Single-image sea surface target feature enhancement anti-interference method

Abstract

The invention discloses a single-image sea surface target feature enhancement anti-interference method, and the method comprises the steps: carrying out the slicing of an SLC image, wherein the background of the SLC image is a sea-land boundary or a pure sea surface scene, and a target exists on the sea surface; fourier transform is carried out on the image matrix of the slices along the distance dimension to obtain a distance frequency domain-azimuth time domain representation matrix, an interference mask is predicted, an interference spectrum in the representation matrix is determined, and a pure interference matrix is obtained; when the pure interference matrix meets the low-rank property of the pure interference matrix, taking the pure interference matrix as an initialized low-rank matrix; otherwise, taking the image matrix as an initialized low-rank matrix; modeling the image matrix as superposition of an unknown quantity target signal matrix and an unknown quantity interference matrix, performing low-rank sparse decomposition on the image matrix, and constructing an optimization model; solving the optimization model to obtain a result of removing the interference of the image matrix; and after the interference of the image matrixes of all the slices is removed, splicing operation is carried out to obtain an SLC image after the interference is removed.

Description

Technical Field

[0001] The invention belongs to the field of radar signal processing, and in particular relates to a single image sea surface target feature enhancement method in an anti-interference scenario, which is suitable for suppressing narrowband and complex modulated broadband interference on the sea surface in a synthetic aperture radar single look complex (SLC) image. Background Art

[0002] Synthetic Aperture Radar (SAR) plays an indispensable role in military imaging and reconnaissance scenarios with its all-weather and high-resolution features. On the one hand, with the widespread application of modern radio technology, electromagnetic spectrum resources are becoming increasingly limited, which may lead to an increase in the probability of radio frequency interference; radio frequency interference mainly comes from ground radars, meteorological detection systems, communication equipment, etc. operating in the same frequency band. On the other hand, in a battlefield environment, the interference generated by the electronic countermeasure process has the characteristics of strong energy, certain coherence with the signal, and high degree of interweaving with the target signal in the time-frequency space domain. Once the SAR system is polluted by interference, it will affect the impulse response of the signal, thereby reducing the quality of the SAR image, and further affecting the detection effect of SAR reconnaissance satellites on surface ship targets.

[0003] At present, the main interference suppression methods are mainly divided into parametric methods, semi-parametric methods and non-parametric methods. The parametric method estimates the interference parameters by establishing an interference parameterization model, thereby suppressing the interference in the echo data. However, the mathematical modeling of the interference signal in the adversarial scenario is relatively complex, and the error in the estimation of the model parameters will lead to inaccurate reconstruction of the broadband interference. The non-parametric method uses the difference in amplitude characteristics between the interference and the real echo to effectively suppress the interference in the transform domain, and the calculation amount is relatively small. However, it is not easy to obtain a suitable mapping transformation that can effectively distinguish the interference from the echo signal in the SAR image affected by the large time-width bandwidth interference in the adversarial scenario. The semi-parametric method transforms the interference suppression problem into a hyperparametric optimization problem. For example, it uses low-rank sparse features to establish an optimization model for interference suppression. Compared with the parametric method, the advantage of the semi-parametric method is that it does not require accurate prior knowledge of the interference signal, and can achieve good interference suppression performance through a relatively simple model. However, the performance of the semi-parametric method is highly dependent on strict assumptions, and often requires the use of multiple images to form assumptions. Summary of the invention

[0004] The purpose of the present invention is to provide a method for enhancing the features of a single-image sea surface target in an anti-interference scenario, so as to overcome the problems existing in the prior art.

