Radar planar imaging sidelobe suppression method, device, electronic equipment and computer program product

Abstract

The present application provides a radar plane imaging sidelobe suppression method, device, electronic device and computer program product, the method comprising: obtaining a constructed training data set, the training data set comprising radar plane image samples containing side lobes and corresponding radar plane image samples without side lobes or after side lobe suppression; training the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model; obtaining a radar plane image to be processed; and adaptively suppressing the radar plane image to be processed based on the adaptive sidelobe suppression model. The solution of the embodiment of the present application is essentially to suppress the sidelobe of the radar by signal processing means, and to implement the direct processing of radar images in a fully deployed radar system without changing the hardware facilities, and dynamically adjusting the sidelobe suppression strategy according to different task requirements and scenarios.

Description

Technical Field

[0001] The present invention relates to the fields of radar signal processing and artificial intelligence, and in particular to a radar planar imaging sidelobe suppression method, device, electronic equipment and computer program product. Background Art

[0002] The multiple-input multiple-output (MIMO) virtual aperture radar imaging system can capture more angular information of the target with its advantages of low cost and high resolution. However, the presence of sidelobes in the radar images generated by the system will interfere with the accuracy of target detection and thus reduce the image quality. Therefore, the study of sidelobe suppression technology is of vital importance in the field of MIMO radar image processing.

[0003] Common sidelobe suppression algorithms are mainly discussed from the perspective of array optimization design. The sidelobe suppression technology based on antenna array optimization design mainly suppresses the sidelobes in the antenna radiation pattern by improving the antenna array layout, thereby improving the detection performance and anti-interference ability of the system. Usually, this can be achieved by reducing the number of array elements, adjusting the antenna spacing, or changing the antenna pointing direction. However, achieving sidelobe suppression by optimizing the antenna array actually requires changes to the hardware facilities, which limits the flexibility and adjustability of the solution, and thus affects the adaptability and maintainability of the radar system. Summary of the invention

[0004] In order to solve the above technical problems, the present application provides a radar planar imaging sidelobe suppression method, device, electronic device and computer program product to at least solve or alleviate the problems existing in the above-mentioned prior art.

[0005] In a first aspect, a radar planar imaging sidelobe suppression method is provided, comprising:

[0006] Acquire a constructed training data set, wherein the training data set includes radar plane image samples containing side lobes and corresponding radar plane image samples without side lobes or after side lobes are suppressed;

[0007] The adaptive neural network model is trained based on the training data set to obtain an adaptive sidelobe suppression model;

[0008] Acquire the radar plane image to be processed;

[0009] Adaptive sidelobe suppression is performed on the radar plane image to be processed based on the adaptive sidelobe suppression model.

[0010] Optionally, the adaptive neural network model includes an input layer, a first convolution layer, a second convolution layer, a pooling layer, an activation layer, a fully connected layer, and an output layer, wherein the first convolution layer has a deformable convolution kernel, and the second convolution layer has an offset addition convolution layer;

[0011] The step of training the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model includes:

[0012] Converting the radar plane image samples into an input feature map based on the input layer;

[0013] Performing deformable convolution processing on the input feature map based on the first convolution layer to obtain a first convolution layer output;

[0014] Based on the second convolutional layer, the output of the first convolutional layer is subjected to an offset addition and a convolution process to obtain an output of the second convolutional layer;

[0015] Performing downsampling processing on the output of the second convolutional layer based on the pooling layer to obtain a pooling layer output;

[0016] Based on the activation layer, a nonlinear transformation process is performed on the output of the pooling layer to obtain an activation layer output;

[0017] Performing linear transformation processing on the activation layer output based on the fully connected layer to obtain a fully connected layer output;

[0018] Performing classification processing on the output of the fully connected layer based on the output layer to obtain a prediction result of sidelobe suppression;

[0019] Based on the prediction result of the sidelobe suppression, a loss value is calculated to adjust the adaptive neural network model until the adaptive sidelobe suppression model is obtained.

[0020] Optionally, when performing deformable convolution processing on the input feature map based on the first convolution layer to obtain the first convolution layer output, the sampling position is adaptively adjusted according to the local features of the input feature map to capture the geometric transformation information of the target on the radar plane image sample and obtain the first convolution layer output accordingly.

[0021] Optionally, the second convolution layer includes an offset addition convolution layer, a standard convolution kernel and an offset convolution kernel;

[0022] The offset is added and convolution is performed on the output of the first convolution layer based on the second convolution layer to obtain the output of the second convolution layer, the offset of the sampling position is learned based on the offset addition convolution layer, and the offset is added to the sampling point of the standard convolution kernel to obtain the deformed sampling point; based on the deformed sampling point, the offset convolution kernel performs feature extraction on the output of the first convolution layer to obtain the output of the second convolution layer.

[0023] Optionally, when the offset-based convolution layer is added to learn the offset of the sampling position, the learned offset is reshaped into a position matrix, and the position matrix is ​​added to a sampling area consisting of sampling points on the input feature map corresponding to the standard convolution kernel, and the sampling area is dynamically adjusted according to the local features of the input feature map.

[0024] Optionally, when reshaping the learned offset into a position matrix, the learned offset is represented as an offset tensor, and the offset tensor is dimensionally adjusted according to the hyperparameters of the standard convolution kernel to obtain the position matrix, in which each element represents the new position of the sampling point of the standard convolution kernel.

