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

Abstract

The application discloses a radar image denoising method and system based on a complex perception double-branch Unet, and belongs to the field of image denoising. The method comprises the following steps: firstly, obtaining historical echo signals of a radar used for target detection, and obtaining radar images corresponding to the historical echo signals to form a training set; then, constructing a neural network based on the complex perception double-branch Unet, and training the neural network by using the training set; finally, obtaining echo signals of the radar, and obtaining a radar image corresponding to the echo signals, performing denoising processing on the radar image by using the trained neural network, obtaining a denoised radar image, and realizing the denoising of the radar image. The application can realize the denoising of the radar image under different signal-to-clutter ratio environments.

Description

Technical Field

[0001] This invention belongs to the fields of signal processing, image denoising and target detection, and specifically relates to a radar image denoising method and system based on complex sensing dual-branch Unet. Background Technology

[0002] Radar is one of the most effective tools for detecting unmanned aerial vehicle (UAV) targets. Radar-based target detection refers to the detection and localization of targets by preprocessing and using target detection algorithms based on the echo signals received by the radar, thereby obtaining relevant target information. With its advantage of operating in all weather conditions and around the clock, radar can capture target distance and speed information, providing an effective means for target detection and tracking.

[0003] Pulse Doppler radar typically detects targets based on range Doppler images, using methods such as statistical model-based and feature-based approaches. Range Doppler images (i.e., radar images) are obtained by signal accumulation and Fourier transform of echo signals. When environmental clutter and noise interference are strong, target signals are easily overwhelmed, leading to a decline in the performance of detection algorithms. Therefore, denoising range Doppler images and improving image quality is a crucial step to enhance detection efficiency and performance. Traditional radar image denoising methods include singular value decomposition (SVD), empirical mode decomposition (EMD), and wavelet transform-based methods. However, these methods require prior assumptions about background clutter, which drastically reduces performance in today's complex and dynamic environments. Deep learning, on the other hand, can automatically extract latent features of targets through training, achieving better image denoising and reconstruction. It exhibits strong robustness and generalization ability, thus deep neural networks are increasingly being used for radar image denoising. Current deep learning denoising algorithms based on range Doppler images cannot fully exploit the inherent characteristics of radar data due to their simple network architecture. This leads to a decline in network performance when faced with weak target signals or target aliasing.

[0004] Denoising algorithms based on range Doppler images can enhance image quality and thus improve detection efficiency. However, traditional methods are limited by model assumptions, and currently there are few deep learning-based algorithms that can fully exploit the inherent characteristics of radar data. Summary of the Invention

[0005] To address the problems in the prior art, this invention proposes a radar image denoising method and system based on complex sensing dual-branch Unet.

[0006] The technical solution adopted in this invention is as follows:

[0007] In a first aspect, the present invention discloses a radar image denoising method based on complex sensing dual-branch Unet, comprising the following steps:

[0008] 1) Acquire historical echo signals from the radar used for target detection, and obtain the radar images corresponding to the historical echo signals to form a training set;

[0009] 2) Construct a neural network based on complex-aware dual-branch Unet, which is used for denoising radar images. The neural network includes a complex-aware encoder module, a bottleneck module, and a decoder module. The encoder module includes multiple encoder blocks connected in series. Each encoder block includes a frequency domain enhancement module with an improved decoupling mechanism, a complex-aware fusion strategy module, a convolutional layer, a hybrid attention mechanism module, and a downsampling layer connected in series. The complex-aware fusion strategy module is used to promote feature interaction of radar images in complex form and enhance target feature extraction.

[0010] 3) Train the neural network using the training set obtained in step 1) to obtain a trained neural network;

[0011] 4) Obtain the radar echo signal and obtain the radar image corresponding to the echo signal. Then, use the trained neural network to denoise the radar image to obtain the denoised radar image, thus realizing radar denoising.

[0012] Secondly, this invention discloses a radar image denoising system based on complex sensing dual-branch Unet, comprising:

[0013] The neural network construction and training module is used to acquire historical echo signals of radar used for target detection, obtain radar images corresponding to the historical echo signals, form a training set, construct a neural network based on complex perception dual-branch Unet, and the neural network is used to denoise the radar images. It includes an encoder module, a bottleneck module and a decoder module based on complex perception, and then uses the training set to train the neural network.