[0005] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0006] A single image sea surface target feature enhancement and anti-interference method, comprising:

[0007] Slice the synthetic aperture radar SLC image with the background being the land-sea boundary or a pure sea surface scene and with targets on the sea surface;

[0008] Perform Fourier transform on the slice image matrix along the distance dimension to obtain the distance frequency domain-azimuth time domain representation matrix; predict the interference mask of the representation matrix by calculating the statistics of the representation matrix;

[0009] Using the interference mask, determine the interference spectrum in the characterization matrix, perform inverse Fourier transform on the interference spectrum to obtain a pure interference matrix; when the pure interference matrix satisfies the low rank property of the pure interference matrix, use it as an initialization low rank matrix; otherwise, use the image matrix as an initialization low rank matrix;

[0010] The image matrix is ​​modeled as a superposition of an unknown target signal matrix and an unknown interference matrix, and the image matrix is ​​subjected to low-rank sparse decomposition to construct an optimization model; wherein the initialization low-rank matrix is ​​an initialization value of the unknown interference matrix;

[0011] The optimization model is solved to obtain the result of the image matrix after the interference is removed; the image matrices of all slices are combined after the interference is removed to obtain the SLC image after the interference is removed.

[0012] Furthermore, the statistics of the calculation characterization matrix are expressed as:

[0013]

[0014] Where X fr (i,j) is X fr The i-th row and j-th column element in Z score (i, j) is the corresponding statistic; mean(·) is to find the mean of the two-dimensional matrix, std(·) is to calculate the standard deviation of the two-dimensional matrix; X fr It is the distance frequency domain-azimuth time domain representation matrix obtained after Fourier transform of the image matrix.

[0015] Further, the predicting of the interference mask for the characterization matrix includes:

[0016] Interference Mask m The elements in can be represented as:

[0017]

[0018] Where normcdf(·) is the cumulative standard normal distribution, |·| is the absolute value operation, and α is the confidence level.

[0019] Furthermore, the interference mask is used to determine the interference spectrum in the characterization matrix, and the interference spectrum is inversely Fourier transformed to obtain a pure interference matrix, including:

[0020] The representation matrix X fr Corresponding Mask m All elements that are 0 are set to 0, and the matrix X fr The remaining part is the interference spectrum, and the inverse Fourier transform is performed on it to obtain the pure interference matrix I m .

[0021] Furthermore, the judgment criterion of low rank is:

[0022] When the rank percentage of the matrix is ​​less than 15%, it is considered to meet the low rank property; when the rank percentage is greater than 15%, it is considered not to meet the low rank property; the rank percentage calculation formula is:

[0023] r=rank(I m ) / Nr

[0024] Where rank(·) is the rank of the matrix, and Nr represents the total number of fast-time discrete sampling points of the image matrix.

[0025] Furthermore, the image matrix is ​​modeled as a superposition of an unknown target signal matrix and an unknown interference matrix, and the image matrix is ​​subjected to low-rank sparse decomposition to construct an optimization model, which is expressed as:

[0026]

[0027] Among them, X represents the image matrix, S represents the unknown target signal matrix, and L represents the unknown interference matrix, where ||·|| * is the nuclear norm, ||·||1 is the 1 norm, λ is a hyperparameter, L0 is the initialization low rank matrix, that is, the initialization value of L; the initialization value of S is set to 0, I m is a pure interference matrix.

[0028] Furthermore, solving the optimization model to obtain the result after removing interference from the image matrix includes:

[0029] The optimization model is decomposed into the following sub-optimization models to obtain the k-th generation optimization results of S and L:

[0030] Step 4.1, use the following sub-optimization model to solve the current k-th generation optimal sparse matrix S k :

[0031]

[0032] where Y k The matrix is ​​the kth generation Lagrange multiplier, μ is the Lagrange penalty coefficient, ||·||F is the F norm, L k-1 is the low-rank matrix solved for the k-1th generation;

[0033] Step 4.2, use the following sub-optimization model to solve the optimal low-rank matrix L of the current kth generation k :

[0034]

[0035] Step 4.3, update the Lagrange multiplier:

[0036] Y k =Y k-1 +(XS k-1 -L k-1 )

[0037] Step 4.4, determine whether to stop the iteration or continue to execute 4.1 according to whether the iteration termination condition is met; the termination condition is set to the reconstruction error is less than the set value err:

[0038] XS k -L k <err

[0039] If the iteration is stopped, the optimal solution of S and L can be obtained. and is the result of removing interference from the sliced ​​image matrix, and is the final output image with restored quality. For slices with only interference.

[0040] Furthermore, the sub-optimization model is solved using a classical soft threshold operator method.

