Roadside ISAR vehicle target imaging micro-Doppler interference suppression method based on ISPA-Net
By constructing the ISPA-Net architecture and utilizing recursive neural networks and the LeakyReLU activation function, the micro-Doppler interference in inverse synthetic aperture radar imaging is effectively suppressed, the imaging quality of vehicle targets is improved, the problem of imaging quality degradation caused by the micro-Doppler effect in existing technologies is solved, and high-quality target recognition and classification are achieved.
Patent Information
- Application Number
- CN202510716276.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-30
- Publication Date
- 2025-09-16
AI Technical Summary
In existing technology, in inverse synthetic aperture radar imaging, the micro-Doppler effect of vehicle targets causes a decrease in imaging quality, affecting target recognition and classification. Existing methods are difficult to effectively suppress micro-Doppler interference.
An ISPA-Net-based method is used to construct a training dataset, introduce a recursive neural network mechanism and LeakyReLU activation function, design a combined loss function, and build an ISPA-Net architecture to suppress micro-Doppler interference. The spatial attention block and residual block are used to eliminate interference in ISAR images.
It effectively suppresses micro-Doppler interference, improves the readability and recognizability of ISAR images, and provides high-quality target recognition and classification information for smart transportation.
Smart Images
Figure CN120652402A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of radar imaging technology, and in particular relates to a method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net. Background Art
[0002] Inverse Synthetic Aperture Radar (ISAR) is a radar imaging technology used to acquire high-resolution images of moving targets. It has important applications in military and civilian fields such as space early warning and target surveillance. Roadside ISAR uses radars placed on both sides of the road to achieve all-day, all-weather, high-resolution ISAR imaging of vehicle targets, thereby providing information support for vehicle-road collaboration and intelligent transportation. However, in practical applications, vehicle targets have their own particularities: in addition to overall translational motion, the wheels also rotate. When using radar to perform ISAR imaging of moving vehicle targets, the ISAR image is affected by the micro-Doppler (mD) effect generated by the wheel rotation, which seriously degrades the ISAR imaging quality. The main target image is interfered with by the micro-Doppler effect, further affecting target recognition and classification based on ISAR images.
[0003] At present, the methods for suppressing micro-Doppler interference mainly achieve micro-Doppler interference suppression in the signal domain, which can be divided into three categories.
[0004] The first type is to suppress the micro-motion part and the main part by taking advantage of the different characteristics of the micro-motion part in different domains. Although this method is simple and efficient, it is difficult to implement for the main body and micro-motion signals that are difficult to distinguish in each domain.
[0005] The second category is based on the idea of signal decomposition, which introduces empirical mode decomposition to remove micro-Doppler interference. However, the empirical mode decomposition algorithm has problems such as mode aliasing, and further theoretical proof is needed.
[0006] The third category is based on compressed sensing (CS), which uses the joint sparsity of the main signal to suppress the micro-Doppler effect. However, if the measured data error of this method is large, the time-frequency spectrum may not have very obvious joint sparsity, which will affect the subsequent imaging results. Summary of the Invention
[0007] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a micro-Doppler interference suppression method for roadside ISAR vehicle target imaging based on ISPA-Net, which can suppress micro-Doppler interference in ISAR images of moving vehicle targets, improve the readability and recognizability of ISAR images, and provide high-quality ISAR image information for target recognition and classification of intelligent transportation roadside perception units.
[0008] In order to achieve the above object, the technical solution adopted by the present invention is:
[0009] A method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net, comprising the following steps:
[0010] Step 1: construct a training dataset by simulating roadside ISAR vehicle image data, or construct a training dataset by measuring roadside ISAR vehicle images with millimeter-wave radar and preprocessing them;
[0011] Step 2: Introduce a recurrent neural network mechanism and use the LeakyReLU activation function to build the ISPA-Net architecture for ISAR image micro-Doppler interference suppression. Then design a combined loss function for training to obtain the optimal ISAR image micro-Doppler interference suppression model.
[0012] Step 3: Generate an interference-suppressed ISAR image using the optimal ISAR image micro-Doppler interference suppression model.
[0013] In step 1, when constructing a training data set through actual measurement, the collected millimeter-wave radar measured data is preprocessed by pulse compression, moving target display, and two-dimensional filtering to prepare for subsequent micro-Doppler interference suppression;
[0014] Millimeter-wave radars are set up on both sides of the road; ISAR imaging of vehicle targets is performed; and the data collection environment is normal sunny weather.
