A SAR radio frequency interference image domain detection method and device

Abstract

The application provides a SAR radio frequency interference image domain detection method and device, wherein a SAR complex image is obtained by performing SAR imaging on an original echo signal, and the SAR complex image is converted into a gray scale image; an improved U2Net-based intelligent interference detection network is used to detect the gray scale image to obtain a predicted mask, so as to determine an interference area of the SAR complex image. Since the prior art can only identify whether RFI exists in a SAR image, the application further identifies specific parts of the SAR image contaminated by RFI, that is, pixel-level detection, so that the SAR data containing RFI does not need to be discarded directly, the integrity of the data is maintained, and the interference in the image can be processed specifically instead of processing the whole image; a large number of non-interference comparison images of the same scene are not needed, the implementation complexity is reduced; and no threshold and other hyperparameters need to be predefined, so that the robustness is good and the application is effective for various scenes.

Description

Technical Field

[0001] This invention relates to the field of SAR interference detection technology, and in particular to a method and apparatus for detecting SAR radio frequency interference in the image domain. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing tool that provides crucial data for observing and understanding changes in the Earth's surface environment. Its high resolution and all-weather image acquisition capabilities make SAR a valuable method for numerous scientific research, commercial, and defense applications. Currently, the surge in electronic device usage and the increasing complexity of the electromagnetic environment mean that signals such as television broadcasts, wireless communications, active jamming signals, and other non-cooperative SAR signals can all interfere with SAR observations. These incoherent signals manifest as various bright linear stripes or blocky artifacts in the SAR image domain, severely impacting SAR's operational capabilities. From a computer vision perspective, active radio frequency interference (RFI) in SAR images is more of an anomaly. Therefore, SAR image data can be directly used to locate SAR images with anomalies and segment regions containing RFI artifacts, enabling RFI marking and localization.

[0003] In the field of RFI detection, a series of research results have emerged in recent years. Agnieszka Chojka et al. proposed an image processing-based RFI detection method in the paper "RFI artefacts detection in Sentinel-1 level-1 SLC data based on image processing techniques" (Sensors, Vol. 13, 2020). This method uses feature extraction techniques, including pixel convolution, thresholding, and nearest neighbor structure filtering, for preprocessing, and achieves RFI detection through CNN classification. However, the image processing method relies on experience in threshold selection and has low robustness. Li, N. et al. proposed a method for simultaneously screening and detecting RFI from massive SAR images in the paper "Simultaneous Screening and Detection of RFI From Massive SAR Images: A Case Study on European Sentinel-1" (IEEE Transactions on Geoscience and Remote Sensing, Vol. 60, 2022). This method utilizes previously acquired preprocessed time-series SAR images to construct an approximately radio frequency-free background image. Then, differential images are generated based on image change detection methods, and adaptive thresholds are used to analyze them. The entropy of all differential images is calculated to complete the initial screening of radio frequency (RF) data. However, this method requires a large number of interference-free images of the same scene at different times, which places high demands on data collection and storage. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a method and apparatus for SAR radio frequency interference image domain detection. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] In a first aspect, the present invention provides a SAR radio frequency interference image domain detection method comprising:

[0006] S100 utilizes a radar system to periodically transmit pulse signals and receive raw echo signals;

[0007] S200, the original echo signal is subjected to SAR imaging to obtain a SAR complex image, and the SAR complex image is converted into a grayscale image;

[0008] S300: Using a trained intelligent interference detection network based on improved U2Net, the grayscale image formed by the original echo signal is detected to obtain a prediction mask on the grayscale image.

[0009] S400, based on the prediction mask, determine the interference region on the SAR complex image.

