A method for ISAR image compensation under low signal-to-noise ratio conditions

Abstract

The present invention relates to the technical field of ISAR image processing, and discloses a method and system for ISAR image compensation under low signal-to-noise ratio conditions. The method comprises the following steps: step S1, data acquisition; step S2, collecting ISAR data; step S3, extracting translational information; step S4, establishing an adaptive compensation model; step S5, implementing adaptive translational compensation; step S6, training a neural network; step S7, evaluating and tuning; and step S8, iterative optimization. The method extracts translational information by preprocessing the data, establishes an adaptive compensation model, performs motion compensation through the neural network, and adjusts the parameters of the adaptive compensation model according to image quality indicators. The method solves the problem of poor visualization and resolution of ISAR images under low signal-to-noise ratio conditions, making it difficult to distinguish the morphological structure of the observed target.

Description

Technical Field

[0001] The present invention relates to the technical field of ISAR image processing, and in particular to an ISAR image compensation method under low signal-to-noise ratio conditions. Background Art

[0002] Inverse Synthetic Aperture Radar (ISAR) is capable of high-resolution two-dimensional imaging of non-cooperative targets at all times, in all weather conditions, and at long distances, providing information on the target's morphological structure. As the ISAR observation range increases, the signal-to-noise ratio (SNR) of the echo decreases. Existing ISAR imaging translational motion compensation methods struggle to accurately estimate and compensate for the translational component, making it impossible to obtain focused images of the target. Furthermore, ISAR images with low SNRs exhibit poor visualization and resolution, making it difficult to discern the target's morphological structure.

[0003] The signal-to-noise ratio (SNR) refers to the ratio of signal to noise. A low SNR indicates a weak signal or high interference. In radar applications, a low SNR can be caused by a target's long range, small cross-sectional area, or weak scattering properties. This can lead to reduced image quality and make target identification difficult. In ISAR, the SNR can be particularly degraded when the target is moving at high speeds. The motion of the target and / or radar platform causes Doppler shift and blurring effects in the ISAR imaging process. To obtain clear ISAR images, these motions must be compensated. Motion compensation adjusts the received echo data to counteract the effects of target and platform motion. This can be achieved by knowing the target or platform motion parameters and then processing the echo data accordingly. Motion compensation is crucial for obtaining high-quality ISAR images.

[0004] Existing ISAR translational compensation methods typically implement this compensation in two steps: envelope alignment and autofocusing. In low signal-to-noise ratio (SNR) conditions, the target envelope is comparable to or even submerged in the noise, degrading envelope alignment performance. This, in turn, makes it difficult for autofocusing to accurately estimate and compensate for initial phase errors, resulting in an inability to obtain focused imaging results. To address ISAR image compensation under low SNR conditions, a low SNR ISAR image compensation method is needed to address the poor visualization and resolution of ISAR images under these conditions, making it difficult to distinguish the morphological structure of the observed target. Summary of the Invention

[0005] To solve the problems raised in the above background technology, the present invention provides an ISAR image compensation method under low signal-to-noise ratio conditions. In order to solve the problem that the existing ISAR translation compensation method is difficult to accurately estimate and compensate the initial phase error by self-focusing, and cannot obtain focused imaging, the method first extracts translation information through data preprocessing, then establishes an adaptive compensation model, performs motion compensation through a neural network, and adjusts the parameters of the adaptive compensation model according to image quality indicators.

[0006] To achieve the above object, the first aspect of the present invention provides an ISAR image compensation method under low signal-to-noise ratio conditions, comprising:

[0007] Step S1, data acquisition: collecting ISAR data, including radar echo data, translational data, and system and environmental information;

[0008] Step S2, data preprocessing: performing denoising and / or interference removal processing on the radar echo data;

[0009] Step S3, extracting translation information: extracting the target's translation data, including velocity, acceleration, and position, through radar echo data;

[0010] Step S4, establishing an adaptive compensation model: establishing an adaptive translation compensation model based on the target motion model and actual translation information, taking the ISAR pre-processed radar echo data, translation data, system and environment information as input, and taking the corresponding adjusted translation parameters as target output;

[0011] Step S5, implementing adaptive translation compensation: determining the loss function and neural network architecture for adaptive translation compensation, and using the difference between the neural network output and the actual translation parameters as compensation;

[0012] Step S6, training the neural network: inputting data into the neural network, adjusting the weights and biases of the neural network so that the network can approximate the mapping relationship between input and output;

[0013] Step S7, evaluation and tuning: evaluating the compensation effect and adjusting the adaptive compensation model parameters according to the image quality index;

[0014] Step S8, iterative optimization: perform iterative optimization based on real-time feedback from the system, improve the adaptive rules, and improve the effect of translation compensation through different filter parameter combinations.

