from at least one of the sub-time-frequency diagrams to obtain corresponding implicit neural representation features, perform interference separation based on the original time-frequency diagram and the implicit neural representation features, and output the target signal.

Images