[0010] Secondly, the present invention provides a SAR radio frequency interference image domain detection device comprising:

[0011] The receiving module is configured to periodically transmit pulse signals using the radar system and receive raw echo signals;

[0012] The conversion module is configured to perform SAR imaging on the original echo signal to obtain a SAR complex image, and convert the SAR complex image into a grayscale image;

[0013] The detection module is configured to use a trained intelligent interference detection network based on improved U2Net to detect the grayscale image formed by the original echo signal and obtain a prediction mask on the grayscale image.

[0014] The determination module is configured to determine the interference region on the SAR complex image based on the prediction mask.

[0015] Beneficial effects:

[0016] This invention provides a method and apparatus for detecting radio frequency interference (RFI) in the image domain of SAR. The method involves obtaining a complex SAR image by performing SAR imaging on the original echo signal, and then converting the complex SAR image into a grayscale image. A trained intelligent interference detection network based on an improved U2Net is used to detect interference in the grayscale image, obtaining a prediction mask to determine the interference region on the complex SAR image. While existing technologies can only identify the presence of RFI in SAR images, this invention proposes to further identify specific RFI-contaminated parts of the SAR image, i.e., pixel-level detection. This eliminates the need to discard SAR data containing RFI, maintaining data integrity, and allows for targeted processing of interference within the image rather than processing the entire image. It also eliminates the need for numerous interference-free comparison images of the same scene, reducing implementation complexity. Furthermore, it eliminates the need for predefined thresholds and other hyperparameters, exhibiting good robustness and widespread effectiveness across various scenarios.

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of a SAR radio frequency interference image domain detection method provided in an embodiment of the present invention.

[0019] Figure 2 A schematic diagram of the training process of the intelligent interference detection network based on the improved U2Net provided in the embodiments of the present invention;

[0020] Figure 3This is a schematic diagram of SAR imaging using the Chirp Scaling imaging algorithm in an embodiment of the present invention;

[0021] Figure 4 Examples of dataset sample images and labeled images from embodiments of the present invention;

[0022] Figure 5 This is a schematic diagram of the ASPP-d module structure according to an embodiment of the present invention;

[0023] Figure 6 This is a schematic diagram of the channel weighting module according to an embodiment of the present invention;

[0024] Figure 7 This is a schematic diagram of the improved U2Net structure according to an embodiment of the present invention;

[0025] Figure 8 This is a schematic diagram comparing the results of interference detection experiments in an embodiment of the present invention. Detailed Implementation

[0026] 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.

[0027] refer to Figure 1 The present invention provides a SAR radio frequency interference image domain detection method, comprising:

[0028] S100 utilizes a radar system to periodically transmit pulse signals and receive raw echo signals;

[0029] S200, the original echo signal is subjected to SAR imaging to obtain a SAR complex image, and the SAR complex image is converted into a grayscale image;

[0030] S300: Using a trained intelligent interference detection network based on improved U2Net, the grayscale image formed by the original echo signal is detected to obtain a prediction mask on the grayscale image.

[0031] The trained intelligent interference detection network based on the improved U2Net is obtained by constructing a training set and introducing the Focal loss function for constraint during the training process.

[0032] S400, based on the prediction mask, determine the interference region on the SAR complex image.

[0033] This invention provides a SAR radio frequency interference (RFI) image domain detection method. It obtains a SAR complex image by performing SAR imaging on the original echo signal and converts the SAR complex image into a grayscale image. A trained intelligent interference detection network based on an improved U2Net is used to detect interference in the grayscale image, obtaining a prediction mask to determine the interference region on the SAR complex image. While existing technologies can only identify the presence of RFI in SAR images, this invention further identifies specific RFI-contaminated parts of the SAR image, i.e., pixel-level detection. This eliminates the need to discard SAR data containing RFI, maintaining data integrity, and allows for targeted processing of interference within the image rather than processing the entire image. It also eliminates the need for numerous interference-free comparison images of the same scene, reducing implementation complexity. Furthermore, it eliminates the need for predefined thresholds and other hyperparameters, exhibiting good robustness and widespread effectiveness across various scenarios.

