One-dimensional synthetic aperture depth convolutional neural network and image reconstruction method

Abstract

The application discloses a one-dimensional synthetic aperture depth convolutional neural network and an image reconstruction method, the method is characterized in that: a deep convolutional neural network structure for a one-dimensional synthetic aperture radiometer is designed, a visibility function measured by a one-dimensional synthetic aperture system is taken as an input of the convolutional neural network, a microwave brightness temperature image is taken as an output, and one-dimensional synthetic aperture microwave brightness temperature image reconstruction is realized. Compared with a traditional brightness temperature reconstruction method, the method can obviously inhibit Gibbs oscillation without reducing spatial resolution, and improves the quality of the reconstructed brightness temperature image; the application is simple and reliable, can effectively improve the imaging quality of the synthetic aperture radiometer, improves the effective utilization rate of satellite remote sensing data, and has a wide market application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to a one-dimensional synthetic aperture depth convolutional neural network and an image reconstruction method, and belongs to the technical field of space microwave remote sensing. BACKGROUND

[0002] A one-dimensional synthetic aperture radiometer front end adopts a one-dimensional antenna array arranged sparsely, and each two antennas in the array can form a baseline pair to perform interference measurement on a scene signal and output a visibility function (each baseline pair corresponds to a visibility function). A traditional one-dimensional synthetic aperture image reconstruction method utilizes a Fourier transform relationship between a visibility function and a scene brightness temperature to reconstruct a brightness temperature image from the visibility function.

[0003] For a one-dimensional synthetic aperture radiometer, due to the limited number of antennas of the system, the corresponding sampling baseline is limited, so that the radiometer appears to be truncated sampling in the spatial frequency domain, resulting in obvious Gibbs oscillation error in the area with high brightness temperature contrast in the imaging result. The oscillation error will affect the entire brightness temperature image, resulting in large error in the reconstructed brightness temperature image, and ultimately affecting the detection accuracy of the radiometer. In the existing synthetic aperture radiometer imaging algorithm, a windowing method is generally used to reduce the Gibbs oscillation error, but the windowing will reduce the spatial resolution of the image. Therefore, a method for more effectively reducing the Gibbs oscillation error of the image is needed. SUMMARY

[0004] The technical problem solved by the present application is that, for the current one-dimensional synthetic aperture microwave brightness temperature image reconstruction problem, a one-dimensional synthetic aperture depth convolutional neural network and an image reconstruction method are proposed. The present application realizes one-dimensional synthetic aperture microwave brightness temperature image reconstruction by designing a depth convolutional neural network structure for a one-dimensional synthetic aperture radiometer. The method is simple and reliable, can effectively improve the image quality, and improve the inversion accuracy of the synthetic aperture radiometer.

[0005] The technical solution of the present application is:

[0006] The present application discloses a one-dimensional synthetic aperture image reconstruction method based on a depth convolutional neural network, comprising:

[0007] (1) collecting color image training samples of different scenes;

[0008] (2) mapping and interpolating the color image training samples to one-dimensional brightness temperature distribution map training samples with the same pixel points;

[0009] (3) calculating the visibility function by using the one-dimensional brightness temperature distribution map training samples, and simulating to obtain the visibility function training samples corresponding to all the color image training samples;

[0010] (4) constructing a one-dimensional synthetic aperture depth convolutional neural network, initializing network parameters;

[0011] (5) inputting the visibility function training sample into the one-dimensional synthetic aperture depth convolutional neural network for image reconstruction to obtain a network reconstructed brightness temperature image for training;

[0012] (6) calculating an image error between the one-dimensional brightness temperature distribution training sample and the network reconstructed brightness temperature image for training corresponding thereto;

[0013] (7) obtaining a curve of image error changing with training iteration number according to the image error result, judging whether the image error tends to be stable, if yes, completing training to obtain a trained one-dimensional synthetic aperture depth convolutional neural network; otherwise, updating network parameters by using a standard backpropagation stochastic gradient algorithm and returning to step (5);

