A SAR image denoising method based on sparse representation and blind estimation combination

Abstract

The application relates to a SAR image denoising method based on sparse representation and blind estimation combination, which comprises the following steps: S1, selecting a dictionary and initializing a dictionary parameter; S2, obtaining a SAR image to be denoised, and estimating the image noise of the SAR image to be denoised by using a blind estimation algorithm; S3, calculating sparse coding by using an OMP algorithm according to the image noise, iteratively updating the dictionary, and obtaining an updated dictionary and sparse coding; and S4, carrying out sparse decomposition and reconstruction on the SAR image to be denoised by using the updated dictionary and sparse coding, and obtaining a denoised image. The method can well remove the noise in the SAR image and retain the details of the image.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of SAR image processing, and particularly relates to a SAR image denoising method based on joint sparse representation and blind estimation. BACKGROUND

[0002] SAR imaging has the operation ability of all-time, all-region and all-weather, and the imaging effect is not affected by weather, region and climate. Moreover, SAR imaging is independent of distance, and the resolution thereof does not decrease with the increase of distance. SAR imaging is widely applied in the military field and the civil fields such as agriculture, forestry, ocean research and geological survey.

[0003] However, when obtaining the information of the SAR image, various external or internal interferences will inevitably occur, so that the image is mixed with many noises, and the obtained information is incomplete or even wrong. In a general sense, noise is useless information. In the image processing process, the noise must be effectively suppressed to improve the image quality and visual effect, so that the correct judgment of the image can be improved, and the subsequent processing of the image has much significance.

[0004] At present, there are many methods for image denoising, for example, traditional denoising methods such as wavelet threshold method and total variation (TV) norm minimum denoising method, and frequency domain denoising methods such as Fourier transform method, wavelet transform method and multi-scale set analysis. The wavelet threshold method calculates the threshold value by estimating the noise variance; the total variation (TV) norm minimum denoising method can retain most of the edge texture information of the image, but the actual denoising effect of the two methods is not ideal. The frequency domain method mainly utilizes the characteristics that the frequency spectrum distribution of most of the useful information in the image is relatively concentrated, and the noise frequency spectrum distribution is infinitely wide, and by setting a threshold, the useless noise information is filtered out, so as to improve the signal-to-noise ratio of the useful information. However, since the image details and edge information are often mixed in the noise spectrum, the details of the image are also filtered out to a certain extent by the method. SUMMARY

[0005] In order to solve the above problems in the prior art, the present application provides a SAR image denoising method based on joint sparse representation and blind estimation. The technical problem to be solved by the present application is solved by the following technical scheme:

[0006] The present application provides a SAR image denoising method based on joint sparse representation and blind estimation, comprising:

[0007] S1: selecting a dictionary and initializing a dictionary parameter;

[0008] S2: obtaining a SAR image to be denoised, and estimating the image noise of the SAR image to be denoised by using a blind estimation algorithm;

[0009] S3: calculating sparse coding by using OMP algorithm according to the image noise, iteratively updating the dictionary, and obtaining an updated dictionary and sparse coding;

[0010] S4: performing sparse decomposition and reconstruction on the SAR image to be denoised by using the updated dictionary and sparse coding, and obtaining a denoised image.

[0011] In an embodiment of the present application, in step S1, the dictionary is a DCT dictionary, a training dictionary or an adaptive dictionary.

[0012] In an embodiment of the present application, in step S2, the image noise of the SAR image to be denoised is estimated according to the following formula:

[0013]

[0014] wherein, denotes the image noise level, and γ min denotes the minimum eigenvalue of the matrix, and Σy denotes the covariance matrix of the contaminated patch y.

[0015] In an embodiment of the present application, the S3 comprises:

[0016] S31: performing block processing on the SAR image to be denoised, decomposing it into a plurality of 8*8 matrix blocks, and sequentially converting each 8*8 matrix block into a 64*1 column vector in the column and row mode to generate a new matrix Blocks;

[0017] S32: updating each element of each column in the matrix Blocks according to the following mode:

[0018] Blocks[i][j] = Blocks[i][j] - mean(Blocks[:, j]),

[0019] wherein, Blocks[i][j] denotes the (i, j) element of the matrix Blocks, and mean(Blocks[:, j]) denotes the average value of the column in which the element (i, j) is located.

[0020] S33: solving the sparse coding of the matrix Blocks under the dictionary by using the OMP algorithm.