[0034] In a specific embodiment of the present invention, the process of constructing the training set includes:

[0035] a. The radar system periodically transmits pulse signals and receives target echo signals;

[0036] b. Add an interference baseband signal to the target echo signal to obtain a mixed signal; the mixed signal includes the target echo signal, the interference baseband signal, and additive observation noise.

[0037] This invention establishes an interference model by superimposing a radio frequency interference signal onto the echo signal. The mixed signal received by the SAR receiver is a superposition of the target echo signal, the interfering baseband signal, and additive observation noise. The mixed signal is represented as follows:

[0038] S m (x,r)=S(x,r)+J(x,r)+N(x,r);

[0039] Where x represents the azimuth position of the SAR platform at the nth azimuth pulse moment, r represents the spatial position in the slant range direction corresponding to the fast time variable, and S m (x,r) represents the mixed signal, S(x,r) represents the target echo signal, J(x,r) represents the interfering baseband signal, and N(x,r) represents the additive observation noise;

[0040] The target echo signal includes multiple point target baseband echoes, each point target baseband echo being represented as follows:

[0041]

[0042] The interference baseband echo signal is represented as follows:

[0043]

[0044] In the formula, A J (·,·) represents the effective radiated power of the jammer; R J (·) represents the propagation distance of the interference signal; a J (·) represents the envelope of the interference signal, a J (·)=a T (·); Phase term It is the Doppler phase term, the distance delay function R J (x) constitutes the azimuth phase history function of the interfering baseband signal.

[0045] Assuming the synthetic aperture center time t = 0, the range history of the interference signal can be represented by the SAR echo Doppler parameters, i.e.:

[0046]

[0047] Where R0 is the slant range when the beam center illuminates the target, and Δt J f is the moment when the center of the beam illuminates the jammer. dc and f dr These are the Doppler center frequency and Doppler modulation frequency of the target, respectively.

[0048] The most significant difference between the interference signal received by SAR and the target echo signal lies in their azimuth phase history. The azimuth phase history of the target echo signal is expressed as:

[0049]

[0050] Since the radio frequency interference signal is transmitted directly to the SAR by the jammer, its azimuth phase history is expressed as follows:

[0051]

[0052] As can be seen from the above formula, SAR radio frequency interference signals do not maintain the characteristics of SAR signals within or between radar pulses, and therefore cannot be imaged. In the image, they appear as abnormal blocks with range and azimuth spread, i.e., linear or blocky artifacts.

[0053] The interference baseband signals of this invention include radio frequency noise interference signals, noise amplitude modulation interference signals, noise frequency modulation interference signals, linear frequency modulation interference signals, and sinusoidal frequency modulation interference signals. The time-domain mathematical expression for typical SAR radio frequency interference is as follows:

[0054] ① Radio frequency noise interference:

[0055] J1(t)=a n (t)exp(j2πf j t+φ n (t));

[0056] Where J1(t) follows a normal distribution, and the envelope function a n (t) follows a Rayleigh distribution, and the phase function φ n (t) follows a distribution in [0, 2π] and is related to a n (t) are independent, carrier frequency f j It is a constant and much larger than the spectral width of J1(t).

[0057] ② Noise amplitude modulation interference:

[0058] J2(t)=[a0+a n (t)]exp(j2πf j t+φ n (t));

[0059] J2(t) is a generalized stationary random process, where a0 is the carrier voltage, and an(t) is a random process with a mean of 0 and a variance of σ. n 2 Gaussian band-limited white noise.

[0060] ③ Noise FM interference:

[0061]

[0062] J3(t) is a generalized stationary random process, where k FM is a proportionality coefficient, representing the frequency change caused by a unit modulation signal strength.

[0063] ④ Linear frequency modulation interference:

[0064]

[0065] Among them, a k (t),γ k ,f k ,φ k These represent the amplitude, frequency modulation, carrier frequency, and phase of the k-th component of the interference, respectively.