[0014] (8) obtaining a one-dimensional brightness temperature distribution verification sample and a corresponding visibility function verification sample for network verification by simulation or measurement;

[0015] (9) inputting the visibility function verification sample into the trained one-dimensional synthetic aperture depth convolutional neural network to obtain a network reconstructed brightness temperature image for verification; and reconstructing a brightness temperature by using a traditional one-dimensional synthetic aperture brightness temperature reconstruction method to obtain a traditional reconstructed brightness temperature image;

[0016] (10) comparing and analyzing the network reconstructed brightness temperature image for verification and the traditional reconstructed brightness temperature image to realize effect evaluation of the network reconstructed image.

[0017] In the above reconstruction method, the color image training sample is mapped and interpolated to a one-dimensional brightness temperature distribution training sample with the same pixel points, and the specific method is as follows: the color image is converted into a gray image value with a value range of 0-255, the gray image value is mapped into a brightness temperature value of 0-300K, and the image size is adjusted to a fixed pixel number by using a linear interpolation method to complete creation of the one-dimensional brightness temperature distribution.

[0018] In the above reconstruction method, the visibility function is as follows:

[0019]

[0020] In the formula, V(u) is the visibility function, T B (θ) is a brightness temperature distribution corresponding to the one-dimensional brightness temperature distribution, F nk (θ), and F nj (θ) are antenna directional patterns of units k and j, respectively, k , jThe antenna solid angle of the unit k, j, respectively, is θ, the azimuth angle, u is the spatial frequency, u=(x j -x k ) / λ0, wherein x k , x j are the geometric position coordinates of the unit k, j, and λ0 is the wavelength.

[0021] In the reconstruction method, the one-dimensional synthetic aperture depth convolutional neural network comprises a feature enhancement module and an image reconstruction module.

[0022] The feature enhancement module sorts out effective feature information from the visibility function data and outputs the feature information to the image reconstruction module, and is composed of a feature extraction unit and a reverse mapping unit; the feature extraction unit comprises two convolutional layers and extracts feature information from the visibility function data; the reverse mapping unit is composed of two transposed convolutional layers and performs up-sampling on the output of the convolutional layer; a batch normalization layer and a linear rectifier function are added after each convolutional layer and transposed convolutional layer.

[0023] The image reconstruction module is composed of two fully connected layers and six convolutional layers; receives the feature information input by the feature enhancement module, extracts the feature information to reconstruct a brightness temperature image, and outputs a one-dimensional brightness temperature image reconstructed by the network; the two fully connected layers adopt a nonlinear activation function to map the feature information to the subsequent convolutional layers; the six convolutional layers further extract high-order features, perform dimension conversion and image reconstruction using the feature information, and output a one-dimensional brightness temperature image reconstructed by the network.

[0024] In the reconstruction method, the method for determining whether the image error tends to be stable is as follows: the image error of the network reconstructed brightness temperature image can be calculated in each training iteration, the curve of the image error changing with the number of iterations can be obtained according to the image error result corresponding to each iteration, and when the change value of the image error of any adjacent two times of continuous multiple iterations is within 0.1K, it is determined that the image error tends to be stable.

[0025] In the reconstruction method, the standard reverse stochastic gradient algorithm has the formula:

[0026]

[0027] In the formula, W p is the network parameter of the pth iteration, p is the number of times of updating the parameter, α is a learning rate, is the partial derivative of the network parameter with respect to the image error, that is, the gradient of the image error.

[0028] In the above reconstruction method, the network reconstructed brightness temperature image for verification is compared and analyzed with the traditional reconstructed brightness temperature image, the effect of network reconstruction is realized, and the specific method is: the network reconstructed brightness temperature image for verification is compared and analyzed with the one-dimensional brightness temperature distribution map verification sample, the network reconstructed image error is calculated; the traditional reconstructed brightness temperature image is compared and analyzed with the one-dimensional brightness temperature distribution map verification sample, the traditional reconstructed image error is calculated; whether the network reconstructed image error is less than the traditional reconstructed image error is judged, if yes, the network reconstructed brightness temperature image is used, if not, the network parameter is adjusted again, and step (5) is returned.

