Error correction method and system for synthetic aperture radiometer, terminal equipment and medium

Abstract

The invention provides a synthetic aperture radiometer error correction method and system, terminal equipment and a medium, and the method comprises the steps: obtaining an original brightness temperature image of a target near-field imaging scene, and obtaining a corresponding initial near-field visibility function; inputting the initial near-field visibility function into a pre-constructed near-field error compensation network model, extracting frequency domain feature information, and compensating a near-field error to obtain a new near-field visibility function; calculating an error between the new near field visibility function and the target far field visibility function, and if the error is less than or equal to a preset error threshold, performing Fourier transform inversion on the new near field visibility function to obtain a corrected reconstructed brightness temperature image; otherwise, constructing a loss function according to the error, carrying out the back propagation of the near-field error compensation network model, obtaining a trained near-field error compensation network model, and obtaining a corrected reconstructed brightness temperature image based on the trained near-field error compensation network model. The near-field error can be corrected, and the quality of the reconstructed brightness temperature image is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of microwave remote sensing and detection, and particularly relates to a method, system, terminal device and medium for error correction of a synthetic aperture radiometer. Background Art

[0002] In recent years, applications such as security imaging for detecting hidden dangerous goods on the human body, close-range target imaging under harsh visual conditions, and detection of buried objects underground have also put forward an urgent need for high-resolution passive microwave near-field imaging. Therefore, the research on near-field imaging technology is of great significance. Under near-field conditions, the target signals received by the synthetic aperture radiometer are spherical waves, rather than plane wave signals. This difference makes it impossible to describe the relationship between the visibility function of the system and the target brightness temperature distribution with a simple Fourier transform, thus causing the problem of overlapping blur in the near-field reconstructed brightness temperature image. Summary of the Invention

[0003] Aiming at the above defects or improvement requirements of the prior art, the present invention provides a method, system, terminal device and medium for error correction of a synthetic aperture radiometer, aiming to correct near-field errors and improve the quality of the reconstructed brightness temperature image. Thus, the technical problem of overlapping blur in the near-field reconstructed brightness temperature image caused by near-field errors in the prior art is solved.

[0004] In a first aspect, the present invention provides a method for error correction of a synthetic aperture radiometer, the method comprising the following steps:

[0005] Obtain the original brightness temperature image of the target near-field imaging scene, and obtain the initial near-field visibility function corresponding to the target near-field imaging scene according to the original brightness temperature image;

[0006] Input the initial near-field visibility function into a pre-constructed near-field error compensation network model to obtain a new near-field visibility function; the near-field error compensation network model is used to extract the frequency-domain feature information of the initial near-field visibility function, and based on the frequency-domain feature information, compensate for the near-field errors of the initial near-field visibility function, and the near-field errors include phase errors and amplitude errors;

[0007] Calculate the error between the new near-field visibility function and a preset target far-field visibility function. If the error is less than or equal to a preset error threshold, perform an inverse Fourier transform on the new near-field visibility function to obtain the corrected reconstructed brightness temperature image of the target near-field imaging scene; otherwise, construct a loss function according to the error, and use the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, obtain the trained near-field error compensation network model, and input the initial near-field visibility function into the trained near-field error compensation network model. After performing an inverse Fourier transform on the new near-field visibility function output by the trained near-field error compensation network model, obtain the corrected reconstructed brightness temperature image.

[0008] Optionally, the near-field error compensation network model includes an information extraction module, a first feature extraction module, a second feature extraction module, an anti-overfitting module, a feature recombination module, and a feature aggregation module. The information extraction module is used to extract the amplitude information and phase information of the visibility function. The first feature extraction module is used to extract the shallow frequency-domain features of the visibility function based on the amplitude information and phase information. The shallow frequency-domain features characterize the overall structure and distribution law of the frequency-domain data, mainly reflecting the change trends of the near-field and far-field visibility functions in terms of global amplitude and phase. The second feature extraction module is used to extract the deep frequency-domain features of the visibility function from the shallow frequency-domain features. The deep frequency-domain features characterize the local detail differences of the frequency-domain data, depicting the fine structure deviation and local distortion caused by the near-field effect. The anti-overfitting module is used to randomly deactivate some neurons in the deep frequency-domain features to prevent the model from overfitting. The feature recombination module is used to recombine the deep frequency-domain features output by the anti-overfitting module to obtain a multi-channel feature map. The feature aggregation module is used to perform average pooling on the multi-channel feature map to achieve near-field error compensation.

[0009] Optionally, both the first feature extraction module and the second feature extraction module are composed of five sequentially connected residual convolution blocks;

[0010] The shallow frequency-domain features are expressed as:

[0011] F1 = X5(F0) + conv(F0)

[0012] X i (F0) = ReLU(BN(conv2(ReLU(conv1(F0))))) + shortcut, i = 1

[0013] X i (F0) = ReLU(BN(conv2(ReLU(conv1(X i-1 (F0))))) + shortcut, i = 2, 3, 4, 5

[0014] where, F1 represents the shallow frequency-domain features, F0 represents the amplitude information and phase information of the visibility function output by the information extraction module, F0 = Normalize(stack(Flatten(Re(X)), Flatten(Im(X)))), Re(X) represents the amplitude information extracted from the real part of the visibility function X, Im(X) represents the phase information extracted from the imaginary part of the visibility function X, Flatten(·) represents the flattening operation, stack(·) represents the concatenation operation, Normalize(·) represents the normalization operation, conv(F0) represents the convolution operation on F0, used to extract the original features of F0, X irepresents the output of the i-th residual convolution block, conv1 represents the first convolutional layer in the residual convolution block, ReLU(·) represents the activation function, conv2 represents the second convolutional layer in the residual convolution block, BN(·) represents batch normalization, and shortcut represents the skip connection;

[0015] The deep frequency domain feature is expressed as:

[0016] F2 = X5(F1) + conv(F1)

[0017] X i (F1) = ReLU(BN(conv2(ReLU(conv1(F1))))) + shortcut, i = 1

[0018] X i (F1) = ReLU(BN(conv2(ReLU(conv1(X i-1 (F1))))) + shortcut, i = 2, 3, 4, 5

[0019] Among them, F2 represents the deep frequency domain feature.

[0020] Optionally, the deep frequency domain features output by the anti-overfitting module are recombined to obtain a multi-channel feature map, including:

[0021] Through the calculation formula

[0022] F4 = conv4(LeakyReLU(conv3(F3)))

[0023] F3 = dropout(F2)

[0024] to obtain the multi-channel feature map F4; conv3 represents the third convolutional layer in the feature recombination module, LeakyReLU(·) represents the activation function, which is used to retain partial gradient responses for negative values and improve the network's ability to capture subtle features, conv4 represents the fourth convolutional layer in the feature recombination module, and dropout(F2) represents the result of randomly inactivating some neurons in F2 by the anti-overfitting module.