[0025] Optionally, if the center position of the convolution kernel in the first convolution layer is , the positions of other sampling points on the convolution kernel in the first convolution layer are , then the relative position matrix R between the center position of the acquisition area on the standard convolution kernel and the positions of other sampling points on it is expressed as: .

[0026] Optionally, when the first convolution layer performs deformable convolution processing on the input feature map to obtain the first convolution layer output, the center position of the convolution kernel in the first convolution layer is Pixel value It is expressed as: , after the sampling position is adaptively adjusted according to the local features of the input feature map, the center position of the convolution kernel in the first convolution layer Pixel value Expressed as ,in Represents the offset of other sampling point positions on the convolution kernel in the first convolution layer.

[0027] Optionally, based on the deformed sampling points, the offset convolution kernel performs feature extraction on the first convolution layer output to obtain the second convolution layer output, and linear interpolation is performed on the pixel values ​​of the deformed sampling points to obtain deformed sampling pixel values, the deformed sampling pixel values ​​are convolved with the offset convolution kernel, and feature extraction is performed on the first convolution layer output to obtain the second convolution layer output.

[0028] Optionally, linear interpolation processing is performed on the pixel values ​​of the deformed sampling points based on the following formula to obtain deformed sampling pixel values: , , ,in, represents the deformed sampled pixel value, represents any sampling point after the deformation, represents the sampling position of the pixel value in the output of the first convolutional layer with an integer value, is the two-dimensional linear interpolation kernel, represents the linear interpolation weight in the x direction, (q x ,p x ) represents the coordinates of q and p in the x direction, (q y ,p y ) represents the coordinates of q and p in the y direction.

[0029] In a second aspect, a radar planar imaging sidelobe suppression device is provided, comprising:

[0030] A first program unit is used to obtain a constructed training data set, wherein the training data set includes radar plane image samples containing side lobes and corresponding radar plane image samples without side lobes or after side lobes are suppressed;

[0031] A second program unit is used to train the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model;

[0032] A third program unit is used to obtain a radar plane image to be processed;

[0033] The fourth program unit is used to perform adaptive sidelobe suppression on the radar plane image to be processed based on the adaptive sidelobe suppression model.

[0034] In a third aspect, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer executable program, and the processor executes the computer executable program to implement the method described in any one of the embodiments of the present application.

[0035] In a fourth aspect, a computer program product is provided, on which computer executable instructions are stored, and the computer executable instructions are executed to implement the method described in any one of the embodiments of the present application.

[0036] The solution of the embodiment of the present application is essentially to suppress the sidelobe of the radar through signal processing means, which can be implemented in an already deployed radar system. The radar image can be directly processed without changing the hardware facilities, and the sidelobe suppression strategy can be dynamically adjusted according to different task requirements and scenarios. Therefore, a more flexible and later adjustable solution makes the radar system more adaptable and maintainable. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The schematic embodiment drawings and their descriptions of this application are used to explain this application and do not constitute an improper limitation on this application. In the drawings:

[0038] Figure 1 It is a flow chart of a radar planar imaging sidelobe suppression method according to an embodiment of the present application.

[0039] Figure 2 The antenna array layout of a specific radar system according to an embodiment of the present application.

[0040] Figure 3 (a)-(b) shows the schematic diagram of the simulated point target position and the range-azimuth spectrum of the simulated target, respectively.

[0041] Figure 4 (a)-(d) are the original range-azimuth spectrum, the range-azimuth spectrum based on Hamming window sidelobe suppression, the range-azimuth spectrum based on SVA sidelobe suppression, and the range-azimuth spectrum based on DCN-SVA sidelobe suppression, respectively.

[0042] Figure 5 (a)-(b) are the azimuth profile (amplitude value) and the azimuth profile (dB), respectively.

[0043] Figure 6 It is the PSLR of different sidelobe suppression algorithms under different SNR conditions in the range-azimuth spectrum according to the embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the solution of the present application, the following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0045] Figure 1 FIG. 1 is a flow chart of a radar planar imaging sidelobe suppression method according to an embodiment of the present application. Figure 1 As shown, the radar planar imaging sidelobe suppression method comprises:

[0046] Acquire a constructed training data set, wherein the training data set includes radar plane image samples containing side lobes and corresponding radar plane image samples without side lobes or after side lobes are suppressed;

[0047] The adaptive neural network model is trained based on the training data set to obtain an adaptive sidelobe suppression model;

[0048] Acquire the radar plane image to be processed;

[0049] Adaptive sidelobe suppression is performed on the radar plane image to be processed based on the adaptive sidelobe suppression model.

[0050] The solution of the embodiment of the present application is essentially to suppress the sidelobe of the radar through signal processing means, which can be implemented in an already deployed radar system. The radar image can be directly processed without changing the hardware facilities, and the sidelobe suppression strategy can be dynamically adjusted according to different task requirements and scenarios. Therefore, a more flexible and later adjustable solution makes the radar system more adaptable and maintainable.