[0021] S34: updating each element of each column in the matrix Blocks according to the following mode according to the sparse coding:

[0022] Blocks[i][j] = Blocks[i][j] * Coefs + mean(Blocks[:, j]),

[0023] wherein, Blocks[:,j] represents the jth column, mean(Blocks[:,j] represents the mean value of the jth column, and Coefs represents sparse coding;

[0024] S35: repeating S32-S34 until a preset iteration stopping condition is reached, to obtain an updated dictionary and sparse coding, the iteration stopping condition being that a preset iteration number is reached or a residual value is less than the image noise.

[0025] In an embodiment of the present application, S4 comprises:

[0026] S41: reconstructing each of the matrix blocks using the updated dictionary and sparse coding;

[0027] S42: superimposing the reconstructed matrix blocks by weighting, to obtain the denoised image.

[0028] Compared with the prior art, the present application has the beneficial effects that:

[0029] The SAR image denoising method based on joint sparse representation and blind estimation of the present application firstly uses sparse coding and dictionary learning algorithm to represent the main information of an image with as few atomic information as possible, which can well maintain the structure and texture information of the original image, secondly estimates the real noise level of the image using a blind estimation algorithm, and finally filters out the image noise and retains the useful information of the original image. The method can well remove the noise in the SAR image and retain the details of the image.

[0030] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the content of the specification can be implemented, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a schematic diagram of a SAR image denoising method based on joint sparse representation and blind estimation provided by an embodiment of the present application;

[0032] Figure 2 is a flowchart of a SAR image denoising method based on joint sparse representation and blind estimation provided by an embodiment of the present application;

[0033] Figure 3 is an experimental simulation image provided by an embodiment of the present application.

[0034] Figure 4 is a denoising effect diagram of scene 1;

[0035] Figure 5is a denoising effect drawing of scene 2; DETAILED DESCRIPTION

[0036] In order to further clarify the technical means and effects taken by the present application to achieve the predetermined object, a kind of SAR image denoising method based on sparse representation and blind estimation combination according to the present application is described in detail below in conjunction with the specific embodiments and the drawings.

[0037] The foregoing and other technical contents, features and effects of the present application can be clearly presented in the specific embodiment description below in conjunction with the drawings. Through the description of specific embodiments, the technical means and effects taken by the present application to achieve the predetermined object can be more deeply and specifically understood, however, the attached drawings are only provided for reference and illustration, and are not used to limit the technical solutions of the present application.

[0038] Embodiment one

[0039] Please see Figure 1 and Figure 2 , Figure 1 is a schematic diagram of a kind of SAR image denoising method based on sparse representation and blind estimation combination provided by the embodiment of the present application, Figure 2 is a flow chart of a kind of SAR image denoising method based on sparse representation and blind estimation combination provided by the embodiment of the present application. As shown in the figure, the SAR image denoising method based on sparse representation and blind estimation combination of the present embodiment includes:

[0040] S1: select dictionary and initialize dictionary parameters;

[0041] In the present embodiment, the selected dictionary is an overcomplete dictionary, which can be selected as a DCT dictionary, a training dictionary or an adaptive dictionary.

[0042] S2: obtain the SAR image to be denoised, and estimate the image noise of the SAR image to be denoised using a blind estimation algorithm;

[0043] In the present embodiment, the image noise of the SAR image to be denoised is estimated according to formula (1):

[0044]

[0045] wherein, denotes the image noise level, γ min denotes the minimum eigenvalue of the matrix, and Σy denotes the covariance matrix of the contaminated patch y.

[0046] S3: according to the image noise, the sparse coding is calculated using OMP algorithm, the dictionary is iteratively updated, and the updated dictionary and sparse coding are obtained;

[0047] In particular, S3 comprises:

[0048] S31: block processing is performed on the to-be-noise-reduced SAR image, the to-be-noise-reduced SAR image is decomposed into a plurality of 8*8 matrix blocks, and each 8*8 matrix block is converted into a 64*1 column vector in a column and row manner to generate a new matrix Blocks;

[0049] S32: each element of each column in the matrix Blocks is updated in the following manner:

[0050] Blocks[i][j] = Blocks[i][j] - mean(Blocks[:, j]) (2),

[0051] wherein Blocks[i][j] represents an i-th row and j-th column element in the matrix Blocks, and mean(Blocks[:, j]) represents an average value of a column in which the element (i, j) is located;

