Satellite-borne multi-band synthetic aperture radar signal processing optimization method and system

Abstract

The invention provides a satellite-borne multi-band synthetic aperture radar signal processing optimization method and system, and relates to the technical field of signal processing, and the method comprises the steps: obtaining original echo data of a satellite-borne multi-band synthetic aperture radar, carrying out Doppler center frequency estimation, carrying out two-dimensional matrix product operation, and obtaining compensation echo data; distance direction position information and scattering intensity information are obtained based on a distance direction matching filtering algorithm of sub-band decomposition; determining spatial distribution density, performing adaptive block processing, and performing phase error modeling in each block to generate a continuous phase correction function; performing wavelet domain convolution operation to obtain corrected echo data, obtaining a multi-band two-dimensional image, extracting a multi-dimensional feature vector through a deep neural network, performing weight learning, performing nonlinear weighted fusion, and performing adaptive histogram equalization enhancement to determine a synthetic aperture radar imaging result.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to a satellite-borne multi-band synthetic aperture radar signal processing optimization method and system. Background Art

[0002] Spaceborne multi-band synthetic aperture radar is an efficient remote sensing imaging technology that is widely used in ground monitoring, resource exploration, disaster assessment and other fields. By transmitting microwave signals and receiving reflected echoes, a virtual long aperture is synthesized in combination with a mobile platform (such as a satellite) to improve imaging resolution. The multi-band synthetic aperture radar system can simultaneously acquire data in different frequency bands, and utilize the different characteristics of each frequency band to provide more accurate and reliable imaging results in complex terrain and changing environments.

[0003] Although synthetic aperture radar technology has made significant progress in imaging quality, existing signal processing methods still face multiple challenges. First, the phase error caused by Doppler frequency changes will affect the image quality, especially when the target is moving with uniform or variable acceleration. The modeling and compensation of this error are very complicated. Second, current phase compensation technology often targets a single frequency band or specific conditions and lacks the ability to comprehensively process multi-frequency band data. In addition, existing feature extraction and data fusion methods still need to be improved in terms of computational efficiency and image quality.

[0004] In summary, there is an urgent need for an adaptive piecewise polynomial fitting algorithm to achieve accurate Doppler center frequency estimation, combined with phase error modeling based on spline interpolation, a composite optimization objective function of image entropy and contrast, and adaptive learning of multi-dimensional features by deep learning methods, which significantly improves the clarity, accuracy and processing efficiency of imaging. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a satellite-borne multi-band synthetic aperture radar signal processing optimization method and system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for optimizing signal processing of a spaceborne multi-band synthetic aperture radar is provided, comprising:

[0008] The original echo data of the spaceborne multi-band synthetic aperture radar is obtained, and the Doppler center frequency of the original echo data is estimated by using an adaptive piecewise polynomial fitting algorithm, and a frequency domain adaptive compensation factor matrix is ​​constructed. The frequency domain adaptive compensation factor matrix is ​​subjected to a two-dimensional matrix product operation with the original echo data to obtain compensated echo data. The compensated echo data is range-compressed based on a range-matched filtering algorithm of sub-band decomposition to generate a range image, and the range position information and scattering intensity information of the target scattering point are calculated by using a maximum likelihood estimation method based on the range image.

[0009] Based on the distance position information and scattering intensity information of the target scattering points, the spatial distribution density of the target scattering points is determined. Based on the spatial distribution density, the compensation echo data is adaptively divided into blocks, and the block size is dynamically adjusted. In each block, the phase error of the compensation echo data is modeled through a generalized polynomial function containing a quadratic phase error term, a cubic phase error term, and a high-order cross term. A composite optimization objective function that fuses image entropy and contrast is constructed. The composite optimization objective function is quickly iterated through a conjugate gradient algorithm to solve the phase error estimate, and a continuous phase correction function is generated based on the phase error estimate through piecewise spline interpolation.

[0010] The phase correction function and the compensated echo data are convolved in the wavelet domain to obtain the corrected echo data. The corrected echo data are processed based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, the local correlation coefficient, edge intensity and texture features are extracted through a deep neural network to construct a multi-dimensional feature vector, and the multi-dimensional feature vector is weighted learned. Based on the adaptive weight coefficients obtained by learning, the two-dimensional images of different frequency bands are nonlinearly weighted fused, and the synthetic aperture radar imaging results are determined through adaptive histogram equalization enhancement.

[0011] In an optional embodiment,

[0012] The Doppler center frequency estimation of the original echo data using the adaptive piecewise polynomial fitting algorithm includes:

[0013] Segmenting the original echo data, dividing the original echo data into a plurality of sub-data segments according to a preset window length, with a fifty percent overlap rate between adjacent sub-data segments, to obtain sub-data segments to be processed;

[0014] Perform residual evaluation on the sub-data segment to be processed, calculate normalized residual values ​​of polynomial fittings of different orders, compare the ratio of the normalized residual value to the normalized residual value of the previous order polynomial fitting with a preset threshold, and determine the optimal polynomial order when the ratio is less than the preset threshold;

[0015] Constructing a Vandermonde matrix based on an optimal polynomial order, wherein the elements of the Vandermonde matrix are composed of power terms of different orders of time variables, constructing a Hanning weighted matrix, wherein the diagonal elements of the Hanning weighted matrix are calculated by a cosine function, solving the polynomial coefficients determined by the Vandermonde matrix and the Hanning weighted matrix by a least squares method, and constructing a fitting polynomial function according to the polynomial coefficients;

[0016] Calculate the data signal-to-noise ratio of the sub-data segment to be processed, and adaptively adjust the preset window length according to the data signal-to-noise ratio. When the data signal-to-noise ratio is greater than a first preset data signal-to-noise ratio threshold, reduce the preset window length to half of the original length. When the data signal-to-noise ratio is less than a second preset data signal-to-noise ratio threshold, increase the preset window length to twice the original length. Re-fit the data according to the adjusted preset window length and the fitting polynomial function to obtain an optimized fitting polynomial function.

[0017] The error function between the sub-data segment to be processed and the optimized fitting polynomial function is calculated, the mean and variance of the error function are obtained, a compensation term based on a Gaussian function is constructed, and the compensation term is added to the initial frequency estimation value calculated according to the optimized fitting polynomial function to obtain the final Doppler center frequency estimation result.

[0018] In an optional embodiment,

[0019] Based on the range matched filtering algorithm of sub-band decomposition, the compensated echo data is range compressed to generate a range image. Based on the range image, the maximum likelihood estimation method is used to calculate the range position information and scattering intensity information of the target scattering point, including:

[0020] Dividing the spectrum of the compensated echo data into a plurality of frequency sub-bands, with an overlapping area between adjacent frequency sub-bands, the bandwidth of the overlapping area being 20% ​​of the sub-band bandwidth, and using a Hamming window function to weight the spectrum of each frequency sub-band;

[0021] Generate a reference signal based on the transmission signal parameters, the frequency spectrum of the reference signal includes the antenna pattern and the influence of propagation attenuation, and multiply the reference signal by the signal of each frequency sub-band in the frequency domain to obtain a filtering result;

[0022] Calculating the phase difference between the frequency sub-bands, correcting the phase difference using a linear phase compensation method, calculating the signal-to-noise ratio of the signal corresponding to each frequency sub-band, and weighting and summing the filtering results using the signal-to-noise ratio as a weight coefficient to obtain a synthetic range image;

[0023] Performing energy normalization processing on the synthetic range image, dividing the range units according to the system resolution, establishing a mapping relationship between amplitude and distance, searching for the peak position of the synthetic range image, and using the peak position as a rough positioning result of the scattering point;

[0024] In the neighborhood of the rough positioning result of the scattering point, the parabolic interpolation method is used to subdivide the distance unit tenfold, and the peak value search is performed on the subdivided distance unit to obtain the precise position of the scattering point;

[0025] Constructing a likelihood function based on a Gaussian noise model, wherein the likelihood function includes a distance position parameter and a scattering intensity parameter, optimizing and solving the likelihood function in an iterative manner, and taking the parameter corresponding to the maximum value of the likelihood function as an initial estimation result;

[0026] The Cramer-Rao bound of the initial estimation result is calculated to determine whether the initial estimation result reaches the predetermined theoretical accuracy limit. If not, a new search space is constructed in the neighborhood of the initial estimation result, and the parameter iterative optimization is continued to finally obtain the target scattering point location information and scattering intensity information.

[0027] In an optional embodiment,

[0028] In each block, the phase error modeling of the compensated echo data is performed through a generalized polynomial function containing quadratic phase error terms, cubic phase error terms and high-order cross terms. The composite optimization objective function of the fusion image entropy and contrast is constructed, including:

[0029] In each block, a quadratic term with the range position information of the target scattering point as a variable is established as a quadratic phase error term to compensate for the phase error caused by the uniform acceleration motion of the target; a cubic term with the range position information of the target scattering point as a variable is established as a cubic phase error term to compensate for the phase error caused by the variable acceleration motion of the target; a product term of the range position information and the azimuth position information of the target scattering point is established as a cross-phase error term to determine the corresponding relationship between the phase errors caused by the coupling in the range and azimuth directions; the quadratic phase error term, the cubic phase error term and the high-order cross-phase error term are combined to construct a generalized polynomial function;

[0030] Apply the generalized polynomial function to the compensated echo data for phase compensation to obtain a compensated image; calculate the intensity distribution of the compensated image, normalize the intensity distribution to obtain a probability distribution, and calculate the image entropy based on the probability distribution; extract the maximum intensity and the minimum intensity of the compensated image to calculate the image contrast; calculate the difference between the image contrast and 1 to obtain the image contrast loss; determine a first weight coefficient based on the spatial distribution density of the target area, subtract the first weight coefficient from 1 to obtain a second weight coefficient; multiply the image entropy by the first weight coefficient to obtain a first weighted term, multiply the image contrast loss by the second weight coefficient to obtain a second weighted term, and add the first weighted term to the second weighted term to construct a composite optimization objective function.