[0025] Optionally, the error is the mean absolute error.

[0026] Optionally, the expression of the error loss function is:

[0027]

[0028] Among them, Loss represents the error loss value, n = 1,.., N, N represents the total number of tiny patches in the original brightness temperature image, V label represents the preset target far-field visibility function, Vout Represents the near-field visibility function.

[0029] Optionally, the original brightness temperature image includes multiple microfacets.

[0030] In a second aspect, the present invention provides a synthetic aperture radiometer error correction system, including:

[0031] A data acquisition module, configured to acquire the original brightness temperature image of the target near-field imaging scene, and obtain the initial near-field visibility function corresponding to the target near-field imaging scene according to the original brightness temperature image;

[0032] An error compensation module, configured to input the initial near-field visibility function into a pre-constructed near-field error compensation network model to obtain the new near-field visibility function; the near-field error compensation network model is used to extract the frequency-domain feature information of the initial near-field visibility function, and based on the frequency-domain feature information, compensate for the near-field errors of the initial near-field visibility function, and the near-field errors include phase errors and amplitude errors;

[0033] A reconstruction module, configured to calculate the error between the new near-field visibility function and a preset target far-field visibility function. If the error is less than or equal to a preset error threshold, perform an inverse Fourier transform on the new near-field visibility function to obtain the corrected reconstructed brightness temperature image of the target near-field imaging scene; otherwise, construct a loss function according to the error, use the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, obtain the trained near-field error compensation network model, and input the initial near-field visibility function into the trained near-field error compensation network model. After performing an inverse Fourier transform on the new near-field visibility function output by the trained near-field error compensation network model, obtain the corrected reconstructed brightness temperature image.

[0034] In a third aspect, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above method is implemented.

[0035] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above method is implemented.

[0036] The beneficial effects of the present invention are:

[0037] The error correction method for synthetic aperture radiometer provided by the present invention extracts the frequency-domain characteristic information of the initial near-field visibility function, and compensates for the near-field error of the initial near-field visibility function based on the frequency-domain characteristic information, which can correct the near-field error and improve the imaging quality. A loss function is constructed according to the error, and the near-field error compensation network model is backpropagated using the loss function until the error is less than or equal to the preset error threshold. This can not only realize the mapping from the near-field visibility function to the target far-field visibility function in the frequency domain, but also flexibly adjust the compensation strength according to the target far-field visibility function, further improving the imaging quality. Description of the Drawings

[0038] Figure 1 It is the imaging principle diagram of the synthetic aperture radiometer;

[0039] Figure 2 It is the flowchart of the error correction method for synthetic aperture radiometer in one implementation manner of the present application;

[0040] Figure 3 It is the model structure diagram of the near-field error compensation network model in one implementation manner of the present application;

[0041] Figure 4a It is the reconstructed image before correcting the near-field error in one implementation manner of the present application;

[0042] Figure 4b It is the reconstructed image corresponding to the preset target far-field visibility function in one implementation manner of the present application;

[0043] Figure 4c It is the reconstructed image corrected by using the error correction method for synthetic aperture radiometer in one implementation manner of the present application;

[0044] Figure 5a It is the real part spectral image in the u direction of the brightness temperature image corresponding to different visibility functions in one implementation manner of the present application;

[0045] Figure 5b It is the real part spectral image in the v direction of the brightness temperature image corresponding to different visibility functions in one implementation manner of the present application;

[0046] Figure 5c It is the imaginary part spectral image in the u direction of the brightness temperature image corresponding to different visibility functions in one implementation manner of the present application;

[0047] Figure 5d It is the imaginary part spectral image in the v direction of the brightness temperature image corresponding to different visibility functions in one implementation manner of the present application;

[0048] Figure 6aThe IFFT reconstruction image after correction by the error correction method of the synthetic aperture radiometer in one embodiment of the present application;

[0049] Figure 6b The IFFT reconstruction image after correction by the traditional method in one embodiment of the present application;

[0050] Figure 7 The structural schematic diagram of the synthetic aperture radiometer error correction system in one embodiment of the present application;

[0051] Figure 8 The structural schematic diagram of the terminal device in one embodiment of the present application. Specific embodiments

[0052] Aiming at the technical problem that the near-field error in the prior art leads to the overlapping and blurring of the near-field reconstructed brightness temperature image, the present invention discloses a synthetic aperture radiometer error correction method, system, terminal device and medium. The method extracts the frequency domain feature information of the initial near-field visibility function, and based on the frequency domain feature information, compensates for the near-field error of the initial near-field visibility function, which can correct the near-field error and improve the imaging quality; constructs a loss function according to the error, and uses the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, which can not only realize the mapping from the near-field visibility function to the target far-field visibility function in the frequency domain, but also flexibly adjust the compensation intensity according to the target far-field visibility function, further improving the imaging quality.

[0053] For the convenience of understanding, the imaging principle of the synthetic aperture radiometer is first described.

[0054] Specifically as Figure 1 shown, there is a radiation source S on the target scene plane, and the radiation source is divided into multiple tiny surface elements ds. OXY is the antenna plane, located in the plane of z = 0, P1 and P2 are two receiving antennas, and D represents the distance between the two antennas P1 and P2. The distance from the target scene plane to the antenna plane is d, and the distances from the radiation surface element to the two antenna elements are R1 and R2 respectively. The distances from the antennas to the origin of coordinates are d1 and d2 respectively, and the distance from the radiation surface element to the origin is R S . The zenith angle of the position where the radiation surface element is located is θ, and the azimuth angle is

[0055] Assume that the coordinates of the tiny surface element ds in the Cartesian coordinate system are (x, y, z), and the two antennas are (x1, y1, 0) and (x2, y2, 0) respectively. The radiation field strengths received by the two antenna elements are respectively expressed as E1(R1, t) and E2(R2, t), and the visibility function obtained after the complex correlation processing of the two signals is:

[0056]

[0057] In the formula, is the angle formed by the radiation surface element to each antenna element, and T r is the system noise, is the Fringewashing function, and D i is the directivity coefficient of the antenna, and F n is the field strength pattern of the antenna. The exponential part represents the phase difference caused by the different distances from the incident electromagnetic wave to the two antenna elements that form the baseline, and T B represents the bright temperature image of the spatial domain scene, and * represents the conjugate complex number.