[0015] As a further description of the above technical solution: Step S1 includes: radar echo and translation data are acquired according to the following steps:

[0016] Step S101: The radar echo data in the three-dimensional time domain is:

[0017] x(t)=A(t)e jφ(t) +n(t) (1)

[0018] Where: A(t) is the complex amplitude of the radar signal changing with time, j is the imaginary unit j 2 = -1, φ(t) is the phase of the radar signal that varies with time, and n(t) is the noise of the radar signal that varies with time;

[0019] Step S102: The three-dimensional translation data is:

[0020]

[0021] in: is the three-dimensional position vector, is the three-dimensional velocity vector, is the three-dimensional acceleration vector, r0 is the initial position vector, v0 is the initial velocity vector, and a is the acceleration vector;

[0022] Step S103, including the time domain signal data of the three-dimensional time domain radar echo data of the translation:

[0023] x(t)=A(t)e jφ(t) s(t-τ(t))+n(t) (3)

[0024] Where: s(t-τ(t)) is the time delay caused by target motion, and τ(t) is the time delay.

[0025] As a further description of the above technical solution: Step S2 includes: performing denoising on the radar echo data according to the following steps:

[0026] Step S201: De-noise the radar echo data by time domain filtering:

[0027] x denoised (t)=Filter(x(t)) (4)

[0028] Wherein, Filter(x(t)) is a filter, which is one of the mean filter, median filter, and Gaussian filter;

[0029] Step S202: De-noise the radar echo data by frequency domain filtering:

[0030] X(f)=F(x(t)) (5)

[0031] X denoised (f)=Filter(X(f)) (6)

[0032] x denoised (t) = F -1 (X denoised(f)) (7)

[0033] Where F is the Fourier transform, X(f) is the spectrum of the signal, and Filter(X(f)) is the frequency domain filter spectrum, which is one of the mean filter, median filter, and Gaussian filter.

[0034] As a further description of the above technical solution: Step S2 includes: radar echo data is subjected to interference removal processing according to the following steps:

[0035] Step S203: Remove interference from radar echo data through time domain filtering:

[0036] x clean (t)=x(t)-Interference(t) (8)

[0037] Among them, Interference(t) is the identified interference part;

[0038] Step S204: Remove interference from the radar echo data through frequency domain filtering:

[0039] X clean (f)=X(f)-Interference(f) (9)

[0040] x clean (t) = F -1 (X clean (f)) (10)

[0041] Where Interference(f) is the frequency domain part of the identified interference.

[0042] As a further description of the above technical solution: in step S4, the adaptive compensation model is established as follows:

[0043]

[0044] Among them, α, β, γ are weight parameters, MotionEstimate(t) is the real-time estimation of translation information, EnvironmentInfo(t) is the environmental information,

[0045]

[0046] Here, τ(t) is the time delay caused by target motion.

[0047] As a further description of the above technical solution: Step S4, which uses the ISAR pre-processed radar echo data, translation data, system and environmental information as input and uses the corresponding adjusted translation parameters as target output, further includes the following steps:

[0048] Step S401: Input can be expressed as:

[0049] X=[X radar ,X motion ,X system / env ] (13)

[0050] Among them, X radar is the feature vector of radar echo data, X motion is the eigenvector of the translation data, X system / env is the feature vector of system and environment information;

[0051] Step S402, the output can be expressed as:

[0052] Y=[y1,y2,y3] (14)

[0053] Where y1 is the target position information that needs to be adjusted, y2 is the target speed information that needs to be adjusted, and y3 is the target acceleration information that needs to be adjusted;

[0054] As a further description of the above technical solution: Step S5, which uses the difference between the neural network output and the actual translation parameter as compensation, further includes the following steps:

[0055] Step S501: Using the ISAR complex image I as the input X of the neural network mapping function, the complex image I has an image size of h×w×2, where h is the image height, w is the image width, 2 is the number of channels, and the real part and the imaginary part each occupy one channel;

[0056] Step S502: Construct a feature extraction module, including convolution blocks Conv_block1 and Conv_block2, and maximum pooling layers Max_pool1 and Max_pool2. Conv_block1 includes two convolution layers Conv1 and Conv2 with activation function ReLU, and the convolution kernel size is 3×3. It takes the ISAR image I as input and outputs a feature map I0 with a size of h×w×64. The maximum pooling layer Max_pool1 is used for dimensionality reduction and outputs a feature map I1 with a size of h1×w1×64, where h1=h / 2 and w1=w / 2.

[0057]

[0058] Conv_block2 includes two convolutional layers Conv3 and Conv4 with activation function ReLU. The convolution kernel size is 3×3. Conv_block2 takes feature map I1 as input and outputs feature map I2 with size h1×w1×128. It uses the maximum pooling layer Max_pool2 to reduce the dimension and output feature map I3 with size h2×w2×256, where h2=h1 / 2 and w2=w1 / 2.

[0059]

[0060] The feature map I3 output by the feature extraction module is used as the input of the feature encoding module;

[0061] Step S503: construct a feature encoding module, including two self-attention networks Transformer1 and Transformer2, as well as a batch normalization layer BN5 and a convolutional layer Conv5. Transformer1 includes batch normalization BN1, window-based multi-head self-attention W-MSA, batch normalization BN2 and multi-layer perceptron MLP1. The input of BN1 is residually connected to the output of W-MSA, and the input of BN2 is residually connected to the output of MLP1. Transformer1 takes feature map I3 as input and outputs feature map I3.