[0031] In an optional embodiment,

[0032] The composite optimization objective function is quickly iterated by the conjugate gradient algorithm to obtain the phase error estimate, and based on the phase error estimate, a continuous phase correction function is generated by piecewise spline interpolation, including:

[0033] A numerical differentiation method is used to calculate the gradient of the composite optimization objective function with respect to each coefficient in the generalized polynomial function, and the negative direction of the gradient is used as the initial search direction. An optimal step length is determined by a one-dimensional search in the initial search direction. The coefficients in the generalized polynomial function are updated based on the optimal step length to obtain updated coefficients. An updated gradient of the composite optimization objective function is calculated based on the updated coefficients. The inner product ratio of the updated gradient to the gradient before the update is calculated to obtain a conjugate coefficient. The conjugate combination of the negative direction of the updated gradient and the initial search direction is used as a new search direction. Iteration is repeated until the difference between two adjacent iterations of the composite optimization objective function is less than a preset convergence threshold, and a convergence coefficient is obtained. The convergence coefficient is substituted into the generalized polynomial function to obtain a phase error estimate.

[0034] The center position of each block is determined as a reference interpolation node, and supplementary interpolation nodes are added in the overlapping area of ​​adjacent blocks; the cubic spline basis function is selected as the interpolation basis function, and a spline coefficient equation group is constructed based on the reference interpolation node and the supplementary interpolation node; the boundary condition that the second-order derivative at the endpoint is zero is set for the spline coefficient equation group, the spline coefficient is solved to obtain the spline coefficient, the spline coefficient is combined with the cubic spline basis function, the phase error estimate is segmentedly interpolated, and a continuous phase correction function is generated.

[0035] In an optional embodiment,

[0036] The phase correction function and the compensation echo data are convolved in the wavelet domain to obtain the corrected echo data. The corrected echo data is processed based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, the local correlation coefficient, edge strength and texture features are extracted through a deep neural network to construct a multi-dimensional feature vector, and the multi-dimensional feature vector is weighted. Based on the adaptive weight coefficient obtained by learning, the two-dimensional images of different frequency bands are nonlinearly weighted fused, and enhanced through adaptive histogram equalization to determine the synthetic aperture radar imaging results, including:

[0037] Performing discrete wavelet transform on the phase correction function to obtain a first wavelet coefficient, performing discrete wavelet transform on the compensation echo data to obtain a second wavelet coefficient, performing multiplication operation on the first wavelet coefficient and the second wavelet coefficient in the wavelet domain, and performing inverse wavelet transform on the result of the multiplication operation to obtain correction echo data;

[0038] Determining kernel function parameters according to the scattering characteristics of the corrected echo data, constructing a variable-scale kernel function based on the kernel function parameters, and using the variable-scale kernel function to perform azimuth compression processing on the corrected echo data to obtain a multi-band two-dimensional image;

[0039] A normalized cross-correlation operator is used to calculate the local correlation of the multi-band two-dimensional image to obtain a correlation coefficient matrix, a Canny edge detection operator is used to extract the edge of the multi-band two-dimensional image to obtain an edge intensity matrix, and a Gabor filter group is used to filter the multi-band two-dimensional image to obtain a texture feature matrix; the correlation coefficient matrix, the edge intensity matrix, and the texture feature matrix are cascaded in the feature dimension to obtain a multi-dimensional feature vector;

[0040] Constructing a deep neural network including an input layer, a hidden layer, and an output layer, wherein the number of neurons in the input layer is the same as the dimension of the multidimensional feature vector, and the number of neurons in the output layer is the same as the number of frequency bands, inputting the multidimensional feature vector into the deep neural network for training to obtain an adaptive weight coefficient, and performing nonlinear weighted calculation on the multi-band two-dimensional image according to the adaptive weight coefficient to obtain a fused image;

[0041] The grayscale statistical distribution of the fused image is calculated, a histogram equalization parameter is determined, and the fused image is enhanced using the histogram equalization parameter to obtain a final multi-band synthetic aperture radar image.

[0042] According to a second aspect of the embodiments of the present invention,

[0043] A satellite-borne multi-band synthetic aperture radar signal processing optimization system is provided, comprising:

[0044] The first unit is used to obtain the original echo data of the spaceborne multi-band synthetic aperture radar, use the adaptive piecewise polynomial fitting algorithm to estimate the Doppler center frequency of the original echo data, construct a frequency domain adaptive compensation factor matrix, perform a two-dimensional matrix product operation on the frequency domain adaptive compensation factor matrix and the original echo data to obtain compensated echo data, perform range compression on the compensated echo data based on the range matched filtering algorithm of sub-band decomposition, generate a range image, and use the maximum likelihood estimation method based on the range image to calculate the range position information and scattering intensity information of the target scattering point;

[0045] The second unit is used to determine the spatial distribution density of the target scattering points based on the distance position information and scattering intensity information of the target scattering points, and based on the spatial distribution density, adaptively block the compensation echo data, dynamically adjust the block size, and in each block, use a generalized polynomial function containing a quadratic phase error term, a cubic phase error term, and a high-order cross term to model the phase error of the compensation echo data, construct a composite optimization objective function that fuses image entropy and contrast, and quickly iterate the composite optimization objective function through a conjugate gradient algorithm to solve the phase error estimate, and generate a continuous phase correction function through piecewise spline interpolation based on the phase error estimate;

[0046] The third unit is used to perform wavelet domain convolution operation on the phase correction function and the compensated echo data to obtain the corrected echo data, and process the corrected echo data based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, the local correlation coefficient, edge strength and texture features are extracted through a deep neural network to construct a multi-dimensional feature vector, and the multi-dimensional feature vector is weighted learned. Based on the adaptive weight coefficient obtained by learning, the two-dimensional images of different frequency bands are nonlinearly weighted fused, and the synthetic aperture radar imaging results are determined through adaptive histogram equalization enhancement.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the embodiments of the present invention is:

[0053] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0054] In the embodiment of the present invention, an adaptive piecewise polynomial fitting algorithm is used to reduce errors and improve the accuracy of frequency estimation; a frequency domain adaptive compensation factor matrix is ​​constructed to achieve stability in the signal frequency domain; a range-direction matched filtering algorithm based on subband decomposition is used for effective range compression; a maximum likelihood estimation method can accurately calculate the range position information and scattering intensity information of target scattering points; the data block size is dynamically adjusted according to the spatial distribution density of the target scattering points to improve the efficiency of phase error modeling; a generalized polynomial function containing polynomial terms is used to improve the accuracy of phase error estimation; a composite optimization objective function is constructed by combining image entropy and contrast to improve imaging clarity and contrast; a continuous phase correction function is generated by a conjugate gradient algorithm and piecewise spline interpolation; an azimuth compression algorithm based on a variable scale kernel function ensures that the details and feature information of multi-band data are retained; a deep neural network is used to extract features to achieve efficient fusion of multi-band images; nonlinear weighted fusion and adaptive histogram equalization technology are used to improve image contrast and detail prominence. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of a satellite-borne multi-band synthetic aperture radar signal processing optimization method according to an embodiment of the present invention;

[0056] Figure 2 It is a schematic diagram of the structure of a satellite-borne multi-band synthetic aperture radar signal processing optimization system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0059] Figure 1 FIG. 1 is a flow chart of a method for optimizing signal processing of a spaceborne multi-band synthetic aperture radar according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] S101. Acquire the original echo data of the spaceborne multi-band synthetic aperture radar, use an adaptive piecewise polynomial fitting algorithm to estimate the Doppler center frequency of the original echo data, construct a frequency domain adaptive compensation factor matrix, perform a two-dimensional matrix product operation on the frequency domain adaptive compensation factor matrix and the original echo data to obtain compensated echo data, perform range compression on the compensated echo data based on a range matched filtering algorithm of sub-band decomposition to generate a range image, and use a maximum likelihood estimation method based on the range image to calculate the range position information and scattering intensity information of the target scattering point;

[0061] S102. Based on the distance position information and scattering intensity information of the target scattering points, determine the spatial distribution density of the target scattering points, perform adaptive block processing on the compensation echo data based on the spatial distribution density, dynamically adjust the block size, and in each block, perform phase error modeling on the compensation echo data through a generalized polynomial function containing a quadratic phase error term, a cubic phase error term, and a high-order cross term, construct a composite optimization objective function that fuses image entropy and contrast, quickly iterate the composite optimization objective function through a conjugate gradient algorithm, solve and obtain a phase error estimate, and generate a continuous phase correction function through piecewise spline interpolation based on the phase error estimate;

[0062] S103. Perform wavelet domain convolution operation on the phase correction function and the compensated echo data to obtain corrected echo data. Process the corrected echo data based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, extract local correlation coefficients, edge strength and texture features through a deep neural network to construct a multi-dimensional feature vector, and perform weight learning on the multi-dimensional feature vector. Based on the adaptive weight coefficients obtained by learning, perform nonlinear weighted fusion on the two-dimensional images of different frequency bands, and determine the synthetic aperture radar imaging result through adaptive histogram equalization enhancement.

[0063] First, the original echo data of the spaceborne multi-band synthetic aperture radar is obtained. These data usually contain echo information of multiple frequency bands, such as X, C, L bands, etc. For each frequency band, the original echo data is usually a two-dimensional complex matrix, one dimension of which represents the range sampling and the other dimension represents the azimuth sampling.