[0029] In the above reconstruction method, the image error is calculated according to the following formula:

[0030]

[0031] In the formula, T k is the brightness temperature of the kth one-dimensional brightness temperature distribution sample, invT k is the brightness temperature of the network reconstructed image corresponding to the kth sample; N is the number of brightness temperature pixels of the one-dimensional brightness temperature distribution map and the reconstructed brightness temperature image, m is the number of samples, k represents the sample serial number, and W={u1,...u i} is the set of all network parameters.

[0032] The application discloses a one-dimensional comprehensive aperture depth convolutional neural network, which comprises a feature enhancement module and an image reconstruction module.

[0033] The feature enhancement module extracts effective feature information from training samples and outputs the feature information to the image reconstruction module, and the feature enhancement module comprises a feature extraction unit and a reverse mapping unit.

[0034] The feature extraction unit comprises two convolutional layers and extracts feature information from the visibility function data; the reverse mapping unit comprises two transposed convolutional layers and performs up-sampling on the output of the convolutional layer to obtain an output with the same size as the input data of the feature enhancement module; a batch normalization layer and a linear rectifier function are added after each convolutional layer and transposed convolutional layer.

[0035] The image reconstruction module receives the feature information output by the feature enhancement module, reconstructs a brightness temperature image through the feature information, and outputs a one-dimensional brightness temperature image reconstructed by the network; the image reconstruction module comprises two fully connected layers and six convolutional layers.

[0036] The two fully connected layers adopt a nonlinear activation function to map the feature information to the subsequent convolutional layers; the six convolutional layers further extract high-order features, realize dimension conversion between the frequency domain and the spatial domain and image reconstruction by using the feature information, and output the one-dimensional brightness temperature image reconstructed by the network.

[0037] In the one-dimensional synthetic aperture depth convolutional neural network, the batch normalization layer and the linear rectifier function are added, and the specific method is as follows:

[0038] The batch normalization layer performs batch normalization processing on the output data of the convolutional layer and the transposed convolutional layer, and the formula is as follows:

[0039]

[0040] In the formula, x represents the input data of the batch normalization layer, BN(x) represents the output after batch normalization processing, E(x) and D(x) are the mean and variance of the input data respectively, and alpha and beta represent the scale factor and the translation factor, which are parameters learned by the network during training.

[0041] The linear rectifier function non-linearizes the data and increases the non-linear expression of the model, and the formula is as follows:

[0042] ReLU(x) = max(0, x)

[0043] In the formula, x represents the input data of the linear rectifier function, and ReLU(x) represents the output after linear rectifier function processing, and the linear rectifier function output is a number greater than or equal to 0.

[0044] In the one-dimensional synthetic aperture depth convolutional neural network, the six-layer convolutional layer further extracts high-order features, uses the feature information to realize dimension conversion and image reconstruction, and outputs the one-dimensional brightness temperature image reconstructed by the network, and the specific method is as follows: The high-order feature information between the visibility function in the frequency domain and the brightness temperature image in the spatial domain and the system characteristics of the synthetic aperture radiometer are obtained through the six-layer convolutional layer, the dimension conversion from the frequency domain to the spatial domain is learned, and thus the brightness temperature image reconstruction is realized.