[0052] S33: an OMP algorithm is used to solve sparse coding of the matrix Blocks under a dictionary;

[0053] S34: each element of each column in the matrix Blocks is updated in the following manner according to the sparse coding:

[0054] Blocks[i][j] = Blocks[i][j] * Coefs + mean(Blocks[:, j]) (3),

[0055] wherein Blocks[:, j] represents the j-th column, mean(Blocks[:, j] represents an average value of the j-th column, and Coefs represents the sparse coding;

[0056] S35: S32-S34 are repeated until a preset iteration stopping condition is reached, to obtain an updated dictionary and sparse coding, and the iteration stopping condition is that a preset number of iterations is reached or a residual value is less than image noise.

[0057] S4: sparse decomposition and reconstruction are performed on the to-be-noise-reduced SAR image by using the updated dictionary and sparse coding, to obtain a denoised image.

[0058] In particular, S4 comprises:

[0059] S41: each matrix block is reconstructed by using the updated dictionary and sparse coding;

[0060] S42: the reconstructed matrix blocks are weighted and superimposed to obtain the denoised image.

[0061] In this embodiment, each column (64*1) in the updated matrix Blocks is converted into 8*8 blocks, which is the inverse operation of step S31. The denoised image is then obtained.

[0062] The SAR image denoising method based on sparse representation and blind estimation in this embodiment first uses sparse coding and dictionary learning algorithms to represent the main information of the image with as little atomic information as possible, which can well preserve the structure and texture information of the original image. Secondly, it uses a blind estimation algorithm to estimate the true noise level of the image. Finally, it filters out image noise and retains the useful information of the original image. This method can effectively remove noise in SAR images and preserve image details.

[0063] Furthermore, the sparse representation, dictionary learning algorithm, and blind estimation algorithm involved in the SAR image denoising method based on the joint sparse representation and blind estimation in this embodiment will be described in detail.

[0064] Sparse representation of signals aims to obtain a more concise signal representation model that can express more complex image information, thereby simplifying the data volume and complexity of image signal processing. Overcomplete redundancy functions, where the number of atomic signals in the dictionary is much greater than the length of the signal, make the signal representation model sparsely represented by an overcomplete dictionary. To obtain a more concise signal representation, the sparse representation model needs to be trained with sample data to achieve a sparse representation of the signal under this redundant dictionary. The model is then solved to recover the original signal. Furthermore, constraints are added to the sparse representation model based on the overcomplete dictionary to reduce modeling errors.

[0065] The sparse redundancy representation model based on overcomplete dictionaries assumes that any signal x can be represented in a dictionary D∈R of atoms. (n×m) The image (n << m) is represented by a sparse vector, x = Dα. D represents the dictionary matrix, and α represents the sparse vector. By solving for the dictionary and the sparse vector, the image can be reconstructed using x = Dα. The goal of sparse representation is to obtain a sufficiently sparse vector α.

[0066] By using signal coding methods, given a complete dictionary D, a sparse vector α is obtained. The obtained α vector has three non-zero terms, corresponding to three atoms in D. The image signal can be reconstructed and the original signal recovered using the formula: x = Dα.

[0067] As analyzed above, the image signal can be reconstructed by linear combination of a small number of atoms, which achieves the purpose of sparse representation of data. The key problem is to require better sparse coefficients, and the solution to this problem is to solve the equation. Since the overcomplete dictionary is redundant, many decomposition forms of the signal will be obtained, so the key is to find the optimal sparse vector to approximate the original signal. Therefore, a constraint condition is needed to obtain a better solution, so the above model can be represented as:

[0068] α = arg α min‖α‖0s.t.x = Dα (4),

[0069] where ‖α‖0 represents the l0 norm, which is used to constrain the number of non-zero elements in the sparse vector and characterize the sparsity.

[0070] However, due to noise, x≈Dα is more suitable for describing the signal. Through prior knowledge, given x and D, the optimal sparse coefficient α is obtained as much as possible, so the formula can be equivalent to solving the following optimization problem:

[0071]

[0072] where ε represents the error constraint constant, represents the residual of sparse decomposition. Therefore, after obtaining α, the reconstructed image signal x can be obtained from x = Dα by combining the dictionary D.