[0014] The encoder module includes multiple encoder blocks connected in series. Each encoder block includes a frequency domain enhancement module with an improved decoupling mechanism, a complex perception fusion strategy module, a convolutional layer, a hybrid attention mechanism module, and a downsampling layer, all connected in series. The complex perception fusion strategy module is used to promote feature interaction of radar images in complex form and enhance target feature extraction.

[0015] The denoising module is used to denoise the radar image corresponding to the echo signal using a trained neural network to obtain a denoised radar image, thereby achieving denoising of the radar image.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1) This invention proposes a neural network for denoising range-Doppler images in the radar field. The real and imaginary parts of the complex-form range-Doppler image (radar image) are input into the dual-branch encoder module of the neural network, fully exploiting the inherent characteristics of the complex-form range-Doppler image. This overcomes the problems of insufficient data information mining and poor algorithm robustness in existing technologies, thereby achieving effective denoising and target reconstruction of range-Doppler images. This method effectively solves the problem of poor denoising performance caused by weak target signals and target aliasing in high signal-to-clutter environments.

[0018] 2) This invention proposes a frequency domain enhancement module with an improved decoupling mechanism, which utilizes the frequency domain characteristics of the target to achieve frequency domain enhancement, overcoming the problem of poor denoising effect caused by weak target signal and target aliasing, and introduces a frequency domain loss function to further improve network performance.

[0019] 3) This invention proposes a complex number perception fusion strategy. The complex number perception fusion strategy module can promote the interaction and fusion between the real and imaginary parts of complex number data, enabling the network to learn the salient features of the target.

[0020] 4) This invention proposes a hybrid attention mechanism module, which consists of a convolutional block attention module and a strip convolution module. Since targets are usually presented as stripes in range Doppler images, this module can effectively extract target features. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the basic steps of the radar image denoising method based on complex sensing dual-branch Unet of the present invention.

[0022] Figure 2 This is a schematic diagram of the encoder block of the present invention;

[0023] Figure 3 This is a schematic diagram of the frequency domain enhancement module of the improved decoupling mechanism of the present invention;

[0024] Figure 4 This is a schematic diagram of the complex number perception fusion strategy module of the present invention;

[0025] Figure 5 This is a schematic diagram of the hybrid attention mechanism module of the present invention;

[0026] Figure 6 This is a schematic diagram of the decoder block of the present invention;

[0027] Figure 7 This is a schematic diagram of the radar image denoising system based on complex sensing dual-branch Unet according to the present invention;

[0028] Figure 8The images show the experimental results after denoising using the simulated distance Doppler image dataset and different methods according to the embodiments of the present invention. Detailed Implementation

[0029] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0030] The object of this invention is the down-converted radar echo signal. As a well-known technology in the field, down-conversion is the process of converting a high-frequency signal into a lower-frequency signal. By reducing the frequency and retaining bandwidth and phase information, the purpose is to facilitate subsequent processing and analysis.

[0031] like Figure 1 The diagram shown is a flowchart illustrating the basic steps of the radar image denoising method based on complex sensing dual-branch Unet of the present invention. The method of the present invention mainly includes the following steps:

[0032] Step 1): Constructing the training set

[0033] First, historical radar echo signals for target detection are acquired. Then, the historical echo signals are preprocessed to obtain the corresponding range Doppler images (i.e., radar images). Finally, all range Doppler images are used to construct a training set. Here, the historical echo signals refer to the down-converted historical radar echo signals.

[0034] In this embodiment, signal preprocessing includes pulse compression and Doppler processing methods, that is, pulse compression and Doppler processing methods are used to obtain a distance Doppler image.

[0035] make This represents a one-dimensional sequence of historical echo signals. The formula for calculating pulse compression, indicating the time, is as follows:

[0036]

[0037] in, The signal obtained by conjugating the radar transmitted signal after deconvolution; This represents a one-dimensional sequence of historical echo signals after pulse compression.

[0038] For multiple pulse echoes, the time... writing ,in Indicates slow time; This indicates the number of received pulse echoes (i.e., echo signals in pulse form); Indicates the pulse repetition interval; This represents fast time, that is, the time sequence within one pulse repetition interval. Then, the one-dimensional sequence of the historical echo signal after pulse compression... Represented as a two-dimensional echo sequence .