[0045] The beneficial effects of the present application over the prior art are as follows:

[0046] (1) The present application designs a one-dimensional synthetic aperture depth convolutional neural network, learns the mapping relationship between the visibility function input and the brightness temperature output, realizes one-dimensional synthetic aperture image reconstruction based on the depth convolutional neural network, effectively suppresses the Gibbs oscillation error of the reconstructed brightness temperature image without reducing the spatial resolution, and improves the image quality; and once the network is trained, the depth learning neural network can be used to quickly reconstruct the output image of the one-dimensional synthetic aperture radiometer;

[0047] (2) The present application designs a depth learning neural network for one-dimensional synthetic aperture, trains the network, and uses the trained network to perform one-dimensional synthetic aperture image reconstruction, which can effectively reduce the image Gibbs oscillation error, thereby improving the detection accuracy of the synthetic aperture radiometer. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The structure diagram of the one-dimensional synthetic aperture depth convolutional neural network of the present application;

[0049] Figure 2 The structure diagram of the feature enhancement module of the present application;

[0050] Figure 3 The structure diagram of the image reconstruction module of the present application;

[0051] Figure 4 The brightness temperature reconstruction method based on the deep convolutional neural network of the present application (a) network training process (b) network verification and application process;

[0052] Figure 5 The array arrangement schematic diagram of the present application;

[0053] Figure 6 The point source scene schematic diagram of the present application;

[0054] Figure 7 The brightness temperature reconstruction result of the present application (a) IFFT reconstructed image (b) network reconstructed image. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0056] As shown in Figure 1 The present application discloses a one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network, which comprises:

[0057] Step (1), collecting color image training samples of different scenes;

[0058] Step (2), mapping and interpolating the color image training samples to one-dimensional brightness temperature distribution map training samples with the same pixel points, the specific method being: converting the color image to a gray image value with a value range of 0-255, mapping the gray image value to a brightness temperature value of 0-300K, and adjusting the image size to a fixed pixel point number by linear interpolation method, thereby completing the creation of the one-dimensional brightness temperature distribution map.

[0059] Step (3), using the one-dimensional brightness temperature distribution map training sample to calculate the visibility function, and simulating to obtain the visibility function training sample corresponding to all color image training samples; the visibility function formula being:

[0060]

[0061] In the formula, V(u) is the visibility function, TB (θ) is a brightness temperature distribution corresponding to the one-dimensional brightness temperature distribution, F nk (θ), F nj (θ) are the antenna patterns of units k, j, respectively, Ω k , Ω j are the antenna solid angles of units k, j, respectively, θ is an azimuth angle, u is a spatial frequency, u = (x j -x k ) / λ0, wherein x k , x j are the geometric position coordinates of units k, j, and λ0 is a wavelength.

[0062] Step (4), constructing a one-dimensional synthetic aperture depth convolutional neural network, initializing network parameters;

[0063] Step (5), inputting the visibility function training sample into the one-dimensional synthetic aperture depth convolutional neural network for image reconstruction to obtain a network-reconstructed brightness temperature image for training;

[0064] The one-dimensional synthetic aperture depth convolutional neural network comprises a feature enhancement module and an image reconstruction module.

[0065] The feature enhancement module sorts out effective feature information from the visibility function data and outputs the feature information to the image reconstruction module, and is composed of a feature extraction unit and a reverse mapping unit; the feature extraction unit comprises two convolutional layers and extracts feature information from the visibility function data; the reverse mapping unit is composed of two transposed convolutional layers and performs up-sampling on the output of the convolutional layers; a batch normalization layer and a linear rectifier function are added after each convolutional layer and transposed convolutional layer.

[0066] The image reconstruction module is composed of two fully connected layers and six convolutional layers; receives the feature information input by the feature enhancement module, and reconstructs a brightness temperature image through the feature information to output a one-dimensional network-reconstructed brightness temperature image; the two fully connected layers adopt a nonlinear activation function to map the feature information to the subsequent convolutional layers; the six convolutional layers further extract high-order features, perform dimension conversion and image reconstruction using the feature information, and output the one-dimensional network-reconstructed brightness temperature image.

[0067] Step (6), calculating an image error between a one-dimensional brightness temperature distribution training sample and a network-reconstructed brightness temperature image corresponding to the one-dimensional brightness temperature distribution training sample for training;

[0068] Step (7), obtaining a curve of the image error changing with the number of training iterations according to the image error result, judging whether the image error tends to be stable, if yes, completing the training to obtain a trained one-dimensional synthetic aperture depth convolutional neural network; otherwise, updating the network parameters using a standard backpropagation stochastic gradient algorithm and returning to step (5); as shown in FIG. Figure 4 (a).