[0058] For the entire synthetic aperture measurement system, the visibility function expression obtained from the correlation output of the entire observation field plane is:

[0059]

[0060] The distance R from the antenna element to the point source i is:

[0061]

[0062] Under the far-field condition, that is Approximating R i with the first-order Taylor series, that is:

[0063]

[0064] where Also So the above formula can be expressed as:

[0065]

[0066] Therefore

[0067]

[0068] where D x = x j - x i , D y = y j - y i , under the far-field condition Let and be defined as the difference between the positions of the two antenna elements and the wavelength normalized, and let represent the direction cosine. Substituting the above expressions into the formula, we can get

[0069]

[0070] Define the corrected bright temperature as

[0071]

[0072] Ignoring the Fringe washing function and assuming the identity of all antenna elements, ideally, the relationship between the visibility function and the bright temperature data can be expressed as:

[0073]

[0074] It can be found that under far-field conditions, the visibility function and the bright temperature data have a Fourier transform relationship.

[0075] However, in near-field imaging scenarios (such as concealed dangerous goods security inspection imaging of the human body), when the radiation wave of the radiation source in the near-field region reaches the antenna, it cannot be regarded as a plane wave but a spherical wave. Therefore, the distance difference between the target source and different antennas cannot be ignored. Thus, the visibility function with near-field error is expressed as:

[0076]

[0077] It is found that compared with the far-field, the biggest difference in the near-field visibility function is the addition of this near-field error term. The existence of the near-field error term causes the use of traditional synthetic aperture imaging methods directly to result in aberration problems such as blurred and defocused imaging results. Moreover, the closer the target is to the system, the more serious the deterioration effect is.

[0078] To solve the above problems, the present invention provides a method for correcting the error of a synthetic aperture radiometer, as Figure 2 shown, the method includes the following steps:

[0079] Step 21, obtaining the original bright temperature image of the target near-field imaging scene, and obtaining the initial near-field visibility function corresponding to the target near-field imaging scene according to the original bright temperature image.

[0080] It should be understood that the synthetic aperture radiometer emits signals through electromagnetic waves of a series of frequencies (usually in the microwave or millimeter-wave band) and receives the echoes of the target area. These signals include the radiation reflected from the target object and the radiation emitted from the target surface. By measuring the intensity and spectral characteristics of these radiations, the radiometer can deduce the temperature distribution of the target surface. The bright temperature data collected by the radiometer and processed can be presented by an imaging method. These images are usually displayed in the form of a pseudo-color map or a grayscale map, and each pixel in the bright temperature image represents the surface temperature at a specific position.

[0081] Obtaining the initial near-field visibility function corresponding to the target near-field imaging scene according to the original bright temperature image has been described in detail in the content of the imaging principle part above and will not be elaborated here.

[0082] Step 22: Input the initial near-field visibility function into the pre-constructed near-field error compensation network model to obtain a new near-field visibility function.

[0083] In an embodiment of the present invention, the near-field error compensation network model is used to extract the frequency-domain characteristic information of the initial near-field visibility function, and based on the frequency-domain characteristic information, compensate for the near-field errors of the initial near-field visibility function. The above-mentioned near-field errors include phase errors and amplitude errors.

[0084] The structure of the near-field error compensation network model will be described below.

[0085] Specifically, as Figure 3 shown, the near-field error compensation network model includes an information extraction module 31, a first feature extraction module 32, a second feature extraction module 33, an anti-overfitting module 34, a feature recombination module 35, and a feature aggregation module 36 that are connected in sequence.

[0086] Among them, the information extraction module 31 is used to extract the amplitude information and phase information of the visibility function.

[0087] Specifically, in an embodiment of the present invention, considering that the visibility function includes a real part and an imaginary part, the real part contains amplitude information, and the imaginary part contains phase information. Directly inputting it into the near-field error compensation network model cannot effectively decompose the inherent characteristics of the visibility function. Therefore, before inputting into the network, the data is first flattened and normalized to obtain a 1*165 matrix. Then, the real part and the imaginary part of the visibility function are extracted respectively, and the data is combined into a 2-channel data of 2*165 using the stack function. This can not only effectively decompose the inherent characteristics of the visibility function but also optimize the learning process of the network. This process is specifically expressed as F0 = Normalize(stack(Flatten(Re(X)), Flatten(Im(X)))), where Re(X) represents the amplitude information extracted from the real part of the visibility function X, Im(X) represents the phase information extracted from the imaginary part of the visibility function X, Flatten(·) represents the flattening operation, stack(·) represents the splicing operation, and Normalize(·) represents the normalization operation.

[0088] The first feature extraction module 32 is used to extract the shallow frequency-domain characteristics of the visibility function according to the amplitude information and phase information.

[0089] Specifically, in an embodiment of the present invention, both the first feature extraction module 32 and the second feature extraction module 33 are composed of five residual convolutional blocks connected in sequence (residual convolutional block 3A ( Figure 3 3A in Figure 3 ), residual convolutional block 3B (Figure 3 in 3C), residual convolutional block 3D( Figure 3 in 3D), residual convolutional block 3E( Figure 3 in 3E)), and each residual convolutional block includes two layers of one-dimensional convolution (conv1, conv2), residual connection, batch normalization (BN), and ReLU activation operation. Moreover, two skip connections are used between the modules. The first skip connection maps the amplitude information and phase information of the visibility function output by the information extraction module 31 into multi-channel data, adds it to the output of the first feature extraction module 32, retains the feature details of the input data, and enhances feature extraction. The second skip connection maps the amplitude information and phase information of the visibility function output by the information extraction module 31 into more channels, adds it to the deep frequency domain features output by the second feature extraction module 33, and realizes the effective fusion of shallow frequency domain features and deep frequency domain features, improving the feature expression ability.

[0090] Specifically, the shallow frequency domain features extracted by the first feature extraction module 32 are expressed as:

[0091] F1 = X5(F0) + conv(F0)

[0092] X i (F0) = ReLU(BN(conv2(ReLU(conv1(F0))))) + shortcut, i = 1

[0093] X i (F0) = ReLU(BN(conv2(ReLU(conv1(X i-1 (F0))))) + shortcut, i = 2, 3, 4, 5

[0094] Among them, F1 represents the shallow frequency domain features, F0 represents the amplitude information and phase information of the visibility function output by the information extraction module 31, conv(F0) represents the convolution operation on F0 for extracting the original features of F0, X i represents the output of the i-th residual convolutional block, conv1 represents the first convolutional layer in the residual convolutional block, ReLU(·) represents the activation function, conv2 represents the second convolutional layer in the residual convolutional block, BN(·) represents batch normalization, and shortcut represents the skip connection.

[0095] The second feature extraction module 33 is used to extract the deep frequency domain features of the visibility function from the shallow frequency domain features. The deep frequency domain features extracted by the second feature extraction module 33 are expressed as:

[0096] F2 = X5(F1) + conv(F1)

[0097] Xi (F1) = ReLU(BN(conv2(ReLU(conv1(F1))))) + shortcut, i = 1

[0098] X i (F1) = ReLU(BN(conv2(ReLU(conv1(X i-1 (F1)))))) + shortcut, i = 2, 3, 4, 5

[0099] Among them, F2 represents the deep - frequency - domain feature.

