Broadband radar echo denoising method and device based on U-shaped self-attention model

Through the broadband radar echo denoising method based on the U-shaped self-attention model, the problem of loss of HRRP detailed information in the prior art is solved, and a higher recognition rate and denoising performance are achieved.

CN120161429APending Publication Date: 2025-06-17XIDIAN UNIV

Patent Information

Application Number
CN202510367708.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-03-26
Publication Date
2025-06-17

AI Technical Summary

Technical Problem

The prior art ignores the weak scattering center of the target in radar target recognition, resulting in the loss of HRRP details, affecting the recognition rate and denoising performance.

Method used

A broadband radar echo denoising method based on the U-shaped self-attention model is adopted to generate a training data set by adding Gaussian white noise, and a distance image denoising network including multiple encoders, intermediate modules and multiple decoders is constructed. The first target multi-head self-attention layer and wavelet denoising down-noising sampling layer are used to denoise the noise-containing high-resolution distance image echo data.

Benefits of technology

While ensuring the denoising performance, the distance image detail information is retained, which improves the recognition rate of the subsequent recognition network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120161429A_ABST
    Figure CN120161429A_ABST
Patent Text Reader

Abstract

The invention provides a broadband radar echo denoising method and device based on a U-shaped self-attention model, and relates to the technical field of radar signal processing. Comprising the following steps: acquiring high-resolution range profile echo data; gaussian white noise is added to the high-resolution range profile echo data to obtain a training data set; a range profile denoising network based on a U-shaped self-attention model is constructed, the range profile denoising network comprises a plurality of encoders, and each encoder comprises a first target multi-head self-attention layer added with a range profile supporting area mask and a wavelet denoising down-sampling layer; training the range profile denoising network by using the training data set to obtain a trained range profile denoising network; and inputting the noisy high-resolution range profile echo data into the trained range profile denoising network to denoise the noisy high-resolution range profile echo data, and outputting the denoised high-resolution range profile echo data. Therefore, the detail information of the range profile is reserved, and the recognition rate of the subsequent recognition network is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and particularly to a wideband radar echo denoising method and device based on a U-shaped self-attention model. Background Art

[0002] Radar target recognition is a key technology. Its core lies in analyzing the electromagnetic wave signals reflected by the target to extract its physical and geometric features, so as to realize the classification and attribute determination of the target. The high-resolution range profile (HRRP) of a radar is the projection vector sum of the sub-echoes of the target scattering points obtained by a high-resolution radar in the radar ray direction. The HRRP contains rich structural and feature information of the target, so it is widely used in the recognition task of target models. In a modern battlefield environment, a radar needs to identify non-cooperative targets at a long distance in the first time. When a non-cooperative target unit enters the limit detection distance of the radar, the radar echo power is significantly attenuated due to long-distance transmission, resulting in a significant reduction in the signal-to-noise ratio of the target's HRRP, seriously weakening the recognition accuracy of the radar and possibly causing misjudgment, which will have an adverse impact on subsequent decision-making and actions. Therefore, it is particularly important to ensure the accuracy of radar target recognition.

[0003] Currently, to ensure the accuracy of radar target recognition, a wideband radar complex echo denoising method based on a variational encoding-decoding network is usually adopted. This existing method takes the scattering center model theory as the core and believes that the strong scattering points of the target are sparse; based on its sparsity, the HRRP that retains the strong scattering centers of the target is obtained through a variational encoding-decoding network. However, this existing method ignores the weak scattering centers of the target, resulting in the loss of detailed information of the target HRRP, affecting the recognition rate of the variational encoding-decoding network and resulting in poor denoising performance. Summary of the Invention

[0004] The purpose of the embodiments of the present invention is to provide a wideband radar echo denoising method and device based on a U-shaped self-attention model, so as to solve the problem that the loss of detailed information of the target HRRP affects the recognition rate of the variational encoding-decoding network and results in poor denoising performance.