[0051] By acquiring radar plane image samples containing side lobes and corresponding image samples without side lobes or after side lobe suppression, a comprehensive and specific training data set is constructed. This data set not only covers various forms and intensities of side lobes, but also provides ideal suppression effects as target outputs, providing rich learning materials for the neural network model. Traditional algorithms often rely on specific mathematical models or assumptions and are difficult to adapt to complex and changeable side lobe situations. However, neural network training based on large data sets can automatically learn the complex mapping relationship between side lobes and ideal images, thereby overcoming the limitations of traditional algorithms. Furthermore, the adaptive neural network model has strong nonlinear fitting and self-learning capabilities, and can automatically adjust internal parameters to adapt to different input data. In the side lobe suppression task, this means that the model can dynamically adjust the suppression strategy according to the side lobe characteristics of different radar images to achieve personalized side lobe suppression. In addition, when processing side lobes, traditional algorithms usually use fixed filters or parameter settings, which makes it difficult to flexibly respond to side lobe changes in different images. The adaptive neural network model can automatically adjust the suppression strategy according to the characteristics of the input image, thereby suppressing the side lobes more accurately and retaining image details. By applying the adaptive neural network model to the sidelobe suppression task, this method can comprehensively utilize the information in the training data set, learn the complex mapping relationship between the sidelobe and the ideal image, and dynamically adjust it according to the characteristics of the input image in practical applications. This comprehensive processing method has significantly improved the sidelobe suppression capability. When processing sidelobes, traditional algorithms can often only optimize for a certain type of specific situation and find it difficult to take other situations into account. This method achieves comprehensive coverage and effective suppression of multiple sidelobe situations through the training and application of the adaptive neural network model, thereby improving the comprehensive ability of sidelobe suppression.

[0052] Optionally, the adaptive neural network model includes an input layer, a first convolution layer, a second convolution layer, a pooling layer, an activation layer, a fully connected layer, and an output layer, wherein the first convolution layer has a deformable convolution kernel, and the second convolution layer has an offset addition convolution layer;

[0053] The step of training the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model includes:

[0054] Converting the radar plane image samples into an input feature map based on the input layer;

[0055] Performing deformable convolution processing on the input feature map based on the first convolution layer to obtain a first convolution layer output;

[0056] Based on the second convolutional layer, the output of the first convolutional layer is subjected to an offset addition and a convolution process to obtain an output of the second convolutional layer;

[0057] Performing downsampling processing on the output of the second convolutional layer based on the pooling layer to obtain a pooling layer output;

[0058] Based on the activation layer, a nonlinear transformation process is performed on the output of the pooling layer to obtain an activation layer output;

[0059] Performing linear transformation processing on the activation layer output based on the fully connected layer to obtain a fully connected layer output;

[0060] Performing classification processing on the output of the fully connected layer based on the output layer to obtain a prediction result of sidelobe suppression;

[0061] Based on the prediction result of the sidelobe suppression, a loss value is calculated to adjust the adaptive neural network model until the adaptive sidelobe suppression model is obtained.

[0062] To this end, by combining innovative points such as deformable convolution kernels and offset-added convolution layers, an efficient and adaptive neural network model is constructed for radar planar imaging sidelobe suppression, which has the following technical benefits:

[0063] 1. Deformable convolution kernel: A deformable convolution kernel is used in the first convolution layer, which can adaptively adjust the sampling position according to the local features of the input feature map. This mechanism enables the model to more flexibly capture the geometric transformation information of the target on the radar plane image sample, thereby improving the accuracy and robustness of feature extraction. This is crucial for the sidelobe suppression task, because the sidelobes are often closely related to the geometry and position of the target. Offset-added convolution layer: The offset-added convolution layer is introduced in the second convolution layer to further enhance the model's ability to suppress sidelobes. By learning the offset of the sampling position and adding it to the sampling point of the standard convolution kernel, the model is able to more accurately locate and suppress the sidelobes while retaining the target information in the image.

[0064] 2. Hierarchical network structure: From the input layer to the output layer, the model gradually extracts and transforms image features through a progressive processing flow, and finally obtains the prediction results of sidelobe suppression. This hierarchical network structure enables the model to gradually learn complex mapping relationships, thereby improving the accuracy and stability of sidelobe suppression. Diversified processing layers: The model contains multiple processing layers such as convolutional layers, pooling layers, activation layers, and fully connected layers. Each processing layer has its own unique functions and effects. The combination of these processing layers enables the model to process image features more comprehensively and improve the comprehensive ability of sidelobe suppression.

[0065] 3. The model performs adaptive learning through training data sets and can automatically adjust internal parameters to adapt to different input data. This adaptive learning mechanism enables the model to more flexibly cope with various complex sidelobe situations, improving the robustness and generalization of sidelobe suppression. During the training process, the model calculates the loss value based on the predicted results of sidelobe suppression and optimizes the model parameters through the back propagation algorithm. This optimization mechanism enables the model to continuously approach the ideal sidelobe suppression effect, thereby improving the performance and accuracy of the model.

[0066] 4. By effectively suppressing side lobes and retaining image details, this method can significantly improve the quality of radar flat images. This is crucial for radar target detection and recognition applications, as high-quality images can provide more accurate and reliable information support. Applying this method to radar systems can enhance the overall performance and practicality of the system. By improving the sidelobe suppression capability, the radar system can detect and identify targets more accurately, thereby providing better quality and more efficient services for military, civilian and other fields.

[0067] Optionally, when performing deformable convolution processing on the input feature map based on the first convolution layer to obtain the first convolution layer output, the sampling position is adaptively adjusted according to the local features of the input feature map to capture the geometric transformation information of the target on the radar plane image sample and obtain the first convolution layer output accordingly.

[0068] In the radar plane image target detection task, the deformable convolution processing based on the first convolution layer is used to adaptively adjust the sampling position according to the local features of the input feature map, which has the following technical benefits:

[0069] (1) Adaptive sampling position adjustment makes the convolution kernel no longer limited to fixed grid positions, but can be flexibly moved to more informative positions. This flexibility helps the model better capture the geometric features of the target, especially when the target is deformed or changes its posture, so that it can more accurately locate and identify the target.