[0069] To determine whether the image error has stabilized, the specific method is as follows: the image error of the brightness temperature image reconstructed by the training network can be calculated in each training iteration. Based on the image error result corresponding to each iteration, the curve of the image error changing with the number of iterations can be obtained. When the change value of any two adjacent image errors in multiple consecutive iterations is within 0.1K, it is determined that the image error has stabilized. In this embodiment, the number of iterations is 5, and the change value of the image error between any two adjacent iterations is within 0.1K.

[0070] The standard inverse stochastic gradient algorithm has the following formula:

[0071]

[0072] In the formula, W p Let be the network parameters in the p-th iteration, where p is the number of times the parameters are updated, and α is the learning rate. This is the partial derivative of the network parameters with respect to the image error, i.e., the gradient of the image error.

[0073] Step (8): Obtain a one-dimensional brightness temperature distribution map verification sample and its corresponding visibility function verification sample for network verification through simulation or measurement;

[0074] Step (9): Input the visibility function verification sample into the trained one-dimensional integrated aperture depth convolutional neural network to obtain the brightness temperature image reconstructed by the network for verification; for the visibility function verification sample, use the traditional one-dimensional integrated aperture brightness temperature reconstruction method to reconstruct the brightness temperature and obtain the traditional reconstructed brightness temperature image.

[0075] Step (10): Compare and analyze the brightness temperature image reconstructed by the verification network with the traditionally reconstructed brightness temperature image to evaluate the effect of the network-reconstructed image. Figure 4 As shown in (b), the specific method is as follows: the brightness temperature image reconstructed by the verification network and the one-dimensional brightness temperature distribution map are used to verify the sample, and the network reconstruction image error is calculated; the traditional reconstruction image error is calculated by the traditional reconstruction image and the one-dimensional brightness temperature distribution map; it is determined whether the network reconstruction image error is less than the traditional reconstruction image error. If so, the network is used to reconstruct the brightness temperature image; if not, the network parameters are readjusted and the process returns to step (5).

[0076] Image error, calculated using the following formula:

[0077]

[0078] In the formula, T k Let invT be the brightness temperature of the k-th one-dimensional brightness temperature distribution sample. kis the brightness temperature of the network reconstructed image corresponding to the kth sample; N is the number of brightness temperature pixel points of the one-dimensional brightness temperature distribution map and the reconstructed brightness temperature image, m is the number of samples, k represents the sample serial number, and W={u1,...u i} is the set of all network parameters.

[0079] As shown in Figure 2 , the application discloses a one-dimensional synthetic aperture depth convolutional neural network, which comprises a feature enhancement module and an image reconstruction module.

[0080] The feature enhancement module extracts effective feature information from training samples and outputs the feature information to the image reconstruction module, and is composed of a feature extraction unit and a reverse mapping unit.

[0081] The feature extraction unit comprises two convolution layers (C1-C2) and extracts feature information from the visibility function data; the reverse mapping unit is composed of two transposed convolution layers (TC1-TC2) and performs up-sampling on the output of the convolution layer to obtain an output with the same size as the input data of the feature enhancement module; a batch normalization layer and a linear rectifier function are added after each convolution layer and transposed convolution layer.

[0082] The image reconstruction module receives the feature information output by the feature enhancement module, reconstructs a brightness temperature image through the feature information, and outputs a one-dimensional brightness temperature image reconstructed by the network; the image reconstruction module is composed of two fully connected layers (FC1 and FC2 layers) and six convolution layers; as shown in Figure 3 .