[0100] The anti - overfitting module 34 is used to randomly inactivate some neurons in the deep - frequency - domain feature to prevent the model from overfitting.

[0101] Specifically, the process of randomly inactivating some neurons in the deep - frequency - domain feature can be expressed as F3 = dropout(F2). This module can improve the generalization ability of the network.

[0102] The feature recombination module 35 is used to recombine the deep - frequency - domain feature output by the anti - overfitting module 34 to obtain a multi - channel feature map.

[0103] Specifically, through the calculation formula

[0104] F4 = conv4(LeakyReLU9conv3(F3)))

[0105] F3 = dropout(F2)

[0106] Obtain the multi - channel feature map F4; conv3 represents the third convolutional layer in the feature recombination module 35, which is used to reduce the number of channels, screen out the key features in the multi - channels and reduce redundancy, LeakyReLU(·) represents the activation function, which is used to retain partial gradient responses for negative values and improve the network's ability to capture subtle features, conv4 represents the fourth convolutional layer in the feature recombination module 35, which is used to recombine the dimension - reduced features. This combination can integrate the multi - scale features extracted by the residual module and provide rich feature information for subsequent global pooling and output estimation.

[0107] The feature aggregation module 36 is used to perform average pooling on the multi - channel feature map to achieve near - field error compensation.

[0108] Step 23: Calculate the error between the newly obtained near-field visibility function and the preset target far-field visibility function. If the error is less than or equal to the preset error threshold, perform an inverse Fourier transform on the newly obtained near-field visibility function to obtain the reconstructed brightness temperature image after correction of the target near-field imaging scene; otherwise, construct a loss function based on the error, and use the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, obtain the trained near-field error compensation network model, and input the initial near-field visibility function into the trained near-field error compensation network model. After performing an inverse Fourier transform on the newly obtained near-field visibility function output by the trained near-field error compensation network model, obtain the reconstructed brightness temperature image after correction.

[0109] Specifically, in the embodiment of the present invention, the above-mentioned error is the mean absolute error.

[0110] The expression of the loss function constructed according to the error is:

[0111]

[0112] where Loss represents the error loss value, n = 1,.., N, N represents the total number of tiny surface elements in the original brightness temperature image, V label represents the preset target far-field visibility function, and V out represents the newly obtained near-field visibility function.

[0113] In the embodiment of the present invention, the process of inputting the initial near-field visibility function into the trained near-field error compensation network model to obtain the reconstructed brightness temperature image after correction includes:

[0114] Input the initial near-field visibility function into the trained near-field error compensation network model to obtain the corrected newly obtained near-field visibility function;

[0115] Perform an inverse Fourier transform on the corrected newly obtained near-field visibility function to obtain the reconstructed brightness temperature image after correction of the target near-field imaging scene.

[0116] To verify the effectiveness of the present invention, the following comparative experiments were conducted.

[0117] Example 1

[0118] In this embodiment, the original brightness temperature image is selected from the NWPU-RESISC45 remote sensing image scene classification dataset. The NWPU-RESISC45 dataset is a large-scale publicly available dataset for remote sensing image scene classification, including 45 types of scenes such as airports, basketball courts, deserts, industrial areas, forests, and rivers. Each scene has 700 remote sensing images, for a total of 31,500 images. The images are rich in information and cover a wide range of contents.

[0119] After obtaining the original scene brightness temperature image dataset, the visibility function is obtained through an ideal uniform sampling synthetic aperture system simulation program. That is, each antenna pair receives the radiation signals from the radiation sources in the original scene, and through complex correlation calculation and redundancy removal operations, the visibility functions under near-field and far-field conditions are generated respectively. The generated simulation data is divided into training dataset, validation dataset, and test training set according to a certain ratio. The antenna array used in the ideal uniform sampling synthetic aperture system simulation program is a matrix antenna array composed of 24 antenna elements.

[0120] After determining the parameters of the simulation program, the steps to generate the simulation dataset are as follows:

[0121] 1) Select any original scene brightness temperature image (P) from the original scene brightness temperature image dataset, and input P into the ideal synthetic aperture radiometer simulation program;

[0122] 2) Set the distances between the scene plane and the antenna plane to 0.8 m and 5 m respectively, and generate the ideal simulation synthetic aperture visibility function V(u, v);

[0123] 3) After processing each original scene image through steps 1 and 2 respectively, combine all the obtained samples together to form the simulation dataset (D);

[0124] 4) Divide the simulation dataset D into training, validation, and test datasets according to a certain ratio.

[0125] Among them, the visibility function at 0.8 m is the input dataset of the network, and the visibility function at 5 m is used as the label dataset of the network. In order to remove the irrelevant DC components and make the network more focused on the meaningful frequency domain information, the values at the zero frequency of the dataset are set to zero. After obtaining the corrected visibility function through testing, then extract the values at the zero frequency of the input dataset and fill them into the zero frequency of the corrected visibility function.

[0126] The experimental results are as Figure 4a , Figure 4b , Figure 4c shown, where Figure 4a is the reconstructed image before correcting the near-field error, Figure 4b is the reconstructed image corresponding to the preset target far-field visibility function, Figure 4c is the reconstructed image corrected by using the synthetic aperture radiometer error correction method provided by this application. As can be seen from Figures 4a to 4c , the brightness temperature image ([[]] Figure 4c ) obtained after near-field error correction is better than the image before correction ([[]] Figure 4a) is much clearer. The image directly obtained by inverting the near-field frequency-domain data shows serious blurring and distortion. This is because under near-field conditions, the electromagnetic wave is a spherical wave, and the intensity of the electromagnetic wave received at different positions is different. Moreover, due to the short distance, interference and diffraction phenomena may occur, resulting in distortion and distortion of the near-field brightness temperature image. Intuitively, the brightness temperature image corresponding to the corrected visibility function ( Figure 4c ) and the image corresponding to the visibility function of 5m used as the training label in the experiment ( Figure 4b ) have similar effects, and even reduce the background error to a certain extent. Therefore, the error correction method of the synthetic aperture radiometer proposed in this paper has a good correction effect and obtains a corresponding high-quality image in the spatial domain.

[0127] After qualitatively comparing the images obtained by inverting the visibility function before and after correction, in another embodiment of the present invention, RMSE, peak signal-to-noise ratio PSNR, and SSIM are used to quantitatively compare the quality of different images. The calculations of RMSE and PSNR are both based on the original scene brightness temperature image corresponding to the data. Among them, RMSE is an overall measure of the error between the inverted brightness temperature image and the original scene brightness temperature image. The smaller the RMSE, the closer the image is to the original scene. PSNR is also an overall measure of the quality of the inverted brightness temperature image. The larger the value, the smaller the distortion of the reconstructed brightness temperature image.