[0070] (2) By capturing the geometric transformation information of the target, the deformable convolution can simulate the actual motion trajectory of the target on the radar plane image, thereby extracting the target's features more accurately. This helps to improve the accuracy and robustness of feature extraction, allowing the model to maintain high detection performance when facing different scenes and different types of radar plane images.

[0071] (3) Deformable convolution can also reduce the amount of computation and memory usage, and improve processing efficiency. Traditional convolutional neural networks often require more computing resources and memory space when processing deformable targets. Deformable convolution can more effectively utilize the information of the input feature map by adaptively adjusting the sampling position, thereby reducing unnecessary computation and memory usage.

[0072] Optionally, the second convolution layer includes an offset addition convolution layer, a standard convolution kernel and an offset convolution kernel;

[0073] The offset is added and convolution is performed on the output of the first convolution layer based on the second convolution layer to obtain the output of the second convolution layer, the offset of the sampling position is learned based on the offset addition convolution layer, and the offset is added to the sampling point of the standard convolution kernel to obtain the deformed sampling point; based on the deformed sampling point, the offset convolution kernel performs feature extraction on the output of the first convolution layer to obtain the output of the second convolution layer.

[0074] In the radar planar imaging sidelobe suppression task, introducing a second convolution layer including an offset addition convolution layer, a standard convolution kernel, and an offset convolution kernel has the following benefits:

[0075] 1. The offset-adding convolution layer can adaptively adjust the sampling position of the standard convolution kernel by learning the offset of the sampling position. This mechanism enables the model to locate the sidelobes more accurately and separate them from the radar plane image. The learning of the offset is based on the local features of the input feature map, so it can capture the subtle differences between the sidelobes and the image background and improve the accuracy of sidelobe suppression.

[0076] The offset is added to the sampling points of the standard convolution kernel to obtain deformed sampling points. These sampling points can surround the sidelobe area more closely, thereby more effectively extracting sidelobe features. At the same time, due to the flexibility of the sampling position, the model can cope with sidelobes of different shapes and sizes, improving the robustness of sidelobe suppression.

[0077] Based on the deformed sampling points, the offset convolution kernel extracts features from the output of the first convolution layer. Due to the precise adjustment of the sampling position, the offset convolution kernel can more accurately capture the feature information of the sidelobe and provide more valuable input for subsequent processing.

[0078] 2. The adaptive learning mechanism of adding convolutional layers and offset convolution kernels enables the model to adjust the sampling position and convolution kernel parameters according to different input data. This flexibility helps the model cope with complex and changing radar plane images and improves the generalization ability of the model. By adaptively adjusting the sampling position and convolution kernel parameters, the model can better capture the essential characteristics of the input data instead of just memorizing the training data. This helps reduce the risk of overfitting and improves the performance of the model on unseen data.

[0079] 3. By adaptively adjusting the sampling position, the model can avoid unnecessary calculations for insignificant areas. This helps reduce the amount of calculations and improves processing efficiency. Due to the precise adjustment of the sampling position, the offset convolution kernel can extract sidelobe features more quickly. This helps speed up the entire processing flow and improve the real-time performance of radar planar imaging sidelobe suppression.

[0080] Optionally, when the offset-based convolution layer is added to learn the offset of the sampling position, the learned offset is reshaped into a position matrix, and the position matrix is ​​added to a sampling area consisting of sampling points on the input feature map corresponding to the standard convolution kernel, and the sampling area is dynamically adjusted according to the local features of the input feature map.

[0081] To this end, in the radar planar imaging sidelobe suppression task, an offset-based convolution layer is used to learn the offset of the sampling position, and the learned offset is reshaped into a position matrix and added to the sampling area of ​​the input feature map corresponding to the standard convolution kernel. This has the following technical benefits:

[0082] (1) The sampling area is dynamically adjusted according to the local features of the input feature map, which means that the model can adaptively determine the best sampling position for each local feature area. This dynamic adjustment helps to capture the characteristics of the sidelobe more accurately, thereby improving the accuracy of sidelobe suppression.

[0083] The offset-adding convolutional layer can learn accurate offsets by learning the local features of the input feature map. These offsets reflect the difference between the side lobes and the image background, which helps to separate the side lobes more accurately and preserve image details.

[0084] (2) By dynamically adjusting the sampling area and adding the learned offset, the model can adaptively cope with sidelobes of different shapes, sizes, and positions. This adaptive sampling mechanism enhances the robustness of the model, enabling it to handle complex and variable radar planar images. Since the model can learn the mapping relationship between the local features of the input feature map and the offset, it can show good generalization ability on unseen data. This is particularly important for the radar planar imaging sidelobe suppression task, because radar images in practical applications are often highly diverse and complex.

[0085] (3) By dynamically adjusting the sampling area and adding precise offsets, the model is able to avoid unnecessary calculations for insignificant areas. This helps reduce the amount of calculation and improves processing efficiency.

[0086] Due to the precise adjustment of the sampling position, the model can extract sidelobe features more quickly, which helps to speed up the entire processing flow and improve the real-time performance of radar planar imaging sidelobe suppression.

[0087] (4) This technical solution can improve the quality of radar planar images by accurately suppressing side lobes and preserving image details. This is crucial for subsequent image analysis and processing tasks, as they often rely on high-quality input images to extract useful information.