[0041] A terminal device comprises a processor, a memory and a computer program stored in the memory; when the processor is executed by a computer, the single-image sea surface target feature enhancement and anti-interference method is implemented.

[0042] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the single-image sea surface target feature enhancement and anti-interference method is implemented.

[0043] Compared with the prior art, the present invention has the following technical features:

[0044] According to the texture features of interference in the image domain and the local sparseness of ship targets, this paper proposes a single-image sea surface target feature enhancement method in the anti-interference scenario. This method can not only remove narrowband and broadband interference above the target, but also effectively suppress the defocusing phenomenon of the ship target, improve the signal-to-noise ratio in the area, and improve the feature distinction between the target and the background to improve the subsequent target detection performance. In addition, based on the traditional sparse decomposition algorithm, the completeness of target information recovery can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flowchart of an embodiment of the method of the present invention;

[0046] Figure 2 The image slices of the SLC containing the ship target are interfered, where (a) is the measured data scene of scene A superimposed with ultra-wideband sinusoidal frequency modulation interference; (b) is the scene of scene B with actual narrowband interference;

[0047] Figure 3 is the scene range frequency domain-azimuth time domain matrix, where (a) is the range frequency domain-azimuth time domain matrix of scene A; (b) is the range frequency domain-azimuth time domain matrix of scene B;

[0048] Figure 4 To extract the pure interference matrix; (a) is the pure interference matrix extracted for scene A; (b) is the pure interference matrix extracted for scene B;

[0049] Figure 5 The processing result of scene A, where (a) is the low-rank matrix after decomposition; (b) is the sparse matrix after decomposition; (c) is the original imaging result without interference;

[0050] Figure 6 The processing results of scene B, where (a) is the low-rank matrix after decomposition by the original robust principal component analysis method; (b) is the low-rank matrix after decomposition by the principal component analysis method of the present invention; (c) is the sparse matrix after decomposition by the present invention; and (d) is the frequency domain notch result. DETAILED DESCRIPTION

[0051] The present invention provides a single-image sea surface target feature enhancement and anti-interference method. First, a full-size SAR image is processed in blocks to construct a target sparse interference low-rank matrix to be decomposed; secondly, the interference roughly extracted by a frequency domain notch filter is used as the initial low-rank matrix value to guide its decomposition process, improve the model accuracy, and retain more target features; finally, a robust principal component analysis method is performed to decompose and reconstruct the sparse matrix as a target detection slice for removing interference, defocusing, and sidelobe tailing phenomena, and the low-rank matrix is ​​the separated interference matrix.

[0052] Referring to the accompanying drawings, a method for enhancing the anti-interference of sea surface target features in a single image includes the following steps:

[0053] Step 1: Slice a large-size synthetic aperture radar SLC image with a background of land-sea boundary or pure sea scene and targets on the sea surface into small slices. The slice size can be set to 256×256. Suppose the slice image matrix Represents, where X(i,j) represents the i-th row and j-th column element in the image matrix X of the slice, where 1≤i≤Na and 1≤j≤Nr, i represents the slow time discrete sampling point, j represents the fast time sampling point, Na represents the total number of slow time discrete sampling points, Nr represents the total number of fast time discrete sampling points, the size of Na and Nr depends on the selected slice size, and can be set to 256 respectively, Represents a complex set.

[0054] Step 2: Perform Fourier transform on X along the distance dimension, that is, transform each row element of X into the frequency domain to obtain the distance frequency domain-azimuth time domain representation matrix X fr , the dimension of the representation matrix is ​​also Na×Nr; calculate X fr The statistic Z score , to quantify the specificity of interference relative to the target signal:

[0055]

[0056] Where X fr (i,j) is X fr The i-th row and j-th column element in Z score (i, j) is the corresponding statistic, mean(·) is to find the mean of the two-dimensional matrix, and std(·) is to calculate the standard deviation of the two-dimensional matrix;

[0057] By calculating Z score (i, j) and set the confidence level to design the frequency domain notch filter, so as to fr Perform a two-dimensional interference mask m The prediction of m The elements in can be represented as:

[0058]

[0059] Among them, normcdf(·) is the cumulative standard normal distribution, |·| is the absolute value operation, and α is the confidence level, which can be set according to experience.