[0039] Along the fast time dimension Perform a Fast Fourier Transform to obtain the range-Doppler image matrix corresponding to the historical echo signal. The calculation formula is as follows:

[0040]

[0041] in, This indicates the total number of pulses processed. In this embodiment, =256; Indicates frequency; Indicates the index of the pulse; Indicates the imaginary part.

[0042] The distance Doppler image matrix As a range Doppler image (i.e., radar image) It consists of the real part and the virtual part composition.

[0043] Step 2): Construct a neural network based on a complex-aware dual-branch Unet. The neural network includes a complex-aware encoder module, a bottleneck module, and a decoder module. The complex-aware encoder module consists of an improved frequency domain enhancement module (i.e., a frequency domain enhancement module with an improved decoupling mechanism), a complex-aware fusion strategy module, convolutional layers, a hybrid attention mechanism module, and downsampling layers. It is used to parse the input range-Doppler image and perform feature encoding on the range-Doppler image to obtain latent features. The bottleneck module is used to perform dimensionality transformation and feature mapping on the latent features. The decoder module includes multiple decoder blocks, each consisting of an upsampling layer, a skip connection layer, a convolutional layer, and a hybrid attention mechanism module. It is used to reconstruct the image and output a denoised range-Doppler image. In this embodiment, random parameters are used as the initial weights of the network.

[0044] The specific working process of each module in the complex-aware dual-branch Unet neural network is as follows:

[0045] Step 21): The complex-aware encoder module aims to perform feature encoding on the input range Doppler image, extract target feature information for subsequent image reconstruction. See [link to encoder block structure] for details. Figure 2The complex-aware encoder module consists of two branch encoder modules. Each branch encoder module includes multiple encoder blocks connected in series. Each encoder block consists of an improved frequency domain enhancement module, a complex-aware fusion strategy module, a convolutional layer, a hybrid attention mechanism module, and a downsampling layer.

[0046] For the input complex form of the distance Doppler image Decompose it into real part and the virtual part The data are input into different branch encoder modules. The number of decoder blocks is equal to the number of encoder blocks in a single branch encoder module.

[0047] In each branch encoder module, the real (or imaginary) part of the range Doppler image first passes through an improved frequency domain enhancement module. This module enhances weak and aliased targets in the spatial domain through frequency domain processing, improving the poor reconstruction effect caused by weak target signals and target aliasing. Its structure is as follows: Figure 3 As shown, Figure 3 (a) in the diagram represents the original frequency domain enhancement module, which will... Figure 3 The gray dashed box in (a) of the text is modified to... Figure 3 The content of (b) in the above is the improved frequency domain enhancement module.

[0048] The input to the improved frequency domain enhancement module (i.e., the real or imaginary part of the range Doppler image) is denoted as... First, perform layer normalization; the calculation formula is as follows:

[0049]

[0050] in, This represents the output of the layer normalization function before the Fast Fourier Transform; The representation layer normalization function.

[0051] Then, a Fast Fourier Transform is performed on the normalized output to transform the features to the frequency domain, obtaining a frequency domain complex tensor. Subsequently, the obtained frequency domain complex tensor is decoupled, and its amplitude components are extracted. and phase components The amplitude component typically contains texture and energy information of an image, while the phase component contains structural and contour information. Global max pooling and global average pooling are used to obtain salient features and global statistical background from the amplitude component, respectively, resulting in a first vector describing the amplitude features. Second vector Then, the two vectors describing the magnitude features are concatenated by channels and fed into a transformation module consisting of convolutional layers and a sigmoid activation function to generate an adaptive attention weight coefficient that matches the dimension of the original magnitude component. The generated adaptive attention weight coefficients are multiplied element-wise with the original amplitude components to enhance important frequency features and suppress noise frequencies. To preserve the stability of the original information, the processed result... The residuals of the original amplitude component are summed to obtain the enhanced amplitude component. The calculation formula is as follows:

[0052]

[0053]

[0054]

[0055]

[0056] in, Indicates Fast Fourier Transform; Indicates global average pooling; Indicates global max pooling; Indicates the convolution operation; This represents the Sigmoid activation function; This represents the dot product operation.