[0088] The suppression of side lobes helps reduce interference and noise in the image, thereby improving the readability of the image. This allows radar operators to more easily identify targets and make accurate judgments.

[0089] Optionally, when reshaping the learned offset into a position matrix, the learned offset is represented as an offset tensor, and the offset tensor is dimensionally adjusted according to the hyperparameters of the standard convolution kernel to obtain the position matrix, in which each element represents the new position of the sampling point of the standard convolution kernel.

[0090] To this end, in the radar planar imaging sidelobe suppression task, reshaping the learned offsets into a position matrix and resizing the offset tensor through the hyperparameters of the standard convolution kernel has the following technical benefits:

[0091] (1) The learned offset is represented in the form of a tensor, which provides a multi-dimensional description of the offset. The tensor can more accurately capture the differences in local features in the input feature map, thereby providing a personalized offset for each sampling point.

[0092] Through accurate offset characterization, the model can more accurately locate the sidelobe area and effectively suppress it, which helps to improve the quality of radar plane images and reduce the interference of sidelobes on target detection.

[0093] (2) Adjustment based on standard convolution kernel hyperparameters: The dimension of the offset tensor is adjusted based on the hyperparameters of the standard convolution kernel (such as kernel size, step size, etc.). This flexibility ensures that the offset can be accurately mapped to the sampling points of the standard convolution kernel, thereby achieving precise adjustment of the sampling position.

[0094] The flexible dimension adjustment mechanism enables the model to adapt to standard convolution kernels of different sizes and shapes, thereby improving the adaptability and generalization ability of the model. This is especially important for processing radar plane images of different resolutions and sizes.

[0095] (3) Through precise offset characterization and flexible dimensionality adjustment, the model can avoid unnecessary calculations for insignificant areas. This helps reduce computational redundancy and improves processing efficiency. Due to the precise adjustment of the sampling position, the model can extract image features more quickly. This helps speed up the entire processing flow and improve the real-time performance of radar planar imaging sidelobe suppression.

[0096] (4) Through precise offset adjustment and sampling position adjustment, the model can effectively suppress sidelobe interference and improve the signal-to-noise ratio and contrast of the image. Sidelobe suppression helps reduce noise and interference in the image, making the image clearer and easier to interpret. This is crucial for subsequent image analysis and processing tasks.

[0097] Optionally, if the center position of the convolution kernel in the first convolution layer is , the positions of other sampling points on the convolution kernel in the first convolution layer are , then the relative position matrix R between the center position of the acquisition area on the standard convolution kernel and the positions of other sampling points on it is expressed as: .

[0098] The values ​​of each matrix element in the position matrix R represent the relative position relationship between the center position of the acquisition area on the standard convolution kernel and the positions of other sampling points. The specific description is as follows:

[0099] (-1,-1): Indicates that the center position moves 1 unit to the left and 1 unit to the top.

[0100] (-1,0): Indicates that the center position moves only 1 unit to the left.

[0101] (-1,1): Indicates that the center position moves 1 unit to the left and 1 unit upward.

[0102] (0,-1): Indicates that the center position moves up only 1 unit.

[0103] (0,0): represents the center position itself, that is, the origin.

[0104] (0,1): Indicates that the center position moves down only 1 unit.

[0105] (1,-1): Indicates that the center position moves 1 unit to the right and 1 unit upward (but this description also needs to consider the coordinate system and be understood as a description when the relative position is in the next right grid).

[0106] (1,0): Indicates that the center position moves only 1 unit to the right.

[0107] (1,1): Indicates that the center position moves 1 unit to the right and 1 unit to the bottom.

[0108] Optionally, the first convolution layer output is obtained by performing deformable convolution processing on the input feature map based on the first convolution layer.

[0109] Optionally, based on the deformed sampling points, the offset convolution kernel performs feature extraction on the first convolution layer output to obtain the second convolution layer output, and linear interpolation is performed on the pixel values ​​of the deformed sampling points to obtain deformed sampling pixel values, the deformed sampling pixel values ​​are convolved with the offset convolution kernel, and feature extraction is performed on the first convolution layer output to obtain the second convolution layer output.

[0110] To this end, the deformed sampling points are linearly interpolated and convolved with the offset convolution kernel to extract features from the first convolution layer output and obtain the second convolution layer output, which has technical benefits:

[0111] (1) Linear interpolation can more accurately reflect the details and texture of the image, making the pixel values ​​of the deformed sampling points smoother and more reasonable. This helps the offset convolution kernel to more accurately capture the key features in the image, thereby improving the accuracy of feature extraction and the generalization ability of the model.

[0112] (2) When processing deformed, rotated or non-rigidly transformed images, linear interpolation can reduce the impact of image deformation on feature extraction. By adjusting the position of the sampling points and performing linear interpolation, the model can better adapt to different types of image data and improve the robustness and adaptability of the model.

[0113] (3) Linear interpolation processing enables the deformed sampling points to more accurately locate the key areas in the image. This helps the model to more accurately identify and locate targets when performing tasks such as target detection and segmentation, thereby improving positioning accuracy and task accuracy.

[0114] Optionally, linear interpolation processing is performed on the pixel values ​​of the deformed sampling points based on the following formula to obtain deformed sampling pixel values: , , ,in, represents the deformed sampled pixel value, represents any sampling point after the deformation, represents the sampling position of the pixel value in the output of the first convolutional layer with an integer value, is the two-dimensional linear interpolation kernel, represents the linear interpolation weight in the x direction, (q x ,p x )express , the coordinate of p in the x direction, (q y ,p y )express , the coordinate of p in the y direction.