[0062]

[0063] Among them, I'3 is the intermediate output feature map, Transformer2 includes batch normalization BN3, multi-head self-attention SW-MSA based on shift window, deep convolution DWConv, batch normalization BN4 and multi-layer perceptron MLP2, among which the input of BN3 is residually connected to the output of SW-MSA, the input and output of DWConv are residually connected, and the input of BN4 is residually connected to the output of MLP2. Transformer2 takes feature map I"3 as input and outputs feature map I"4.

[0064]

[0065] Among them, I4 and I'4 are intermediate output feature maps, the kernel size of the last convolution layer Conv5 of the feature encoding module is 3×3, and the final output of the feature encoding module is the feature map I5 with a size of h2×w2×256.

[0066] I5=Conv5[BN5(I4”)]+I3 (19)

[0067] The feature map I5 output by the feature encoding module is used as the input of the feature decoding module;

[0068] Step S504: construct a feature decoding module, which includes two convolution blocks Conv_block3 and Conv_block4, two sub-pixel convolution layers SP_Conv1 and SP_Conv2, and a convolution 1×1Conv with a kernel size of 1. The feature map I5 output by the feature encoding module is used as input, and the sub-pixel convolution layer SP_Conv1 is used to implement upsampling. The size of the output feature map is h1×w1×128, and then it is integrated with the feature map I2 by pixel-by-pixel addition to output a feature map I6 with a size of h1×w1×256.

[0069] I6=SP_Conv1(I5)+I2 (20)

[0070] And as the input of Conv_block3, Conv_block3 includes 2 convolutional layers Conv6 and Conv7 with activation function ReLU, convolution kernel size is 3×3, and output feature map I7 with size h1×w1×128,

[0071] I7=Conv7[Conv6(I6) ReLU ] ReLU (twenty one)

[0072] The sub-pixel convolution layer SP_Conv2 is used to achieve upsampling, and the output feature map of size h×w×64 is integrated with the feature map I0 by pixel-by-pixel addition, and the output feature map I8 of size h×w×128 is output.

[0073] I8=SP_Conv2(I7)+I0 (22)

[0074] And as the input of Conv_block4, Conv_block4 includes 2 convolutional layers Conv8 and Conv9 with activation function ReLU, the convolution kernel size is 3×3, Conv_block4 takes feature map I8 as input and outputs feature map I9 with size h×w×64,

[0075] I9=Conv9[Conv8(I8) ReLU ] ReLU (twenty three)

[0076] Reuse the convolution Conv with a kernel size of 1 1×1 Transform and finally output feature map I 10 ,

[0077] I 10 =Conv 1×1 (I9) (24)

[0078] The size is h×w×2, with feature map I 10 As output Y.

[0079] As a further description of the above technical solution: Step S6, the step of adjusting the weights and biases of the neural network is:

[0080] Step S601, input layer to hidden layer:

[0081] H 1 =σ(W 1 X+b 1 ) (25)

[0082] Among them, H 1 is the output of the hidden layer, σ is the activation function, which is a nonlinear function such as ReLU (Rectified Linear Unit) or Sigmoid, W 1 is the weight matrix from the input layer to the hidden layer, b 1 is the bias vector from the input layer to the hidden layer;

[0083] Step S602, hidden layer to output layer:

[0084] Y=σ(W 2 H 1 +b 2 ) (26)

[0085] Among them, W 2 is the weight matrix from the hidden layer to the output layer, b 2 is the bias vector from the hidden layer to the output layer;

[0086] Step S603, using a gradient descent algorithm with momentum to update the parameters of the neural network, initializing the weights, biases, and momentum terms of the neural network, where the momentum term is used to record the direction and magnitude of the previous parameter update;

[0087] Step S604: Randomly select a batch of samples from the training data set. During the back propagation process of the current batch, calculate the average values ​​of the gradients dW and db obtained by differentiating the selected samples with respect to the neural network parameters. Calculate the new momentum term v based on the current average gradient and the momentum term of the previous parameter update. dW and v db , and use the new momentum term to update the neural network parameters, weights W and bias b,

[0088]

[0089] Among them, α is the learning rate and β is the hyperparameter;

[0090] Step S606: Repeat steps S604 to S605 until the preset number of training epochs is reached. The loss function L(W,b) is expressed as follows:

[0091]

[0092] Among them, N is the number of samples, M is the number of output features, and Y ij Output the predicted value for the network, is the true value of the label, the optimization goal is to minimize the loss function value, and the neural network parameters are trained as compensation parameters.

[0093] As a further description of the above technical solution: Step S7, adjusting the adaptive compensation model parameters according to the image quality index, including one or more of peak signal-to-noise ratio, contrast, image standard deviation, and image average gradient, is achieved by one or more of the following steps:

[0094] Step S701, peak signal-to-noise ratio:

[0095]

[0096] Among them, PSNR represents the ratio between the energy of the maximum signal in the image and the background noise, I, K, M, N represent the image size of the current image I and the reference image K is M×N, MAX1 is the maximum grayscale value, usually 255;

[0097] Step S702, contrast:

[0098]

[0099] Among them, δ(i,j)=|ij| represents the grayscale difference between adjacent pixels, P δ (i, j) represents the distribution probability of pixels with grayscale difference between adjacent pixels;

[0100] Step S703, image standard deviation:

[0101]

[0102] Where f(i,j) is the image pixel value and μ is the image mean;

[0103] Step S704, image average gradient value:

[0104]

[0105] Among them, G x and G y The gradient values ​​of the image in the horizontal and vertical directions.