[0128] The calculation formulas of RMSE, PSNR, and SSIM are as follows:

[0129]

[0130]

[0131] Where M and N are the number of pixels of the reconstructed image in two directions, T i is the brightness temperature image obtained by inverting the near-field, far-field, and near-field corrected visibility functions, T is the original scene brightness temperature image, and S is the dynamic range of the brightness temperature values in the original scene image. and μ T are the average values of T i and T respectively; and are the variances of T i and T respectively; is the covariance between T i and T; c1 and c2 are constants used to stabilize the formula.

[0132] The brightness temperature images obtained by inverting all the near-field visibility functions, far-field visibility functions, and corrected visibility functions are compared with their corresponding original target scenes, and the results are shown in Table 1. Among them, the average RMSE values are 35.73K, 29.07K, and 28.20K respectively, the average PSNR values are 12.34dB, 15.92dB, and 15.81dB respectively, and the average SSIM values are 0.1628, 0.2331, and 0.2287 respectively.

[0133] Table 1

[0134]

[0135] It can be seen from Table 1 that the brightness temperature image after near-field error correction has an obvious advantage in RMSE, indicating that the quality of the corrected image is high and the error is significantly reduced. In terms of PSNR and SSIM, the corrected image is significantly better than the near-field image before correction and is very close to the index of the far-field image. The improvement of PSNR indicates that the corrected image is closer to the real target scene in the overall brightness temperature distribution, and the improvement of SSIM shows that the corrected image has been improved to a certain extent in terms of structural similarity, and the edges and texture details in the image have been restored to a certain extent. In addition, the indexes of the corrected image in PSNR and SSIM are very close to those of the far-field image, further indicating that the synthetic aperture radiometer error correction method provided by this application can accurately learn and correct the differences between the near-field and far-field frequency domains.

[0136] In order to more intuitively evaluate the correction effect of the MSR-NFECNet method (the synthetic aperture radiometer error correction method provided by this application) in the frequency domain, in another embodiment of this application, after normalizing the near-field, far-field, and corrected visibility functions, a normalized spectral image is obtained. Since the difference in the two-dimensional spectral image is not very obvious, the maximum value points of the two-dimensional spectral image are extracted, and one-dimensional curves are intercepted along the u and v directions, and a locally enlarged spectrum is drawn in the region near zero frequency. By comparing the RMSE values of the real and imaginary parts of these curves, the correction ability of the synthetic aperture radiometer error correction method provided by this application for frequency domain errors can be further verified, as specifically Figures 5a to 5d shown. It can be seen from Figures 5a to 5d that there are obvious gaps between the near-field, far-field, and corrected spectra and the spectrum of the original image. In the imaginary part, whether in the u direction or the v direction, the corrected spectrum is significantly closer to the spectra of the far-field image and the original scene image. However, from the real part, the difference may not be very obvious. In another embodiment of this application, RMSE is also used to quantitatively compare the near-field, far-field, and corrected frequency domains with the original scene spectrum. The average result values obtained are shown in Table 2.

[0137] Table 2

[0138] 0.8m spectrum 5m spectrum Corrected spectrum Real(u) 0.0689 0.0605 0.1132 Real(v) 0.0681 0.0621 0.1232 Imag(u) 0.2145 0.1894 0.1965 Imag(v) 0.2162 0.2091 0.2125

[0139] As can be seen from Table 2, the corrected spectrum has significant optimization in the imaginary part. Compared with the near-field spectrum, the RMSE of the imaginary part after correction is reduced to 0.1965 and 0.2125 in the u and v directions respectively, which is closer to the trend of the far-field spectrum. This indicates that the correction method has obvious advantages in the error correction of the imaginary part spectrum. However, in terms of the RMSE value of the real part spectrum, the corrected result is not better than the near-field spectrum, and the real part RMSE in the u and v directions increases compared with the near-field spectrum. Although the real part correction effect is not as expected, from the overall imaging effect, the image quality after correction is significantly better than that before correction. The clarity and detail restoration of the target in the reconstructed image have both improved, indicating that the correction method has successfully optimized the error propagation in the imaging process while performing spectrum correction. The improvement of the imaginary part spectrum may have a more significant impact on the imaging quality, which to a certain extent makes up for the deficiency of the real part correction.

[0140] In order to comprehensively evaluate the generalization ability and performance stability of the model, in another embodiment of the present application, near-field data at different distances are used for testing and combined for training. By the verification effect of the model under different distance conditions, it is judged whether the error correction method of this synthetic aperture radiometer can adapt to the near-field error characteristics that have not appeared before.

[0141] Specifically, data at different distances are input for testing. Test set data with target scenes and array distances of 0.5 m, 1.5 m, 2 m, 3 m, and 4 m are respectively generated and input into the network with the visibility function of 0.8 m as the training set. The obtained output results are inversely transformed by the inverse Fourier transform, and the inverse-transformed image is compared with the original scene image. The calculated average RMSE is shown in Table 3.

[0142] Table 3

[0143] 0.5m 1.5m 2m 3m 4m RMSE 37.28 38.23 40.62 41.13 41.18 PSNR 12.7727 13.5469 13.2861 13.3480 13.3990 SSIM 0.1222 0.1447 0.1349 0.1452 0.1518

[0144] As can be seen from Table 3, the closer the data is to 0.8 m in distance, the higher the quality of the inverse-transformed image. However, there is still room for improvement in the overall effect. Therefore, in order to further verify the adaptability of the network to visibility functions at different distances, an attempt is made to generate a training set using data combinations at different distances.

[0145] Visibility functions at different distances are combined for training. The visibility functions of 0.8 m and 1.5 m are combined to generate a training set for model training. During testing, data of 0.8 m, 1.5 m, and the combined training set (0.8_1.5 m) are respectively input, and the obtained results are shown in Table 4.

[0146] Table 4

[0147] 0.8m 1.5m 0.8_1.5m RMSE 28.8881 28.8974 28.8928 PSNR 15.8061 15.8183 15.8122 SSIM 0.2293 0.2296 0.2295

[0148] As can be seen from Table 4, after training with a combination of visibility functions at two different distances, the test results are significantly improved. This indicates that the synthetic aperture radiometer error correction method provided by the present application can learn more extensive and diverse error characteristics from multi-distance data, thereby enhancing the correction ability for near-field errors. The way of data combination effectively enhances the generalization ability of the model, making it show higher robustness when dealing with different scenarios.