[0060] Step 3, X fr Corresponding Mask m All elements that are 0 are set to 0, and the matrix X fr The remaining part is the interference spectrum, and the inverse Fourier transform is performed on it to obtain the pure interference matrix I m :

[0061]

[0062] The interference matrix extracted for broadband interference scenarios is not accurate. The validity can be judged based on the rank percentage of the extracted interference matrix. When the rank percentage is less than 15%, it is considered to meet the low rank property. The extracted interference matrix I m Valid, it is used as the low-rank matrix initialized in step 4; when the rank percentage is greater than 15%, it is considered that the low-rank property is not satisfied, and the extracted interference matrix is ​​invalid. At this time, let I m =X.

[0063] The rank percentile is calculated as:

[0064] r=rank(I m ) / Nr (4)

[0065] Where rank(·) is the rank of the matrix.

[0066] Step 4: Model the image matrix X as the superposition of the unknown target signal matrix S and the unknown interference matrix L.

[0067] X=L+S (5)

[0068] The matrix X is subjected to low-rank sparse decomposition using an improved robust principal component analysis algorithm, and the following optimization model is iteratively solved to obtain the target optimization variables S and L:

[0069]

[0070] where ||·|| * is the nuclear norm used to constrain the low-rank condition, ||·||1 is the 1 norm used to constrain the sparse condition, and λ is the algorithm hyperparameter, which is generally taken as max(·) is the maximum value of the two values ​​in the brackets, L0 is the initialization value of the low-rank matrix, that is, the initialization value of L; the initialization value of S is set to 0.

[0071] After initializing the low-rank matrix, the process of solving the optimization model by the Lagrangian method in the original robust principal component analysis method needs to be adjusted, and the optimization model (6) is decomposed into the following sub-optimization models. The specific steps for calculating the k-th generation optimization results S and L are as follows:

[0072] Step 4.1, use the following sub-optimization model to solve the current k-th generation optimal sparse matrix S k :

[0073]

[0074] where Y k The matrix is ​​the k-th generation Lagrange multiplier, μ is the Lagrange penalty coefficient, which can be 10λ, ||·||F is the F norm, L k-1 is the low-rank matrix solved for the k-1th generation; the optimization model is solved using the classic soft threshold operator method, which has the same steps as the classic robust principal component analysis method and will not be described in detail here.

[0075] Step 4.2, use the following sub-optimization model to solve the current optimal low-rank matrix L k :

[0076]

[0077] The optimization model is solved using the singular value decomposition algorithm and the classic soft threshold operator method, which has the same steps as the classic robust principal component analysis method and will not be elaborated here.

[0078] Step 4.3, update the Lagrange multiplier:

[0079] Y k =Y k-1 +(XS k-1 -L k-1 ) (9)

[0080] Step 4.4, determine whether to stop the iteration or continue to execute 4.1 according to whether the iteration termination condition is met; the termination condition can generally be set to that the reconstruction error is less than the set value err, which can be set to 1e-5, and the specific size can be adjusted according to the actual value.

[0081] XS k -L k <err (10)

[0082] If the iteration is stopped, the optimal solution of S and L can be obtained. and is the result of removing interference from the sliced ​​image matrix, and is the final output image with restored quality. For slices with only interference; first optimize S k Matrix post-optimization L k The matrix can be used to represent S containing the complete target and residual interference. k The interference is gradually eliminated based on the matrix, which is opposite to the optimization process of the original method where the matrix is ​​initialized to 0 and gradually increases the sparse pixel points. This ensures the integrity of the target. Steps 1 to 4 are performed on all slices. After stitching, the original SLC image with anti-interference feature enhanced is obtained.

[0083] Example:

[0084] Step 1: Cut the large-size synthetic aperture radar SLC image with the background of the land-sea boundary or the pure sea scene and the target on the sea surface into small slices, and set the slice size to 256×256. Suppose one of the slices is the matrix Where X(i,j) represents the i-th row and j-th column element in the slice matrix X, where 1≤i≤Na and 1≤j≤Nr, i represents the slow-time discrete sampling point, j represents the fast-time sampling point, Na represents the number of slice orientation points, and Nr represents the number of slice fast-time sampling points. The size of Na and Nr depends on the selected slice size and is set to 256 respectively. Here, two scenarios are used as implementation examples for illustration. Scenario A is the imaging result of measured echo data superimposed with simulated ultra-wideband sinusoidal modulation interference, and scenario B is the data affected by actual narrowband radio frequency interference, such as Figure 2 .

