A method for SAR image speckle suppression based on machine learning

Abstract

The present invention discloses a method for suppressing coherent speckle in SAR images based on machine learning, which comprises the following steps: firstly, artificially adding noise to a noise-free optical image, and performing preliminary speckle suppression using a SAR-BM3D method to obtain a preliminary speckle suppression image set; then, performing a wavelet transform on the preliminary speckle suppression image, and performing wavelet reconstruction on the extracted high-frequency part to obtain a high-frequency feature image of the preliminary speckle suppression image; and finally, using the high-frequency feature image to guide training and test an ID-CNN to obtain a final speckle suppression image. The present invention improves the ability of a neural network to restore image structure texture information, and improves the efficiency of performing speckle suppression processing on SAR images.

Description

Technical Field

[0001] The invention belongs to the technical field of radar image processing, and in particular to a SAR image coherent speckle suppression method based on machine learning. Background Art

[0002] In SAR (synthetic aperture radar) image processing technology, the suppression of coherent speckle noise is an extremely important preprocessing step. Since SAR satellites belong to coherent imaging systems, the coherent speckle noise caused by the interference effect between the coherent echoes of each scattering point in the resolution unit is inevitable. In order to improve the accuracy of subsequent important applications such as target classification and detection, the coherent speckle noise of SAR images must be effectively suppressed first.

[0003] At present, the mainstream SAR speckle reduction methods include inter-domain adaptive filtering, non-local filtering, wavelet filtering and deep learning. With the continuous development of machine learning and the explosive increase in computing resources, deep convolutional neural networks have been widely used in fields such as SAR image speckle reduction due to their high precision and high efficiency. Although the SAR image speckle reduction method based on deep learning has been able to achieve good results on SAR images with different noise intensities, the excessive smoothing caused by it has lost many important image textures, greatly affecting the quality of SAR images. Therefore, how to improve the ability of convolutional neural networks to preserve image texture and efficiently remove speckle noise is an urgent problem to be solved in SAR image speckle reduction methods. Summary of the invention

[0004] The purpose of the present invention is to provide a SAR image coherent speckle suppression method based on machine learning, to improve the restoration of SAR image texture details, and to efficiently perform speckle suppression on SAR images.

[0005] The technical solution to achieve the purpose of the present invention is: a SAR image coherent speckle suppression method based on machine learning, comprising the following steps:

[0006] Step 1, artificially add noise to the noise-free optical grayscale image, and use the SAR-BM3D method to perform preliminary spot suppression to obtain a preliminary spot suppression image; wherein SAR-BM3D is a three-dimensional block matching filter method for SAR image denoising;

[0007] Step 2, performing discrete Haar wavelet transform on the preliminary anti-spot image, and performing wavelet reconstruction on the extracted high-frequency part to obtain a high-frequency feature image of the preliminary anti-spot image;

[0008] Step 3: Use the high-frequency feature image to guide the training and testing of ID-CNN to obtain the final despeckle image, where ID-CNN is a residual neural network used for SAR image denoising.

[0009] Compared with the prior art, the present invention has the following significant advantages: (1) firstly, artificial noise is added to the noise-free optical image, and the SAR-BM3D method is used to perform preliminary spot suppression to obtain a preliminary spot suppression image set; then, the preliminary spot suppression image is subjected to wavelet transform, and the extracted high-frequency part is subjected to wavelet reconstruction to obtain a high-frequency feature image of the preliminary spot suppression image, which can better preserve the image texture details compared with the simple deep neural network method; (2) the high-frequency feature image is used to guide the training and testing of the ID-CNN to obtain the final spot suppression image, which can obtain higher peak signal-to-noise ratio, structural similarity and other indicators compared with the traditional filtering method; (3) the ability of the neural network to restore the image structure texture information is improved, and the efficiency of spot suppression processing for SAR images is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flow chart of the SAR image spot suppression method based on machine learning of the present invention.

[0011] Figure 2 3 is a comparison chart of the spot suppression results with different weights in the embodiments of the present invention.

[0012] Figure 3 It is a speckle suppression result diagram for a SAR image in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] Combination Figure 1 The present invention provides a SAR image speckle suppression method based on machine learning, comprising the following steps:

[0014] Step 1, artificially add noise to the noise-free optical grayscale image, and use the SAR-BM3D method to perform preliminary spot suppression to obtain a preliminary spot suppression image; wherein SAR-BM3D is a three-dimensional block matching filter method for SAR image denoising;

[0015] Step 2, performing discrete Haar wavelet transform on the preliminary anti-spot image, and performing wavelet reconstruction on the extracted high-frequency part to obtain a high-frequency feature image of the preliminary anti-spot image;

[0016] Step 3: Use the high-frequency feature image to guide the training and testing of ID-CNN to obtain the final despeckle image, where ID-CNN is a residual neural network used for SAR image denoising.

[0017] As a specific implementation, the noise-free optical grayscale image described in step 1 is artificially noised, and the SAR-BM3D method is used to perform preliminary spot suppression to obtain a preliminary spot suppression image, as follows:

[0018] Step 1.1, using a multiplicative noise model to artificially add noise to the noise-free optical grayscale image to obtain a noisy simulated SAR image data set;

[0019] Step 1.2: The images in the noisy simulated SAR image dataset are randomly flipped horizontally or vertically and randomly scaled by multiples of 1, 0.9, 0.8, and 0.7, and then 40×40 image blocks are extracted using a 10-pixel step size for training;

[0020] Step 1.3, using the SAR-BM3D method to denoise the extracted noisy image block to obtain a preliminary despeckle image block;

[0021] Step 1.4: fuse the preliminary anti-spot image blocks to their original positions respectively, and then make a final estimate to obtain a preliminary anti-spot image.