[0015] The step 1 is specifically as follows:
[0016] (1a) Based on the radar parameters under simulation conditions, Matlab is used to randomly generate target ISAR data. In the target ISAR data, the clean image represents the ISAR image with only the main scattering points, and the micro-Doppler interference image represents the ISAR image with both the main scattering points and the rotation scattering points. The clean image is used for comparison, to supervise the model training process and to calculate the loss function size of the clean image and the image processed by ISPA-Net.
[0017] (1b) A scattering point P(x P ,y P) has an angular velocity ω around the center of rotation relative to the radar P , and a scattering point Q(x Q ,y Q ) has a rotation center of O′ and an angular velocity ω Q After translation and rotation compensation, for the scattering point P, the signal received by the radar is
[0018]
[0019] Among them, f r and t m Represent the frequency in range and the slow time in azimuth, x k and y k Respectively represent the position of the kth scattering point on the X-axis and Y-axis, σ' k represents the scattering coefficient received by the kth scattering point, T p ,γ=B / T p ,λ=c / f c Respectively represent pulse width, modulation frequency, and wavelength; B and f c represent bandwidth and carrier frequency respectively;
[0020] For the rotating scattering point Q, the signal received by the radar is
[0021]
[0022] in, ω O′ =ω P , r Q is the rotation radius of the scattering point Q around O′, β is the complementary angle between the radar line of sight and the rotation plane of the scattering point Q, is the initial phase of the scattering point Q, c is the speed of light;
[0023] (1c) The vehicle ISAR image is written as follows:
[0024] I cm (f,g)=I r (f,g)+I m (f,g) Formula 3
[0025] Where f (f = 1, 2, ..., M) represents the fth equally spaced fast time unit in the range domain, g (g = 1, 2, ..., N) represents the gth sampling pulse of the azimuth slow time, I cm (f, g) represent ISAR images with micro-Doppler interference, I r (f, g) represents the image of the subject’s scattering point, I m (f, g) Micro-Doppler images of rotating scattering points;
[0026] Based on the ISAR image model in formula 3, the micro-Doppler interference image I cm (f, g) Restore the main scattered point image I r (f,g), as shown in Formula 4:
[0027]
[0028] stI cm (f,g)=I r (f,g)+I m (f,g) Formula 4
[0029] in represents the two-norm, I out (f, g) represent ISAR images processed by ISPA-Net network;
[0030] (1d) According to Formula 1 and Formula 2, main scattering points and rotation scattering points are randomly generated. Then, the ISAR image with only main scattering points is considered as a clean image, and the ISAR image with both main scattering points and rotation scattering points is considered as a micro-Doppler interference image.
[0031] In step 2, the ISPA-Net architecture includes four parts: standard residual block, spatial attention block, spatial attention residual block and spatial attention model;
[0032] The ISPA-Net architecture includes a standard residual block, a spatial attention block, a spatial attention residual block, and a spatial attention model. The standard residual block is used to extract ISAR image features from roadside vehicle images. The spatial attention block is used to eliminate interference information in the ISAR image features to obtain more accurate ISAR image features. The spatial attention residual block is combined with a micro-Doppler attention map to eliminate micro-Doppler interference in the ISAR feature map. The spatial attention model can notice the micro-Doppler interference information of the roadside vehicle image and generate a micro-Doppler attention map. The spatial attention residual block and the spatial attention model together constitute the spatial attention block.
[0033] Each standard residual block consists of three 3×3 convolutional layers, and each convolutional layer is followed by an activation layer. First, the input ISAR image passes through a 3×3 convolutional layer and an activation layer to obtain a rough ISAR feature map, and then passes through two standard residual blocks to further extract the ISAR image features.
[0034] The processed ISAR feature map is input into the spatial attention block, and the four spatial attention blocks are used to identify and eliminate the micro-Doppler interference information of the ISAR feature map to obtain more accurate ISAR image features.
[0035] The spatial attention block consists of a spatial attention model and three spatial attention residual blocks. The spatial attention model can notice the micro-Doppler interference information of the ISAR image and generate a micro-Doppler attention map to guide the subsequent spatial attention residual block processing. The spatial attention residual block also uses a residual block and combines it with the micro-Doppler attention map to better eliminate the micro-Doppler interference in the ISAR feature map through the learned negative residual.
[0036] Finally, the ISAR feature map output by the fourth spatial attention block is passed to two standard residual blocks for reconstructing and outputting the final ISAR image, that is, the ISAR image after micro-Doppler interference suppression.