[0064] Next, the Doppler center frequency of the original echo data is estimated using an adaptive piecewise polynomial fitting algorithm. In specific implementation, the echo data can be divided into multiple sub-blocks in the azimuth direction, and Fourier transform is performed on each sub-block to find the frequency corresponding to the spectrum peak as the Doppler center frequency estimate of the sub-block. Then, these estimates are fitted with a polynomial function to obtain the Doppler center frequency change curve of the entire data in the azimuth direction. Taking the X-band as an example, the Doppler center frequency estimation result that may be obtained is a change from -320Hz to 320Hz.

[0065] Based on the estimated Doppler center frequency, a frequency domain adaptive compensation factor matrix is ​​constructed. Each element of the matrix corresponds to a sampling point of the original echo data. The compensated echo data can be obtained by performing a two-dimensional matrix product operation on the compensation factor matrix and the original echo data. This step can correct the phase error caused by the Doppler frequency change.

[0066] The compensated echo data is compressed in the range direction, and a range-direction matched filtering algorithm based on sub-band decomposition is used. First, the echo data is Fourier transformed in the range direction, and then the spectrum is divided into several sub-bands. Each sub-band is matched filtered separately, and finally the results of each sub-band are synthesized to obtain the range image. This method can effectively suppress range ambiguity. Taking the X-band 150MHz bandwidth signal as an example, the spectrum can be equally divided into 10 15MHz sub-bands for processing.

[0067] According to the obtained range image, the maximum likelihood estimation method is used to calculate the range position information and scattering intensity information of the target scattering point. Specifically, the position of the scattering point can be determined by finding the local peak in the range image, and the peak amplitude is the scattering intensity. For example, it may be obtained that a certain target has 3 main scattering points, with distances of 1000m, 1002m and 1005m, and relative scattering intensities of 1, 0.8 and 0.6.

[0068] Based on the spatial distribution of the target scattering points, the compensated echo data is adaptively divided into blocks. A smaller block size is used in areas with dense target scattering points, and a larger block size is used in sparse areas. For example, the data can be divided into sub-blocks of varying sizes from 64×64 to 256×256. Within each block, a generalized polynomial function is used to model the phase error, which contains quadratic phase error terms, cubic phase error terms, and high-order cross terms, and can more accurately describe complex phase errors.

[0069] A composite optimization objective function integrating image entropy and contrast is constructed. Image entropy reflects the degree of image focus, and contrast reflects the clarity of the image. Combining the two can obtain a more comprehensive image quality evaluation. The conjugate gradient algorithm is used to quickly iterate and optimize the objective function to obtain the phase error estimate. Usually, 10-20 iterations are required to converge.

[0070] Based on the obtained discrete phase error estimate, a continuous phase correction function is generated by cubic spline interpolation. The corrected echo data can be obtained by performing wavelet domain convolution operation on the function and the compensated echo data. Wavelet domain processing can suppress noise while retaining image details.

[0071] The corrected echo data is compressed in azimuth using an algorithm based on a variable-scale kernel function. This algorithm can adapt to the Doppler frequency variation characteristics of different regions by adjusting the scale of the kernel function, thereby achieving a better focusing effect. The processing result is a two-dimensional image with multiple frequency bands.

[0072] For the obtained multi-band images, a deep convolutional neural network is constructed to extract local correlation coefficients, edge strengths, texture features, etc. The network contains multiple convolutional layers, pooling layers, and fully connected layers. The input is a multi-band image, and the output is an extracted feature vector. For example, a 256-dimensional feature vector can be extracted to characterize various attributes of the image.

[0073] The weights of the extracted multi-dimensional feature vectors can be learned by using algorithms such as stochastic gradient descent to optimize the weight coefficients. The learned weights reflect the importance of different features in image fusion. For example, the weight of edge features may be large, 0.4, while the weight of texture features may be small, 0.1.

[0074] Based on the learned weight coefficients, nonlinear weighted fusion is performed on images of different frequency bands. During fusion, the pixel values ​​of each frequency band image are multiplied by the corresponding weights, and then the results are added to obtain the fused image. Nonlinear fusion can better preserve the detailed information of each frequency band image.

[0075] Finally, the fused image is enhanced by adaptive histogram equalization. This method can effectively improve the contrast of the image and highlight the target features. During processing, the image is divided into multiple sub-blocks, and histogram equalization is performed in each sub-block. Then, the results of each sub-block are smoothly fused using the bilinear interpolation method to obtain the final enhanced image. This is the final imaging result of the spaceborne multi-band synthetic aperture radar.

[0076] Through the above series of processing steps, the imaging quality of spaceborne multi-band synthetic aperture radar can be effectively improved to obtain high-contrast radar images. This method comprehensively considers key technologies such as Doppler frequency estimation, adaptive blocking, phase error correction, multi-band feature extraction and fusion, and can better adapt to complex imaging environments.

[0077] In this embodiment, the Doppler center frequency of the original echo data can be accurately estimated through an adaptive piecewise polynomial fitting algorithm, thereby reducing errors and improving the accuracy of frequency estimation; a frequency domain adaptive compensation factor matrix is ​​constructed, and a matrix product operation is performed with the original echo data to achieve efficient frequency compensation and ensure the stability of the signal in the frequency domain; a range-direction matched filtering algorithm based on subband decomposition is used to effectively perform range compression and generate a high-resolution range image; the range-direction position information and scattering intensity information of the target scattering point can be accurately calculated through the maximum likelihood estimation method, thereby improving the reliability of positioning and intensity information; the data block size is dynamically adjusted according to the spatial distribution density of the target scattering point to ensure more flexible data processing and effectively improve the efficiency of phase error modeling; a generalized polynomial function containing quadratic and cubic phase error terms and high-order cross terms is used to achieve more sophisticated phase error modeling and improve the accuracy of phase error estimation; combined with image entropy and contrast, The construction of a composite optimization objective function can comprehensively improve the clarity and contrast of imaging; the phase error estimate is obtained through rapid iterative optimization using the conjugate gradient algorithm, and a continuous phase correction function is generated through piecewise spline interpolation, thus achieving high-precision phase correction of compensated echo data; the azimuth compression algorithm based on the variable scale kernel function makes the compression and fusion of multi-band data more accurate, ensuring that the details and feature information in the two-dimensional image are preserved; the local correlation coefficient, edge strength and texture features are extracted through a deep neural network, and the weights of the multi-dimensional feature vectors are learned to automatically adapt to the differences between different frequency bands, thus achieving efficient fusion of multi-band images; the nonlinear weighted fusion is used in combination with an adaptive weight coefficient to ensure that the best target features are retained in the multi-band image, further improving the clarity and resolution of the image; the adaptive histogram equalization enhancement technology can improve the image contrast, highlight the details of key areas, and make the imaging results of the synthetic aperture radar clearer and more interpretable.

[0078] In an optional implementation, using an adaptive piecewise polynomial fitting algorithm to estimate the Doppler center frequency of the original echo data includes:

[0079] Segmenting the original echo data, dividing the original echo data into a plurality of sub-data segments according to a preset window length, with a fifty percent overlap rate between adjacent sub-data segments, to obtain sub-data segments to be processed;

[0080] Perform residual evaluation on the sub-data segment to be processed, calculate normalized residual values ​​of polynomial fittings of different orders, compare the ratio of the normalized residual value to the normalized residual value of the previous order polynomial fitting with a preset threshold, and determine the optimal polynomial order when the ratio is less than the preset threshold;

[0081] Constructing a Vandermonde matrix based on an optimal polynomial order, wherein the elements of the Vandermonde matrix are composed of power terms of different orders of time variables, constructing a Hanning weighted matrix, wherein the diagonal elements of the Hanning weighted matrix are calculated by a cosine function, solving the polynomial coefficients determined by the Vandermonde matrix and the Hanning weighted matrix by a least squares method, and constructing a fitting polynomial function according to the polynomial coefficients;

[0082] Calculate the data signal-to-noise ratio of the sub-data segment to be processed, and adaptively adjust the preset window length according to the data signal-to-noise ratio. When the data signal-to-noise ratio is greater than a first preset data signal-to-noise ratio threshold, reduce the preset window length to half of the original length. When the data signal-to-noise ratio is less than a second preset data signal-to-noise ratio threshold, increase the preset window length to twice the original length. Re-fit the data according to the adjusted preset window length and the fitting polynomial function to obtain an optimized fitting polynomial function.

[0083] The error function between the sub-data segment to be processed and the optimized fitting polynomial function is calculated, the mean and variance of the error function are obtained, a compensation term based on a Gaussian function is constructed, and the compensation term is added to the initial frequency estimation value calculated according to the optimized fitting polynomial function to obtain the final Doppler center frequency estimation result.

[0084] The Vandermonde matrix specifically refers to a matrix used to represent the power relationship between polynomial coefficients and input variables. In the Vandermonde matrix, the elements of each row are arranged in sequence by different powers of the same variable. This matrix is ​​often used in polynomial fitting problems. By arranging the power terms of different orders into a matrix form, it is convenient to solve the coefficients of the polynomial through matrix operations. The structure of the Vandermonde matrix is ​​very regular and is widely used in the fields of data fitting and signal processing, especially in solving linear equations and least squares problems, which can help construct polynomial expressions.

[0085] The Hanning weighted matrix specifically refers to a diagonal matrix used for weighted processing of signals, and is usually used for window functions in signal processing. The diagonal elements of the matrix are calculated by the Hanning window function, which can smooth the boundaries of the signal and reduce the impact of spectral leakage. The Hanning window function can effectively reduce the discontinuity of the data at the boundary by using the cosine function for weighting, ensuring that the data is smoother and more stable during the processing process. The Hanning weighted matrix is ​​widely used in signal processing and filtering tasks, especially in scenarios where spectrum analysis or frequency estimation is required. Weighting can reduce calculation errors and improve the reliability of the results.