[0085] Step 2: The matrix X is modeled as the superposition of the unknown target signal matrix S and the unknown interference matrix L.

[0086] X=L+S (1)

[0087] Perform Fourier transform on each row of X and transform X into distance frequency domain-azimuth time domain X fr like Figure 3 , we can see that complex modulated ultra-wideband interference forms dense bright lines in the frequency domain, while narrowband interference forms interference strips with relatively concentrated energy. fr Z score The statistic quantifies the specificity of the interference relative to the target signal:

[0088]

[0089] Where mean(·) is used to find the mean of a two-dimensional matrix, and std(·) is used to calculate the standard deviation of a two-dimensional matrix. score The confidence interval of X is set to design the frequency domain notch filter fr Perform a two-dimensional interference mask m The prediction is obtained. m Final detection mask Mask m :

[0090]

[0091] Among them, normcdf(·) is the cumulative standard normal distribution, |·| is the absolute value operation, and α is the confidence level, which is set to 0.2.

[0092] Step 3, X fr Corresponding Mask m All 0 values ​​are set to 0, and the remaining X frThe matrix is ​​the interference spectrum, and the inverse Fourier transform is performed on it to obtain the pure interference matrix I m :

[0093]

[0094] The interference matrix extracted for broadband interference scenarios is not accurate, such as Figure 4 (b), the extracted interference matrix rank percentage is 18.83%, which does not meet the low rank assumption, so this matrix is ​​not used as the initial low rank matrix. Let I m = X. The interference matrix of narrowband interference scenario is as follows Figure 4 (a), its rank percentage is 8.56%, which satisfies the low rank assumption. m The matrix is ​​valid.

[0095] Step 4: Use the improved robust principal component analysis algorithm to perform low-rank sparse decomposition on the matrix X and solve the following optimization model:

[0096]

[0097] where ||·|| * is the nuclear norm used to constrain the low-rank condition, ||·||1 is the 1-norm used to constrain the sparse condition, and λ is the algorithm hyperparameter, which is the default in the sample. max(·) is the maximum value of the two values ​​in brackets, and L0 is the initial low-rank matrix.

[0098] Finally, the two scenarios get the optimal results as follows Figure 5 , the defocus linearity of the ship target is reduced. The narrowband interference scenario is compared with the original robust principal component analysis method and the frequency domain notch result. Figure 6 It can be seen that the present invention suppresses the side lobes and reduces the noise compared with the frequency domain notch method. Figure 6 (c) and (d) are the comparison results. Compared with the traditional robust principal component analysis method, more target information is extracted, and there are no residual target pixels in the separated interference matrix, such as Figure 6 Comparison results of (a) and (b).

[0099] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A single image sea surface target feature enhancement and anti-interference method, characterized in that: include: Slice the synthetic aperture radar SLC image with the background being the land-sea boundary or a pure sea surface scene and with targets on the sea surface; Perform Fourier transform on the image matrix of the slice along the distance dimension to obtain the distance frequency domain-azimuth time domain representation matrix; By calculating the statistics of the characterization matrix, the interference mask of the characterization matrix is ​​predicted; Using the interference mask, determine the interference spectrum in the characterization matrix, perform inverse Fourier transform on the interference spectrum to obtain a pure interference matrix; when the pure interference matrix satisfies the low rank property of the pure interference matrix, use it as an initialization low rank matrix; otherwise, use the image matrix as an initialization low rank matrix; The image matrix is ​​modeled as a superposition of an unknown target signal matrix and an unknown interference matrix, and the image matrix is ​​subjected to low-rank sparse decomposition to construct an optimization model; wherein the initialization low-rank matrix is ​​an initialization value of the unknown interference matrix; The optimization model is solved to obtain the result of the image matrix after the interference is removed; the image matrices of all slices are combined after the interference is removed to obtain the SLC image after the interference is removed.