[0149] To verify the contribution of each module in the near-field error compensation network model of the present application to the performance improvement, in another embodiment of the present application, an ablation experiment was also conducted. The specific experimental results are shown in Table 5. In this ablation experiment, the second feature extraction module, skip connection, overfitting prevention module, and feature aggregation module in the near-field error compensation network model were removed respectively, and the RMSE values of the model on the test set under different combinations were recorded.

[0150] Table 5

[0151] Residual block 2 Skip connection Anti-overfitting module Feature aggregation module RMSE Res-CNN × √ √ √ 28.67 Res-CNN √ × √ √ 28.82 Res-CNN √ √ × √ 29.21 Res-CNN √ √ √ × 29.01 Res-CNN √ √ √ √ 28.2

[0152] As can be seen from Table 5, when the second feature extraction module is removed, the RMSE value of the model rises to 28.82, indicating that the residual block helps to improve the prediction accuracy of the model. When the skip connection is removed, the RMSE value slightly increases to 28.67, indicating that the skip connection is helpful for the efficient transmission of information and the stable update of gradients. After removing the overfitting prevention module, the RMSE further rises to 29.21, which shows that the overfitting prevention module plays an important role in reducing overfitting. When the feature aggregation module is removed, the RMSE value increases to 29.01, indicating that the feature aggregation module has a positive effect on extracting effective features and reducing computational complexity.

[0153] Overall, under the complete near-field error compensation network model, the RMSE value of the model is the lowest, which is 28.2. This indicates that each component in the network can significantly improve the performance of the model under the synergistic effect, verifying the rationality and effectiveness of the network structure design.

[0154] To verify the robustness of the method proposed in this paper and further evaluate the near-field error correction effect of the synthetic aperture radiometer error correction method provided by the present application under different error intensities, based on the ideal simulation data set, the visibility functions in the training set and the test set were respectively subjected to noise addition processing. The added Gaussian noise has a mean of 0 and standard deviations of 0.05, 0.1, 0.15, 0.2, 0.25, and 0.3.

[0155] In the case of training with noise-free data and testing with noisy data, Gaussian noise with different intensities was added to the 0.8m visibility function in the test set data. The noise intensities ranged from 0.05 to 0.3 (with a mean of 0 and standard deviations of 0.05, 0.1, 0.15, 0.2, 0.25, and 0.3). At different noise intensities, the synthetic aperture radiometer error correction method provided by the present application was used to correct the noisy data, and then the corresponding spatial domain image was inversely obtained. At the same time, the near-field noise data was directly inversely obtained. Finally, the images obtained by the two methods were compared with the original scene image. Two samples were randomly selected from the test dataset. At different noise intensities, all the data in the test set was processed by two different methods, and the average RMSE, PSNR, and SSIM of the reconstructed image compared with the original scene image were calculated. The results are shown in Tables 6(a) and 6(b). Table 6(a) shows the values calculated after correcting the visibility function using the synthetic aperture radiometer error correction method provided by the present application, and Table 6(b) shows the corresponding index values of the image inversely obtained by directly using the inverse Fourier transform for the uncorrected visibility function.

[0156] Table 6(a)

[0157] 0.05 0.1 0.15 0.2 0.25 0.3 RMSE 28.5107 28.7079 29.0049 29.4123 29.8541 30.3695 PSNR 15.6370 15.5428 15.4097 15.2283 15.0700 14.8789 SSIM 0.2252 0.2182 0.2103 0.2012 0.1939 0.1864

[0158] Table 6(b)

[0159] 0.05 0.1 0.15 0.2 0.25 0.3 RMSE 35.7352 35.7566 35.7920 35.8398 35.9028 35.9780 PSNR 12.3414 12.3463 12.3569 12.3723 12.3871 12.4047 SSIM 0.1624 0.1616 0.1604 0.1590 0.1576 0.1560

[0160] Combining Tables 6(a) to 6(b), it can be seen that after adding noise to the test dataset, the quality of the image deteriorates as the noise intensity increases. However, the synthetic aperture radiometer error correction method provided by the present application has better noise resistance. Compared with the traditional method, for all noise intensities, the synthetic aperture radiometer error correction method provided by the present application can always obtain higher PSNR, SSIM, and lower RMSE. The RMSE of the brightness temperature image reconstructed from the visibility function corrected by the synthetic aperture radiometer error correction method provided by the present application is always lower than 31K, and the PSNR is always higher than 14dB. However, the RMSE of the brightness temperature image obtained by the traditional method is always higher than 35.5K, and the PSNR is always lower than 12dB. Moreover, the SSIM value corresponding to the synthetic aperture radiometer error correction method provided by the present application is also always higher than that of the traditional method.

[0161] In the case of training with noisy data and testing with noisy data, the training set data with a noise intensity of 0.3 was used for training, and the synthetic aperture radiometer error correction method provided by the present application was also tested using datasets with various different noise intensities. The experimental results are shown in Table 7.

[0162] Table 7

[0163] 0.05 0.1 0.15 0.2 0.25 0.3 0.3 RMSE 27.9293 27.9548 28.0167 28.1029 28.2331 28.4182 30.3695 PSNR 15.0293 15.0221 14.9980 14.9579 14.9024 14.8266 14.8789 SSIM 0.2250 0.2236 0.2212 0.2178 0.2132 0.2074 0.1864

[0164] Based on Table 7, by comparing the test results with a noise intensity of 0.3 in Table 6, it can be found that after training with noisy data, both the RMSE and PSNR of the near-field correction image have certain improvements. Especially for the RMSE index, it has dropped from 30.37K to 28.42K. When testing with data of other noise intensities, it can be found that the smaller the noise intensity, the better the results in all aspects, and it is almost better than the results of training with noise-free data. This may be because the data measured by the radiometer in actual work is essentially a result of a mixture of target signals and various noises. Training with noisy data can more realistically simulate the characteristics of the radiometer measurement data, making the model more conform to the actual data distribution. In addition, adding noisy training data can also enhance the model's adaptability and robustness to noise, prompting it to focus on more universal and significant features in the data, thereby effectively avoiding overfitting.

[0165] In another embodiment of the present application, in order to verify the near-field error correction ability of the synthetic aperture radiometer error correction method provided by the present application for the actual measurement scenario of the system, multiple observation scenarios are selected as the extended source scenarios for experiments. Based on the obtained trained MSR-NFECNet network, the visibility functions obtained by the system are corrected using different methods respectively. Finally, the reconstructed results of the brightness temperature images of the extended source scenarios under different near-field error correction methods are as Figures 6a - 6b shown. Figure 6a is the reconstructed image of the synthetic aperture radiometer after near-field error correction by the synthetic aperture radiometer error correction method provided by the present application, Figure 6b is the reconstructed image of the synthetic aperture IFFT transform after near-field error correction using the traditional method. Combining Figures 6a to 6b it can be seen that after correcting the system-measured visibility using the traditional near-field error correction method and then performing brightness temperature reconstruction using the IFFT transform method, the background error of its imaging is still relatively large. However, after correcting the system-measured visibility using the synthetic aperture radiometer error correction method provided by the present application and then using the IFFT transform method for brightness temperature reconstruction, the background is purer, the Gibbs oscillation is smaller, and the imaging quality is better than that after correction using the traditional near-field error correction method, verifying the ability of the synthetic aperture radiometer error correction method provided by the present application to correct the system-measured visibility of the extended source scenario.