[0083] The two fully connected layers adopt a nonlinear activation function to map the feature information to the subsequent convolution layers; the six convolution layers further extract high-order features, utilize the feature information to realize dimension conversion from the frequency domain to the spatial domain and image reconstruction, and output a one-dimensional brightness temperature image reconstructed by the network, and the specific method is as follows: the six convolution layers are used to obtain high-order feature information between the visibility function in the frequency domain and the brightness temperature image in the spatial domain and the system characteristics of the synthetic aperture radiometer, learn the dimension conversion from the frequency domain to the spatial domain, and thus realize brightness temperature image reconstruction.

[0084] The batch normalization layer and the linear rectifier function are added, and the specific method is as follows:

[0085] The batch normalization layer performs batch normalization processing on the output data of the convolution layer and the transposed convolution layer, and the formula is as follows:

[0086]

[0087] In the formula, x represents the input data of the batch normalization layer, BN(x) represents the output after batch normalization processing, E(x) and D(x) are respectively the mean and variance of the input data, a and β represent a scale factor and a translation factor, and and are parameters learned by the network during training.

[0088] The linear rectification function non-linearizes the data, and increases the non-linear expression of the model, and the formula is:

[0089] ReLU(x)=max(0,x)

[0090] In the formula, x represents the input data of the linear rectification function, and ReLU(x) represents the output after the linear rectification function processing, and the linear rectification function output is a number greater than or equal to 0.

[0091] Embodiment 1

[0092] In this embodiment, a 6-unit one-dimensional linear array is simulated, Figure 5 The array used for simulation is shown in the figure, the minimum spacing of the array elements is 0.5 times the wavelength, and the maximum baseline of the feed array is 13 times the wavelength. Based on the array design, a one-dimensional synthetic aperture depth convolutional neural network is designed, and 1000 pictures of different scenes are selected to convert into original scene brightness temperature image samples for network training according to the foregoing network training steps. The target observation scene is set to a point source scene as shown in Figure 6 The corresponding visibility function is obtained from the point source scene.

[0093] The traditional one-dimensional synthetic aperture brightness temperature reconstruction method (inverse Fourier transform IFFT) is used to reconstruct the brightness temperature of the visibility function of the point source scene, and the result is shown in Figure 7 (a). The visibility function of the point source scene is input into the trained network, and the brightness temperature is reconstructed, and the result is shown in Figure 7 (b). The brightness temperature image reconstructed by the one-dimensional synthetic aperture depth convolutional neural network is closer to the original scene, and the sidelobe gibbs oscillation is reduced by 70% compared with the brightness temperature image reconstructed by IFFT, and the zero beam width of the images reconstructed by the two methods is basically the same, which indicates that the spatial resolution is not degraded.

[0094] Therefore, the one-dimensional synthetic aperture image reconstruction method based on the depth convolutional neural network provided by the application can obviously suppress gibbs oscillation without degrading spatial resolution, and improve the quality of the reconstructed brightness temperature image.

[0095] The application learns the mapping relationship between the visibility function input and the brightness temperature output by designing a one-dimensional synthetic aperture depth convolutional neural network composed of a feature enhancement module and an image reconstruction module. The network is trained by brightness temperature image samples and the network performance is verified. Thus, the problem of obvious gibbs oscillation error in the imaging result due to the limited sampling baseline of the radiometer is solved, the quality of the one-dimensional synthetic aperture image reconstruction is improved, and the reconstructed brightness temperature is more consistent with the original scene brightness temperature distribution.

[0096] The application can effectively improve the detection precision of a synthetic aperture radiometer, and is a one-dimensional synthetic aperture depth convolutional neural network and image reconstruction method.

[0097] The above-described embodiments are only the preferred specific embodiments of the application, and the common changes and replacements made by those skilled in the art within the technical scheme range of the application shall be included in the protection scope of the application.