[0037] The spatial attention model includes a recurrent neural network initialized based on the ReLU activation function and the identity matrix. Two rounds of four-way recurrent neural networks use context information to generate a globally perceived feature map. The recurrent neural network is used to project micro-Doppler interference information into four main directions: up, down, left, and right.
[0038] By adding a branch to capture spatial context information, the projected micro-Doppler features are selectively highlighted, and an additional convolutional layer and sigmoid activation function are used to generate a micro-Doppler attention map to guide the subsequent micro-Doppler interference suppression process.
[0039] In step 2, the combined loss function is defined as:
[0040] L total =L1+L SSIM +L MSE Formula 5
[0041] in, P represents the ISAR image after micro-Doppler interference suppression output by ISPA-Net, and C represents the clean image in the training set. represents a norm; L SSIM =1-SSIM(P,C), used to constrain the structural similarity between the processed image and the clean image; A represents the attention map from the first SAM in the network, M represents the binary map of micro-Doppler interference, represents the two-norm.
[0042] In step 2:
[0043] (2a) The trained ISPA-Net architecture is the weight matrix, i.e., the network parameters W, and the ISP A-Net imaging network parameters W are randomly initialized;
[0044] (2b) Input the ISAR image I with micro-Doppler effect interference after preprocessing in (1a) into the ISPA-Net architecture, and calculate layer by layer according to the network cascade order to obtain the predicted output result P = WI;
[0045] (2c) Use the Adam optimization algorithm to update and optimize the network parameters W. The update formula is:
[0046] W new =W-ηΔL Formula 6
[0047] Among them, W new is the updated network parameter, W is the network parameter before the update, η is the learning rate, set to 0.001, ΔL is the partial derivative of the loss function with respect to W, that is
[0048] (2d) Use the updated weight W new Repeat the calculation process of (2b)-(2c), perform iterative updates, and save the network parameter W with the minimum loss during the iteration process best , that is, the ISPA-Net model that optimally suppresses micro-Doppler in ISAR images is obtained.
[0049] The step 3 is specifically implemented as follows:
[0050] (3a) Pulse compression and two-dimensional filtering are performed on the simulated / measured data to obtain the ISAR image I with micro-Doppler interference used in the test. test ;
[0051] (3b) I test Input into the trained optimal ISPA-Net architecture, and the ISAR image P after suppressing micro-Doppler interference can be obtained. test :
[0052] P test =W best I test Formula 7.
[0053] Beneficial effects of the present invention:
[0054] The present invention uses a two-round four-way IRNN architecture to effectively learn and distinguish micro-Doppler features in a local-to-global attention manner, enhances the recognition of micro-Doppler interference in ISAR images of moving vehicles, and uses the LeakyReLU activation function to replace the ReLU activation function to maintain the activity of neurons, thereby improving the training stability and the performance of the ISAR image micro-Doppler interference suppression model.
[0055] The loss function of the present invention during training includes L1 loss function, MSE loss function and custom L1 loss function for constraining structural similarity.SSIM Loss function is used to construct a combined loss function training network model for ISAR image micro-Doppler interference suppression, which further improves the model's effect on ISAR image micro-Doppler interference suppression. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 This is a flow chart of the micro-Doppler interference suppression method for roadside ISAR vehicle target imaging based on ISPA-Net of the present invention.
[0057] Figure 2 Schematic diagram of the roadside ISAR imaging model for vehicle targets.
[0058] Figure 3 This is the ISPA-Net structure diagram constructed by the present invention.
[0059] Figure 4 This is a sample diagram of the simulation training data set. Figure 4 (a) is a clean ISAR image. Figure 4 (b) is the ISAR image with micro-Doppler interference.
[0060] Figure 5 The micro-Doppler interference suppression results of simulated point target ISAR images based on different algorithms, where (a), (b), and (c) represent the simulation data results of different groups, respectively.
[0061] Figure 6 Schematic diagram of the vehicle target model used in the simulation experiment.
[0062] Figure 7 To simulate the micro-Doppler interference suppression results of vehicle ISAR images based on different algorithms.
[0063] Figure 8 This is a sample diagram of the measured training data set. Figure 8 (a) is the optical image of the vehicle, Figure 8 (b) is the ISAR image of the moving vehicle after fine filtering. Figure 8 (c) is the ISAR image of a real moving vehicle after being processed by pulse compression, moving target display and other methods.