[0086] Specifically, the method for estimating the Doppler center frequency of the original echo data using the adaptive piecewise polynomial fitting algorithm comprises the following steps:

[0087] First, the original echo data is segmented. The original echo data is divided into multiple sub-segments according to a preset window length (e.g., 1024 sampling points), and adjacent sub-segments overlap by 50%. For example, for the original echo data with a length of 10,000 sampling points, 18 sub-segments can be divided, each sub-segment contains 1024 sampling points, and adjacent sub-segments overlap by 512 sampling points. In this way, a sequence of sub-segments to be processed can be obtained.

[0088] Next, a residual evaluation is performed on each sub-data segment to be processed to determine the optimal polynomial fitting order. Starting from the 1st order polynomial, the order is gradually increased, and the normalized residual values ​​of polynomial fittings of different orders are calculated. The normalized residual value of the current order is compared with the normalized residual value of the previous order, and the ratio of the two is calculated. When the ratio is less than a preset threshold (for example, 0.95), it is considered that the improvement in fitting accuracy brought about by increasing the order is not obvious, and the order at this time is the optimal polynomial order. For example, for a sub-data segment, the normalized residual values ​​of the 1st to 6th order polynomial fittings are 0.8, 0.6, 0.4, 0.35, 0.33, and 0.32, respectively, then the 4th order polynomial is the optimal order.

[0089] Then, according to the determined optimal polynomial order, the Vandermonde matrix and the Hanning weighting matrix are constructed. The elements of the Vandermonde matrix are composed of different orders of power terms of the time variable, and the diagonal elements of the Hanning weighting matrix are calculated by the cosine function. The least squares method is used to solve the polynomial coefficients determined by these two matrices to construct the fitting polynomial function. For example, for a 4th-order polynomial fit, a 0 +a 1 t+a 2 t 2 +a 3 t 3 +a 4 t4 The fitting function of .

[0090] Afterwards, the data signal-to-noise ratio of the sub-data segment to be processed is calculated, and the preset window length is adaptively adjusted according to the data signal-to-noise ratio. When the data signal-to-noise ratio is greater than a first preset threshold (e.g., 20 dB), the preset window length is reduced to half of the original, i.e., 512 sampling points. When the data signal-to-noise ratio is less than a second preset threshold (e.g., 5 dB), the preset window length is increased to twice the original, i.e., 2048 sampling points. The data is refitted according to the adjusted window length and the fitting polynomial function to obtain an optimized fitting polynomial function.

[0091] Finally, the error function between the sub-data segment to be processed and the optimized fitting polynomial function is calculated to obtain the mean and variance of the error function. A compensation term is constructed based on the Gaussian function, and the compensation term is added to the initial frequency estimate calculated based on the optimized fitting polynomial function to obtain the final Doppler center frequency estimation result. For example, the initial frequency estimate of a sub-data segment is 1000Hz, the mean of the error function is 5Hz, and the variance is 2Hz. 2 , the final frequency estimation result is about 1005Hz.

[0092] Through the above steps, adaptive piecewise polynomial fitting of the original echo data can be achieved to obtain accurate Doppler center frequency estimation results. This method can adaptively adjust processing parameters according to data characteristics, effectively improving the accuracy and robustness of frequency estimation.

[0093] In this embodiment, the optimal polynomial order is dynamically selected through residual evaluation to ensure the best accuracy of polynomial fitting, thereby improving the estimation accuracy of the Doppler center frequency; based on the adaptive adjustment mechanism of the data signal-to-noise ratio, the window length of data processing is flexibly adjusted, so that the best fitting effect can be achieved in high-noise or low-noise environments, further optimizing the data processing efficiency and accuracy; by constructing a compensation term based on a Gaussian function and adding it to the initial frequency estimate, the error in data fitting is significantly reduced, and the reliability of the frequency estimation result is improved; combined with the use of the Vandermonde matrix and the Hanning weighting matrix, it is ensured that the data weighted processing during polynomial fitting is smoother, further improving the stability and robustness of the fitting results; the least squares method combined with the adaptive window adjustment and compensation mechanism can dynamically optimize the polynomial fitting process, so that data processing can remain efficient and accurate under various noise conditions.

[0094] In an optional implementation, based on the range matched filtering algorithm of sub-band decomposition, the compensated echo data is range compressed to generate a range image, and the range position information and scattering intensity information of the target scattering point are calculated using the maximum likelihood estimation method according to the range image, including:

[0095] Dividing the spectrum of the compensated echo data into a plurality of frequency sub-bands, with an overlapping area between adjacent frequency sub-bands, the bandwidth of the overlapping area being 20% ​​of the sub-band bandwidth, and using a Hamming window function to weight the spectrum of each frequency sub-band;

[0096] Generate a reference signal based on the transmission signal parameters, the frequency spectrum of the reference signal includes the antenna pattern and the influence of propagation attenuation, and multiply the reference signal by the signal of each frequency sub-band in the frequency domain to obtain a filtering result;

[0097] Calculating the phase difference between the frequency sub-bands, correcting the phase difference using a linear phase compensation method, calculating the signal-to-noise ratio of the signal corresponding to each frequency sub-band, and weighting and summing the filtering results using the signal-to-noise ratio as a weight coefficient to obtain a synthetic range image;

[0098] Performing energy normalization processing on the synthetic range image, dividing the range units according to the system resolution, establishing a mapping relationship between amplitude and distance, searching for the peak position of the synthetic range image, and using the peak position as a rough positioning result of the scattering point;

[0099] In the neighborhood of the rough positioning result of the scattering point, the parabolic interpolation method is used to subdivide the distance unit tenfold, and the peak value search is performed on the subdivided distance unit to obtain the precise position of the scattering point;

[0100] Constructing a likelihood function based on a Gaussian noise model, wherein the likelihood function includes a distance position parameter and a scattering intensity parameter, optimizing and solving the likelihood function in an iterative manner, and taking the parameter corresponding to the maximum value of the likelihood function as an initial estimation result;

[0101] The Cramer-Rao bound of the initial estimation result is calculated to determine whether the initial estimation result reaches the predetermined theoretical accuracy limit. If not, a new search space is constructed in the neighborhood of the initial estimation result, and the parameter iterative optimization is continued to finally obtain the target scattering point location information and scattering intensity information.

[0102] The Hamming window function specifically refers to a weighting function, which is commonly used in spectrum analysis in signal processing. Its main function is to reduce spectrum leakage, especially when the signal is processed in segments, it can smooth the edge of the signal, thereby reducing distortion in the frequency domain. Compared with other window functions, the Hamming window introduces a specific weighting coefficient to smooth the transition of the signal at the boundary, making the spectrum leakage smaller and better highlighting the main frequency components in the signal. The weighted value of the Hamming window function is larger in the middle of the window function and gradually decreases at the edge, presenting a shape similar to a parabola.

[0103] The Cramer-Rao bound specifically refers to a concept in parameter estimation theory, which is used to represent the theoretical minimum variance of parameter estimation under given noise conditions. In other words, it defines a limit that the variance of parameter estimation cannot be lower than in any unbiased estimation. CRB can provide a reference standard for evaluating the accuracy of estimation for the system. By comparing the actual estimated variance with CRB, the efficiency of the estimation method or the potential room for improvement can be judged. In signal processing and statistics, the Cramer-Rao bound is often used to measure whether the parameter estimation results have achieved the optimal theoretical accuracy.

[0104] The specific implementation method of performing range compression on the compensated echo data based on the range matched filtering algorithm of sub-band decomposition to generate a range image, and using the maximum likelihood estimation method according to the range image to calculate the range position information and scattering intensity information of the target scattering point is as follows:

[0105] First, the spectrum of the compensated echo data is divided into multiple frequency sub-bands. Specifically, the entire spectrum can be equally divided into 8 sub-bands, and a 20% overlap area is set between adjacent sub-bands. For example, for an echo signal with a bandwidth of 500MHz, it can be divided into 8 sub-bands, each with a bandwidth of 62.5MHz, and an overlap bandwidth of 12.5MHz between adjacent sub-bands. Then, the spectrum of each frequency sub-band is weighted using a Hamming window function to reduce spectrum leakage.

[0106] Next, a reference signal is generated based on the transmit signal parameters. The spectrum of the reference signal needs to include the effects of the antenna pattern and propagation attenuation. For example, for a system operating in the X-band with a center frequency of 10 GHz, the frequency domain expression of the reference signal can be generated based on the antenna gain model and the free space propagation loss model. The reference signal is multiplied in the frequency domain with the signal of each frequency sub-band to obtain the filtering result.

[0107] Then, the phase difference between the frequency sub-bands is calculated. Since sub-band division will introduce phase discontinuity, the phase difference needs to be corrected. The linear phase compensation method can be used to achieve phase correction by calculating the average phase difference in the overlapping area of ​​adjacent sub-bands and performing linear interpolation in the entire sub-band. At the same time, the signal-to-noise ratio of the corresponding signal of each frequency sub-band is calculated, and the signal-to-noise ratio is used as a weight coefficient to perform weighted summation on the filtering results to obtain a synthetic range image.

[0108] Perform energy normalization on the synthetic range image so that the maximum amplitude value is normalized to 1. Divide the range units according to the system range resolution and establish a mapping relationship between amplitude and distance. For example, for a system with a range resolution of 0.3m, the range axis can be divided into multiple range units at intervals of 0.3m. Search for the peak position of the synthetic range image and use the peak position as the coarse positioning result of the scattering point.

[0109] In the neighborhood of the rough positioning result of the scattering point, the parabolic interpolation method is used to subdivide the distance unit tenfold. For example, if the rough positioning result is the 100th distance unit, 100 new distance points are inserted between the 99th and 101st distance units. The peak value of the subdivided distance unit is searched to obtain the precise position of the scattering point.

[0110] Construct a likelihood function based on the Gaussian noise model, which includes the distance position parameter and the scattering intensity parameter. Use an iterative method to optimize the likelihood function, and you can use an optimization algorithm such as the gradient descent method or the Newton method. The parameter corresponding to the maximum value of the likelihood function is used as the initial estimate result.