[0057] At the same time, through a learnable weight matrix The original phase components are weighted and mapped, and the structure weights in the phase space are adjusted to obtain the adjusted phase components. This ensures that the spatial structure consistency of phase information is maintained while the amplitude is enhanced; the calculation formula is as follows:

[0058]

[0059] The enhanced amplitude component With the adjusted phase components The components are then recombined, restored to a complex frequency domain tensor using Euler's formula, and then mapped back from the frequency domain to the time domain using the inverse fast Fourier transform (IFFT). The calculation formula is as follows:

[0060]

[0061] in, This represents the inverse fast Fourier transform.

[0062] Finally, after layer normalization and multilayer perceptron processing, the processed result is added to the original input as a residual to obtain the output of the improved frequency domain enhancement module. The calculation formula is as follows:

[0063]

[0064]

[0065] in, This represents the output of the layer normalization function after the Fast Fourier Transform; This indicates the output of the improved frequency domain enhancement module; The function representing a multilayer perceptron.

[0066] Then, the output of the improved frequency domain enhancement module will be sent to the complex sensing fusion strategy module. This module achieves the intrinsic interaction and feature mining of radar data by fusing the feature data of the real and imaginary branches. Its structure is as follows: Figure 4 As shown.

[0067] First, obtain the output features from the previous encoder block of the opposing branch. The background information is then locally averaged to extract smooth background information, followed by feature mapping through a convolutional layer to obtain smooth features. The output features are then... By subtracting the smoothed feature element by element, the output features of the opposing branch can be extracted. High-frequency components The calculation formula is as follows:

[0068]

[0069] in, This indicates local average pooling.

[0070] Then, the output of the improved frequency domain enhancement module in the current encoder block of the current branch is... With high frequency components Element-wise addition is performed, and the result is then fed into a convolutional layer for inter-channel feature compression and mapping. Finally, the output of the convolutional layer is mapped to the (0, 1) interval using a sigmoid activation function, generating an adaptive attention map with spatial and channel constraints. This enables the interaction of feature information between different branches, utilizing high-frequency details of opposing branches to supplement the representational capabilities of the current branch; its calculation formula is as follows:

[0071]

[0072] The generated adaptive attention map Compared to the output of the improved frequency domain enhancement module in the current encoder block Element-wise multiplication and residual connection are performed to obtain the final output features of the complex number perception fusion strategy module. The calculation formula is as follows:

[0073]

[0074] The final output features obtained The features are fed into the convolutional layer of the current encoder block for feature mapping to obtain the features. The calculation formula is as follows:

[0075]

[0076] Subsequently, features The hybrid attention mechanism module, fed into the current encoder block, further extracts target features and improves target saliency through an attention mechanism. Its structure is as follows: Figure 5 As shown. Features output by the convolutional layer. As input to this module, the input first passes through the Convolutional Block Attention (CBAM) module. The CBAM module processes the input features by performing parallel computations of channel attention and spatial attention. Perform preliminary global importance calibration and output intermediate feature tensors. The calculation formula is as follows:

[0077]

[0078] in, This represents a function that represents the attention module of a convolutional block.

[0079] Then, the intermediate feature tensor The data is fed into two parallel strip convolution modules, each using a 1×5 horizontal or vertical convolution kernel on the intermediate feature tensor. Feature extraction is performed. This step aims to capture narrow strip features with long-range dependencies on the target in the distance-Doppler image. Then, the Sigmoid activation function is applied to the features extracted from the horizontal and vertical branches respectively to generate the corresponding horizontal attention maps. and vertical attention map Subsequently, the two attention maps are added element-wise and merged to obtain a bar attention map that covers both the vertical and horizontal dimensions. The calculation formula is as follows:

[0080]

[0081]

[0082] in, and These represent convolution operations with kernels of 1×5 and 5×1, respectively.

[0083] The resulting bar attention map With intermediate feature tensor Element-wise multiplication is performed to obtain the final output features of the hybrid attention mechanism module of the current encoder block. The calculation formula is as follows:

[0084]

[0085] In the above hybrid attention mechanism module, from the features To output features The mapping is defined as Function, that is:

[0086]

[0087] The obtained output features It will be used as the feature output for skip links in the decoder module. The input is the feature output of the branch encoder module, which is the real part of the distance from the Doppler image. Real part jump connection feature The input is the feature output of the branch encoder module, which is the imaginary part of the distance from the Doppler image. Imaginary part skip connection feature .