[0106] The second aspect of the present invention is an ISAR image compensation system under low signal-to-noise ratio conditions, characterized in that it is used for an ISAR image compensation method under low signal-to-noise ratio conditions as described above, comprising: a data acquisition module, a data recording system, a data preprocessing module, a translational information extraction module, an adaptive compensation model establishment module, a neural network training platform, an evaluation and tuning module, and an iterative optimization module. The data acquisition module includes an ISAR radar system, a translational data collector, and an environmental monitoring sensor. The radar system, the translational data collector, and the environmental monitoring sensor are connected to the data recording system through a communication interface, and then sequentially connected to the data preprocessing module, the translational information extraction module, the adaptive compensation model establishment module, the neural network training platform, the evaluation and tuning module, and the iterative optimization module through a data interface.

[0107] Preferably, the data preprocessing module, translation information extraction module, adaptive compensation model establishment module, evaluation and tuning module, and iterative optimization module are high-performance computing servers or data analysis workstations.

[0108] Preferably, the neural network training platform is a deep learning workstation.

[0109] The present invention provides an ISAR image compensation method under low signal-to-noise ratio conditions, which solves the problem that the existing ISAR translation compensation method is difficult to accurately estimate and compensate the initial phase error through self-focusing, and cannot obtain focused imaging. First, translation information is extracted through data preprocessing, and then an adaptive compensation model is established. Motion compensation is performed through a neural network, and the parameters of the adaptive compensation model are adjusted according to image quality indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0111] Figure 1 Flowchart of the ISAR image compensation method under low signal-to-noise ratio conditions proposed by the present invention DETAILED DESCRIPTION

[0112] The present invention is further described in detail below by specific embodiments in conjunction with the accompanying drawings. In the following embodiments, many detailed descriptions are intended to enable the present application to be better understood. However, those skilled in the art can readily appreciate that some of the features may be omitted in different circumstances, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the present application are not shown or described in the specification. This is to avoid overwhelm the core of the present application with excessive descriptions, and for those skilled in the art, it is not necessary to describe these related operations in detail. They can fully understand the related operations based on the description in the specification and the general technical knowledge in the art.

[0113] In addition, the features, operations, or characteristics described in the specification may be combined in any appropriate manner to form various embodiments. Furthermore, the steps or actions in the method description may be reordered or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various sequences in the specification and drawings are provided solely for the purpose of clearly describing a particular embodiment and are not intended to be mandatory, unless otherwise specified.

[0114] The serial numbers assigned to the components in this document, such as "first", "second", etc., are only used to distinguish the objects described and do not have any order or technical meaning. Specific implementation method 1

[0116] A first aspect of the present invention provides an ISAR image compensation method under low signal-to-noise ratio conditions, comprising:

[0117] Step S1, data acquisition: collecting ISAR data, including radar echo data, translational data, and system and environmental information;

[0118] Step S2, data preprocessing: performing denoising and / or interference removal processing on the radar echo data;

[0119] Step S3, extracting translation information: extracting the target's translation data, including velocity, acceleration, and position, through radar echo data;

[0120] Step S4, establishing an adaptive compensation model: Based on the target motion model and actual translation information, an adaptive translation compensation model is established, with ISAR pre-processed radar echo data, translation data, system and environment information as input, and corresponding adjusted translation parameters as target output, further comprising the following steps:

[0121] Step S401: The input can be expressed as:

[0122] X=[X radar ,X motion ,Xsystem / env ] (33)

[0123] Among them, X radar is the feature vector of radar echo data, X motion is the eigenvector of the translation data, X system / env is the feature vector of system and environment information;

[0124] Step S402, the output can be expressed as:

[0125] Y=[y1,y2,y3] (34)

[0126] Where y1 is the target position information that needs to be adjusted, y2 is the target speed information that needs to be adjusted, and y3 is the target acceleration information that needs to be adjusted;

[0127] Step S5, implementing adaptive translation compensation: determining the loss function and neural network architecture for adaptive translation compensation, and using the difference between the neural network output and the actual translation parameters as compensation, further comprising the following steps:

[0128] Step S501: Using the ISAR complex image I as the input X of the neural network mapping function, the complex image I has an image size of h×w×2, where h is the image height, w is the image width, 2 is the number of channels, and the real part and the imaginary part each occupy one channel;

[0129] Step S502: Construct a feature extraction module, including convolution blocks Conv_block1 and Conv_block2, and maximum pooling layers Max_pool1 and Max_pool2. Conv_block1 includes two convolution layers Conv1 and Conv2 with activation function ReLU, and the convolution kernel size is 3×3. It takes the ISAR image I as input and outputs a feature map I0 with a size of h×w×64. The maximum pooling layer Max_pool1 is used for dimensionality reduction and outputs a feature map I1 with a size of h1×w1×64, where h1=h / 2 and w1=w / 2.