[0111] Calculate the Cramer-Rao bound of the initial estimation result to determine whether the initial estimation result meets the predetermined theoretical accuracy limit. For example, the distance position estimation accuracy can be set to be no less than 1 / 10 of the system resolution, and the scattering intensity estimation accuracy can be set to be no less than 3dB. If the accuracy requirement is not met, construct a new search space in the neighborhood of the initial estimation result and continue to perform iterative parameter optimization. The search range can be narrowed to ±5% of the initial estimation result, and the step size can be reduced to 1 / 10 of the original. Repeat the optimization process until the accuracy requirement is met or the maximum number of iterations is reached, and finally the location information and scattering intensity information of the target scattering point are obtained.

[0112] Through the above steps, the range-oriented matched filtering and maximum likelihood estimation based on subband decomposition can be realized to obtain high-precision target scattering characteristic information. This method makes full use of frequency domain information, improves anti-interference ability through subband processing, and further improves parameter estimation accuracy by using maximum likelihood estimation, which can be effectively applied to imaging radar systems.

[0113] In this embodiment, by using the Hamming window function to perform weighted processing on the spectrum of each frequency sub-band, spectrum leakage can be effectively reduced, the resolution and signal fidelity of the sub-band frequency can be improved, and the filtering effect can be enhanced; linear phase compensation is performed on the phase difference between the frequency sub-bands, the phase difference of the frequency sub-bands is corrected, the phase error is reduced, and the accuracy of the synthetic range image is ensured; based on the signal-to-noise ratio as a weight coefficient, the filtering results are weighted and summed, so that the weights of different frequency bands are adaptively adjusted during the signal processing process, and the signal-to-noise ratio and image quality of the synthetic range image are improved; the range unit is subdivided by parabolic interpolation, combined with peak search, the precise positioning of the scattering point is achieved, and the spatial resolution of radar imaging is improved; the Gaussian noise model is used to construct the likelihood function, and the iterative optimization method is combined to optimize and solve the position information and scattering intensity information of the scattering point to ensure more accurate parameter estimation; by calculating the Cramer-Rao bound of the initial estimation result, it is judged whether the estimation reaches the theoretical accuracy limit, and the estimation accuracy of the scattering point position information and scattering intensity is further guaranteed. If the accuracy requirement is not met, further iterative optimization is performed to ensure the accuracy and robustness of the final result.

[0114] In an optional implementation, in each block, phase error modeling is performed on the compensated echo data by a generalized polynomial function including a quadratic phase error term, a cubic phase error term, and a high-order cross term, and a composite optimization objective function of fusion image entropy and contrast is constructed, including:

[0115] In each block, a quadratic term with the range position information of the target scattering point as a variable is established as a quadratic phase error term to compensate for the phase error caused by the uniform acceleration motion of the target; a cubic term with the range position information of the target scattering point as a variable is established as a cubic phase error term to compensate for the phase error caused by the variable acceleration motion of the target; a product term of the range position information and the azimuth position information of the target scattering point is established as a cross-phase error term to determine the corresponding relationship between the phase errors caused by the coupling in the range and azimuth directions; the quadratic phase error term, the cubic phase error term and the high-order cross-phase error term are combined to construct a generalized polynomial function;

[0116] Apply the generalized polynomial function to the compensated echo data for phase compensation to obtain a compensated image; calculate the intensity distribution of the compensated image, normalize the intensity distribution to obtain a probability distribution, and calculate the image entropy based on the probability distribution; extract the maximum intensity and the minimum intensity of the compensated image to calculate the image contrast; calculate the difference between the image contrast and 1 to obtain the image contrast loss; determine a first weight coefficient based on the spatial distribution density of the target area, subtract the first weight coefficient from 1 to obtain a second weight coefficient; multiply the image entropy by the first weight coefficient to obtain a first weighted term, multiply the image contrast loss by the second weight coefficient to obtain a second weighted term, and add the first weighted term to the second weighted term to construct a composite optimization objective function.

[0117] Specifically, first, a phase error model is established in each block. Specifically, the distance position information of the target scattering point is used as a variable, and a quadratic term is constructed as a quadratic phase error term to compensate for the phase error caused by the uniform acceleration of the target. For example, (r-r0) can be used. 2 A quadratic term of the form, where r represents the distance position of the target scattering point and r0 represents the reference position. At the same time, a cubic term (r-r0) is constructed 3 As a third-order phase error term, it is used to compensate for the phase error caused by the variable acceleration motion of the target. In addition, it is necessary to establish the product term (r-r0)(a-a0) of the range position information and the azimuth position information of the target scattering point as a cross-phase error term, where a represents the azimuth position of the target scattering point and a0 represents the reference position. This term is used to determine the corresponding relationship between the phase errors caused by the coupling in the range and azimuth directions.

[0118] Next, the above quadratic phase error term, cubic phase error term and high-order cross-phase error term are combined to construct a generalized polynomial function. For example, the polynomial function can be obtained:

[0119] Among them, α, β, and γ are the coefficients to be estimated.

[0120] Then, the constructed generalized polynomial function is applied to the compensated echo data to perform phase compensation to obtain a compensated image. Multiply by , where j is the imaginary unit, to achieve phase compensation.

[0121] Next, the intensity distribution of the compensated image is calculated. The intensity image can be obtained by modulo the compensated complex image, and then the intensity image is normalized to obtain the probability distribution. Based on the probability distribution, the image entropy can be calculated. For example, assuming that the normalized intensity values ​​are 0.1, 0.2, 0.3, and 0.4, respectively, the image entropy can be calculated as -(0.1log0.1+0.2log0.2+0.3log0.3+0.4log0.4).

[0122] At the same time, the maximum intensity and minimum intensity of the compensated image are extracted to calculate the image contrast. For example, if the maximum intensity value is 100 and the minimum intensity value is 10, the contrast can be calculated as (100-10) / (100+10). The difference between the image contrast and 1 is then calculated to obtain the image contrast loss. In this example, the contrast loss is 1-(100-10) / (100+10).

[0123] In addition, the first weight coefficient needs to be determined based on the spatial distribution density of the target area. For example, the proportion of non-zero pixels in the target area can be counted. If the proportion is 60%, the first weight coefficient can be set to 0.6. Subtract the first weight coefficient from 1 to obtain the second weight coefficient of 0.4.

[0124] Finally, the image entropy is multiplied by the first weight coefficient to obtain the first weighted term; the image contrast loss is multiplied by the second weight coefficient to obtain the second weighted term. The first weighted term is added to the second weighted term to construct a composite optimization objective function. For example, assuming that the calculated image entropy is 2.5 and the contrast loss is 0.2, the composite optimization objective function can be expressed as 0.6×2.5+0.4×0.2=1.58.

[0125] Through the above steps, the phase error compensation method based on the generalized polynomial function can be realized, and the composite optimization objective function integrating image entropy and contrast can be constructed. In practical applications, the coefficients of the polynomial function can be continuously adjusted through iterative optimization to make the composite optimization objective function optimal, thereby obtaining the best phase error compensation effect.

[0126] In this embodiment, by constructing quadratic, cubic and cross-phase error terms, the phase errors caused by uniform acceleration and variable acceleration motion of the target and distance-azimuth coupling are effectively compensated, thereby greatly improving the accuracy and quality of imaging. Generalized polynomial functions are used to uniformly model and compensate for different types of phase errors, providing a flexible error correction method suitable for complex target motion models, ensuring accurate compensation for various errors. By combining image entropy and image contrast loss to construct a composite optimization objective function, it is ensured that the imaging result has both information richness (high entropy) and good visual clarity (high contrast), thereby improving the overall quality of the image. Based on the spatial distribution density of the target area, the weight coefficients of image entropy and contrast loss are dynamically adjusted to achieve adaptive weighted optimization, ensuring that the image quality optimization is more balanced and adaptable in different imaging scenarios. Through the normalization of the intensity distribution after phase compensation and the contrast optimization, the details and edge features in the image can be better presented, thereby improving the resolution and fineness of radar imaging.

[0127] In an optional implementation, the composite optimization objective function is quickly iterated by a conjugate gradient algorithm to obtain a phase error estimate, and a continuous phase correction function is generated by piecewise spline interpolation based on the phase error estimate, including:

[0128] A numerical differentiation method is used to calculate the gradient of the composite optimization objective function with respect to each coefficient in the generalized polynomial function, and the negative direction of the gradient is used as the initial search direction. An optimal step length is determined by a one-dimensional search in the initial search direction. The coefficients in the generalized polynomial function are updated based on the optimal step length to obtain updated coefficients. An updated gradient of the composite optimization objective function is calculated based on the updated coefficients. The inner product ratio of the updated gradient to the gradient before the update is calculated to obtain a conjugate coefficient. The conjugate combination of the negative direction of the updated gradient and the initial search direction is used as a new search direction. Iteration is repeated until the difference between two adjacent iterations of the composite optimization objective function is less than a preset convergence threshold, and a convergence coefficient is obtained. The convergence coefficient is substituted into the generalized polynomial function to obtain a phase error estimate.

[0129] The center position of each block is determined as a reference interpolation node, and supplementary interpolation nodes are added in the overlapping area of ​​adjacent blocks; the cubic spline basis function is selected as the interpolation basis function, and a spline coefficient equation group is constructed based on the reference interpolation node and the supplementary interpolation node; the boundary condition that the second-order derivative at the endpoint is zero is set for the spline coefficient equation group, the spline coefficient is solved to obtain the spline coefficient, the spline coefficient is combined with the cubic spline basis function, the phase error estimate is segmentedly interpolated, and a continuous phase correction function is generated.