[0088] Secondly, the downsampling layer of the current encoder block is used to refine the output features. Perform downsampling to obtain the output features of the current encoder block. This output feature is then fed into the next encoder module of the current branch; its calculation formula is as follows:

[0089]

[0090]

[0091] in, This represents the downsampling function.

[0092] Output characteristics of the upper-level cascaded encoder block The input is fed to the next-level cascaded encoder blocks. After processing by M cascaded encoder blocks, the complex-aware encoder module finally outputs the features. In this embodiment, M=3.

[0093] Step 22): The bottleneck module mainly performs feature mapping on the extracted feature vectors to obtain the output features. The output feature is used for subsequent decoding and reconstruction. The bottleneck module consists of a convolutional layer and a hybrid attention mechanism module; its calculation formula is as follows:

[0094]

[0095] Step 23): The decoder module reconstructs the image through upsampling and convolution operations. During the reconstruction process, clutter and noise are suppressed, and the denoised distance Doppler image is output. The structure of the decoder block is as follows: Figure 6 As shown, the decoder module consists of an upsampling layer, a skip connection layer, a convolutional layer, and a hybrid attention mechanism module, ultimately outputting a denoised range Doppler image.

[0096] Output characteristics of the bottleneck module As input to the decoder module, the first decoder block in the decoder module first processes the output features. Upsampling is performed to increase the spatial resolution to a size that matches the current decoding stage; the calculation formula is as follows:

[0097]

[0098] in, This represents the upsampling function.

[0099] Then the upsampled features Real skip connection features corresponding to encoder blocks from the encoding phase sequence and the feature of imaginary skip connections Concatenate along the channel dimension to obtain the concatenated feature vector. The calculation formula is as follows:

[0100]

[0101] Then, the concatenated feature vectors Convolutional operations are applied for channel compression and feature integration, which are then fed into a hybrid attention mechanism module to output the final decoded features of the current decoder block. The calculation formula is as follows:

[0102]

[0103] Output features of the upper-level cascaded decoder block The input is fed to the next-level cascaded decoder block. After processing by M cascaded decoder blocks, the decoder module finally outputs... , This is the final image output by the neural network. In this embodiment, M=3. The height of the output image; This is the width of the output image.

[0104] Step 3): Neural Network Training

[0105] Complex data from the distance-Doppler images obtained through Fourier transform are used as training samples. The neural network employs supervised training. To alleviate the imbalance between the target and background and the small proportion of the target in the image, Dice and focal loss functions are used as loss functions for image reconstruction. Simultaneously, to ensure the effective operation of the improved frequency domain enhancement module, a frequency domain MSE loss function is introduced. Total loss function. It is expressed as follows:

[0106]

[0107]

[0108]

[0109]

[0110] in, Represents the Dice loss function; Represents the focal loss function; The MSE loss function in the frequency domain; and These represent the denoised distance-Doppler image output by the neural network and the ground truth image corresponding to the distance-Doppler image input to the neural network, respectively. It is a smoothing factor; and All of these are used to control the weights of the samples; This represents the mean square error function. This indicates that the absolute value is being calculated. and These are all weighting factors, in this embodiment , .

[0111] This invention updates network weights using Adam gradient descent with adaptive learning rate adjustment. In this embodiment, the training is iterated 300 times.

[0112] Step 4): Acquire the radar echo signal and obtain the range Doppler image corresponding to the echo signal. Then, use the trained neural network to denoise the range Doppler image to obtain the denoised range Doppler image, thus achieving denoising of the range Doppler image.

[0113] The method for obtaining the range Doppler image corresponding to the echo signal is the same as the method for obtaining the range Doppler image corresponding to the historical echo signal in step 1).

[0114] Figure 7 This is a block diagram of a radar image denoising system based on complex sensing dual-branch Unet, according to an embodiment. The system includes:

[0115] The neural network construction and training module is used to acquire historical echo signals of radar used for target detection and obtain range Doppler images corresponding to the historical echo signals to form a training set. It constructs a neural network based on complex sensing dual-branch Unet, which is used to denoise the range Doppler images. The neural network includes an encoder module, a bottleneck module, and a decoder module based on complex sensing. The neural network based on complex sensing dual-branch Unet is trained using the training set to obtain a trained neural network.