[0066] ⑤ Sine wave frequency modulation interference:

[0067]

[0068] Where, β k These represent the modulation factors that interfere with the k-th component.

[0069] c. Perform SAR imaging on both the target echo signal and the mixed signal to obtain a SAR complex image, and convert the SAR complex image into a grayscale image;

[0070] This invention performs SAR imaging based on the Chirp Scaling algorithm on both the target echo signal and the mixed signal, obtaining complex images after SAR imaging, denoted as S(τ,t). a ) and S m (τ,t a ); The target echo signal and the mixed signal complex image are respectively moduloed: X1=|S(τ,t) a )|,X2=|S m (τ,t a Normalize the magnitude of SAR complex images with interfering baseband signals to form grayscale images;

[0071] The specific formula for normalization is as follows:

[0072]

[0073] In the above formula, X'2 represents the normalized data, and X2 represents the data to be normalized. Let X2 be the minimum value of the data to be normalized. The maximum value of the data X2 to be normalized.

[0074] d. Compare the grayscale image formed by the target echo signal with the grayscale image formed by the mixed signal, mark the changed part in the grayscale image formed by the mixed signal, and use the changed part as an interference artifact;

[0075] e. Use all grayscale images containing interference artifacts to form the training and testing sets.

[0076] This invention allows for manual annotation of interference artifacts in images using the LabelMe tool. A dataset for training and testing is constructed, using only interfered SAR image data. All images in the dataset are 572×572 pixels as the standard input size, totaling 2000 images. By comparing interfered and uninterrupted SAR images, the varying parts of the training set images, i.e., interference artifacts, are manually annotated using LabelMe software. Interference appears as linear or blocky artifacts in the images. The training and testing sets are randomly partitioned at a 9:1 ratio. Examples of dataset sample images and annotated images are shown below. Figure 4 As shown.

[0077] In a specific embodiment of the present invention, the intelligent interference detection network based on the improved U2Net includes: an improved U2Net network and a channel attention module; the improved U2Net network includes multiple scale-variable codecs and three scale-invariant dilated convolutional modules, the multiple codecs and the three dilated convolutional modules forming a U-shaped structure, and each dilated convolutional module uses a dilated spatial pyramid convolutional module to replace its own dilated convolutional layer.

[0078] This invention proposes an improved U2Net that replaces the Cross-entropy loss function used in the U2Net network with a Focal loss function to alleviate the training difficulties caused by imbalanced positive and negative samples. An improved Spatial Pyramid Convolutional Module with Dilated Convolutions (ASPP-d) is introduced to replace the partially dilated convolutional layers at the bottom of U2Net, effectively increasing the receptive field and enhancing the network's feature extraction capabilities to adapt to the multi-scale characteristics of RFI. Furthermore, a channel attention module is introduced before outputting the predicted mask results to assign weights to the predicted mask output by each layer. Combining all these improvements, and using appropriate hyperparameters, the improved U2Net network is trained using a training dataset.

[0079] Because RFI artifacts exhibit multi-scale characteristics, they affect the balance between positive and negative samples. To address this issue, this invention employs the Focal loss function. Focal loss is a variant of binary cross-entropy loss, introducing two new hyperparameters α. t γ. The form of the Focal loss function is as follows:

[0080] Focalloss(p t )=α t (1-p t ) γ log(p t );

[0081] Where, α t The addition of p allows the loss function to balance the relative importance of positive and negative samples; t The predicted class probability, with values ​​ranging from [0,1]; 1-p t It is called the modulation factor; γ controls the rate at which the weights decrease.