[0130] Specifically, the composite optimization objective function is firstly iterated quickly through the conjugate gradient algorithm to obtain the phase error estimation value. The specific steps are as follows:

[0131] The numerical differentiation method is used to calculate the gradient of the composite optimization objective function to each coefficient in the generalized polynomial function. Taking the 5th-order generalized polynomial as an example, the function contains 6 coefficients a0~a5. For each coefficient, the partial derivative is calculated by the central difference method. For example, the partial derivative calculation formula for a0 is: (f(a0+h)-f(a0-h)) / (2h), where h is a small perturbation, generally 1e-6. The partial derivatives of the 6 coefficients are calculated in turn to form a gradient vector.

[0132] The negative direction of the gradient is used as the initial search direction. For example, if the gradient vector is [1.2, -0.5, 0.8, -0.3, 1.5, -0.6], the initial search direction is [-1.2, 0.5, -0.8, 0.3, -1.5, 0.6].

[0133] The optimal step length is determined by one-dimensional search in the initial search direction. The golden section method is used to search for the step length that minimizes the objective function in the interval [0, 1]. For example, the optimal step length obtained by the final search is 0.35.

[0134] Based on the optimal step size, the coefficients in the generalized polynomial function are updated to obtain the updated coefficients. The update formula is: new coefficient = old coefficient + step size * search direction. For example, the update result of a0 is: a0_new = a0_old + 0.35*(-1.2).

[0135] Calculate the updated gradient of the composite optimization objective function based on the updated coefficients. Use the same numerical differentiation method as the first step.

[0136] The conjugate coefficient is obtained by calculating the inner product ratio of the gradient after update and the gradient before update. The formula for calculating the conjugate coefficient is: β = (g new T *g new ) / (g old T *g old ), where g new and g old are the new and old gradient vectors respectively.

[0137] The conjugate combination of the negative direction of the updated gradient and the initial search direction is used as the new search direction. The calculation formula for the new search direction is: new =-g new +β*d old , where d old For the old search direction.

[0138] Repeat the above steps until the difference between two consecutive iterations of the composite optimization objective function is less than the preset convergence threshold, and obtain the convergence coefficient. For example, the convergence threshold is set to 1e-6, and after 50 iterations, the objective function value changes less than 1e-6, and the coefficient obtained at this time is the convergence coefficient.

[0139] Substitute the convergence coefficient into the generalized polynomial function to obtain the phase error estimate. For example, if the convergence coefficient is [1.5, -0.8, 0.6, -0.2, 0.1, -0.05], the phase error estimate is:

[0140] Next, based on the phase error estimate, a continuous phase correction function is generated through piecewise spline interpolation. The specific steps are as follows:

[0141] The center position of each block is determined as the reference interpolation node. For example, for 5 blocks, their center positions are [-0.8, -0.4, 0, 0.4, 0.8] respectively, and these 5 points are used as reference interpolation nodes.

[0142] Supplementary interpolation nodes are added in the overlapping areas of adjacent blocks. For example, one supplementary interpolation node is added at each of the four positions [-0.6, -0.2, 0.2, 0.6], forming a total of 9 interpolation nodes.

[0143] The cubic spline basis function is selected as the interpolation basis function. The expression of the cubic spline basis function is: S(x) = a(x-xi) 3 +b(x-xi) 2 +c(x-xi)+d, where xi is the interpolation node.

[0144] The spline coefficient equations are constructed based on the reference interpolation nodes and the supplementary interpolation nodes. The equations contain interpolation conditions and smooth connection conditions. The interpolation conditions require that the function value of the spline function at the node is equal to the given phase error estimate. The smooth connection condition requires that the first-order derivative and the second-order derivative of adjacent spline segments at the connection are continuous.

[0145] The spline coefficient equations are set with a boundary condition that the second-order derivative at the endpoints is zero. This boundary condition is called a natural boundary condition, which can ensure the smoothness of the interpolation function at the endpoints.

[0146] Solve the spline coefficient equations to get the spline coefficients. Use the chasing method to solve the tridiagonal linear equations to get the 4 coefficients a, b, c, d of each spline segment.

[0147] The spline coefficients are combined with the cubic spline basis functions to perform segmented interpolation on the phase error estimate to generate a continuous phase correction function. In each segmented interval, the corresponding spline function is used for interpolation calculation. The final phase correction function is continuous and smooth in the entire definition domain and passes through all interpolation nodes accurately.

[0148] Through the above steps, the estimation and correction of phase error are completed, and a continuous phase correction function is obtained. This function can be used for subsequent phase correction processing to improve the quality of radar imaging.

[0149] In this embodiment, the gradient is calculated by numerical differentiation, and the conjugate gradient method is used for optimization iteration, so that the optimal estimate of the phase error can be quickly found, the convergence process is accelerated, the computational complexity is reduced, and the processing efficiency is improved; the optimal step size is determined by one-dimensional search in the initial search direction, and the coefficients of the generalized polynomial function are dynamically updated based on the step size, so that each iteration can be closer to the optimal solution, ensuring that the iterative process converges quickly; the conjugate direction of the updated gradient is used to avoid the "sawtooth phenomenon" that occurs in the traditional gradient descent method, improve the efficiency and accuracy of the optimization path, and ensure that the global optimum is reached faster; the phase correction function is generated by the cubic spline interpolation method, and additional interpolation nodes are added in the overlapping area of ​​adjacent blocks to ensure the continuity and smoothness of the phase correction function in the entire area, and reduce the phenomenon of phase jump or discontinuity; the boundary condition that the second-order derivative at the endpoint is zero is set in the spline coefficient equation group, which further improves the smoothness and accuracy of the phase correction function in the edge area and ensures the stability of the overall imaging quality; the adaptive interpolation processing based on the block center and the overlapping area nodes can more accurately correct the phase error, ensuring that the image processing results are more accurate,

[0150] In an optional implementation, a phase correction function and the compensation echo data are convolved in the wavelet domain to obtain corrected echo data, and the corrected echo data is processed based on an azimuth compression algorithm of a variable scale kernel function to obtain a multi-band two-dimensional image, and local correlation coefficients, edge strengths, and texture features are extracted through a deep neural network based on each frequency band in the multi-band two-dimensional image to construct a multi-dimensional feature vector, and weight learning is performed on the multi-dimensional feature vector, and nonlinear weighted fusion is performed on the two-dimensional images of different frequency bands based on the adaptive weight coefficients obtained by learning, and the synthetic aperture radar imaging results are determined by adaptive histogram equalization enhancement, including:

[0151] Performing discrete wavelet transform on the phase correction function to obtain a first wavelet coefficient, performing discrete wavelet transform on the compensation echo data to obtain a second wavelet coefficient, performing multiplication operation on the first wavelet coefficient and the second wavelet coefficient in the wavelet domain, and performing inverse wavelet transform on the result of the multiplication operation to obtain correction echo data;

[0152] Determining kernel function parameters according to the scattering characteristics of the corrected echo data, constructing a variable-scale kernel function based on the kernel function parameters, and using the variable-scale kernel function to perform azimuth compression processing on the corrected echo data to obtain a multi-band two-dimensional image;

[0153] A normalized cross-correlation operator is used to calculate the local correlation of the multi-band two-dimensional image to obtain a correlation coefficient matrix, a Canny edge detection operator is used to extract the edge of the multi-band two-dimensional image to obtain an edge intensity matrix, and a Gabor filter group is used to filter the multi-band two-dimensional image to obtain a texture feature matrix; the correlation coefficient matrix, the edge intensity matrix, and the texture feature matrix are cascaded in the feature dimension to obtain a multi-dimensional feature vector;

[0154] Constructing a deep neural network including an input layer, a hidden layer, and an output layer, wherein the number of neurons in the input layer is the same as the dimension of the multidimensional feature vector, and the number of neurons in the output layer is the same as the number of frequency bands, inputting the multidimensional feature vector into the deep neural network for training to obtain an adaptive weight coefficient, and performing nonlinear weighted calculation on the multi-band two-dimensional image according to the adaptive weight coefficient to obtain a fused image;

[0155] The grayscale statistical distribution of the fused image is calculated, a histogram equalization parameter is determined, and the fused image is enhanced using the histogram equalization parameter to obtain a final multi-band synthetic aperture radar image.

[0156] Firstly, the phase correction function is subjected to discrete wavelet transform to obtain the first wavelet coefficients. Specifically, the phase correction function can be subjected to three-layer wavelet decomposition using Haar wavelet basis function to obtain low-frequency approximate coefficients and high-frequency detail coefficients, which constitute the first wavelet coefficients.

[0157] Then the same discrete wavelet transform is performed on the compensation echo data to obtain the second wavelet coefficient. For the compensation echo data of size 128×128, the Haar wavelet basis function is also used to perform 3-layer decomposition to obtain the wavelet coefficients of 7 sub-bands.

[0158] Then, in the wavelet domain, the first wavelet coefficient is multiplied by the second wavelet coefficient. After the coefficients at corresponding positions are multiplied, a new wavelet coefficient is obtained.

[0159] Finally, the multiplied wavelet coefficients are inversely transformed to reconstruct the corrected echo data. The inverse transformation process uses the same Haar basis function as the forward transformation, and through three inverse transformations, the wavelet coefficients are reconstructed into time domain signals, which are the corrected echo data.

[0160] After obtaining the corrected echo data, it is necessary to perform azimuth compression processing on it to obtain a multi-band two-dimensional image. First, the kernel function parameters are determined according to the scattering characteristics of the corrected echo data. Specifically, the distribution of scattering centers can be calculated by analyzing the amplitude and phase characteristics of the corrected echo data, and the scale parameters and shape parameters of the kernel function can be determined according to the spatial distribution characteristics of the scattering centers.