[0116] The encoder module includes multiple encoder blocks connected in series. Each encoder block includes a frequency domain enhancement module with an improved decoupling mechanism, a complex number perception fusion strategy module, a convolutional layer, a hybrid attention mechanism module, and a downsampling layer, all connected in series. The complex number perception fusion strategy module is used to promote feature interaction of the distance Doppler image in complex form and enhance target feature extraction.

[0117] The denoising module is used to acquire the radar echo signal and obtain the range Doppler image corresponding to the echo signal. Then, the trained neural network is used to denoise the range Doppler image to obtain the denoised range Doppler image, thus realizing the denoising of the range Doppler image (i.e., the radar image).

[0118] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0119] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative. For example, the image preprocessing module can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules can be combined or integrated into another unit. Furthermore, the connection between the modules shown or discussed can be a communication connection through some interfaces, which can be electrical or other forms. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort. The following uses simulated distance Doppler images as an example to illustrate specific implementation methods to demonstrate the technical effects of the present invention. Specific steps in the embodiments will not be repeated.

[0120] Example

[0121] Next, using a simulation dataset as the research object, we will verify the radar image denoising method of this invention. To comprehensively evaluate the denoising effect, we will present it from both visual and quantitative perspectives. The Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Measure (SSIM), and Noise Suppression Ratio (NSR) are used for evaluation, expressed by the following formulas:

[0122]

[0123]

[0124]

[0125]

[0126] in, This represents the distance Doppler image input to the neural network; and They represent and The value of the i-th pixel; It is the maximum pixel value at the distance from the Doppler image; and These represent the mean values ​​of the distance-Doppler image input to the neural network and the denoised distance-Doppler image output by the neural network, respectively. and These represent the variances of the distance-Doppler image input to the neural network and the denoised distance-Doppler image output by the neural network, respectively. This represents the covariance between the distance-Doppler image input to the neural network and the denoised distance-Doppler image output by the neural network. and These are all constants that avoid having a denominator of 0.

[0127] The radar and target parameters for the simulation dataset are set as follows: the transmitted signal is a linear frequency modulated signal with a carrier frequency of 3 GHz, a bandwidth of 5 MHz, a pulse width of 5 μs, a pulse repetition frequency of 2 kHz, and a sampling frequency of 30 MHz; the target distance is between 0.5 km and 5.0 km, and the target speed is between 0 m / s and 40 m / s. Figure 8 The first column shows the noisy range-Doppler images and the ground truth images corresponding to the denoised range-Doppler images with signal-to-clutter ratios (SCR) of -25dB, -15dB, and -5dB, respectively. The second to fifth columns are the denoising results obtained by the simulation dataset using the embodiments of the present invention and different radar image denoising algorithms.

[0128] Table 1 Evaluation Indicators of Detection Results from Simulated Radar Dataset

[0129]

[0130] The contrastive method SVD (Singular Value Decomposition) is derived from: M. Poon, R. Khan, and S. Le-Ngoc, “A singular value decomposition (SVD) based method for suppressing ocean clutter in high frequency radar,” IEEE Transactions on Signal Processing, vol. 41, no. 3, pp. 1421-1425, 1993.

[0131] The comparison method SCS-CNN (Selective Reconstruction of Features based Convolutional Neural Network) is from: L. Wen, C. Zhong, X. Huang, and J. Ding, “Sea clutter suppression based on selective reconstruction of features,” in 2019 6th Asia-Pacific Conference on Synthetic Aperture Radar (APSAR), pp. 1–6, 2019.

[0132] The comparative method CycleGAN (Cycle Generative Adversarial Network) is from: Z. Wu, J. Pei, W. Huo, Y. Huang, Y. Zhang, and H. Yang, “A machine learning approach to clutter suppression for marine surveillance radar,” in 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, pp.3137–3140, 2021.

[0133] The comparison method, Unet-CNN (Unet based Convolutional Neural Network), is from: J. Williams, L. Rosenberg, V. Stamatescu, and T.-T. Cao, “Maritime radar target detection using convolutional neural networks,” in 2022 IEEE Radar Conference (RadarConf22), pp. 1–6, 2022.