[0064] Figure 9 The results of micro-Doppler interference suppression of ISAR images of moving vehicles based on different algorithms are shown in Figure 1, where (a), (b), and (c) represent the measured data results of different batches, respectively. DETAILED DESCRIPTION
[0065] The present invention will be described in further detail below with reference to the accompanying drawings.
[0066] Reference Figures 1-9This embodiment provides a method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net. The specific implementation steps are as follows:
[0067] Step 1: Construction of a training dataset. A training dataset of point target ISAR images is constructed from the simulated data, or a training dataset of vehicle ISAR images is constructed using the collected millimeter-wave radar measured data through pulse compression, moving target display, and two-dimensional filtering preprocessing methods. This is used for subsequent ISPA-Net model training.
[0068] The specific implementation of this step is as follows:
[0069] (1a) According to the radar parameters under the simulation conditions given in Table 1, 300 sets of point target ISAR data are randomly generated using Matlab. The clean image represents the ISAR image with only the main scattering points, and the micro-Doppler effect interference image represents the ISAR image with both the main scattering points and the rotation scattering points. The clean images are used for comparison, to supervise the model training process and to calculate the loss function size of the clean images and the images processed by ISPA-Net.
[0070] Table 1
[0071]
[0072] (1b) Figure 2 As shown in the figure, since the vehicle moves in a straight line along the road, a scattering point P(x P ,y P ) has an angular velocity ω around the center of rotation relative to the radar P , and a scattering point Q(x Q ,y Q ) has a rotation center of O′ and an angular velocity ω Q After translation and rotation compensation, for the scattering point P, the signal received by the radar is
[0073]
[0074] Among them, f r and t m Represent the frequency in range and the slow time in azimuth, x k and y k Respectively represent the position of the kth scattering point on the X-axis and Y-axis, σ' k represents the scattering coefficient received by the kth scattering point, T p ,γ=B / T p ,λ=c / f c Respectively represent pulse width, modulation frequency, and wavelength; B and f c represent bandwidth and carrier frequency respectively;
[0075] Since a certain scattering point Q not only rotates around the entire target center, but also rotates at a greater angular velocity ω Q Rotating around the rotation center O', at this time for the rotating scattering point Q, the signal received by the radar is
[0076]
[0077] in, ω O′ =ω P , r Q is the rotation radius of the scattering point Q around O′, β is the complementary angle between the radar line of sight and the rotation plane of the scattering point Q, is the initial phase of the scattering point Q, and c is the speed of light.
[0078] (1c) By Figure 2 From the roadside ISAR imaging model of vehicle targets, it can be seen that the final vehicle ISAR image can be regarded as the sum of the vehicle body scattering point image and the wheel rotation scattering point image, so it can be written as follows:
[0079] I cm (f,g)=I r (f,g)+I m (f,g) Formula 3
[0080] Where f (f = 1, 2, ..., M) represents the fth equally spaced fast time unit in the range domain, g (g = 1, 2, ..., N) represents the gth sampling pulse of the azimuth slow time, I cm (f, g) represent ISAR images with micro-Doppler interference, I r (f, g) represents the image of the subject’s scattering point, I m (f, g) Micro-Doppler images of rotating scattering points.
[0081] Based on the ISAR image model in formula 3, the present invention aims to obtain the micro-Doppler interference image I cm (f, g) Restore the main scattered point image I r (f, g). The optimization problem of the image matrix can be solved by deep learning methods. Therefore, the task can be regarded as solving the mathematical model of scoliosis optimization with l2 norm constraints, as shown in formula (4):
[0082]
[0083] stI cm (f,g)=I r (f,g)+I m (f,g) Formula 4
[0084] in represents the two-norm, I out (f, g) represent ISAR images processed by the ISPA-Net network.
[0085] (1d) According to Equations 1 and 2, randomly generate 5 to 7 main scattering points and 3 to 4 rotating scattering points. The rotating scattering points have a rotation radius of 0.3 to 1 meter, an angular velocity of 10πrad / s to 20πrad / s, and the distance range of the scattering points on the coordinate axis is randomly set between [-9, 9]. ISAR images containing only main scattering points are considered clean images, while ISAR images containing both main scattering points and rotating scattering points are considered micro-Doppler interference images.
[0086] Step 2: By introducing a recurrent neural network mechanism to generate more discriminative contextual features and using the LeakyReLU activation function, we construct the ISPA-Net architecture for ISAR image micro-Doppler interference suppression.