Claims

1. A one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network, characterized in that, include: (1) Collect color image training samples from different scenes; (2) The color image training samples are mapped and interpolated to form one-dimensional brightness temperature distribution map training samples with the same number of pixels; (3) Using the one-dimensional brightness temperature distribution map training samples, calculate the visibility function and simulate to obtain the visibility function training samples corresponding to all the color image training samples; (4) Construct a one-dimensional synthetic aperture depth convolutional neural network and initialize the network parameters; (5) Input the visibility function training samples into a one-dimensional integrated aperture depth convolutional neural network for image reconstruction to obtain a brightness temperature image reconstructed by the training network. (6) Calculate the image error between the training sample of the one-dimensional brightness temperature distribution map and the corresponding brightness temperature image reconstructed by the training network; (7) Based on the image error results, obtain the curve of image error changing with the number of training iterations, and determine whether the image error tends to be stable. If so, complete the training and obtain the trained one-dimensional integrated aperture depth convolutional neural network; otherwise, update the network parameters using the standard inverse stochastic gradient algorithm and return to step (5). (8) Obtain one-dimensional brightness temperature distribution map verification samples and their corresponding visibility function verification samples for network verification through simulation or measurement; (9) Input the visibility function verification sample into the trained one-dimensional synthetic aperture depth convolutional neural network to obtain the brightness temperature image reconstructed by the verification network; for the visibility function verification sample, use the traditional one-dimensional synthetic aperture brightness temperature reconstruction method to reconstruct the brightness temperature and obtain the traditional reconstructed brightness temperature image. (10) Compare and analyze the brightness temperature image reconstructed by the verification network with the traditional brightness temperature image to evaluate the effect of the network-reconstructed image. The image error is calculated using the following formula: In the formula, T k Let invT be the brightness temperature of the k-th one-dimensional brightness temperature distribution sample. k Let m be the brightness temperature of the network-reconstructed image corresponding to the k-th sample; N is the number of brightness temperature pixels in the one-dimensional brightness temperature distribution map and the reconstructed brightness temperature image, m is the number of samples, k represents the sample index, and W = {u1,...u...} i } represents the set of all network parameters.

2. The one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network according to claim 1, characterized in that: The color image training samples are mapped and interpolated to form one-dimensional brightness temperature distribution map training samples with the same number of pixels. Specifically, the color image is converted into a grayscale image with a value range of 0 to 255. The grayscale image values ​​are mapped to brightness temperature values ​​of 0 to 300K. The image size is adjusted to a fixed number of pixels using linear interpolation to complete the creation of the one-dimensional brightness temperature distribution map.

3. The one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network according to claim 1, characterized in that: The visibility function is defined as follows: In the formula, V(u) is the visibility function, and T B (θ) represents the brightness temperature distribution corresponding to the one-dimensional brightness temperature distribution map, F nk (θ), F nj (θ) represent the antenna patterns of elements k and j, respectively, Ω k Ω j Let θ be the antenna solid angle of elements k and j, θ be the azimuth angle, and u be the spatial frequency, where u = (x j -x k ) / λ0, where x k x j Let λ be the geometric position coordinates of elements k and j, and λ0 be the wavelength.

4. The one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network according to claim 1, characterized in that: The one-dimensional synthetic aperture depth convolutional neural network includes: a feature enhancement module and an image reconstruction module; The feature enhancement module extracts effective feature information from the visibility function data and outputs it to the image reconstruction module. It consists of a feature extraction unit and a reverse mapping unit. The feature extraction unit includes two convolutional layers to extract feature information from the visibility function data. The reverse mapping unit consists of two transposed convolutional layers to upsample the output of the convolutional layers. A batch normalization layer and a linear rectified function are added after each convolutional layer and transposed convolutional layer. The image reconstruction module consists of two fully connected layers and six convolutional layers. It receives feature information from the feature enhancement module, extracts and reconstructs a brightness temperature image using the feature information, and outputs a one-dimensional brightness temperature image reconstructed by the network. The two fully connected layers use a non-linear activation function to map the feature information to subsequent convolutional layers. The six convolutional layers further extract higher-order features, use the feature information for dimensionality transformation and image reconstruction, and output a one-dimensional brightness temperature image reconstructed by the network.