[0115] The following uses an application in a specific scenario as an example to illustrate the technical benefits of the present application solution.

[0116] like Figure 2 As shown in the antenna array layout of a specific radar system, the yellow icons represent 12 transmitting antennas, 9 of which are spaced 2 The transmitting antennas are located in the horizontal dimension, with 3 spacings of 0.5 The transmitting antennas are located in the elevation dimension; the blue icons represent 16 receiving antennas, including three groups of antennas A, B, and C located in the horizontal dimension, with group A and group B separated by 4. , Group B and Group C are separated by 16 , and the antenna spacing in each group is 0.5 Through MIMO virtual aperture technology, the azimuth dimension theoretically contains 144 pixels with a spacing of 0.5 There are actually 86 non-overlapping virtual receiving antennas, so the azimuth angle resolution is 1.33°.

[0117] As shown in Figure 3 (a)-(b), the schematic diagram of the simulated point target position and the range-azimuth spectrum of the simulated target are shown respectively. A "I"-shaped sidelobe appears in each target azimuth dimension, and the weaker target points are masked by the sidelobe. Within the angular resolution range, the stronger target points are synthesized into a single target, which ultimately leads to a decrease in the quality of the radar image. In order to verify the effectiveness of the sidelobe suppression method (hereinafter referred to as DCN-SVA) proposed in this application, the range-azimuth spectrum before and after sidelobe suppression, the azimuth profile, and the sidelobe peak ratio under different signal-to-noise ratio conditions were selected as evaluation indicators, and compared with the windowing method and the traditional SVA algorithm.

[0118] As shown in Figure 4 (a)-(d), they are the original range-azimuth spectrum, the range-azimuth spectrum based on Hamming window sidelobe suppression, the range-azimuth spectrum based on SVA sidelobe suppression, and the range-azimuth spectrum based on DCN-SVA sidelobe suppression. The effect of applying Hamming window, SVA (Spatial Variant Aperture) and the scheme of the present application to the range-azimuth spectrum of Figure 3 (b) is shown. It can be observed that the Hamming window technology effectively suppresses all sidelobes in the azimuth dimension, but at the same time it also weakens weak target signals and causes the main lobe of other targets to widen. Although the SVA algorithm retains weak target signals and maintains the main lobe width, the main lobe energy of some targets is damaged. Figure 4 (d) shows the results of the sidelobe suppression of the scheme of the present application. Compared with the SVA algorithm and the Hamming window technology, the image processed by this scheme suppresses the sidelobes while better maintaining the integrity of the main lobe energy and weak target signals.

[0119] As shown in Figure 5 (a)-(b), the azimuth profile (amplitude value) and the azimuth profile (dB) are obtained by taking the logarithm of the data, as shown in Figure 5 (b). This figure more intuitively reveals the sidelobe level before and after sidelobe suppression. The relevant PSLR (peak sidelobe ratio) and ILSR (integrated sidelobe ratio) data have been recorded in detail in Table 1. By comparing the image and the table data, it can be clearly concluded that the DCN-SVA neural network performs better in sidelobe suppression than the traditional method. Specifically, this method achieves lower PSLR and ILSR values. In the case where two adjacent strong targets are difficult to distinguish due to the merging of sidelobes, the DCN-SVA neural network can effectively distinguish the two strong targets and maintain the integrity of the mainlobe energy. In addition, in the case where a weak target is obscured by a high sidelobe, this method can effectively retain the visibility of the weak target while suppressing the sidelobe.

[0120] Table 1 PSLR and ISLR before and after sidelobe suppression in the range-azimuth spectrum

[0121] Original Image Hamming Window Weighted SVA Weighted DCN-SVA PSLR -3.54 dB -11.04 dB -14.79 dB -30.41 dB ISLR -4.67 dB -11.85 dB -16.23 dB -32.46 dB

[0122] PSLR under different SNR (Signal-to-Noise Ratio) conditions can be used to measure the robustness of each algorithm. For strong targets, Gaussian noise is introduced into the simulated radar echo data under the conditions of -10 dB, -15 dB, -20 dB and -30 dB signal-to-noise ratio. Subsequently, imaging processing is performed on each set of data, and different image processing techniques are used. The obtained peak sidelobe level (PSLR) data are recorded in Table 2. According to the data in the table, Figure 6The sidelobe suppression effects of various algorithms under different signal-to-noise ratio environments are demonstrated. The study found that when the signal-to-noise ratio is lower than -15 dB, the sidelobe level of the unprocessed image is approximately -3.5 dB. After applying the window function and SVA algorithm, the sidelobe suppression effect is relatively weak, and the suppressed sidelobe levels are approximately -11 dB and -20 dB, respectively. According to the study, this level of sidelobes may still lead to false detection of targets. In contrast, the DCN-SVA-based method exhibits stronger sidelobe suppression capabilities under various signal-to-noise ratio conditions, and its suppression effect gradually increases with the decrease of the signal-to-noise ratio, reaching a maximum sidelobe suppression level of approximately -35.92dB. However, observations Figure 6 It can be seen that when the signal-to-noise ratio is greater than or equal to -10 dB, the noise will overwhelm the target, causing all sidelobe suppression techniques to fail.