[0161] Then, a variable-scale kernel function is constructed based on the above parameters. The Gaussian kernel function can be used as the basis, and the scale and shape of the kernel function can be adjusted according to the determined parameters to obtain a variable-scale kernel function that adapts to the current scene. For example, for point target scenes, the kernel function scale is small; for area target scenes, the kernel function scale is relatively large.

[0162] The constructed variable-scale kernel function is used to perform azimuth compression processing on the corrected echo data. Specifically, a two-dimensional convolution operation is performed on the corrected echo data and the variable-scale kernel function to implement azimuth matched filtering and obtain a compressed two-dimensional image. The above processing is performed on the echo data of different frequency bands respectively, and finally a multi-band two-dimensional image is obtained.

[0163] In order to extract features from multi-band 2D images, the normalized cross-correlation operator is first used to calculate the local correlation. Specifically, for the 2D image of each frequency band, a 5×5 sliding window is selected, and the normalized cross-correlation coefficients between the pixels in the window and the central pixel are calculated to obtain the correlation coefficient matrix. For a 128×128 image, a 124×124 correlation coefficient matrix can be obtained.

[0164] Then, the Canny edge detection operator is used to extract edge features. For each frequency band of the two-dimensional image, the image is first smoothed with a Gaussian filter, and then the image gradient amplitude and direction are calculated. The edge intensity matrix is ​​obtained through non-maximum suppression and double threshold detection. Similarly, a 124×124 edge intensity matrix can be obtained.

[0165] Then, the texture features are extracted through the Gabor filter bank. A Gabor filter bank with four directions and three scales is constructed to filter the two-dimensional image in each frequency band to obtain 12 filter responses. These responses are concatenated at each pixel position to obtain a 124×124×12 texture feature matrix.

[0166] The above three feature matrices are cascaded in the feature dimension to obtain a multi-dimensional feature vector. For each pixel position, the feature vector dimension is 1+1+12=14. Therefore, a 124×124×14 multi-dimensional feature vector is finally obtained.

[0167] In order to learn the adaptive weight coefficients, a deep neural network consisting of an input layer, a hidden layer, and an output layer is constructed. The number of neurons in the input layer is 14, which is the same as the dimension of the feature vector. The hidden layer adopts a 3-layer structure, with the number of neurons being 64, 32, and 16 respectively. The number of neurons in the output layer is the same as the number of frequency bands, assuming 4 frequency bands.

[0168] The multi-dimensional feature vector is input into the deep neural network for training. The back propagation algorithm is used, and the mean square error is used as the loss function to optimize the network parameters through multiple rounds of iterations. After the training is completed, the output values ​​of the four neurons in the output layer are the adaptive weight coefficients corresponding to the four frequency bands.

[0169] According to the learned adaptive weight coefficients, a nonlinear weighted calculation is performed on the multi-band two-dimensional image. Specifically, the two-dimensional image of each frequency band is multiplied by the corresponding weight coefficient, and then the weighted images are added to obtain the fused image.

[0170] Finally, the fused image needs to be enhanced by histogram equalization. First, the grayscale statistical distribution of the fused image is calculated to obtain the grayscale histogram. Then, the equalization parameters, such as truncation percentage, contrast limit, etc., are determined according to the histogram distribution. These parameters are used to perform adaptive histogram equalization on the fused image, adjust the pixel grayscale distribution, enhance the image contrast, and obtain the final multi-band synthetic aperture radar image.

[0171] Through the above steps, the fusion enhancement of multi-band SAR images is achieved, which effectively improves the image quality and information richness. This method comprehensively utilizes wavelet transform, variable scale kernel function, multi-feature extraction, deep learning and other technologies, and can adaptively process SAR images of different scenes, with strong robustness and versatility.

[0172] In this embodiment, a phase correction operation is performed in the wavelet domain through discrete wavelet transform, which effectively improves the correction accuracy and reduces the influence of phase error on echo data, thereby improving the quality of subsequent imaging; the kernel function scale is adaptively adjusted according to the scattering characteristics, and the corrected echo data is compressed in azimuth, which can significantly improve the resolution of multi-band two-dimensional images and enhance the ability to express target details; a deep neural network is used to train multi-dimensional feature vectors to obtain adaptive weight coefficients, and nonlinear weighted fusion of multi-band two-dimensional images is realized, which effectively integrates information from different frequency bands and improves the accuracy and clarity of the final imaging; the grayscale distribution of the fused image is optimized through histogram equalization, which enhances the contrast and detail expression of the image, so that the final generated multi-band synthetic aperture radar image has better visual effects and interpretability.

[0173] Figure 2 FIG. 1 is a schematic diagram of the structure of a satellite-borne multi-band synthetic aperture radar signal processing optimization system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0174] The first unit is used to obtain the original echo data of the spaceborne multi-band synthetic aperture radar, use the adaptive piecewise polynomial fitting algorithm to estimate the Doppler center frequency of the original echo data, construct a frequency domain adaptive compensation factor matrix, perform a two-dimensional matrix product operation on the frequency domain adaptive compensation factor matrix and the original echo data to obtain compensated echo data, perform range compression on the compensated echo data based on the range matched filtering algorithm of sub-band decomposition, generate a range image, and use the maximum likelihood estimation method based on the range image to calculate the range position information and scattering intensity information of the target scattering point;

[0175] The second unit is used to determine the spatial distribution density of the target scattering points based on the distance position information and scattering intensity information of the target scattering points, and based on the spatial distribution density, adaptively block the compensation echo data, dynamically adjust the block size, and in each block, use a generalized polynomial function containing a quadratic phase error term, a cubic phase error term, and a high-order cross term to model the phase error of the compensation echo data, construct a composite optimization objective function that fuses image entropy and contrast, and quickly iterate the composite optimization objective function through a conjugate gradient algorithm to solve the phase error estimate, and generate a continuous phase correction function through piecewise spline interpolation based on the phase error estimate;

[0176] The third unit is used to perform wavelet domain convolution operation on the phase correction function and the compensated echo data to obtain the corrected echo data, and process the corrected echo data based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, the local correlation coefficient, edge strength and texture features are extracted through a deep neural network to construct a multi-dimensional feature vector, and the multi-dimensional feature vector is weighted learned. Based on the adaptive weight coefficient obtained by learning, the two-dimensional images of different frequency bands are nonlinearly weighted fused, and the synthetic aperture radar imaging results are determined through adaptive histogram equalization enhancement.

[0177] According to a third aspect of the embodiments of the present invention,

[0178] An electronic device is provided, comprising:

[0179] processor;

[0180] a memory for storing processor-executable instructions;

[0181] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0182] A fourth aspect of the embodiments of the present invention is:

[0183] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0184] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing signal processing of a spaceborne multi-band synthetic aperture radar, characterized in that: include: The original echo data of the spaceborne multi-band synthetic aperture radar is obtained, and the Doppler center frequency of the original echo data is estimated by using an adaptive piecewise polynomial fitting algorithm, and a frequency domain adaptive compensation factor matrix is ​​constructed. The frequency domain adaptive compensation factor matrix is ​​subjected to a two-dimensional matrix product operation with the original echo data to obtain compensated echo data. The compensated echo data is range-compressed based on a range-matched filtering algorithm of sub-band decomposition to generate a range image, and the range position information and scattering intensity information of the target scattering point are calculated by using a maximum likelihood estimation method based on the range image. Based on the distance position information and scattering intensity information of the target scattering points, the spatial distribution density of the target scattering points is determined. Based on the spatial distribution density, the compensation echo data is adaptively divided into blocks, and the block size is dynamically adjusted. In each block, the phase error of the compensation echo data is modeled through a generalized polynomial function containing a quadratic phase error term, a cubic phase error term, and a high-order cross term. A composite optimization objective function that fuses image entropy and contrast is constructed. The composite optimization objective function is quickly iterated through a conjugate gradient algorithm to solve the phase error estimate, and a continuous phase correction function is generated based on the phase error estimate through piecewise spline interpolation. The phase correction function and the compensated echo data are convolved in the wavelet domain to obtain the corrected echo data. The corrected echo data are processed based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, the local correlation coefficient, edge intensity and texture features are extracted through a deep neural network to construct a multi-dimensional feature vector, and the multi-dimensional feature vector is weighted learned. Based on the adaptive weight coefficients obtained by learning, the two-dimensional images of different frequency bands are nonlinearly weighted fused, and the synthetic aperture radar imaging results are determined through adaptive histogram equalization enhancement.

2. The method according to claim 1, characterized in that: The Doppler center frequency estimation of the original echo data using the adaptive piecewise polynomial fitting algorithm includes: Segmenting the original echo data, dividing the original echo data into a plurality of sub-data segments according to a preset window length, with a fifty percent overlap rate between adjacent sub-data segments, to obtain sub-data segments to be processed; Perform residual evaluation on the sub-data segment to be processed, calculate normalized residual values ​​of polynomial fittings of different orders, compare the ratio of the normalized residual value to the normalized residual value of the previous order polynomial fitting with a preset threshold, and determine the optimal polynomial order when the ratio is less than the preset threshold; Constructing a Vandermonde matrix based on an optimal polynomial order, wherein the elements of the Vandermonde matrix are composed of power terms of different orders of time variables, constructing a Hanning weighted matrix, wherein the diagonal elements of the Hanning weighted matrix are calculated by a cosine function, solving the polynomial coefficients determined by the Vandermonde matrix and the Hanning weighted matrix by a least squares method, and constructing a fitting polynomial function according to the polynomial coefficients; Calculate the data signal-to-noise ratio of the sub-data segment to be processed, and adaptively adjust the preset window length according to the data signal-to-noise ratio. When the data signal-to-noise ratio is greater than a first preset data signal-to-noise ratio threshold, reduce the preset window length to half of the original length. When the data signal-to-noise ratio is less than a second preset data signal-to-noise ratio threshold, increase the preset window length to twice the original length. Re-fit the data according to the adjusted preset window length and the fitting polynomial function to obtain an optimized fitting polynomial function. The error function between the sub-data segment to be processed and the optimized fitting polynomial function is calculated, the mean and variance of the error function are obtained, a compensation term based on a Gaussian function is constructed, and the compensation term is added to the initial frequency estimation value calculated according to the optimized fitting polynomial function to obtain the final Doppler center frequency estimation result.