2. The single image sea surface target feature enhancement and anti-interference method according to claim 1 is characterized in that: The calculation characterizes the statistics of the matrix, expressed as: Where X fr (i,j) is X fr The i-th row and j-th column element in Z score (i, j) is the corresponding statistic; mean(·) is to find the mean of the two-dimensional matrix, std(·) is to calculate the standard deviation of the two-dimensional matrix; X fr It is the distance frequency domain-azimuth time domain representation matrix obtained after Fourier transform of the image matrix.

3. The single image sea surface target feature enhancement and anti-interference method according to claim 1 is characterized in that: The predicting of the interference mask for the characterization matrix includes: Interference Mask m The elements in can be represented as: Where normcdf(·) is the standard normal distribution cumulative, |·| is the absolute value operation, and α is the confidence level.

4. The single image sea surface target feature enhancement and anti-interference method according to claim 1 is characterized in that: The interference mask is used to determine the interference spectrum in the characterization matrix, and the interference spectrum is inversely Fourier transformed to obtain a pure interference matrix, including: The representation matrix X fr Corresponding Mask m All elements that are 0 are set to 0, and the matrix X fr The remaining part is the interference spectrum, and the inverse Fourier transform is performed on it to obtain the pure interference matrix I m .

5. The single image sea surface target feature enhancement and anti-interference method according to claim 1 is characterized in that: The criterion for judging low rank is: When the rank percentage of the matrix is ​​less than 15%, it is considered to satisfy the low rank property; when the rank percentage is greater than 15%, it is considered not to satisfy the low rank property; the rank percentage calculation formula is: r=rank(I m ) / Nr Where rank(·) is the rank of the matrix, and Nr represents the total number of fast-time discrete sampling points of the image matrix.

6. The single image sea surface target feature enhancement and anti-interference method according to claim 1 is characterized in that: The image matrix is ​​modeled as a superposition of an unknown target signal matrix and an unknown interference matrix, and the image matrix is ​​subjected to low-rank sparse decomposition to construct an optimization model, which is expressed as: Among them, X represents the image matrix, S represents the unknown target signal matrix, and L represents the unknown interference matrix, where ||·|| * is the nuclear norm, ||·||1 is the 1 norm, λ is a hyperparameter, L0 is the initialization low rank matrix, that is, the initialization value of L; the initialization value of S is set to 0, I m is a pure interference matrix.

7. The single image sea surface target feature enhancement and anti-interference method according to claim 1 is characterized in that: Solving the optimization model to obtain the result of removing interference from the image matrix includes: The optimization model is decomposed into the following sub-optimization models to obtain the k-th generation optimization results of S and L: Step 4.1, use the following sub-optimization model to solve the current k-th generation optimal sparse matrix S k : where Y k The matrix is ​​the kth generation Lagrange multiplier, μ is the Lagrange penalty coefficient, ||·|| F is the F norm, L k-1 is the low-rank matrix solved for the k-1th generation; Step 4.2, use the following sub-optimization model to solve the optimal low-rank matrix L of the current kth generation k : Step 4.3, update the Lagrange multiplier: AND k =And k-1 +(XS k-1 -L k-1 ) Step 4.4, determine whether to stop the iteration or continue to execute 4.1 according to whether the iteration termination condition is met; the termination condition is set to the reconstruction error is less than the set value err: XS k -L k err If the iteration is stopped, the optimal solution of S and L can be obtained. and is the result of removing interference from the sliced ​​image matrix, and is the final output image with restored quality. For slices with only interference.

8. The single image sea surface target feature enhancement and anti-interference method according to claim 7, characterized in that: The sub-optimization model is solved using the classical soft threshold operator method.

9. A terminal device comprising a processor, a memory and a computer program stored in the memory; characterized in that: When the processor is executed by a computer, the single image sea surface target feature enhancement and anti-interference method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium, wherein a computer program is stored in the medium; characterized in that: When the computer program is executed by a processor, the single image sea surface target feature enhancement and anti-interference method according to any one of claims 1 to 8 is implemented.

Images