[0134] The denoising results of the simulation dataset are as follows Figure 8 As shown in Table 1, the quantization results demonstrate that this invention achieves the best performance under various signal-to-noise ratio (SNR) conditions. While all methods perform well in scenarios with high SNR, this method maintains a smaller reconstruction error and higher structural similarity even when the SNR decreases. Furthermore, based on the experimental results with an SCR of -15 dB in the visualization results, this method can clearly distinguish and reconstruct aliased targets, while other methods perform poorly.

[0135] This invention is based on the complex data form of range Doppler images and focuses on exploring the inherent characteristics of radar complex data during the encoding process. It solves the problem of poor image reconstruction effect caused by weak target signals and target aliasing. It specifically designs an improved frequency domain enhancement module, a complex perception fusion strategy module, and a hybrid attention mechanism module. It uses a weighted sum of Dice loss, focal loss, and frequency domain MSE loss as the loss function to achieve efficient denoising and reconstruction of range Doppler images under different signal-to-noise ratio environments.

[0136] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A radar image denoising method based on complex sensing dual-branch Unet, characterized in that, Includes the following steps: 1) Acquire historical echo signals from the radar used for target detection, and obtain the radar images corresponding to the historical echo signals to form a training set; 2) Construct a neural network based on complex-aware dual-branch Unet, which is used for denoising radar images. The neural network includes a complex-aware encoder module, a bottleneck module, and a decoder module. The encoder module includes multiple encoder blocks connected in series. Each encoder block includes a frequency domain enhancement module with an improved decoupling mechanism, a complex-aware fusion strategy module, a convolutional layer, a hybrid attention mechanism module, and a downsampling layer connected in series. The complex-aware fusion strategy module is used to promote feature interaction of radar images in complex form and enhance target feature extraction. 3) Train the neural network using the training set obtained in step 1); 4) Use the trained neural network to denoise the radar image corresponding to the echo signal to obtain the denoised radar image, thus achieving radar image denoising.

2. The radar image denoising method based on complex sensing dual-branch Unet according to claim 1, characterized in that, In step 1), the radar image corresponding to the historical echo signal is obtained, including: First, pulse compression is performed on the historical echo signal to obtain a two-dimensional sequence of the pulse-compressed historical echo signal. Then, a fast Fourier transform is performed on the two-dimensional sequence to obtain the radar image matrix of the historical echo signal. The radar image matrix is ​​used as the radar image corresponding to the historical echo signal. The formula for calculating pulse compression is: ； in, This refers to the signal obtained by conjugating the radar transmitted signal after inversion. A one-dimensional sequence representing the historical echo signal after pulse compression; A one-dimensional sequence representing historical echo signals; Then let This represents a two-dimensional sequence of historical echo signals after pulse compression, where Indicates slow time; Indicates the number of received echo signals in pulse form; It indicates fast time, that is, the time sequence within a pulse repetition interval.

3. The radar image denoising method based on complex sensing dual-branch Unet according to claim 1, characterized in that, In step 2), the encoder module is used to encode the radar image to obtain latent features; the bottleneck module is used to perform dimensional transformation and feature mapping on the latent features to obtain output features. The decoder module is used to reconstruct the image based on the output features obtained from the bottleneck module, thus obtaining a denoised radar image. The encoder module consists of two branch encoder modules, each of which includes multiple encoder blocks connected in series; the bottleneck module includes sequentially connected convolutional layers and a hybrid attention mechanism module; the decoder module includes multiple decoder blocks connected in series, each of which includes sequentially connected upsampling layers, skip connection layers, convolutional layers, and a hybrid attention mechanism module.

4. The radar image denoising method based on complex sensing dual-branch Unet according to claim 3, characterized in that, The radar image is split into real and imaginary parts, and the real and imaginary parts are input into different branch encoder modules respectively; The input of the branch encoder module is the input of the frequency domain enhancement module of the first encoder block in the branch encoder module; the input of the complex sensing fusion strategy module of the a-th encoder block in the branch encoder module is the output of the frequency domain enhancement module of the a-th encoder block in the branch encoder module, and the output of the (a-1)-th encoder block in another branch encoder module.