[0082] The ASPP module uses multiple parallel dilated convolutional layers with different sampling rates. Features extracted at each sampling rate are further processed in separate branches and then fused to generate the final result. This module constructs convolutional kernels with different receptive fields through varying dilation rates to capture multi-scale object information. Consider a two-dimensional dilated convolution: for each location i of the input feature map y and the filter w, a dilated convolution is applied to the input feature map x:

[0083]

[0084] In the formula, k represents the kernel size, r represents the sampling rate, and the above formula means that a new filter is obtained by inserting r-1 zero values ​​along each spatial dimension between two consecutive filter values. The feature map x is then convolved through this filter, and y[i] represents the final feature map. Therefore, dilated convolution can control the receptive field of the filter and the compactness of the network output features by changing the sampling rate without increasing the number of parameters and computation.

[0085] To adapt to the task of RFI detection, this invention improves the ASPP module, calling it the ASPP-d module, by incorporating more branches with different hole rates, thereby enhancing the ability to acquire features at multiple scales. The specific structure of the improved module is as follows: Figure 5 As shown, it includes 5 convolutional layers. The first one is a one-dimensional convolution, and the others are two-dimensional convolutions. The two-dimensional convolution kernels are all 3*3, and the dilatation rates are 6, 12, 18 and 24 respectively.

[0086] The U2Net model borrows the idea of ​​FPN, performing 2D upsampling on each output of the Decoder stage and fusing these temporary outputs together for 2D convolution to output the final result. These temporary outputs are presented as multi-channel prediction masks before the 2D convolution. Since the underlying prediction masks are obtained through multiple upsampling steps, these masks suffer some accuracy loss, which is detrimental to the final output. To address this issue, a channel attention approach is proposed for channel weighting, increasing the output weights of channels with significant features and obtaining a more accurate prediction mask after fusion. The proposed channel weighting module is as follows: Figure 6 As shown, for the input feature map, a dimensionality transformation is first performed, and then the dimensionality-transformed feature map is input into a multilayer perceptron. It uses two multilayer perceptrons to amplify cross-dimensional channel spatial dependencies. Finally, the tensor is converted to its original dimensions and a sigmoid activation is applied to the output.

[0087] Combining all the above improvements, the overall network structure of U2Net is improved as follows: Figure 7 As shown.

[0088] The improved U2Net training process used the PyTorch framework and was trained on an NVIDIA GeForce RTX 3060 GPU with 12GB of VRAM. The Adam optimizer was used to optimize the network parameters. Considering the limited computational resources, the batch size was set to 2, the momentum to 0.9, the weight decay to 1e-4, the initial learning rate to 0.05, and a multinomial descent strategy was used. The entire training process was completed after 50 epochs.

[0089] In an optional embodiment of the present invention, the present invention can also utilize the trained intelligent interference detection network based on improved U2Net to segment interference artifacts in the test set, and use four evaluation metrics to evaluate the test results to obtain the performance of the trained intelligent interference detection network based on improved U2Net.

[0090] To verify the performance of the proposed method, the algorithm will be evaluated qualitatively and quantitatively. For qualitative evaluation, differences in image segmentation results can be visually compared. For quantitative evaluation, mIoU, Accuracy, Precision, and F1 score on the test dataset are used to reflect the network's performance. The expressions for these four metrics are:

[0091]

[0092]

[0093]

[0094]

[0095] In the formula, TP represents the number of correctly detected interference pixels; TN represents the number of correctly detected background pixels; FP represents the number of falsely detected interference pixels; FN represents the number of falsely detected background pixels; Recall represents the recall rate, and the formula is:

[0096] The improved U2Net was evaluated quantitatively. Validation experiments showed that its mIoU, Accuracy, Precision, and F1 scores were 87.37%, 98.01%, 90.89%, and 86.92%, respectively. The test results indicate that the model can fit the training data well and has a strong learning ability on imbalanced datasets, demonstrating better segmentation performance.

[0097] The evaluation was conducted qualitatively. To better demonstrate the good performance of the proposed method in interference detection, two typical scenarios from the test set were selected, and fast views of these scenarios were input into the improved U2Net network for experiments. The obtained output detection results are as follows: Figure 8 As shown, in Figure 8 In the figures, Figures a1-a2 are the original quick view; Figures b1-b2 are the label images; and Figures c1-c2 are the outputs of the network proposed in this embodiment.