[0123] Table 2 Peak sidelobe ratio of different algorithms under different SNR conditions in range-azimuth spectrum

[0124] SNR Original Image Hamming Window Weighted SVA Weighted DCN-SVA -25 dB -3.93 dB -12.73 dB -20.18 dB -35.92 dB -20 dB -3.8 dB -11.47 dB -16.01 dB -34.43 dB -15 dB -3.01 dB -11.04 dB -14.79 dB -30.42 dB -10 dB -2.99 dB -3.13 dB -3.53 dB -1.2 dB

[0125] In a second aspect, a radar planar imaging sidelobe suppression device is provided, comprising:

[0126] A first program unit is used to obtain a constructed training data set, wherein the training data set includes radar plane image samples containing side lobes and corresponding radar plane image samples without side lobes or after side lobes are suppressed;

[0127] A second program unit is used to train the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model;

[0128] A third program unit is used to obtain a radar plane image to be processed;

[0129] The fourth program unit is used to perform adaptive sidelobe suppression on the radar plane image to be processed based on the adaptive sidelobe suppression model.

[0130] In the above device, the exemplary explanation of the technical processing performed by each functional unit can be found in the above Figure 1 Records of.

[0131] In a third aspect, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer executable program, and the processor executes the computer executable program to implement the method described in any one of the embodiments of the present application.

[0132] In a fourth aspect, a computer program product is provided, on which computer executable instructions are stored, and the computer executable instructions are executed to implement the method described in any one of the embodiments of the present application.

[0133] For exemplary explanations of the third and fourth aspects, please refer to the above Figure 1 Records of.

[0134] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A radar planar imaging sidelobe suppression method, characterized in that: include: Acquire a constructed training data set, wherein the training data set includes radar plane image samples containing side lobes and corresponding radar plane image samples without side lobes or after side lobes are suppressed; The constructed adaptive neural network model is trained based on the training data set to obtain an adaptive sidelobe suppression model; Acquire the radar plane image to be processed; Performing adaptive sidelobe suppression on the radar plane image to be processed based on the adaptive sidelobe suppression model; The adaptive neural network model includes an input layer, a first convolution layer, a second convolution layer, a pooling layer, an activation layer, a fully connected layer, and an output layer, wherein the first convolution layer has a deformable convolution kernel, and the second convolution layer has an offset addition convolution layer; The step of training the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model includes: Converting the radar plane image samples into an input feature map based on the input layer; Performing deformable convolution processing on the input feature map based on the first convolution layer to obtain a first convolution layer output; Based on the second convolutional layer, the output of the first convolutional layer is subjected to an offset addition and a convolution process to obtain an output of the second convolutional layer; Performing downsampling processing on the output of the second convolutional layer based on the pooling layer to obtain a pooling layer output; Based on the activation layer, a nonlinear transformation process is performed on the output of the pooling layer to obtain an activation layer output; Performing linear transformation processing on the activation layer output based on the fully connected layer to obtain a fully connected layer output; Performing classification processing on the output of the fully connected layer based on the output layer to obtain a prediction result of sidelobe suppression; Based on the prediction result of the sidelobe suppression, a loss value is calculated to adjust the adaptive neural network model until the adaptive sidelobe suppression model is obtained; When the first convolution layer performs deformable convolution processing on the input feature map to obtain the first convolution layer output, the sampling position is adaptively adjusted according to the local features of the input feature map to capture the geometric transformation information of the target on the radar plane image sample and obtain the first convolution layer output accordingly; Wherein, the second convolution layer includes an offset addition convolution layer, a standard convolution kernel and an offset convolution kernel; When performing offset addition and convolution processing on the output of the first convolution layer based on the second convolution layer to obtain the output of the second convolution layer, the offset of the sampling position is learned by the convolution layer based on the offset addition, and the offset is added to the sampling point of the standard convolution kernel to obtain the deformed sampling point; based on the deformed sampling point, the offset convolution kernel is used to perform feature extraction on the output of the first convolution layer to obtain the output of the second convolution layer; Wherein, when the offset of the sampling position is learned by adding a convolution layer based on the offset, the learned offset is reshaped into a position matrix, and the position matrix is ​​added to the sampling area composed of the sampling points on the input feature map corresponding to the standard convolution kernel, and the sampling area is dynamically adjusted according to the local features of the input feature map; wherein, when the learned offset is reshaped into a position matrix, the learned offset is represented as an offset tensor, and the offset tensor is dimensionally adjusted according to the hyperparameters of the standard convolution kernel to obtain the position matrix, in which each element represents the new position of the sampling point of the standard convolution kernel; Among them, if the center position of the convolution kernel in the first convolution layer is , the positions of other sampling points on the convolution kernel in the first convolution layer are , then the relative position matrix R between the center position of the acquisition area on the standard convolution kernel and the positions of other sampling points on it is expressed as: ; When the first convolution layer output is obtained by performing deformable convolution processing on the input feature map based on the first convolution layer, the center position of the convolution kernel in the first convolution layer is Pixel value It is expressed as: , after the sampling position is adaptively adjusted according to the local features of the input feature map, the center position of the convolution kernel in the first convolution layer Pixel value Expressed as ,in Represents the offset of other sampling point positions on the convolution kernel in the first convolution layer; Wherein, based on the deformed sampling point, the offset convolution kernel performs feature extraction on the output of the first convolution layer, and when obtaining the output of the second convolution layer, linear interpolation processing is performed on the pixel value of the deformed sampling point to obtain a deformed sampled pixel value, the deformed sampled pixel value is convolved with the offset convolution kernel, and feature extraction is performed on the output of the first convolution layer to obtain the output of the second convolution layer; The pixel values ​​of the deformed sampling points are linearly interpolated based on the following formula to obtain deformed sampling pixel values: , , ,in, represents the deformed sampled pixel value, represents any sampling point after the deformation, represents the sampling position of the pixel value in the output of the first convolutional layer with an integer value, is the two-dimensional linear interpolation kernel, represents the linear interpolation weight in the x direction, (q x , p x ) represents the coordinates of q and p in the x direction, ( q y , p y ) represents the coordinates of q and p in the y direction.