[0005] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] The first aspect of the present invention provides a wideband radar echo denoising method based on a U-shaped self-attention model, including:

[0007] Obtain high-resolution range profile echo data;

[0008] Add Gaussian white noise to the high-resolution range profile echo data so that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is within a preset range, and obtain a training data set;

[0009] Construct a range profile denoising network based on the U-shaped self-attention model. The range profile denoising network includes a plurality of encoders, an intermediate module, and a plurality of decoders connected in sequence. Each encoder includes a plurality of feature extraction modules, and each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer. The first target multi-head self-attention layer is a multi-head self-attention layer with a range profile support area mask added;

[0010] Train the range profile denoising network using the training data set to obtain a trained range profile denoising network;

[0011] Input the noisy high-resolution range profile echo data into the trained range profile denoising network, and use the trained range profile denoising network to denoise the noisy high-resolution range profile echo data, and output the denoised high-resolution range profile echo data.

[0012] A second aspect of the present invention provides a wideband radar echo denoising device based on a U-shaped self-attention model, including:

[0013] An acquisition module for acquiring high-resolution range profile echo data;

[0014] An addition module for adding Gaussian white noise to the high-resolution range profile echo data so that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is within a preset range, and obtaining a training data set;

[0015] A construction module for constructing a range profile denoising network based on the U-shaped self-attention model. The range profile denoising network includes a plurality of encoders, an intermediate module, and a plurality of decoders connected in sequence. Each encoder includes a plurality of feature extraction modules, and each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer. The first target multi-head self-attention layer is a multi-head self-attention layer with a range profile support area mask added;

[0016] A training module for training the range profile denoising network using the training data set to obtain a trained range profile denoising network;

[0017] A denoising module for inputting the noisy high-resolution range profile echo data into the trained range profile denoising network, and using the trained range profile denoising network to denoise the noisy high-resolution range profile echo data, and outputting the denoised high-resolution range profile echo data.

[0018] Compared with the prior art, the broadband radar echo denoising method and device based on the U-shaped self-attention model provided by the present invention acquire high-resolution range profile echo data; add Gaussian white noise to the high-resolution range profile echo data so that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is within a preset range to obtain a training data set; construct a range profile denoising network based on the U-shaped self-attention model, the range profile denoising network includes a plurality of encoders, an intermediate module, and a plurality of decoders connected in sequence, each encoder includes a plurality of feature extraction modules, and each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer, and the first target multi-head self-attention layer is a multi-head self-attention layer with a range profile support area mask added; use the training data set to train the range profile denoising network to obtain a trained range profile denoising network; input the noisy high-resolution range profile echo data into the trained range profile denoising network to use the trained range profile denoising network to denoise the noisy high-resolution range profile echo data and output the denoised high-resolution range profile echo data. In this way, the first target multi-head self-attention layer with a range profile support area mask added can assign more attention weights to the data in the target support area of the noisy high-resolution range profile echo data, reduce the attention to noise, and the wavelet denoising downsampling layer decomposes the intermediate features into high-frequency and low-frequency parts and can filter out high-frequency noise, so that while ensuring the denoising performance of the range profile denoising network, the details of the range profile are retained, and the recognition rate of the subsequent recognition network is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown by way of illustration and not limitation, and the same or corresponding reference numerals indicate the same or corresponding parts, wherein:

[0020] Figure 1 Schematically shows a flowchart of a broadband radar echo denoising method based on a U-shaped self-attention model;

[0021] Figure 2 Schematically shows a structural diagram of a range profile denoising network based on a U-shaped self-attention model;

[0022] Figure 3 Schematically shows a structural diagram of a broadband radar echo denoising device based on a U-shaped self-attention model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art.

[0024] It should be noted that: unless otherwise specified, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art to which the present invention pertains.

[0025] The method in the embodiments of the present invention will be described in detail below.

[0026] Figure 1 Schematically shown is a flowchart of a broadband radar echo denoising method based on a U-shaped self-attention model in the embodiments of the present invention. Refer to Figure 1 As shown, the broadband radar echo denoising method based on the U-shaped self-attention model may include:

[0027] S101. Obtain high-resolution range profile echo data.

[0028] The HRRP echo data is data of five types of aircraft acquired by any broadband radar. The HRRP echo data contains the radar one-dimensional high-resolution range profile, and the dimension of the HRRP echo data is 256 range cells.