[0166] Since the true brightness temperature distribution of the extended source scenario is unknown, the relative standard deviation RSD is used to measure the correction effect of different near-field error correction methods on the system-measured visibility function of the extended source scenario and the noise characteristics of the reconstructed brightness temperature image. The formula is as follows:

[0167]

[0168] For a total of 13 groups of measured data in the Zhanyuan scenario, for the visibility functions after calibration by the traditional near-field error correction method and the synthetic aperture radiometer error correction method provided in this application, the relative variances corresponding to the brightness temperature image reconstruction using the IFFT method are 0.5681 and 0.3932 respectively. By comparing the two methods, it can be found that the synthetic aperture radiometer error correction method provided in this application can reduce the relative variance of the inversion image by 30.79%. The results show that for the Zhanyuan scenario, the synthetic aperture radiometer error correction method provided in this application can improve the calibration effect of the visibility function of the ASR-Y system in the near-field measurement, and thus improve the quality of the reconstructed brightness temperature image.

[0169] In summary, the synthetic aperture radiometer error correction method provided by the present invention can correct the near-field error and improve the imaging quality by extracting the frequency-domain characteristic information of the initial near-field visibility function and compensating the near-field error of the initial near-field visibility function based on the frequency-domain characteristic information; constructing a loss function according to the error, and using the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, which can not only realize the mapping of the near-field visibility function to the target far-field visibility function in the frequency domain, but also flexibly adjust the compensation strength according to the target far-field visibility function to further improve the imaging quality. It provides a data basis for applications such as security inspection imaging, concealed item detection, medical imaging, and near-field remote sensing.

[0170] The synthetic aperture radiometer error correction system provided by the present invention will be described below. As Figure 7 shown, the synthetic aperture radiometer error correction system 700 includes:

[0171] A data acquisition module 701, configured to acquire the original brightness temperature image of the target near-field imaging scenario, and acquire the initial near-field visibility function corresponding to the target near-field imaging scenario according to the original brightness temperature image;

[0172] An error compensation module 702, configured to input the initial near-field visibility function into a pre-constructed near-field error compensation network model to obtain a new near-field visibility function; the near-field error compensation network model is used to extract the frequency-domain characteristic information of the initial near-field visibility function, and based on the frequency-domain characteristic information, compensate the near-field error of the initial near-field visibility function, and the near-field error includes phase error and amplitude error;

[0173] A reconstruction module 703 is configured to calculate the error between the newly acquired near-field visibility function and a preset target far-field visibility function. If the error is less than or equal to a preset error threshold, an inverse Fourier transform is performed on the newly acquired near-field visibility function to obtain a reconstructed brightness temperature image after correction of the target near-field imaging scene. Otherwise, a loss function is constructed based on the error, and the loss function is used to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, obtaining a trained near-field error compensation network model. Then, the initial near-field visibility function is input into the trained near-field error compensation network model, and after performing an inverse Fourier transform on the newly acquired near-field visibility function output by the trained near-field error compensation network model, a corrected reconstructed brightness temperature image is obtained.

[0174] It should be noted that for the information interaction, execution process, etc. between the above-mentioned devices / units, since they are based on the same concept as the method embodiments of the present application, for their specific functions and the technical effects brought, reference can be specifically made to the method embodiment part, and details will not be elaborated here. Those skilled in the art can clearly understand that for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of the functional units and modules are only for the convenience of mutual distinction and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, and details will not be elaborated here.

[0175] As Figure 8 shown, an embodiment of the present invention provides a terminal device. As Figure 8 shown, the terminal device D10 in this embodiment includes: at least one processor D100 ( Figure 8 only one processor is shown in the figure), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100. When the processor D100 executes the computer program D102, the steps in any of the above method embodiments are implemented.

[0176] Specifically, when the processor D100 executes the computer program D102, it acquires the original brightness temperature image of the target near-field imaging scene, and obtains the initial near-field visibility function corresponding to the target near-field imaging scene based on the original brightness temperature image; inputs the initial near-field visibility function into a pre-constructed near-field error compensation network model to obtain a new near-field visibility function; calculates the error between the new near-field visibility function and a preset target far-field visibility function. If the error is less than or equal to the preset error threshold, performs an inverse Fourier transform on the new near-field visibility function to obtain the corrected reconstructed brightness temperature image of the target near-field imaging scene; otherwise, constructs a loss function based on the error, and uses the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, obtaining a trained near-field error compensation network model, and inputs the initial near-field visibility function into the trained near-field error compensation network model. After performing an inverse Fourier transform on the new near-field visibility function output by the trained near-field error compensation network model, the corrected reconstructed brightness temperature image is obtained. Among them, by extracting the frequency domain feature information of the initial near-field visibility function and compensating for the near-field error of the initial near-field visibility function based on the frequency domain feature information, the near-field error can be corrected and the imaging quality can be improved; constructing a loss function based on the error and using the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold can not only realize the mapping from the near-field visibility function to the target far-field visibility function in the frequency domain, but also flexibly adjust the compensation intensity according to the target far-field visibility function, further improving the imaging quality.

[0177] The so-called processor D100 may be a central processing unit (CPU, Central Processing Unit), and this processor D100 may also be other general-purpose processors, digital signal processors (DSP, Digital Signal Processor), application specific integrated circuits (ASIC, Application Specific Integrated Circuit), field-programmable gate arrays (FPGA, Field-Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or this processor may also be any conventional processor, etc.

[0178] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk equipped on the terminal device D10, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Further, the memory D101 may also include both the internal storage unit and the external storage device of the terminal device D10. The memory D101 is used to store an operating system, application programs, a BootLoader, data, and other programs, such as the program code of the computer program. The memory D101 may also be used to temporarily store data that has been output or will be output.

[0179] An embodiment of the present application also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0180] An embodiment of the present application provides a computer program product, and when the computer program product runs on a terminal device, the terminal device can execute to implement the steps in the above-mentioned method embodiments.

[0181] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary, and is not intended to imply that the protection scope of the present application is limited to these examples; under the concept of the present application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of one or more embodiments of the present application as described above, and they are not provided in detail for the sake of brevity.