[0073] The key problem of sparse representation is the design of overcomplete dictionary, which directly affects the sparsity of the sparse coefficient α in the sparse decomposition process of the image signal. Therefore, the design of the dictionary is the core problem of sparse representation. The higher the matching degree of the atoms in the overcomplete dictionary and the characteristics of the image signal, the sparser the decomposition result of the signal, and the more conducive to the processing of the image.

[0074] According to different ways of obtaining the dictionary, the dictionary is generally divided into two types: analytical dictionary and learning dictionary. The analytical dictionary is used for sparse decomposition by using a fixed dictionary, which is relatively simple to implement, but is not suitable for all signals and has poor adaptability. The learning dictionary has better adaptability and can adaptively change according to the input image, and has better processing effect, but the complexity is higher than that of the analytical dictionary. The algorithm flow of the learning dictionary is as follows: first, the noise image is divided into blocks, then the dictionary is trained based on the noise image block, and the redundant discrete cosine transform dictionary (DCT) is selected for training. The dictionary obtained by training is used for sparse decomposition of each noise image block, and the sparse decomposition is averaged from local to global, and finally the denoising image is obtained.

[0075] Discrete cosine dictionary (DCT) is obtained by discrete cosine transform, given a sequence x(n), n=0, 1, 2,..., N-1, its discrete cosine transform is:

[0076]

[0077]

[0078] In matrix form:

[0079] X c =C N x (8),

[0080] Where C N is an N x N transform matrix, whose row vectors are cosine bases.

[0081] For the perfect dictionary obtained after DCT transform, the perfect dictionary is extended to an over-complete dictionary by using fractional frequency method, and the specific method is that the perfect dictionary is traversed and sampled more finely in frequency to obtain a new over-complete dictionary.

[0082] OMP algorithm is one of the methods for sparse decomposition of signals, which decomposes signals on a perfect dictionary library. It is assumed that the signal to be represented is y, and the length of y is n. It is assumed that H represents a Hilbert space. In this space H, a dictionary matrix D is formed by a set of vectors {x1, x2,..., xn}, wherein each vector can be referred to as an atom, and the length of each vector is the same as the length n of the signal to be represented y.

[0083] From the dictionary matrix D (also referred to as an over-complete atom library), an atom (also referred to as a column) that best matches the signal y is selected, a sparse approximation is constructed, and a signal residual is calculated. Then, an atom that best matches the signal residual is selected, and the iteration is repeated. In each iteration process, the selected atom is orthogonalized to ensure that the result of each cycle is an optimal solution. The signal y can be represented by the linear sum of the atoms, plus the final residual value. Whether the residual value meets the requirement is determined by a given sigma value, which is the image noise of the SAR image to be denoised estimated by the blind estimation algorithm.

[0084] The SAR image denoising method based on sparse representation and blind estimation combination of the embodiment combines blind estimation noise level, estimates the noise level (sigma) of the SAR image, determines the size of the required residual value by using the estimated sigma in the OMP algorithm residual iteration, and finally reconstructs the image.

[0085] Regarding the blind denoising theory, in the patch-based noise model estimation technique, a set of patches is derived from the contaminated signal in the raster scan. In this algorithm, the windowing method is applied to each patch and then slid pixel by pixel until each patch is covered. Thus, the patches under test overlap and the data model of each patch can be represented as:

[0086] z i = x i + n i , i = 1, 2, 3..., N (9),

[0087] where N represents the patch number, x i is the first original patch of size M x M, each patch is clarified by its central pixel; z i represents the observed vector patch contaminated by independent and identically distributed noise.

[0088] The additive white Gaussian noise vector is represented by the mean and variance σ n 2 n i = 0. In addition, the noise vectors of overlapping patch pairs exhibit correlation. However, non-overlapping patch pairs mostly appear in the generated patches. To simplify the overlapping problem, it is assumed that the noise vectors are completely independent and uncorrelated between all patches. The input noise sub-image can be studied as a dataset in Euclidean form. Then the variance of the noise image is projected onto a specific axis. In addition, the direction of the axis can be defined using the unit vector u in it. According to the previous assumption that the noise signal is considered to be uncorrelated, the variance of the projected image details can be represented as:

[0089]

[0090] where V(u T x i ) reflects the variance of the set of patches represented by x i in the direction of u, σ n is the standard deviation of the AWGN, and u min represents the minimum variance direction,