[0029] S102. Add Gaussian white noise to the high-resolution range profile echo data so that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is within a preset range, thereby obtaining a training data set.

[0030] The preset range is the set range of the range profile signal-to-noise ratio, and the preset range is from 0 dB to 30 dB.

[0031] Specifically, Gaussian white noise is added to the high-resolution range profile echo data in the complex domain to obtain the noise-added high-resolution range profile echo data, such that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is uniformly distributed within 0 dB to 30 dB, and at the same time, the signal-to-noise ratio of the range profile of the noise-added high-resolution range profile echo data is saved.

[0032] After adding Gaussian white noise to the high-resolution range profile echo data so that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is within a preset range to obtain a training data set, the method further includes:

[0033] Step A1: Perform a modulo operation on the training data set to obtain a modulo data set.

[0034] Step A2: Perform centroid alignment on the modulo dataset to obtain the centroid-aligned dataset.

[0035] Step A3: Perform intensity normalization on the centroid-aligned dataset to obtain the target training dataset.

[0036] Use the target training dataset as the training dataset for the range image denoising network based on the U-shaped self-attention model.

[0037] S103. Construct a range image denoising network based on the U-shaped self-attention model.

[0038] Among them, the range image denoising network includes a plurality of encoders, an intermediate module, and a plurality of decoders connected in sequence. Each encoder includes a plurality of feature extraction modules, and each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer. The first target multi-head self-attention layer is a multi-head self-attention layer with a range image support area mask added.

[0039] Specifically, the number of the plurality of encoders is the same as the number of the plurality of decoders, and each encoder is skip-connected to each decoder.

[0040] The range image denoising network based on the U-shaped self-attention model is a U-shaped network in structure. The range image denoising network realizes cross-layer feature fusion through skip connections.

[0041] Exemplarily, the number of the plurality of encoders is 4, and the corresponding number of the plurality of decoders is also 4. The 4 encoders are the first encoder, the second encoder, the third encoder, and the fourth encoder respectively, and the 4 decoders are the first decoder, the second decoder, the third decoder, and the fourth decoder respectively. Among them, the first encoder is skip-connected to the first decoder, the second encoder is skip-connected to the second decoder, the third encoder is skip-connected to the third decoder, and the fourth encoder is skip-connected to the fourth decoder.

[0042] Specifically, each encoder further includes a first convolutional layer. The first convolutional layer is connected to a plurality of feature extraction modules. The plurality of decoders include a plurality of feature decoding modules connected in sequence and a second convolutional layer. The intermediate module is a Transfomer module.

[0043] The first convolutional layer is used to extract shallow features. The convolution kernel size is 1×3, and the moving step size is 1. LeakyReLU is used as the activation function, and the slope parameter is 0.2. The intermediate module is a Transfomer module. The input channel number of the Transfomer module is 128, the number of attention heads is 32, and the feature length is 16.

[0044] The convolution kernel size of the second convolutional layer is 1×3, the moving step size is 1, the input channel number is 8, and the output channel number is 1.

[0045] Continuing with the above example, Figure 2 A schematic diagram of the range image denoising network based on the U-shaped self-attention model is shown, see Figure 2 As shown, the number of multiple feature extraction modules and multiple feature decoding modules can be 4. The first encoder, the second encoder, the third encoder, and the fourth encoder all include a first convolutional layer and 4 feature extraction modules connected in sequence. The first decoder, the second decoder, the third decoder, and the fourth decoder all include 4 feature decoding modules and a second convolutional layer connected in sequence. The first encoder, the second encoder, the third encoder, the fourth encoder, the intermediate module, i.e., the Transfomer module, the first decoder, the second decoder, the third decoder, and the fourth decoder are connected in sequence.

[0046] Specifically, each feature extraction module further includes multiple first Transfomer sub-modules. The multiple first Transfomer sub-modules are connected to the wavelet denoising downsampling layer. Each first Transfomer sub-module includes a first normalization layer, a first target multi-head self-attention layer, a second normalization layer, and a first feed-forward neural sub-network connected in sequence.