[0087] The specific implementation of this step is as follows:
[0088] like Figure 3 As shown in the figure, the ISPA-Net network for ISAR image micro-Doppler interference suppression of the present invention mainly consists of four parts: standard residual block (Residual Block, RB), spatial attention block (Spatial Attentive Block, SAB), spatial attention residual block (Spatial Attentive Residual Block, SARB) and spatial attention model (Spatial Attentive Module, SAM).
[0089] (2a) First, the input ISAR image is passed through a 3×3 convolution layer (conv) and an activation layer (LeakyReLU) to obtain a rough ISAR feature map, and then passed through two standard residual blocks (RB, Figure 3 (b) Further extract ISAR image features. The skip connection of the residual block can skip the intermediate convolutional and activation layers and connect to the output of the third convolutional layer, which can accelerate the training process and alleviate the problems of gradient vanishing and model degradation. Each residual block includes three 3×3 convolutional layers, each followed by an activation layer. In particular, the present invention replaces all ReLU activation functions in the SPA model with LeakyReLU activation functions and adjusts the number of convolutional layers in the model, improving training stability and the performance of removing micro-Doppler interference from ISAR images.
[0090] (2b) The ISAR feature map processed in 2a is then input into the spatial attention block (SAB, Figure 3 In (c), four spatial attention blocks (SABs) are used to identify and eliminate the micro-Doppler interference information in the ISAR feature map, thereby obtaining more accurate ISAR image features.
[0091] The Spatial Attention Block (SAB) consists of a Spatial Attention Model (SAM, Figure 3 (e)) and three spatial attention residual blocks (SARB, Figure 3 (d)). Using SAM, we can detect micro-Doppler interference in ISAR images and generate a micro-Doppler attention map to guide subsequent SARB processing. SARB also uses a residual block combined with a micro-Doppler attention map to better eliminate micro-Doppler interference in ISAR feature images through the learned negative residual.
[0092] The recurrent neural network (IRNN) architecture is part of the spatial attention model (SAM), and the SAM module is part of building the SAB;
[0093] An initialized recurrent neural network (IRNN) is designed based on the ReLU activation function and the identity matrix, which is also the core building block of SAM. A two-round, four-way IRNN architecture utilizes contextual information to generate a globally aware feature map. The IRNN model projects micro-Doppler interference information onto four cardinal directions. Furthermore, a branch is added to capture spatial contextual information, selectively highlighting the projected micro-Doppler features. An additional convolutional layer and a sigmoid activation function are used to generate a micro-Doppler attention map, which guides the subsequent micro-Doppler interference suppression process.
[0094] (2c) Finally, the ISAR feature map output by the fourth SAB is passed to two standard residual blocks for reconstructing and outputting the final ISAR image, that is, the ISAR image after micro-Doppler interference suppression.
[0095] According to the above description, each module is divided into Figure 3 The ISPA-Net model is constructed to suppress the micro-Doppler effect of ISAR images.
[0096] Step 3: Design the corresponding loss function for ISAR imaging micro-Doppler interference suppression based on the constructed ISPA-Net architecture;
[0097] The combined loss function designed by the present invention is defined as:
[0098] L total =L1+L SSIM +L MSE Formula 5
[0099] in, P represents the image output by ISPA-Net, and C represents the clean image in the training set. represents a norm; L SSIM =1-SSIM(P,C), used to constrain the structural similarity between the processed image and the clean image; A represents the attention map from the first SAM in the network, M represents the binary map of micro-Doppler interference, represents the two-norm.
[0100] The loss function designed in the present invention not only improves the micro-Doppler interference suppression effect of ISAR images, but also further improves the structural similarity between the output image and the clean image.
[0101] Step 4: formulate a corresponding ISAR image micro-Doppler interference suppression training strategy to start training, and update the ISAR image micro-Doppler interference suppression model parameters to obtain the optimal ISAR image micro-Doppler interference suppression model;
[0102] The specific implementation of this step is as follows:
[0103] (4a) Randomly initialize the ISPA-Net imaging network parameters W;
[0104] (4b) The ISAR image I with micro-Doppler effect interference after preprocessing in (1a) is input into the ISPA-Net architecture and calculated layer by layer according to the network cascade order to obtain the predicted output result P = WI;
[0105] (4c) Use the Adam optimization algorithm to update and optimize the network parameters W. The update formula is:
[0106] W new =W-ηΔL Formula 6
[0107] Among them, W new is the updated network parameter, W is the network parameter before the update, η is the learning rate, set to 0.001, ΔL is the partial derivative of the loss function with respect to W, that is
[0108] (4d) Use the updated weight W new Repeat the calculation process of (4b)-(4c), perform iterative update, and save the network parameter W with the minimum loss during the iteration process best , that is, the ISPA-Net model that optimally suppresses micro-Doppler in ISAR images is obtained.