[0130]

[0131] Conv_block2 includes two convolutional layers Conv3 and Conv4 with activation function ReLU. The convolution kernel size is 3×3. Conv_block2 takes feature map I1 as input and outputs feature map I2 with size h1×w1×128. It uses the maximum pooling layer Max_pool2 to reduce the dimension and output feature map I3 with size h2×w2×256, where h2=h1 / 2 and w2=w1 / 2.

[0132]

[0133] The feature map I3 output by the feature extraction module is used as the input of the feature encoding module;

[0134] Step S503: construct a feature encoding module, including two self-attention networks Transformer1 and Transformer2, as well as a batch normalization layer BN5 and a convolutional layer Conv5. Transformer1 includes batch normalization BN1, window-based multi-head self-attention W-MSA, batch normalization BN2 and multi-layer perceptron MLP1. The input of BN1 is residually connected to the output of W-MSA, and the input of BN2 is residually connected to the output of MLP1. Transformer1 takes feature map I3 as input and outputs feature map I3.

[0135]

[0136] Among them, I'3 is the intermediate output feature map, Transformer2 includes batch normalization BN3, multi-head self-attention SW-MSA based on shift window, deep convolution DWConv, batch normalization BN4 and multi-layer perceptron MLP2, among which the input of BN3 is residually connected to the output of SW-MSA, the input and output of DWConv are residually connected, and the input of BN4 is residually connected to the output of MLP2. Transformer2 takes feature map I"3 as input and outputs feature map I"4.

[0137]

[0138] Among them, I4 and I'4 are intermediate output feature maps, the kernel size of the last convolution layer Conv5 of the feature encoding module is 3×3, and the final output of the feature encoding module is the feature map I5 with a size of h2×w2×256.

[0139] I5=Conv5[BN5(I4”)]+I3 (39)

[0140] The feature map I5 output by the feature encoding module is used as the input of the feature decoding module;

[0141] Step S504: construct a feature decoding module, which includes two convolution blocks Conv_block3 and Conv_block4, two sub-pixel convolution layers SP_Conv1 and SP_Conv2, and a convolution 1×1Conv with a kernel size of 1. The feature map I5 output by the feature encoding module is used as input, and the sub-pixel convolution layer SP_Conv1 is used to implement upsampling. The size of the output feature map is h1×w1×128, and then it is integrated with the feature map I2 by pixel-by-pixel addition to output a feature map I6 with a size of h1×w1×256.

[0142] I6=SP_Conv1(I5)+I2 (40)

[0143] And as the input of Conv_block3, Conv_block3 includes 2 convolutional layers Conv6 and Conv7 with activation function ReLU, convolution kernel size is 3×3, and output feature map I7 with size h1×w1×128,

[0144] I7=Conv7[Conv6(I6) ReLU ] ReLU (41)

[0145] The sub-pixel convolution layer SP_Conv2 is used to achieve upsampling, and the output feature map of size h×w×64 is integrated with the feature map I0 by pixel-by-pixel addition, and the output feature map I8 of size h×w×128 is output.

[0146] I8=SP_Conv2(I7)+I0 (42)

[0147] And as the input of Conv_block4, Conv_block4 includes 2 convolutional layers Conv8 and Conv9 with activation function ReLU, the convolution kernel size is 3×3, Conv_block4 takes feature map I8 as input and outputs feature map I9 with size h×w×64,

[0148] I9=Conv9[Conv8(I8) ReLU ] ReLU (43)

[0149] Reuse the convolution Conv with a kernel size of 1 1×1 Transform and finally output feature map I 10 ,

[0150] I 10 =Conv 1×1 (I9) (44)

[0151] The size is h×w×2, with feature map I 10 As output Y.

[0152] Step S6, training the neural network: inputting data into the neural network, adjusting the weights and biases of the neural network so that the network can approximate the mapping relationship between input and output, further comprising the following steps:

[0153] Step S601, input layer to hidden layer:

[0154] H 1 =σ(W 1 X+b 1 ) (45)

[0155] Among them, H 1 is the output of the hidden layer, σ is the activation function, which is a nonlinear function such as ReLU (Rectified Linear Unit) or Sigmoid, W 1 is the weight matrix from the input layer to the hidden layer, b 1 is the bias vector from the input layer to the hidden layer;

[0156] Step S602, hidden layer to output layer:

[0157] Y=σ(W 2 H 1 +b 2 ) (46)

[0158] Among them, W 2 is the weight matrix from the hidden layer to the output layer, b 2 is the bias vector from the hidden layer to the output layer;

[0159] Step S603, using a gradient descent algorithm with momentum to update the parameters of the neural network, initializing the weights, biases, and momentum terms of the neural network, where the momentum term is used to record the direction and magnitude of the previous parameter update;

[0160] Step S604: Randomly select a batch of samples from the training data set. During the back propagation process of the current batch, calculate the average values ​​of the gradients dW and db obtained by differentiating the selected samples with respect to the neural network parameters. Calculate the new momentum term v based on the current average gradient and the momentum term of the previous parameter update. dW and v db , and use the new momentum term to update the neural network parameters, weights W and bias b,

[0161]

[0162] Among them, α is the learning rate and β is the hyperparameter;

[0163] Step S605: Repeat steps S604 to S605 until the preset number of training epochs is reached. The loss function L(W,b) is expressed as follows:

[0164]

[0165] Among them, N is the number of samples, M is the number of output features, and Y ij Output the predicted value for the network, is the true value of the label, the optimization goal is to minimize the loss function value, and the neural network parameters are trained as compensation parameters.