[0047] The range image support region mask is a relative position encoding that includes the range image support region, providing an additional bias for the range image support region when calculating attention. The first feed-forward neural sub-network is a convolutional gated linear unit. The wavelet denoising downsampling layer uses wavelet downsampling to decompose the features into high-frequency information and low-frequency information through wavelet transform, and the feature length becomes half of the original. Then, concatenation is performed in the channel dimension, and filtering is performed through a 1×3 convolutional kernel to halve the feature length and double the number of channels.

[0048] Continuing with the above example, the 4 feature extraction modules are the first feature extraction module, the second feature extraction module, the third feature extraction module, and the fourth feature extraction module connected in sequence. The number of multiple first Transfomer sub-modules in the first feature extraction module, the second feature extraction module, the third feature extraction module, and the fourth feature extraction module is 2. The 2 first Transfomer sub-modules are the first preset Transfomer sub-module and the first target Transfomer sub-module respectively. The first preset Transfomer sub-module, the first target Transfomer sub-module, and the wavelet denoising downsampling layer are connected in sequence. The first preset Transfomer sub-module in the first feature extraction module is connected to the first convolutional layer, and the wavelet denoising downsampling layer in the fourth feature extraction module is connected to the Transfomer module.

[0049] Continuing with the above example, the number of input channels of the first feature extraction module, the second feature extraction module, the third feature extraction module, and the fourth feature extraction module are 8, 16, 32, and 64 respectively, and the output feature lengths are 256, 128, 64, and 32 respectively. The number of attention heads of the first target multi-head self-attention layer in each first Transfomer sub-module are 2, 4, 8, and 16 respectively. The window lengths of the range image support area mask are 64, 32, 16, and 8 respectively.

[0050] Specifically, each feature decoding module includes an upsampling layer and multiple second Transfomer sub-modules connected in sequence. Each second Transfomer sub-module includes a third normalization layer, a second target multi-head self-attention layer, a fourth normalization layer, and a second feed-forward neural sub-network connected in sequence. The second target multi-head self-attention layer is a multi-head self-attention layer with a range image support area mask added.

[0051] The upsampling layer uses a transposed convolution with a convolution kernel size of 1×2 and a stride of 2 for upsampling. The input features of each decoder are concatenated with the corresponding encoder output features through skip connections in the channel dimension, then the channel dimension is halved through a 1×1 convolution kernel, and then passed through the upsampling layer and each second Transfomer sub-module.

[0052] Continuing with the above example, the four feature decoding modules are the first feature decoding module, the second feature decoding module, the third feature decoding module, and the fourth feature decoding module connected in sequence. The number of multiple second Transfomer sub-modules in the first feature decoding module, the second feature decoding module, the third feature decoding module, and the fourth feature decoding module are all 2. The two second Transfomer sub-modules are the second preset Transfomer sub-module and the second target Transfomer sub-module respectively, and the upsampling layer, the second preset Transfomer sub-module, and the second target Transfomer sub-module are connected in sequence. The upsampling layer in the first feature decoding module is connected to the Transfomer module, and the second target Transfomer sub-module in the fourth feature decoding module is connected to the second convolutional layer.

[0053] Continuing with the above example, the number of input channels of the first feature decoding module, the second feature decoding module, the third feature decoding module, and the fourth feature decoding module are 16, 32, 64, and 128 respectively, and the output feature lengths are 256, 128, 64, and 32 respectively. The number of attention heads of the second target multi-head self-attention layer in each second Transfomer sub-module are 2, 4, 8, and 16 respectively. The window lengths of the range image support area mask are 64, 32, 16, and 8 respectively.

[0054] Specifically, the process of adding the multi-head self-attention layer for the distance image support area mask is as follows:

[0055] Step B1: Obtain the preset attention score calculation formula.

[0056] The preset attention score calculation formula is an existing attention score calculation formula.

[0057] Step B2: Incorporate the distance image support area mask into the preset attention score calculation formula to obtain the target attention score calculation formula, so as to determine the first target multi-head self-attention layer.