[0098] The experimental results show that the improved U2Net achieved excellent segmentation results.

[0099] Therefore, this invention provides a SAR radio frequency interference (RFI) image domain detection method that distinguishes RFI artifacts from the background by outputting a predicted mask. This mask can be combined with relevant product information for RFI localization. Experimental results show that the network performs well in RFI segmentation on a self-built dataset. It saves researchers the trouble of finding a large amount of interference-free comparison data and achieves preliminary labeling and localization of RFI through image segmentation.

[0100] This invention provides a SAR radio frequency interference image domain detection device, comprising:

[0101] The receiving module is configured to periodically transmit pulse signals using the radar system and receive raw echo signals;

[0102] The conversion module is configured to perform SAR imaging on the original echo signal to obtain a SAR complex image, and convert the SAR complex image into a grayscale image;

[0103] The detection module is configured to use a trained intelligent interference detection network based on improved U2Net to detect the grayscale image formed by the original echo signal and obtain a prediction mask on the grayscale image.

[0104] The determination module is configured to determine the interference region on the SAR complex image based on the prediction mask.

[0105] This invention provides a method and apparatus for detecting radio frequency interference (RFI) in the image domain of SAR. The method involves obtaining a complex SAR image by performing SAR imaging on the original echo signal, and then converting the complex SAR image into a grayscale image. A trained intelligent interference detection network based on an improved U2Net is used to detect interference in the grayscale image formed by the original echo signal, obtaining a prediction mask on the grayscale image. The interference region on the complex SAR image is then determined based on the prediction mask. While existing technologies can only identify the presence of RFI in SAR images, this invention proposes to further identify the specific RFI-contaminated parts of the SAR image, i.e., pixel-level detection. This eliminates the need to discard SAR data containing RFI, maintaining data integrity. Furthermore, for interference mitigation, targeted operations can be performed instead of processing the entire image. Therefore, this invention does not require a large number of interference-free comparison images of the same scene, reducing implementation complexity; it also does not require predefined thresholds or other hyperparameters, exhibiting good robustness and widespread effectiveness across various scenarios.

[0106] Furthermore, 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 technical features indicated. 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.

[0107] Although this application 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, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0108] 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 method for detecting SAR radio frequency interference in the image domain, characterized in that, include: S100 utilizes a radar system to periodically transmit pulse signals and receive raw echo signals; S200, the original echo signal is subjected to SAR imaging to obtain a SAR complex image, and the SAR complex image is converted into a grayscale image; S300: Using a trained intelligent interference detection network based on improved U2Net, the grayscale image formed by the original echo signal is detected to obtain a prediction mask on the grayscale image. S400, the interference region on the SAR complex image is determined according to the prediction mask; the trained intelligent interference detection network based on the improved U2Net is obtained by constructing a training set and introducing the Focal loss function for constraint during the training process; The intelligent interference detection network based on the improved U2Net includes: an improved U2Net network and a channel attention module; The improved U2Net network includes multiple scale-variable codecs and three scale-invariant dilated convolutional modules. The multiple codecs and three dilated convolutional modules form a U-shaped structure. Each dilated convolutional module replaces its own dilated convolutional layer with a dilated spatial pyramid convolutional module.

2. The SAR radio frequency interference image domain detection method according to claim 1, characterized in that, The process of constructing the training set includes: a. The radar system periodically transmits pulse signals and receives target echo signals; b. Add an interference baseband signal to the target echo signal to obtain a mixed signal; the mixed signal includes the target echo signal, the interference baseband signal, and additive observation noise. c. Perform SAR imaging on both the target echo signal and the mixed signal to obtain a SAR complex image, and convert the SAR complex image into a grayscale image; d. Compare the grayscale image formed by the target echo signal with the grayscale image formed by the mixed signal, mark the changed part in the grayscale image formed by the mixed signal, and use the changed part as an interference artifact; e. Use all grayscale images containing interference artifacts to form the training and testing sets.