[0166] Step S7, evaluation and tuning: evaluating the compensation effect and adjusting the adaptive compensation model parameters according to the image quality index;

[0167] Step S8, iterative optimization: perform iterative optimization based on real-time feedback from the system, improve the adaptive rules, and improve the effect of translation compensation through different filter parameter combinations. Specific embodiment 2

[0169] The second aspect of the present invention provides an ISAR image compensation system under low signal-to-noise ratio conditions, including: a data acquisition module, a data recording system, a data preprocessing module, a translational information extraction module, an adaptive compensation model establishment module, a neural network training platform, an evaluation and tuning module, and an iterative optimization module. The data acquisition module includes an ISAR radar system, a translational data collector, and an environmental monitoring sensor. The radar system, the translational data collector, and the environmental monitoring sensor are connected to the data recording system through a communication interface, and then connected in sequence to the data preprocessing module, the translational information extraction module, the adaptive compensation model establishment module, the neural network training platform, the evaluation and tuning module, and the iterative optimization module through a data interface.

[0170] The radar system collects radar echo data. The translational data collector obtains the target's translational data, including velocity, acceleration, and position. The environmental monitoring sensor collects environmental information, including temperature and humidity, and transmits the collected data to the data recording system through a communication interface. The data recording system stores the collected radar echo data, translational data, and environmental information. The data preprocessing module receives preprocessed data storage information from the data recording system and performs specific denoising algorithms and interference removal operations to denoise and interfere with the data. The translational information extraction module receives preprocessed data from the data preprocessing module, extracts translational information such as velocity, acceleration, and position, and transmits the extracted translational data to the adaptive compensation model establishment module. The adaptive compensation model establishment module establishes an adaptive translational compensation model and transmits the compensation model and related data to the neural network training platform. The neural network training platform trains the neural network, adjusts the model weights and biases, and transmits the training results to the evaluation and tuning module. The evaluation and tuning module evaluates the compensation effect, adjusts the model parameters according to image quality indicators, and transmits the evaluation results and adjustment suggestions to the iterative optimization module. The iterative optimization module optimizes the compensation effect based on real-time feedback and transmits the optimization results and adjustment plan back to the adaptive compensation model establishment module for adjustment.

[0171] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

[0172] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into a single processing unit, or each unit can exist physically separately, or two or more units can be integrated into a single unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the purpose of distinguishing them from each other and are not intended to limit the scope of protection of this application. The specific working processes of the units and modules in the above-mentioned wireless terminal can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0173] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0174] In the embodiments provided by the present invention, it should be understood that the disclosed systems / terminal devices and methods can be implemented in other ways. For example, the system / terminal device embodiments described above are merely illustrative. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the system or unit can be electrical, mechanical or other forms.

[0175] The units described as separate components may or may not be physically separate, and the components described as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0176] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for compensating ISAR images under low signal-to-noise ratio conditions, characterized in that: The following steps are involved: Step S1, data acquisition: collecting ISAR data, including radar echo data, translational data, and system and environmental information; Step S2, data preprocessing: performing denoising and / or interference removal processing on the radar echo data; Step S3, extracting translation information: extracting the target's translation data, including velocity, acceleration, and position, through radar echo data; Step S4, establishing an adaptive compensation model: Based on the target motion model and actual translation information, an adaptive translation compensation model is established, with ISAR preprocessed radar echo data, translation data, system and environment information as input, and corresponding adjusted translation parameters as target output, in accordance with the following steps: Step S401: Establish an adaptive compensation model: Among them, α, β, γ are weight parameters, MotionEstimate(t) is the real-time estimation of translation information, EnvironmentInfo(t) is the environmental information, Among them, τ(t) is the time delay caused by target motion; Step S402 uses ISAR pre-processed radar echo data, translational data, system and environment information as input. The input can be expressed as: X=[X radar ,X motion ,X system / env ] (3) Among them, X radar is the feature vector of radar echo data, X motion is the eigenvector of the translation data, X system / env is the feature vector of system and environment information; Step S403, the output can be expressed as: Y=[y1,y2,y3] (4) Where y1 is the target position information that needs to be adjusted, y2 is the target speed information that needs to be adjusted, and y3 is the target acceleration information that needs to be adjusted; Step S5, implementing adaptive translation compensation: determining the loss function and neural network architecture for adaptive translation compensation, and using the difference between the neural network output and the actual translation parameters as compensation; Step S6, training the neural network: inputting data into the neural network, adjusting the weights and biases of the neural network so that the network can approximate the mapping relationship between input and output; Step S7, evaluation and tuning: evaluating the compensation effect and adjusting the adaptive compensation model parameters according to the image quality index; Step S8, iterative optimization: perform iterative optimization based on real-time feedback from the system, improve the adaptive rules, and improve the effect of translation compensation through different filter parameter combinations.