[0058] Specifically, the target attention score calculation formula is:

[0059]

[0060] Among them, Attention is the target attention score, Softmax(·) is the normalization function, Q is the query matrix, KT is the transpose of the key matrix, dk is the query vector dimension, Mask is the distance image support area mask, and V is the value matrix.

[0061] The determination process of the second target multi-head self-attention layer is the same as that of the first target multi-head self-attention layer.

[0062] Before obtaining the preset attention score calculation formula, the method further includes:

[0063] Step C1: Generate a learnable relative position bias parameter table according to the support area window size.

[0064] Step C2: Construct a support area mask matrix according to the learnable relative position bias parameter table.

[0065] Specifically, according to the index matrix corresponding to the support area window size, look up the bias corresponding to the index matrix in the learnable relative position bias parameter table to construct the support area mask matrix.

[0066] Step C3: Embed the support area mask matrix into the attention matrix to obtain the distance image support area mask.

[0067] Embedding the support area mask matrix into the attention matrix and setting the non-support area window region matrix except the support area mask matrix to zero can obtain the distance image support area mask.

[0068] The range image support region mask makes the range image denoising network based on the U-shaped self-attention model pay more attention to the reconstruction of signals in the support region by superimposing an additional learnable attention bias on the range image support region, reducing the influence of the noise region on the signals in the support region during feature extraction. The range image support region mask can constrain the range image denoising network based on the U-shaped self-attention model to assign more attention weights to the signal region and reduce the attention to the noise region.

[0069] S104. Use the training data set to train the range image denoising network to obtain a trained range image denoising network.

[0070] Input the training data set into the range image denoising network to train the range image denoising network using the mean square error loss function to obtain a trained range image denoising network.

[0071] The batch size during training is 64, the initial learning rate is 1e-4, the optimizer is AdamW, and the total number of iterations is set to 100.

[0072] The trained range image denoising network can preserve the echo details.

[0073] S105. Input the noisy high-resolution range image echo data into the trained range image denoising network to use the trained range image denoising network to denoise the noisy high-resolution range image echo data and output the denoised high-resolution range image echo data.

[0074] The trained range image denoising network fits the high-dimensional feature distribution of HRRP. After the range image denoising network based on the U-shaped self-attention model is trained, it can denoise new category samples without retraining, making the denoising generalization performance stronger; compared with the existing methods, the present invention has higher recovery accuracy for range images and better denoising performance.

[0075] Based on the above Figure 1From the implementation method, it can be seen that the embodiments of the present invention obtain high-resolution range profile echo data; add Gaussian white noise to the high-resolution range profile echo data so that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is within a preset range to obtain a training dataset; construct a range profile denoising network based on a U-shaped self-attention model, the range profile denoising network includes a plurality of encoders, an intermediate module, and a plurality of decoders connected in sequence, each encoder includes a plurality of feature extraction modules, and each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer, and the first target multi-head self-attention layer is a multi-head self-attention layer with a range profile support area mask added; use the training dataset to train the range profile denoising network to obtain a trained range profile denoising network; input the noisy high-resolution range profile echo data into the trained range profile denoising network to use the trained range profile denoising network to denoise the noisy high-resolution range profile echo data and output the denoised high-resolution range profile echo data. In this way, the first target multi-head self-attention layer with a range profile support area mask added can allocate more attention weights to the data in the target support area of the noisy high-resolution range profile echo data, reduce the attention to noise, and the wavelet denoising downsampling layer decomposes the intermediate features into high-frequency and low-frequency parts and can filter out high-frequency noise, so that while ensuring the denoising performance of the range profile denoising network, the details of the range profile are retained, and the recognition rate of the subsequent recognition network is improved.