[0091] u min = arg min u V(u T y i ) = arg min u V(u T x i ) (11),

[0092] In addition, the maximum variance in equation (13) is calculated using PCA. Using the eigenvector associated with the smallest eigenvalue in the covariance matrix, the lowest variance direction can be defined, which can be represented as follows:

[0093]

[0094] where M is the original signal, i.e., the whole patch or sub-image. The variance of the target patch sorted along the minimum variance direction is mostly equal to the minimum eigenvalue of the covariance matrix. Therefore, according to the analysis of the minimum variance direction, the following equation can be derived:

[0095]

[0096] where ∑y represents the covariance matrix of the contaminated patch y, ∑z is the covariance matrix of the original patch z i (noise-free patch), γ min represents the minimum eigenvalue of the matrix.

[0097] The decomposition of the minimum eigenvalue of the covariance matrix of the contaminated sub-image is denoted by the equation. Therefore, the noise model can be easily estimated. However, this decomposition problem is an ill-posed problem because the minimum eigenvalue of the covariance matrix of the original sub-image λminz is not available. Despite this drawback, the noise level can be estimated by taking advantage of the features of digital images. That is, by taking advantage of the redundancy and repeated texture of natural digital images. In addition, the data of digital images only spans a low-dimensional framework. If the data of a patch {zi} ∈ RN×N spans a subspace with a dimension less than N×N, such sub-image or patch is called a low-rank patch.

[0098] Therefore, it is assumed that the minimum eigenvalue of the covariance matrix γ min (∑z) is zero. In theory, the additive Gaussian noise model has the same power in each direction, and all eigenvalues and their power spectrum have the same value. By taking advantage of the subspace spanned by the eigenvectors of the covariance matrix y with zero eigenvalues, the noise level can be estimated as follows:

[0099]

[0100] Embodiment Two

[0101] This embodiment illustrates the denoising effect of the SAR image denoising method based on sparse representation and blind estimation combination of embodiment one through a simulation experiment.

[0102] In this embodiment, Matlab is used for experimental simulation, and two SAR images are used as test images, as shown in Figure 3 The size of each image is 256×256. The denoising effect of the sparse dictionary DCT dictionary combined with the blind estimation algorithm is compared. The software used in the experiment is Matlab2018b, and the computer platform selected is Intel(R) Core(TM) i5-10400F CPU 2.90GHz with 8G memory.

[0103] In order to quantitatively analyze the image quality of each group of experiments in the present embodiment, four indicators, i.e. peak signal-to-noise ratio (PSNR), equivalent number of looks (ENL), edge preservation index (EPI) and structural similarity (SSIM), are used to comprehensively analyze the quality of the processed images from various angles. The following are the calculation expressions of the indicators:

[0104] (1) Peak signal-to-noise ratio

[0105]

[0106]

[0107] In the formula, X(i,j) is the original image, Y(i,j) is the image after denoising, and the larger the PSNR is, the better.

[0108] (2) Equivalent number of looks

[0109]

[0110] The equivalent number of looks represents the mean value of the region of interest, and the denominator is the variance.

[0111] (3) Edge preservation index

[0112] The edge preservation index represents the ability of the filter after processing to maintain the horizontal or vertical edges of the image, and the higher the E value is, the stronger the ability to maintain is

[0113]

[0114] In the formula, m is the number of image pixels; G R1 , G R2 are the gray values of the left and right or top and bottom adjacent pixels, respectively.

[0115] (4) Structural similarity

[0116] Structural similarity is an indicator for measuring the similarity between two images:

[0117]

[0118]

[0119]

[0120] In the formula, u X , u Y are the mean values of images X and Y, respectively, σ X , σ Y are the standard deviations of images X and Y, respectively, are the variances of images X and Y, respectively, and σXY Covariance of representative images X and Y. C1, C2 and C3 are constants, which are used to maintain stability to avoid zero denominator. Usually, C1=(K1*L) 2 , C2=(K2*L) 2 , C3=C2 / 2, generally K1=0.01, K2=0.03, L=255 (dynamic range of pixels, generally 255).

[0121] Finally, the SSIM index is:

[0122] SSIM(X,Y)=f(L(X,Y),C(X,Y),S(X,Y))=[L(X,Y)] α [C(X,Y)] β [S(X,Y)] γ (22),

[0123] Wherein, α, β, γ>0, used to adjust the weight of three modules, generally 1.