2. The ISAR image compensation method under low signal-to-noise ratio conditions according to claim 1, characterized in that: The radar echo and translation data in step S1 are obtained according to the following steps: Step S101: The radar echo data in the three-dimensional time domain is: x(t)=A(t)e jφ(t) +n(t) (5) Where: A(t) is the complex amplitude of the radar signal changing with time, j is the imaginary unit j 2 = -1, φ(t) is the phase of the radar signal that varies with time, and n(t) is the noise of the radar signal that varies with time; Step S102: The three-dimensional translation data is: in: is the three-dimensional position vector, is the three-dimensional velocity vector, is the three-dimensional acceleration vector, r0 is the initial position vector, v0 is the initial velocity vector, and a is the acceleration vector; Step S103, including the time domain signal data of the three-dimensional time domain radar echo data of the translation: x(t)=A(t)e jφ(t) s(t-τ(t))+n(t) (7) Where: s(t-τ(t)) is the time delay caused by target motion, and τ(t) is the time delay.

3. The ISAR image compensation method under low signal-to-noise ratio conditions according to claim 1, characterized in that: In step S2, the radar echo data is subjected to denoising according to the following steps: Step S201: De-noise the radar echo data by time domain filtering: x denoised (t)=Filter(x(t)) (8) Wherein, Filter(x(t)) is a filter, which is one of the mean filter, median filter, and Gaussian filter; Step S202: De-noise the radar echo data by frequency domain filtering: X(f)=F(x(t)) (9) X denoised (f)=Filter(X(f)) (10) x denoised (t)=F -1 (X denoised (f)) (11) Where F is the Fourier transform, X(f) is the spectrum of the signal, and Filter(X(f)) is the frequency domain filter spectrum, which is one of the mean filter, median filter, and Gaussian filter. Step S203: Remove interference from radar echo data through time domain filtering: x clean (t)=x(t)-Interference(t) (12) Among them, Interference(t) is the identified interference part; Step S204: Remove interference from the radar echo data through frequency domain filtering: X clean (f)=X(f)-Interference(f) (13) x clean (t)=F -1 (X clean (f)) (14) Where Interference(f) is the frequency domain part of the identified interference.

4. The ISAR image compensation method under low signal-to-noise ratio conditions according to claim 1, characterized in that: Based on the difference between the neural network output and the actual translation parameters as compensation, the following steps are also included: Step S501: Using the ISAR complex image I as the input X of the neural network mapping function, the complex image I has an image size of h×w×2, where h is the image height, w is the image width, 2 is the number of channels, and the real part and the imaginary part each occupy one channel; Step S502: Construct a feature extraction module, including convolution blocks Conv_block1 and Conv_block2, and maximum pooling layers Max_pool1 and Max_pool2. Conv_block1 includes two convolution layers Conv1 and Conv2 with activation function ReLU, and the convolution kernel size is 3×3. It takes the ISAR image I as input and outputs a feature map I0 with a size of h×w×64. The maximum pooling layer Max_pool1 is used for dimensionality reduction and outputs a feature map I1 with a size of h1×w1×64, where h1=h / 2 and w1=w / 2. Conv_block2 includes two convolutional layers Conv3 and Conv4 with activation function ReLU. The convolution kernel size is 3×3. Conv_block2 takes feature map I1 as input and outputs feature map I2 with size h1×w1×128. It uses the maximum pooling layer Max_pool2 to reduce the dimension and output feature map I3 with size h2×w2×256, where h2=h1 / 2 and w2=w1 / 2. The feature map I3 output by the feature extraction module is used as the input of the feature encoding module; Step S503: construct a feature encoding module, including two self-attention networks Transformer1 and Transformer2, as well as a batch normalization layer BN5 and a convolutional layer Conv5. Transformer1 includes batch normalization BN1, window-based multi-head self-attention W-MSA, batch normalization BN2 and multi-layer perceptron MLP1. The input of BN1 is residually connected to the output of W-MSA, and the input of BN2 is residually connected to the output of MLP1. Transformer1 takes feature map I3 as input and outputs feature map I″3. Among them, I′3 is the intermediate output feature map, Transformer2 includes batch normalization BN3, multi-head self-attention SW-MSA based on shift window, deep convolution DWConv, batch normalization BN4 and multi-layer perceptron MLP2, among which the input of BN3 is residually connected to the output of SW-MSA, the input and output of DWConv are residually connected, and the input of BN4 is residually connected to the output of MLP2. Transformer2 takes feature map I″3 as input and outputs feature map I″4. Among them, I4 and I′4 are intermediate output feature maps, the kernel size of the last convolution layer Conv5 of the feature encoding module is 3×3, and the final output of the feature encoding module is the feature map I5 with a size of h2×w2×256. I5=Conv5[BN5(I'4')]+I3 (19) The feature map I5 output by the feature encoding module is used as the input of the feature decoding module; Step S504: construct a feature decoding module, which includes two convolution blocks Conv_block3 and Conv_block4, two sub-pixel convolution layers SP_Conv1 and SP_Conv2, and a convolution 1×1Conv with a kernel size of 1. The feature map I5 output by the feature encoding module is used as input, and the sub-pixel convolution layer SP_Conv1 is used to implement upsampling. The size of the output feature map is h1×w1×128, and then it is integrated with the feature map I2 by pixel-by-pixel addition to output a feature map I6 with a size of h1×w1×256. I6=SP_Conv1(I5)+I2 (20) And as the input of Conv_block3, Conv_block3 includes 2 convolutional layers Conv6 and Conv7 with activation function ReLU, convolution kernel size is 3×3, and output feature map I7 with size h1×w1×128, I7=Conv7[Conv6(I6) ReLU ] ReLU (21) The sub-pixel convolution layer SP_Conv2 is used to achieve upsampling, and the output feature map of size h×w×64 is integrated with the feature map I0 by pixel-by-pixel addition, and the output feature map I8 of size h×w×128 is output. I8=SP_Conv2(I7)+I0 (22) And as the input of Conv_block4, Conv_block4 includes 2 convolutional layers Conv8 and Conv9 with activation function ReLU, the convolution kernel size is 3×3, Conv_block4 takes feature map I8 as input and outputs feature map I9 with size h×w×64, I9=Conv9[Conv8(I8) ReLU ] ReLU (23) Reuse the convolution Conv with a kernel size of 1 1×1 Transform and finally output feature map I 10 , I 10 =Conv 1×1 (I9) (24) The size is h×w×2, with feature map I 10 As output Y.