[0076] Based on the same inventive concept, as an implementation of the above-mentioned broadband radar echo denoising method based on a U-shaped self-attention model, the embodiments of the present invention also provide a broadband radar echo denoising device based on a U-shaped self-attention model. Figure 3 For the structural diagram of the broadband radar echo denoising device based on the U-shaped self-attention model in the embodiments of the present invention, see Figure 3 As shown, the device may include:

[0077] An acquisition module 301, configured to acquire high-resolution range profile echo data;

[0078] An addition module 302, configured to add Gaussian white noise to the high-resolution range profile echo data so that the range profile signal-to-noise ratio of the noise-added high-resolution range profile echo data is within a preset range to obtain a training dataset;

[0079] A construction module 303, configured to construct a range profile denoising network based on a U-shaped self-attention model, the range profile denoising network includes a plurality of encoders, an intermediate module, and a plurality of decoders connected in sequence, each encoder includes a plurality of feature extraction modules, and each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer, and the first target multi-head self-attention layer is a multi-head self-attention layer with a range profile support area mask added;

[0080] A training module 304, configured to train the range image denoising network by using a training data set, so as to obtain a trained range image denoising network;

[0081] A denoising module 305, configured to input the noisy high-resolution range image echo data into the trained range image denoising network, so as to denoise the noisy high-resolution range image echo data by using the trained range image denoising network, and output the denoised high-resolution range image echo data.

[0082] During the process of adding the multi-head self-attention layer with the range image support area mask in the construction module 303: obtaining a preset attention score calculation formula; integrating the range image support area mask into the preset attention score calculation formula to obtain a target attention score calculation formula, so as to determine a first target multi-head self-attention layer.

[0083] In the construction module 303, the target attention score calculation formula is:

[0084]

[0085] Wherein, Attention is the target attention score, Softmax(·) is a normalization function, Q is a query matrix, KT is the transpose of the key matrix, dk is the query vector dimension, Mask is the range image support area mask, and V is the value matrix.

[0086] Before obtaining the preset attention score calculation formula, the construction module 303 is configured to generate a learnable relative position bias parameter table according to the support area window size; construct a support area mask matrix according to the learnable relative position bias parameter table; and embed the support area mask matrix into the attention matrix to obtain the range image support area mask.

[0087] The device further includes: a modulo processing module, configured to perform modulo processing on the training data set after adding Gaussian white noise to the high-resolution range image echo data to make the range image signal-to-noise ratio of the noisy high-resolution range image echo data within a preset range to obtain a modulo-processed data set;

[0088] A centroid alignment module, configured to perform centroid alignment on the modulo-processed data set to obtain a centroid-aligned data set;

[0089] An intensity normalization module, configured to perform intensity normalization on the centroid-aligned data set to obtain a target training data set.

[0090] It should be noted here that the description of the embodiments of the broadband radar echo denoising device based on the U-shaped self-attention model above is similar to the description of the embodiments of the broadband radar echo denoising method based on the U-shaped self-attention model, and has beneficial effects similar to those of the embodiments of the broadband radar echo denoising method based on the U-shaped self-attention model. For the technical details not disclosed in the embodiments of the broadband radar echo denoising device based on the U-shaped self-attention model of the present invention, please refer to the description of the embodiments of the broadband radar echo denoising method based on the U-shaped self-attention model of the present invention for understanding.

[0091] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, and all should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A broadband radar echo denoising method based on a U-type self-attention model, characterized in that: include: Acquire high-resolution range image echo data; Adding Gaussian white noise to the high-resolution range image echo data so that the range image signal-to-noise ratio of the high-resolution range image echo data after the noise is added is within a preset range, thereby obtaining a training data set; Constructing a range image denoising network based on a U-type self-attention model, the range image denoising network includes a plurality of encoders, an intermediate module and a plurality of decoders connected in sequence, each encoder includes a plurality of feature extraction modules, each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer, the first target multi-head self-attention layer is a multi-head self-attention layer to which a range image support area mask is added; Using the training data set to train the range image denoising network to obtain a trained range image denoising network; The noisy high-resolution range image echo data is input into the trained range image denoising network, so as to denoise the noisy high-resolution range image echo data using the trained range image denoising network, and output the denoised high-resolution range image echo data.

2. The broadband radar echo denoising method based on the U-type self-attention model according to claim 1 is characterized in that: The number of the plurality of encoders is the same as the number of the plurality of decoders, and each encoder is jump-connected to each decoder.

3. The broadband radar echo denoising method based on the U-type self-attention model according to claim 1 is characterized in that: Each encoder also includes a first convolution layer, which is connected to the multiple feature extraction modules. The multiple decoders include multiple feature decoding modules and a second convolution layer connected in sequence. The intermediate module is a Transfomer module.