[0109] In step 5, an ISAR image of a simulated or measured target is obtained, and the same preprocessing operation as in step 1 is performed to construct an ISAR image test set. The optimal ISAR image micro-Doppler interference suppression model obtained in step 4 is then used to suppress micro-Doppler interference in the ISAR image, and the interference-suppressed ISAR image is obtained.
[0110] The specific implementation of this step is as follows:
[0111] (5a) Pulse compression and two-dimensional filtering are performed on the simulated / measured data to obtain the ISAR image I with micro-Doppler interference in the test set. test ;
[0112] (5b) I test Input into the trained optimal ISPA-Net model to obtain the ISAR image P after suppressing micro-Doppler interference. test :
[0113] P test =W best I test Formula 7.
[0114] The effect of the present invention can be further illustrated by the following experiments:
[0115] Simulation data experiment:
[0116] 1. Simulation conditions:
[0117] The relevant parameters of the simulation are shown in Table 1. Under the radar parameter conditions given in Table 1, 310 groups of Figure 4 The data shown, Figure 4 (a) represents the ISAR image with only the main scattering points (clean ISAR image), Figure 4 (b) shows an ISAR image with both main scattering points and rotating scattering points (an ISAR image with micro-Doppler interference). 300 sets of ISAR data were randomly selected as the training set, and the remaining 10 sets of ISAR data were used as the test set. After training, the optimal ISAR image micro-Doppler interference suppression model was obtained.
[0118] The comparison algorithms used in the test data include RPCA and SPA-Net. To quantitatively evaluate the quality of ISAR images after micro-Doppler interference suppression, this paper uses three image performance evaluation metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Image Contrast (IC). Furthermore, for a more intuitive observation, Equation 8 is used to express the degree of contrast approximation, as shown below:
[0119]
[0120] Among them, IC a ∈R,IC b ∈R represents the contrast index of the clean ISAR image and the ISAR image processed by different algorithms. ΔIC represents IC a and IC b Obviously, the smaller ΔIC∈R is, the closer the contrast values of the two images are.
[0121] 2. Simulation content:
[0122] Simulation 1:
[0123] Table 2
[0124]
[0125]
[0126] This simulation experiment uses ISAR images generated by random scattering point model, and performs ISAR image micro-Doppler interference suppression based on different algorithms for three sets of data in the test set. Figure 5 The results of ISAR image micro-Doppler interference suppression using different algorithms are given. Figure 5 (a), (b), and (c) show the results of ISAR image micro-Doppler interference suppression for three different sets of simulation data. Comparison of the three sets of data shows that the proposed ISPA-Net algorithm has excellent micro-Doppler interference suppression performance for ISAR images, validating the effectiveness and superiority of the proposed method.
[0127] In order to quantitatively evaluate the ISAR image micro-Doppler interference suppression effect under different algorithms, Table 2 gives Figure 5As can be seen from Table 2, under different simulation data, the ISPA-Net algorithm of the present invention is superior to other algorithms in terms of various evaluation indicators in suppressing micro-Doppler interference in ISAR images. In other words, the ISPA-Net algorithm has a better effect in suppressing micro-Doppler interference in ISAR images.
[0128] Simulation 2:
[0129] In order to further verify the effect of the ISPA-Net algorithm of the present invention on the micro-Doppler interference suppression of vehicle ISAR images, the simulation was carried out as follows: Figure 6 The vehicle target model shown in FIG. 4 is a vehicle target model whose body is composed of main body scattering points and whose wheels are composed of fast-rotating scattering points. Figure 7 The results of different algorithms for suppressing micro-Doppler interference in simulated vehicle ISAR images are given. Figure 7 It can be seen that the ISPA-Net algorithm of the present invention has a better micro-Doppler interference suppression effect on vehicle ISAR images than other algorithms. The evaluation indicators of the micro-Doppler interference suppression results of ISAR images using different algorithms are shown in Table 3.
[0130] Table 3
[0131]
[0132] As shown in Table 3, the ISPA-Net algorithm proposed in this paper performs better than other algorithms in terms of various evaluation indicators after suppressing micro-Doppler interference in vehicle ISAR images. That is, the ISPA-Net algorithm has the best suppression effect on micro-Doppler interference in vehicle ISAR images.