5. The ISAR image compensation method under low signal-to-noise ratio conditions according to claim 1, characterized in that: The step S6, adjusting the weights and biases of the neural network is as follows: Step S601, input layer to hidden layer: H 1 =σ(W 1 X+b 1 ) (25) Among them, H 1 is the output of the hidden layer, σ is the activation function, which is a nonlinear function such as ReLU (RectifiedLinearUnit) or Sigmoid, W 1 is the weight matrix from the input layer to the hidden layer, b 1 is the bias vector from the input layer to the hidden layer; Step S602, hidden layer to output layer: Y=σ(W 2 H 1 +b 2 ) (26) Among them, W 2 is the weight matrix from the hidden layer to the output layer, b 2 is the bias vector from the hidden layer to the output layer; Step S603, using a gradient descent algorithm with momentum to update the parameters of the neural network, initializing the weights, biases, and momentum terms of the neural network, where the momentum term is used to record the direction and magnitude of the previous parameter update; Step S604: Randomly select a batch of samples from the training data set. During the back propagation process of the current batch, calculate the average values ​​of the gradients dW and db obtained by differentiating the selected samples with respect to the neural network parameters. Calculate the new momentum term v based on the current average gradient and the momentum term of the previous parameter update. dW and v db , and use the new momentum term to update the neural network parameters, weights W and bias b, Among them, α is the learning rate and β is the hyperparameter; Step S606: Repeat steps S604 to S605 until the preset number of training epochs is reached. The loss function L(W,b) is expressed as follows: Among them, N is the number of samples, M is the number of output features, and Y ij Output the predicted value for the network, is the true value of the label, the optimization goal is to minimize the loss function value, and the neural network parameters are trained as compensation parameters.

6. The ISAR image compensation method under low signal-to-noise ratio conditions according to claim 1, characterized in that: The step S7, adjusting the adaptive compensation model parameters according to the image quality index, including one or more of peak signal-to-noise ratio, contrast, image standard deviation, and image average gradient, is achieved by one or more of the following steps: Step S701, peak signal-to-noise ratio: Among them, PSNR represents the ratio between the energy of the maximum signal in the image and the background noise, I, K, M, N represent the image size of the current image I and the reference image K is M×N, MAX1 is the maximum grayscale value, usually 255; Step S702, contrast: Among them, δ(i,j)=|ij| represents the grayscale difference between adjacent pixels, P δ (i, j) represents the distribution probability of pixels with grayscale difference between adjacent pixels; Step S703, image standard deviation: Where f(i,j) is the image pixel value and μ is the image mean; Step S704, image average gradient value: Among them, G x and G y The gradient values ​​of the image in the horizontal and vertical directions.

7. An ISAR image compensation system under low signal-to-noise ratio conditions, used to implement the method according to any one of claims 1 to 6, characterized in that: include: A data acquisition module, a data recording system, a data preprocessing module, a translational information extraction module, an adaptive compensation model establishment module, a neural network training platform, an evaluation and tuning module, and an iterative optimization module. The data acquisition module includes an ISAR radar system, a translational data collector, and an environmental monitoring sensor. The radar system, the translational data collector, and the environmental monitoring sensor are connected to the data recording system through a communication interface, and then connected in sequence to the data preprocessing module, the translational information extraction module, the adaptive compensation model establishment module, the neural network training platform, the evaluation and tuning module, and the iterative optimization module through a data interface.

8. The ISAR image compensation system under low signal-to-noise ratio conditions according to claim 7, characterized in that: The data preprocessing module, translation information extraction module, adaptive compensation model establishment module, evaluation and tuning module, and iterative optimization module are high-performance computing servers or data analysis workstations.

9. The ISAR image compensation system under low signal-to-noise ratio conditions according to claim 8, characterized in that: The neural network training platform is a deep learning workstation.

Images