[0182] One or more embodiments of the present application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the present application. Therefore, any omission, modification, equivalent substitution, improvement, etc. made within the spirit and principle of one or more embodiments of the present application shall be included within the protection scope of the present application.

Claims

1. A method for error correction of a synthetic aperture radiometer, characterized in that, Including: Obtain the original brightness temperature image of the target near-field imaging scene, and obtain the initial near-field visibility function corresponding to the target near-field imaging scene according to the original brightness temperature image; Input the initial near-field visibility function into a pre-constructed near-field error compensation network model to obtain a new near-field visibility function; the near-field error compensation network model is used to extract the frequency-domain feature information of the initial near-field visibility function, and based on the frequency-domain feature information, compensate for the near-field error of the initial near-field visibility function, and the near-field error includes phase error and amplitude error; Calculate the error between the new near-field visibility function and a preset target far-field visibility function. If the error is less than or equal to a preset error threshold, perform inverse Fourier transform on the new near-field visibility function to obtain the corrected reconstructed brightness temperature image of the target near-field imaging scene; otherwise, construct a loss function according to the error, and use the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, obtain the trained near-field error compensation network model, and input the initial near-field visibility function into the trained near-field error compensation network model. After performing inverse Fourier transform on the new near-field visibility function output by the trained near-field error compensation network model, obtain the corrected reconstructed brightness temperature image.

2. The error correction method of the synthetic aperture radiometer according to claim 1, wherein The near-field error compensation network model includes an information extraction module, a first feature extraction module, a second feature extraction module, an anti-overfitting module, a feature recombination module, and a feature aggregation module. The information extraction module is used to extract the amplitude information and phase information of the visibility function. The first feature extraction module is used to extract the shallow frequency-domain features of the visibility function according to the amplitude information and the phase information. The shallow frequency-domain features characterize the overall structure and distribution law of the frequency-domain data, and mainly reflect the change trend of the near-field and far-field visibility functions in the global amplitude and phase. The second feature extraction module is used to extract the deep frequency-domain features of the visibility function from the shallow frequency-domain features. The deep frequency-domain features characterize the local detail differences of the frequency-domain data, and depict the fine structure deviation and local distortion caused by the near-field effect. The anti-overfitting module is used to randomly deactivate some neurons in the deep frequency-domain features to prevent the model from overfitting. The feature recombination module is used to recombine the deep frequency-domain features output by the anti-overfitting module to obtain a multi-channel feature map. The feature aggregation module is used to perform average pooling on the multi-channel feature map to achieve near-field error compensation.

3. The error correction method for a synthetic aperture radiometer according to claim 2, wherein Both the first feature extraction module and the second feature extraction module are composed of five sequentially connected residual convolution blocks; The shallow frequency-domain features are expressed as: F1 = X5(F0) + conv(F0) X i (F0) = ReLU(BN(conv2(ReLU(conv1(F0))))) + shortcut, i = 1 X i (F0) = ReLU(BN(conv2(ReLU(conv1(X i-1 (F0)))))) + shortcut, i = 2, 3, 4, 5 Among them, F1 represents the shallow frequency domain features, F0 represents the amplitude information and phase information of the visibility function output by the information extraction module, F0 = Normalize(stack(Flatten(Re(X)), Flatten(Im(X)))), Re(X) represents the amplitude information extracted from the real part of the visibility function X, Im(X) represents the phase information extracted from the imaginary part of the visibility function X, Flatten(·) represents the flattening operation, stack(·) represents the concatenation operation, Normalize(·) represents the normalization operation, conv(F0) represents the convolution operation on F0 for extracting the original features of F0, X i represents the output of the i-th residual convolution block, conv1 represents the first convolutional layer in the residual convolution block, ReLU(·) represents the activation function, conv2 represents the second convolutional layer in the residual convolution block, BN(·) represents batch normalization, and shortcut represents the skip connection; The deep frequency-domain features are expressed as: F2 = X5(F1) + conv(F1) X i (F1) = ReLU(BN(conv2(ReLU(conv1(F1))))) + shortcut, i = 1 X i (F1) = ReLU(BN(conv2(ReLU(conv1(X i-1 (F1)))))) + shortcut, i = 2, 3, 4, 5 where F2 represents the deep frequency-domain features.

4. The radiometer error correction method according to claim 3, wherein, The recombining the deep frequency-domain features output by the anti-overfitting module to obtain a multi-channel feature map includes: By the calculation formula F4 = conv4(LeakyReLU(conv3(F3))) F3 = dropout(F2) Obtain the multi-channel feature map F4; conv3 represents the third convolutional layer in the feature recombination module, and LeakyReLU(·) represents an activation function, which is used to retain partial gradient responses for negative values and improve the network's ability to capture subtle features. conv4 represents the fourth convolutional layer in the feature recombination module, and dropout(F2) represents the result of randomly deactivating some neurons in F2 by the anti-overfitting module.

5. The error correction method of the synthetic aperture radiometer according to claim 4, characterized in that The error is the mean absolute error.

6. The error correction method of the synthetic aperture radiometer according to claim 5, characterized in that The expression of the error loss function is as follows: where Loss represents the error loss value, n = 1,.., N, N represents the total number of small surface elements in the original brightness temperature image, V label represents the preset target far-field visibility function, V out represents the newly measured near-field visibility function.

7. The error correction method of the synthetic aperture radiometer according to claim 6, characterized in that The original brightness temperature image includes multiple tiny facets.

8. An error correction system for a synthetic aperture radiometer, characterized in that Including: A data acquisition module, configured to acquire the original brightness temperature image of the target near-field imaging scene, and obtain the initial near-field visibility function corresponding to the target near-field imaging scene according to the original brightness temperature image; An error compensation module, configured to input the initial near-field visibility function into a pre-constructed near-field error compensation network model to obtain a new near-field visibility function; the near-field error compensation network model is configured to extract the frequency-domain feature information of the initial near-field visibility function, and based on the frequency-domain feature information, compensate for the near-field error of the initial near-field visibility function, and the near-field error includes a phase error and an amplitude error; A reconstruction module, configured to calculate the error between the new near-field visibility function and a preset target far-field visibility function. If the error is less than or equal to a preset error threshold, perform an inverse Fourier transform on the new near-field visibility function to obtain the corrected reconstructed brightness temperature image of the target near-field imaging scene; otherwise, construct a loss function according to the error, and use the loss function to perform backpropagation on the near-field error compensation network model until the error is less than or equal to the preset error threshold, obtain the trained near-field error compensation network model, and input the initial near-field visibility function into the trained near-field error compensation network model. After performing an inverse Fourier transform on the new near-field visibility function output by the trained near-field error compensation network model, obtain the corrected reconstructed brightness temperature image.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, the method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the method described in any one of claims 1 to 7 is implemented.

Images