[0124] First, set the image block size B=8. Second, the measurement matrix uses an orthogonal Gaussian random matrix, and the reconstruction algorithm selects the orthogonal matching pursuit algorithm (OMP), and the DCT dictionary is used to jointly blind estimate the model for denoising. In order to increase the reliability and universality of the experiment, 2 images are respectively processed by 5 different denoising algorithms for comparison, and the SAR image denoising processing results are shown in Figure 4 and Figure 5 , wherein, (a) is a noise image, (b) is a Lee filter image, (c) is a Frost filter image, (d) is a BM3D filter image, (e) is a dictionary joint Bayes estimation filter image, and (f) is a filter image of the method of the application. The quality evaluation results of the SAR image after denoising are shown in Table 1.

[0125] Table 1 Quality evaluation results of SAR image after denoising

[0126]

[0127] By intuitive comparison of the experimental results, it can be seen that the denoising and image restoration effects of the method based on the DCT dictionary combined with the Bayes estimation and the method of the application are better than those of the traditional Lee filter and Frost filter algorithm, whether in the homogeneous area or in the edge feature area with rich texture. In the processing of scene 1, it can be found that the BM3D algorithm is too smooth, whether by human eye observation or as shown in Table 1, the equivalent number of views (ENL) in the homogeneous area is relatively high, but the edge preservation index (EPI) and the structural similarity (SSIM) are obviously lower. It can be found that the method of the application is obviously better than the Lee filter, the Frost filter and the BM3D filter. Since the sparse dictionary is derived from the image itself, it has better representation characteristics for the original image, so the Bayes estimation combined with the DCT dictionary also has good performance, but since the method of the application uses blind estimation, it can better estimate the image noise level, so the final denoising effect is better.

[0128] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that the item or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed. Without more limitations, the element defined by the statement "including one" does not exclude the presence of another identical element in the item or device including the element.

[0129] The above is a further detailed description of the application in combination with specific preferred embodiments, and the specific implementation of the application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the application belongs, without departing from the concept of the application, a number of simple deductions or substitutions can be made, which should be regarded as falling within the protection scope of the application.

Claims

1. A SAR image denoising method based on sparse representation and blind estimation combination, characterized in that, The method comprises the following steps: S1: selecting a dictionary and initializing dictionary parameters; S2: obtaining a SAR image to be denoised and estimating image noise of the SAR image to be denoised by using a blind estimation algorithm; S3: calculating sparse coding by using an OMP algorithm according to the image noise, iteratively updating the dictionary, and obtaining an updated dictionary and sparse coding; the S3 comprises: S31: performing block processing on the SAR image to be denoised, decomposing the SAR image to be denoised into a plurality of 8*8 matrix blocks, and sequentially converting each 8*8 matrix block into a 64*1 column vector in a column and row manner to generate a new matrix Blocks; S32: updating each element of each column in the matrix Blocks in the following manner: ， wherein denotes the matrix the element in the i-th row, j-th column, denotes the average value of the column in which the element (i,j) lies; S33: solving sparse coding of the matrix Blocks under the dictionary by using the OMP algorithm; S34: updating each element of each column in the matrix Blocks in the following manner according to the sparse coding: ， wherein, denotes the jthcolumn, denotes the average value of the jthcolumn, Coefs denotes the sparse encoding; S35: repeating the S32-S34 until a preset iteration stopping condition is reached, obtaining the updated dictionary and sparse coding, and the iteration stopping condition being that a preset iteration number is reached or a residual value is less than the image noise S4: performing sparse decomposition and reconstruction on the SAR image to be denoised by using the updated dictionary and sparse coding, and obtaining a denoised image.

2. The SAR image denoising method based on sparse representation and blind estimation combination of claim 1, wherein, In the step S1, the dictionary is a DCT dictionary, a training dictionary or an adaptive dictionary.

3. The SAR image denoising method based on sparse representation and blind estimation combination of claim 1, wherein, In the step S2, the image noise of the SAR image to be denoised is estimated in the following formula: ， wherein denotes the image noise level, denotes the smallest eigenvalue of the matrix, denotes the covariance matrix of the contaminated patch y.

4. The SAR image denoising method based on sparse representation and blind estimation combination of claim 1, wherein, The S4 comprises: S41: reconstructing each matrix block by using the updated dictionary and sparse coding; S42: performing weighted superposition on the reconstructed matrix blocks, and obtaining the denoised image.

Images