3. The method according to claim 1, characterized in that Based on the range matched filtering algorithm of sub-band decomposition, the compensated echo data is range compressed to generate a range image. Based on the range image, the maximum likelihood estimation method is used to calculate the range position information and scattering intensity information of the target scattering point, including: Dividing the spectrum of the compensated echo data into a plurality of frequency sub-bands, with an overlapping area between adjacent frequency sub-bands, the bandwidth of the overlapping area being 20% ​​of the sub-band bandwidth, and using a Hamming window function to weight the spectrum of each frequency sub-band; Generate a reference signal based on the transmission signal parameters, the frequency spectrum of the reference signal includes the antenna pattern and the influence of propagation attenuation, and multiply the reference signal by the signal of each frequency sub-band in the frequency domain to obtain a filtering result; Calculating the phase difference between the frequency sub-bands, correcting the phase difference using a linear phase compensation method, calculating the signal-to-noise ratio of the signal corresponding to each frequency sub-band, and weighting and summing the filtering results using the signal-to-noise ratio as a weight coefficient to obtain a synthetic range image; Performing energy normalization processing on the synthetic range image, dividing the range units according to the system resolution, establishing a mapping relationship between amplitude and distance, searching for the peak position of the synthetic range image, and using the peak position as a rough positioning result of the scattering point; In the neighborhood of the rough positioning result of the scattering point, the parabolic interpolation method is used to subdivide the distance unit tenfold, and the peak value search is performed on the subdivided distance unit to obtain the precise position of the scattering point; Constructing a likelihood function based on a Gaussian noise model, wherein the likelihood function includes a distance position parameter and a scattering intensity parameter, optimizing and solving the likelihood function in an iterative manner, and taking the parameter corresponding to the maximum value of the likelihood function as an initial estimation result; The Cramer-Rao bound of the initial estimation result is calculated to determine whether the initial estimation result reaches the predetermined theoretical accuracy limit. If not, a new search space is constructed in the neighborhood of the initial estimation result, and the parameter iterative optimization is continued to finally obtain the target scattering point location information and scattering intensity information.

4. The method according to claim 1, characterized in that: In each block, the phase error modeling of the compensated echo data is performed through a generalized polynomial function containing quadratic phase error terms, cubic phase error terms and high-order cross terms. The composite optimization objective function of the fusion image entropy and contrast is constructed, including: In each block, a quadratic term with the range position information of the target scattering point as a variable is established as a quadratic phase error term to compensate for the phase error caused by the uniform acceleration motion of the target; a cubic term with the range position information of the target scattering point as a variable is established as a cubic phase error term to compensate for the phase error caused by the variable acceleration motion of the target; a product term of the range position information and the azimuth position information of the target scattering point is established as a cross-phase error term to determine the corresponding relationship between the phase errors caused by the coupling in the range and azimuth directions; the quadratic phase error term, the cubic phase error term and the high-order cross-phase error term are combined to construct a generalized polynomial function; Apply the generalized polynomial function to the compensated echo data for phase compensation to obtain a compensated image; calculate the intensity distribution of the compensated image, normalize the intensity distribution to obtain a probability distribution, and calculate the image entropy based on the probability distribution; extract the maximum intensity and the minimum intensity of the compensated image to calculate the image contrast; calculate the difference between the image contrast and 1 to obtain the image contrast loss; determine a first weight coefficient based on the spatial distribution density of the target area, subtract the first weight coefficient from 1 to obtain a second weight coefficient; multiply the image entropy by the first weight coefficient to obtain a first weighted term, multiply the image contrast loss by the second weight coefficient to obtain a second weighted term, and add the first weighted term to the second weighted term to construct a composite optimization objective function.

5. The method according to claim 1, characterized in that The composite optimization objective function is quickly iterated by the conjugate gradient algorithm to obtain the phase error estimate, and based on the phase error estimate, a continuous phase correction function is generated by piecewise spline interpolation, including: A numerical differentiation method is used to calculate the gradient of the composite optimization objective function with respect to each coefficient in the generalized polynomial function, and the negative direction of the gradient is used as the initial search direction. An optimal step length is determined by a one-dimensional search in the initial search direction. The coefficients in the generalized polynomial function are updated based on the optimal step length to obtain updated coefficients. An updated gradient of the composite optimization objective function is calculated based on the updated coefficients. The inner product ratio of the updated gradient to the gradient before the update is calculated to obtain a conjugate coefficient. The conjugate combination of the negative direction of the updated gradient and the initial search direction is used as a new search direction. Iteration is repeated until the difference between two adjacent iterations of the composite optimization objective function is less than a preset convergence threshold, and a convergence coefficient is obtained. The convergence coefficient is substituted into the generalized polynomial function to obtain a phase error estimate. The center position of each block is determined as a reference interpolation node, and supplementary interpolation nodes are added in the overlapping area of ​​adjacent blocks; the cubic spline basis function is selected as the interpolation basis function, and a spline coefficient equation group is constructed based on the reference interpolation node and the supplementary interpolation node; the boundary condition that the second-order derivative at the endpoint is zero is set for the spline coefficient equation group, the spline coefficient is solved to obtain the spline coefficient, the spline coefficient is combined with the cubic spline basis function, the phase error estimate is segmentedly interpolated, and a continuous phase correction function is generated.

6. The method according to claim 1, characterized in that The phase correction function and the compensation echo data are convolved in the wavelet domain to obtain the corrected echo data. The corrected echo data is processed based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, the local correlation coefficient, edge strength and texture features are extracted through a deep neural network to construct a multi-dimensional feature vector, and the multi-dimensional feature vector is weighted. Based on the adaptive weight coefficient obtained by learning, the two-dimensional images of different frequency bands are nonlinearly weighted fused, and enhanced through adaptive histogram equalization to determine the synthetic aperture radar imaging results, including: Performing discrete wavelet transform on the phase correction function to obtain a first wavelet coefficient, performing discrete wavelet transform on the compensation echo data to obtain a second wavelet coefficient, performing multiplication operation on the first wavelet coefficient and the second wavelet coefficient in the wavelet domain, and performing inverse wavelet transform on the result of the multiplication operation to obtain correction echo data; Determining kernel function parameters according to the scattering characteristics of the corrected echo data, constructing a variable-scale kernel function based on the kernel function parameters, and using the variable-scale kernel function to perform azimuth compression processing on the corrected echo data to obtain a multi-band two-dimensional image; A normalized cross-correlation operator is used to calculate the local correlation of the multi-band two-dimensional image to obtain a correlation coefficient matrix, a Canny edge detection operator is used to extract the edge of the multi-band two-dimensional image to obtain an edge intensity matrix, and a Gabor filter group is used to filter the multi-band two-dimensional image to obtain a texture feature matrix; the correlation coefficient matrix, the edge intensity matrix, and the texture feature matrix are cascaded in the feature dimension to obtain a multi-dimensional feature vector; Constructing a deep neural network including an input layer, a hidden layer, and an output layer, wherein the number of neurons in the input layer is the same as the dimension of the multidimensional feature vector, and the number of neurons in the output layer is the same as the number of frequency bands, inputting the multidimensional feature vector into the deep neural network for training to obtain an adaptive weight coefficient, and performing nonlinear weighted calculation on the multi-band two-dimensional image according to the adaptive weight coefficient to obtain a fused image; The grayscale statistical distribution of the fused image is calculated, a histogram equalization parameter is determined, and the fused image is enhanced using the histogram equalization parameter to obtain a final multi-band synthetic aperture radar image.

7. A satellite-borne multi-band synthetic aperture radar signal processing optimization system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain the original echo data of the spaceborne multi-band synthetic aperture radar, use the adaptive piecewise polynomial fitting algorithm to estimate the Doppler center frequency of the original echo data, construct a frequency domain adaptive compensation factor matrix, perform a two-dimensional matrix product operation on the frequency domain adaptive compensation factor matrix and the original echo data to obtain compensated echo data, perform range compression on the compensated echo data based on the range matched filtering algorithm of sub-band decomposition, generate a range image, and use the maximum likelihood estimation method based on the range image to calculate the range position information and scattering intensity information of the target scattering point; The second unit is used to determine the spatial distribution density of the target scattering points based on the distance position information and scattering intensity information of the target scattering points, and based on the spatial distribution density, adaptively block the compensation echo data, dynamically adjust the block size, and in each block, use a generalized polynomial function containing a quadratic phase error term, a cubic phase error term, and a high-order cross term to model the phase error of the compensation echo data, construct a composite optimization objective function that fuses image entropy and contrast, and quickly iterate the composite optimization objective function through a conjugate gradient algorithm to solve the phase error estimate, and generate a continuous phase correction function through piecewise spline interpolation based on the phase error estimate; The third unit is used to perform wavelet domain convolution operation on the phase correction function and the compensated echo data to obtain the corrected echo data, and process the corrected echo data based on the azimuth compression algorithm of the variable scale kernel function to obtain a multi-band two-dimensional image. Based on each frequency band in the multi-band two-dimensional image, the local correlation coefficient, edge strength and texture features are extracted through a deep neural network to construct a multi-dimensional feature vector, and the multi-dimensional feature vector is weighted learned. Based on the adaptive weight coefficient obtained by learning, the two-dimensional images of different frequency bands are nonlinearly weighted fused, and the synthetic aperture radar imaging results are determined through adaptive histogram equalization enhancement.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images