2. A radar planar imaging sidelobe suppression device, characterized in that: include: A first program unit is used to obtain a constructed training data set, wherein the training data set includes radar plane image samples containing side lobes and corresponding radar plane image samples without side lobes or after side lobes are suppressed; A second program unit is used to train the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model; A third program unit is used to obtain a radar plane image to be processed; A fourth program unit, configured to perform adaptive sidelobe suppression on the radar plane image to be processed based on the adaptive sidelobe suppression model; The adaptive neural network model includes an input layer, a first convolution layer, a second convolution layer, a pooling layer, an activation layer, a fully connected layer, and an output layer, wherein the first convolution layer has a deformable convolution kernel, and the second convolution layer has an offset addition convolution layer; The step of training the constructed adaptive neural network model based on the training data set to obtain an adaptive sidelobe suppression model includes: Converting the radar plane image samples into an input feature map based on the input layer; Performing deformable convolution processing on the input feature map based on the first convolution layer to obtain a first convolution layer output; Based on the second convolutional layer, the output of the first convolutional layer is subjected to an offset addition and a convolution process to obtain an output of the second convolutional layer; Performing downsampling processing on the output of the second convolutional layer based on the pooling layer to obtain a pooling layer output; Based on the activation layer, a nonlinear transformation process is performed on the output of the pooling layer to obtain an activation layer output; Performing linear transformation processing on the activation layer output based on the fully connected layer to obtain a fully connected layer output; Performing classification processing on the output of the fully connected layer based on the output layer to obtain a prediction result of sidelobe suppression; Based on the prediction result of the sidelobe suppression, a loss value is calculated to adjust the adaptive neural network model until the adaptive sidelobe suppression model is obtained; When the first convolution layer performs deformable convolution processing on the input feature map to obtain the first convolution layer output, the sampling position is adaptively adjusted according to the local features of the input feature map to capture the geometric transformation information of the target on the radar plane image sample and obtain the first convolution layer output accordingly; Wherein, the second convolution layer includes an offset addition convolution layer, a standard convolution kernel and an offset convolution kernel; When performing offset addition and convolution processing on the output of the first convolution layer based on the second convolution layer to obtain the output of the second convolution layer, the offset of the sampling position is learned by the convolution layer based on the offset addition, and the offset is added to the sampling point of the standard convolution kernel to obtain the deformed sampling point; based on the deformed sampling point, the offset convolution kernel is used to perform feature extraction on the output of the first convolution layer to obtain the output of the second convolution layer; Wherein, when the offset of the sampling position is learned by adding a convolution layer based on the offset, the learned offset is reshaped into a position matrix, and the position matrix is ​​added to a sampling area consisting of sampling points on the input feature map corresponding to the standard convolution kernel, and the sampling area is dynamically adjusted according to the local features of the input feature map; Wherein, when the learned offset is reshaped into a position matrix, the learned offset is represented as an offset tensor, and the offset tensor is dimensionally adjusted according to the hyperparameters of the standard convolution kernel to obtain the position matrix, in which each element represents a new position of a sampling point of the standard convolution kernel; Among them, if the center position of the convolution kernel in the first convolution layer is , the positions of other sampling points on the convolution kernel in the first convolution layer are , then the relative position matrix R between the center position of the acquisition area on the standard convolution kernel and the positions of other sampling points on it is expressed as: ; When the first convolution layer output is obtained by performing deformable convolution processing on the input feature map based on the first convolution layer, the center position of the convolution kernel in the first convolution layer is Pixel value It is expressed as: , after the sampling position is adaptively adjusted according to the local features of the input feature map, the center position of the convolution kernel in the first convolution layer Pixel value Expressed as ,in Represents the offset of other sampling point positions on the convolution kernel in the first convolution layer; Wherein, based on the deformed sampling point, the offset convolution kernel performs feature extraction on the output of the first convolution layer, and when obtaining the output of the second convolution layer, linear interpolation processing is performed on the pixel value of the deformed sampling point to obtain a deformed sampled pixel value, the deformed sampled pixel value is convolved with the offset convolution kernel, and feature extraction is performed on the output of the first convolution layer to obtain the output of the second convolution layer; The pixel values ​​of the deformed sampling points are linearly interpolated based on the following formula to obtain deformed sampling pixel values: , , ,in, represents the deformed sampled pixel value, represents any sampling point after the deformation, represents the sampling position of the pixel value in the output of the first convolutional layer with an integer value, is the two-dimensional linear interpolation kernel, represents the linear interpolation weight in the x direction, (q x , p x ) represents the coordinates of q and p in the x direction, ( q y , p y ) represents the coordinates of q and p in the y direction.

3. An electronic device, characterized in that: The invention comprises a memory and a processor, wherein a computer executable program is stored in the memory, and the processor executes the computer executable program to implement the method according to claim 1.

4. A computer program product, characterized in that Computer executable instructions are stored thereon, and the computer executable instructions are executed to implement the method of claim 1.

Images