3. The SAR radio frequency interference image domain detection method according to claim 2, characterized in that, The mixed signal is represented as: ； in, Indicates the SAR platform's first The azimuth position at each azimuth pulse moment This represents the spatial position along the slant distance direction of the corresponding fast time variable. Indicates a mixed signal. Indicates the target echo signal. This indicates interference with the baseband signal. Indicates additive observation noise; The target echo signal includes multiple point target baseband echoes, each point target baseband echo being represented as follows: ； The interference baseband signal is represented as follows: ； In the formula, This represents the effective radiated power of the jammer. The distance at which the interference signal propagates; For the interference signal envelope, Phase term It is the Doppler phase term, the distance delay function. The azimuth phase history function that constitutes the interfering baseband signal.

4. The SAR radio frequency interference image domain detection method according to claim 2, characterized in that, The interference baseband signals include radio frequency noise interference signals, noise amplitude modulation interference signals, noise frequency modulation interference signals, linear frequency modulation interference signals, and sinusoidal frequency modulation interference signals.

5. The SAR radio frequency interference image domain detection method according to claim 2, characterized in that, Step c includes: Both the target echo signal and the mixed signal are subjected to SAR imaging based on the Chirp Scaling algorithm, and the resulting complex images are represented as follows: and ; Modulus extraction is performed on both the target echo signal and the mixed signal complex image: , ; Normalize the magnitude of the complex SAR image with interfering baseband signal to form a grayscale image; The specific formula for normalization is as follows: ； In the above formula, For the normalized data, For data to be normalized, Data to be normalized The minimum value, Data to be normalized The maximum value.

6. The SAR radio frequency interference image domain detection method according to claim 2, characterized in that, After obtaining the trained intelligent interference detection network based on the improved U2Net, the SAR radio frequency interference image domain detection method further includes: The trained intelligent interference detection network based on the improved U2Net was tested using the test set, and the test results were evaluated using four evaluation metrics to obtain the performance of the trained intelligent interference detection network based on the improved U2Net; the expressions for the four evaluation metrics are as follows: ； ； ； ； In the formula, The number of correctly detected interfering pixels; The number of correctly detected background pixels; This represents the number of falsely detected interfering pixels. It represents the number of falsely detected background pixels; Recall rate is expressed by the formula: .

7. The SAR radio frequency interference image domain detection method according to claim 1, characterized in that, The Focal loss function is expressed as follows: （9）； in, The addition of this allows the loss function to balance the relative importance of positive and negative samples; The predicted class probability has a value range of 1. ; This is called the modulation factor; Control the rate at which the weights decrease.

8. A SAR radio frequency interference image domain detection device, characterized in that, include: The receiving module is configured to periodically transmit pulse signals using the radar system and receive raw echo signals; The conversion module is configured to perform SAR imaging on the original echo signal to obtain a SAR complex image, and convert the SAR complex image into a grayscale image; The detection module is configured to use a trained intelligent interference detection network based on improved U2Net to detect the grayscale image formed by the original echo signal and obtain a prediction mask on the grayscale image. The determination module is configured to determine the interference region on the SAR complex image based on the prediction mask; the trained intelligent interference detection network based on the improved U2Net is obtained by constructing a training set and introducing the Focal loss function for constraint during the training process. The intelligent interference detection network based on the improved U2Net includes: an improved U2Net network and a channel attention module; The improved U2Net network includes multiple scale-variable codecs and three scale-invariant dilated convolutional modules. The multiple codecs and three dilated convolutional modules form a U-shaped structure. Each dilated convolutional module replaces its own dilated convolutional layer with a dilated spatial pyramid convolutional module.

Images