[0133] Measured data experiment:
[0134] 1. Experimental conditions:
[0135] The present invention uses the measured data collected by the millimeter wave radar AWR2243 to verify the performance of the ISPA-Net algorithm. The radar parameters corresponding to the data are shown in Table 4. The optical image of the measured vehicle is shown in Table 4. Figure 8 As shown in (a), after ISAR imaging and processing, Figure 8 (b) and Figure 8 (c) A set of ISAR images of moving vehicles.
[0136] Table 4
[0137]
[0138]
[0139] 2. Experimental content:
[0140] This experiment uses ISAR images of moving vehicles. Based on different algorithms, three groups of different batches of measured data are used to suppress ISAR image micro-Doppler interference. The results are as follows: Figure 9 As shown in the figure. In the measured data, while the RPCA algorithm can suppress micro-Doppler interference in ISAR images, it processes the entire data set on a one-dimensional range image, which results in a more concentrated target body and an increase in the surrounding amplitude, thus affecting the imaging results. The ISPA-Net algorithm of the present invention, on the other hand, is more accurate in identifying micro-Doppler interference in ISAR images of moving vehicles and achieves better micro-Doppler interference suppression than the SPA-Net algorithm. Furthermore, to quantitatively demonstrate the effectiveness of the three algorithms in suppressing micro-Doppler interference, Table 5 presents the evaluation metrics for each algorithm after suppressing micro-Doppler interference in ISAR images of moving vehicles.
[0141] As shown in Table 5, the measured moving vehicle ISAR images processed by the ISPA-Net algorithm of the present invention outperform the other two algorithms in terms of PSNR, SSIM, and ΔIC evaluation metrics. This indicates that the processed moving vehicle ISAR images have higher quality, are closer to clean moving vehicle ISAR images, and have smaller contrast deviations. This demonstrates the effectiveness and superiority of the proposed ISPA-Net algorithm in suppressing micro-Doppler interference on measured moving vehicle ISAR data.
[0142] Table 5
[0143]
Claims
1. A method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net, characterized in that: The following steps are included: Step 1: construct a training dataset by simulating roadside ISAR vehicle image data, or construct a training dataset by measuring roadside ISAR vehicle images with millimeter-wave radar and preprocessing them; Step 2: Introduce a recurrent neural network mechanism and use the LeakyReLU activation function to build the ISPA-Net architecture for ISAR image micro-Doppler interference suppression. Then design a combined loss function for training to obtain the optimal ISAR image micro-Doppler interference suppression model. Step 3: Generate an interference-suppressed ISAR image using the optimal ISAR image micro-Doppler interference suppression model.
2. The method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net according to claim 1, characterized in that: In step 1, when constructing a training data set through actual measurement, the collected millimeter-wave radar measured data is preprocessed by pulse compression, moving target display, and two-dimensional filtering to prepare for subsequent micro-Doppler interference suppression; Millimeter-wave radars are set up on both sides of the road; ISAR imaging of vehicle targets was performed; data collection was performed under normal sunny weather conditions.
3. The method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net according to claim 2, characterized in that: The step 1 is specifically as follows: (1a) Based on the radar parameters under simulation conditions, Matlab is used to randomly generate target ISAR data. In the target ISAR data, the clean image represents the ISAR image with only the main scattering points, and the micro-Doppler interference image represents the ISAR image with both the main scattering points and the rotation scattering points. The clean image is used for comparison, to supervise the model training process and to calculate the loss function size of the clean image and the image processed by ISPA-Net. (1b) A scattering point P(x P ,y P ) has an angular velocity ω around the center of rotation relative to the radar P , and a scattering point Q(x Q ,y Q ) has a rotation center of O′ and an angular velocity ω Q After translation and rotation compensation, for the scattering point P, the signal received by the radar is Among them, f r and t m Represent the frequency in range and the slow time in azimuth, x k and y k Respectively represent the position of the kth scattering point on the X-axis and Y-axis, σ' k represents the scattering coefficient received by the kth scattering point, T p ,γ=B / T p ,λ=c / f c Respectively represent pulse width, modulation frequency, and wavelength; B and f c represent bandwidth and carrier frequency respectively; For the rotating scattering point Q, the signal received by the radar is in, ω O′ =ω P , r Q is the rotation radius of the scattering point Q around O′, β is the complementary angle between the radar line of sight and the rotation plane of the scattering point Q, is the initial phase of the scattering point Q, c is the speed of light; (1c) The vehicle ISAR image is written as follows: I cm (f,g) = I r (f,g) + I m (f,g) Formula 3 Where f (f = 1, 2, ..., M) represents the fth equally spaced fast time unit in the range domain, g (g = 1, 2, ..., N) represents the gth sampling pulse of the azimuth slow time, I cm (f, g) represent ISAR images with micro-Doppler interference, I r (f, g) represents the image of the subject’s scattering point, I m (f, g) Micro-Doppler images of rotating scattering points; Based on the ISAR image model in formula 3, the micro-Doppler interference image I cm (f, g) Restore the main scattered point image I r (f,g), as shown in Formula 4: in represents the two-norm, I out (f, g) represent ISAR images processed by ISPA-Net network; (1d) According to Formula 1 and Formula 2, main scattering points and rotation scattering points are randomly generated. Then, the ISAR image with only main scattering points is considered as a clean image, and the ISAR image with both main scattering points and rotation scattering points is considered as a micro-Doppler interference image.