4. The broadband radar echo denoising method based on the U-type self-attention model according to claim 3 is characterized in that: Each feature extraction module also includes multiple first Transfomer sub-modules, which are connected to the wavelet denoising downsampling layer. Each first Transfomer sub-module includes a first normalization layer, the first target multi-head self-attention layer, a second normalization layer and a first feedforward neural sub-network connected in sequence.

5. The broadband radar echo denoising method based on the U-type self-attention model according to claim 3 is characterized in that: Each feature decoding module includes an upsampling layer and multiple second Transfomer sub-modules connected in sequence. Each second Transfomer sub-module includes a third normalization layer, a second target multi-head self-attention layer, a fourth normalization layer and a second feedforward neural sub-network connected in sequence. The second target multi-head self-attention layer is a multi-head self-attention layer to which the range image support area mask is added.

6. The broadband radar echo denoising method based on the U-type self-attention model according to claim 1 is characterized in that: The process of adding a multi-head self-attention layer of range image support area mask is: Get the preset attention score calculation formula; The range image support area mask is integrated into the preset attention score calculation formula to obtain the target attention score calculation formula to determine the first target multi-head self-attention layer.

7. The broadband radar echo denoising method based on the U-type self-attention model according to claim 6 is characterized in that: The target attention score calculation formula is: Among them, Attention is the target attention score, Softmax(·) is the normalization function, Q is the query matrix, KT is the transpose of the key matrix, dk is the query vector dimension, Mask is the range image support area mask, and V is the value matrix.

8. The broadband radar echo denoising method based on the U-type self-attention model according to claim 6 is characterized in that: Before obtaining the preset attention score calculation formula, the method further includes: Generate a learnable relative position offset parameter table based on the support area window size; Constructing a support region mask matrix according to the learnable relative position bias parameter table; The support region mask matrix is ​​embedded into the attention matrix to obtain the range image support region mask.

9. The broadband radar echo denoising method based on the U-type self-attention model according to claim 1, characterized in that: After adding Gaussian white noise to the high-resolution range image echo data so that the range image signal-to-noise ratio of the high-resolution range image echo data after the noise is added is within a preset range and a training data set is obtained, the method further includes: Performing a modulo process on the training data set to obtain a modulo processed data set; Performing centroid alignment on the data set after modulo taking to obtain a data set after centroid alignment; The center-aligned data set is subjected to intensity normalization to obtain a target training data set.

10. A broadband radar echo denoising device based on a U-type self-attention model, characterized in that: include: An acquisition module is used to acquire high-resolution range image echo data; An adding module is used to add Gaussian white noise to the high-resolution range image echo data so that the range image signal-to-noise ratio of the high-resolution range image echo data after the noise is added is within a preset range, thereby obtaining a training data set; A construction module is used to construct a range image denoising network based on a U-type self-attention model, wherein the range image denoising network includes a plurality of encoders, an intermediate module and a plurality of decoders connected in sequence, each encoder includes a plurality of feature extraction modules, each feature extraction module includes a first target multi-head self-attention layer and a wavelet denoising downsampling layer, wherein the first target multi-head self-attention layer is a multi-head self-attention layer to which a range image support area mask is added; A training module, used to train the range image denoising network using the training data set to obtain a trained range image denoising network; The denoising module is used to input the noisy high-resolution range image echo data into the trained range image denoising network, so as to denoise the noisy high-resolution range image echo data using the trained range image denoising network, and output the denoised high-resolution range image echo data.

Citation Information

Patent Citations

  • Broadband radar complex echo denoising method based on variational coding-decoding network

    CN110118958A

  • HRRP sample generation method based on conditional denoising diffusion probability model

    CN116304701A

  • Radar target identification method and device based on moving window and attention model, and medium

    CN116794626A

  • Radar image denoising training method and device

    CN118584481A

Cited By

  • Echo enhancement and low signal-to-noise ratio robust identification method based on signal structure constraint

    CN121721592A

  • Radar image denoising method and system based on complex number perception double-branch Unet

    CN122151005A