5. The one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network according to claim 1, characterized in that: The method for determining whether the image error tends to stabilize is as follows: the image error of the brightness temperature image reconstructed by the training network can be calculated in each training iteration. Based on the image error result corresponding to each iteration, the curve of the image error changing with the number of iterations can be obtained. When the change value of the image error between any two adjacent iterations in multiple consecutive iterations is within 0.1K, it is determined that the image error tends to stabilize.

6. The one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network according to claim 1, characterized in that: The standard inverse stochastic gradient algorithm is defined as follows: In the formula, W p Let be the network parameters in the p-th iteration, where p is the number of times the parameters are updated, and α is the learning rate. This is the partial derivative of the network parameters with respect to the image error, i.e., the gradient of the image error.

7. The one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network according to claim 1, characterized in that: The brightness temperature image reconstructed by the verification network is compared and analyzed with the traditional brightness temperature image to evaluate the effect of network reconstruction. The specific method is as follows: the network reconstruction image error is calculated by comparing the brightness temperature image reconstructed by the verification network with the one-dimensional brightness temperature distribution map verification sample; the traditional reconstruction image error is calculated by comparing the traditional reconstruction image with the one-dimensional brightness temperature distribution map verification sample; it is determined whether the network reconstruction image error is less than the traditional reconstruction image error. If so, the network is used to reconstruct the brightness temperature image; if not, the network parameters are readjusted and the process returns to step (5).

8. A one-dimensional synthetic aperture depth convolutional neural network, characterized in that, A one-dimensional synthetic aperture image reconstruction method based on a deep convolutional neural network as described in any one of claims 1 to 7, comprising: a feature enhancement module and an image reconstruction module; The feature enhancement module extracts effective feature information from the training samples and outputs it to the image reconstruction module. It consists of a feature extraction unit and a reverse mapping unit. The feature extraction unit includes two convolutional layers to extract feature information from the visibility function data; the reverse mapping unit consists of two transposed convolutional layers to upsample the output of the convolutional layers to obtain an output with the same size as the input data of the feature enhancement module; a batch normalization layer and a linear rectified function are added after each convolutional layer and transposed convolutional layer. The image reconstruction module receives feature information output by the feature enhancement module, extracts and reconstructs a brightness temperature image through the feature information, and outputs a one-dimensional brightness temperature image reconstructed by the network; it consists of two fully connected layers and six convolutional layers. The two fully connected layers employ nonlinear activation functions to map the feature information to subsequent convolutional layers; the six convolutional layers further extract higher-order features, and use the feature information to achieve dimensionality transformation between the frequency domain and the spatial domain and image reconstruction, outputting a one-dimensional brightness temperature image reconstructed by the network.

9. A one-dimensional synthetic aperture depth convolutional neural network according to claim 8, characterized in that: The specific method for adding the batch normalization layer and the linear rectification function is as follows: The batch normalization layer performs batch normalization on the output data of the convolutional layer and the transposed convolutional layer, using the following formula: In the formula, x represents the input data of the batch normalization layer, BN(x) represents the output after batch normalization, E(x) and D(x) are the mean and variance of the input data, respectively, α and β represent the scaling factor and translation factor, which are parameters learned by the network during training. The linear rectifier function nonlinearizes the data, increasing the nonlinear expression of the model. The formula is: ReLU(x) = max(0,x) In the formula, x represents the input data of the linear rectifier function, and ReLU(x) represents the output after processing by the linear rectifier function. The output of the linear rectifier function is a number greater than or equal to 0.

10. A one-dimensional synthetic aperture depth convolutional neural network according to claim 8, characterized in that: The six convolutional layers further extract higher-order features, and use the feature information to achieve dimensional transformation and image reconstruction, outputting a one-dimensional brightness temperature image reconstructed by the network. The specific method is as follows: the six convolutional layers obtain the higher-order feature information between the visibility function in the frequency domain and the brightness temperature image in the spatial domain, as well as the system features of the integrated aperture radiometer, and learn the dimensional transformation from the frequency domain to the spatial domain, thereby realizing the brightness temperature image reconstruction.

Images