4. The method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net according to claim 3, characterized in that: In step 2, the ISPA-Net architecture includes four parts: standard residual block, spatial attention block, spatial attention residual block and spatial attention model; The ISPA-Net architecture includes a standard residual block, a spatial attention block, a spatial attention residual block, and a spatial attention model. The standard residual block is used to extract ISAR image features from roadside vehicle images. The spatial attention block is used to eliminate interference information in the ISAR image features to obtain more accurate ISAR image features. The spatial attention residual block is combined with a micro-Doppler attention map to eliminate micro-Doppler interference in the ISAR feature map. The spatial attention model can notice the micro-Doppler interference information of the roadside vehicle image and generate a micro-Doppler attention map.
5. The method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net according to claim 4, characterized in that: The spatial attention residual block and the spatial attention model together constitute a spatial attention block; Each standard residual block consists of three 3×3 convolutional layers, and each convolutional layer is followed by an activation layer. First, the input ISAR image passes through a 3×3 convolutional layer and an activation layer to obtain a rough ISAR feature map, and then passes through two standard residual blocks to further extract the ISAR image features.
6. The ISPA-Net-based roadside ISAR vehicle target imaging micro-Doppler interference suppression method according to claim 5, wherein the processed ISAR feature map is input into a spatial attention block, and the micro-Doppler interference information of the ISAR feature map is identified and eliminated through four spatial attention blocks to obtain more accurate ISAR image features; The spatial attention block consists of a spatial attention model and three spatial attention residual blocks; The spatial attention model can notice the micro-Doppler interference information of ISAR images and generate a micro-Doppler attention map to guide the subsequent spatial attention residual block processing. The spatial attention residual block also uses a residual block and combines it with the micro-Doppler attention map to better eliminate the micro-Doppler interference in the ISAR feature map through the learned negative residual.
7. According to the ISPA-Net-based roadside ISAR vehicle target imaging micro-Doppler interference suppression method of claim 6, the ISAR feature map output by the fourth spatial attention block is passed to two standard residual blocks for reconstructing and outputting the final ISAR image, i.e., the ISAR image after micro-Doppler interference suppression.
8. The method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net according to claim 7, characterized in that: The spatial attention model includes a recurrent neural network initialized based on the ReLU activation function and the identity matrix. Two rounds of four-way recurrent neural networks use context information to generate a globally perceived feature map. The recurrent neural network is used to project micro-Doppler interference information into four main directions: up, down, left, and right. By adding a branch to capture spatial context information, the projected micro-Doppler features are selectively highlighted, and an additional convolutional layer and sigmoid activation function are used to generate a micro-Doppler attention map to guide the subsequent micro-Doppler interference suppression process.
9. The method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net according to claim 8, characterized in that: In step 2, the combined loss function is defined as: L total =L1+L SSIM +L MSE Formula 5 in, P represents the ISAR image after micro-Doppler interference suppression output by ISPA-Net, and C represents the clean image in the training set. represents a norm; L SSIM =1-SSIM(P,C), used to constrain the structural similarity between the processed image and the clean image; A represents the attention map from the first SAM in the network, M represents the binary map of micro-Doppler interference, represents the two-norm.
10. The method for suppressing micro-Doppler interference in roadside ISAR vehicle target imaging based on ISPA-Net according to claim 9, characterized in that: The step 3 is specifically implemented as follows: (3a) Pulse compression and two-dimensional filtering are performed on the simulated / measured data to obtain the ISAR image I with micro-Doppler interference used in the test. test ; (3b) I test Input into the trained optimal ISPA-Net architecture, and the ISAR image P after suppressing micro-Doppler interference can be obtained. test : P test =W best I test Formula 7.