5. The radar image denoising method based on complex sensing dual-branch Unet according to claim 1, characterized in that, After receiving the input, the frequency domain enhancement module first performs layer normalization and Fourier transform on the input sequentially to obtain a frequency domain complex tensor. Then, it decouples the frequency domain complex tensor to obtain amplitude and phase components. The amplitude components are then processed using global max pooling and global average pooling respectively. The processing results are then concatenated into channels and sequentially input into the convolutional layer and activation function layer to generate adaptive attention weight coefficients. These adaptive attention weight coefficients are multiplied element-wise with the amplitude components, and the residuals are summed to obtain the enhanced amplitude components. Simultaneously, the phase components are weighted using a learnable weight matrix to obtain adjusted phase components. The enhanced amplitude components and adjusted phase components are combined and then subjected to an inverse fast Fourier transform. The results are then sequentially input into the layer normalization layer and the multilayer perceptron. Finally, the output of the multilayer perceptron is summed with the residuals of the input to the frequency domain enhancement module to obtain the output of the frequency domain enhancement module.

6. The radar image denoising method based on complex sensing dual-branch Unet according to claim 4, characterized in that, After receiving the input, the complex-aware fusion strategy module of the a-th encoder block in the branch encoder module performs local average pooling and convolution on the output of the (a-1)-th encoder block in another branch encoder module to obtain smooth features. Then, the output of the (a-1)-th encoder block is subtracted element-wise from the smooth features to obtain high-frequency components. Next, the output of the frequency domain enhancement module of the a-th encoder block in the current branch encoder module is added element-wise to the high-frequency components. After addition, the high-frequency components are input into the convolutional layer and activation function layer to obtain an adaptive attention map. Finally, the adaptive attention map is multiplied element-wise with the output of the frequency domain enhancement module of the a-th encoder block in the current branch encoder module and a residual connection is made to obtain the output of the complex-aware fusion strategy module.

7. The radar image denoising method based on complex sensing dual-branch Unet according to claim 1, characterized in that, After receiving the input, the hybrid attention mechanism module first processes the input using the convolutional block attention module to obtain an intermediate feature tensor. The intermediate feature tensor is then fed into two parallel strip convolution modules for feature extraction. The Sigmoid activation function is applied to the extraction results, and the results are then fused element-wise to obtain a strip attention map. Finally, the strip attention map is multiplied element-wise with the intermediate feature tensor to obtain the final output of the hybrid attention mechanism module. The two strip convolution modules each have a 1×5 horizontal convolution kernel and a vertical convolution kernel.

8. The radar image denoising method based on complex sensing dual-branch Unet according to claim 3, characterized in that, The input of the first decoder block in the decoder module is the output of the bottleneck module. After receiving the input, the a-th decoder block in the decoder module first upsamples the input, and then concatenates the upsampled result with the output of the hybrid attention mechanism module of the M-a+1-th encoder block in each branch encoder module to obtain the concatenated feature vector. The concatenated feature vector passes through the convolutional layer and the hybrid attention mechanism module in sequence to obtain the output of the a-th decoder block; M represents the number of encoder blocks in the branch encoder module.

9. The radar image denoising method based on complex sensing dual-branch Unet according to claim 1, characterized in that, In step 3), the neural network is trained with the goal of minimizing the total loss function, which includes the Dice loss, focal loss, and frequency domain MSE loss.

10. A radar image denoising system based on complex sensing dual-branch Unet that implements the method of claim 1, characterized in that, include: The neural network construction and training module is used to acquire historical echo signals of radar used for target detection, obtain radar images corresponding to the historical echo signals, form a training set, construct a neural network based on complex perception dual-branch Unet, and the neural network is used to denoise the radar images. It includes an encoder module, a bottleneck module and a decoder module based on complex perception, and then uses the training set to train the neural network. The encoder module includes multiple encoder blocks connected in series. Each encoder block includes a frequency domain enhancement module with an improved decoupling mechanism, a complex perception fusion strategy module, a convolutional layer, a hybrid attention mechanism module, and a downsampling layer, all connected in series. The complex perception fusion strategy module is used to promote feature interaction of radar images in complex form and enhance target feature extraction. The denoising module is used to denoise the radar image corresponding to the echo signal using a trained neural network to obtain a denoised radar image, thereby achieving denoising of the radar image.

Images