[0022] As a specific implementation, the multiplicative noise model described in step 1.1 is as follows:

[0023] Y=XN

[0024] Where Y represents the simulated SAR image, X represents the noise-free optical grayscale image, N represents the multiplicative noise, and N obeys the Gamma distribution. The distribution function is as shown in formula (2):

[0025]

[0026] Where L represents the visibility coefficient.

[0027] As a specific implementation, the SAR-BM3D method described in step 1.3 is used to denoise the extracted noisy image block to obtain a preliminary despeckled image block, as follows:

[0028] Step 1.3.1, starting from the first K×K image block in the extracted noisy image block (K is a settable parameter, the larger K is, the more complicated the calculation is, in this example, K is set to 5), K is a set natural number, and a raster scanning method is adopted, each time translating 3 columns until the entire image is traversed, and finally N K×K reference blocks are selected, and the value of N is determined according to the size of the noisy image block and the raster scanning step size; and a sliding search with a step size of 1 in an area of ​​size n×n (for example, n is set to 8) centered on the reference block is performed for a difference less than a first threshold τ hard The comparison blocks of the same size as the reference block are integrated into a three-dimensional matrix. The first two dimensions are the two-dimensional pixel values ​​of the comparison blocks, and the third dimension is the number of comparison blocks. The process of finding the comparison blocks G(P) is as shown in formula (3):

[0029] G(P)={Q:d(P,Q)≤τ hard} (3)

[0030] In the formula, τ hard represents the first threshold, d(P,Q) represents the Euclidean distance between the comparison block P and the reference block Q;

[0031] Step 1.3.2, after forming N (the number of three-dimensional matrices is equal to the number of reference blocks) three-dimensional matrices, perform a two-dimensional DCT transformation on the two-dimensional pixel values ​​of the comparison block, and then perform an Hadamard transform on the third dimension of the transformed matrix; perform hard threshold processing on the three-dimensional matrix after the Hadamard transform, and set the pixels less than the second threshold τ γ The coefficients are set to 0; the three-dimensional matrix after hard threshold processing is subjected to one-dimensional inverse transformation and two-dimensional inverse transformation in the third dimension to obtain the processed image block. The process is as shown in formula (4):

[0032]

[0033] Where Q(P) hard represents the processed image block, represents the normalized 3D linear transformation, represents its inverse transform, and γ represents the hard threshold operation.

[0034] As a specific implementation, in step 2, the preliminary anti-spot image is subjected to discrete Haar wavelet transform, and the extracted high-frequency part is subjected to wavelet reconstruction to obtain a high-frequency feature image of the preliminary anti-spot image, as follows:

[0035] Step 2.1, using discrete Haar wavelet transform on the data after preliminary spot suppression, and obtaining 4 wavelet components for each image;

[0036] Step 2.2: Set all components except the high-frequency components to 0, and then use the discrete Haar wavelet inverse transform to fuse all components into a high-frequency feature image.

[0037] As a specific implementation, the data after preliminary spot suppression in step 2.1 is subjected to discrete Haar wavelet transform, and 4 wavelet components are obtained for each image, as follows:

[0038] Step 2.1.1: For the two-dimensional image y after preliminary spot suppression, 1 ; Y 2 ;...;Y n ), Y n Represents a single row matrix, using the discrete Haar wavelet as a low-pass or high-pass filter on Y n Transform, as shown in formula (5):

[0039] Y nL =Y n *L,Y nH =Y n *H (5)

[0040] Where * represents a two-dimensional convolution operation, L is a low-pass filter [1,1], H is a high-pass filter [-1,1], and Y nL Y n The low-frequency component, Y nH Y n The high frequency component of

[0041] Step 2.1.2, for y=(Y 1 ; Y 2 ;...;Y n ) in each single-row matrix Y n Both are subjected to discrete wavelet transform to obtain the low-frequency component y of y L =(Y 1L ; Y 2L ;...;Y nL ) and high frequency component y H =(Y 1H ; Y 2H ;...;Y nH );

[0042] Step 2.1.3: For the separated low-frequency component y L and high frequency component y H Further perform discrete Haar wavelet transform along the columns to obtain the low-frequency component y L The low frequency component y LL and high frequency component y LH , high frequency component y H The low frequency component y HL and high frequency component y HH .

[0043] Furthermore, in step 2.2, all components except the high-frequency component are set to 0, and then all components are fused into a high-frequency feature image using the discrete Haar wavelet inverse transform, as follows:

[0044] Remove the high frequency component y HH The other three components y LL ,y LH ,y HL Set all to 0, and then use the discrete Haar wavelet inverse transform to fuse all components into a high-frequency feature image.

[0045] As a specific implementation method, the high-frequency feature image is used to guide the training and testing of ID-CNN in step 3 to obtain the final anti-spot image, as follows:

[0046] Step 3.1: The noisy image x noisy With high-frequency feature image x HFPerform element-by-element superposition to obtain the training input x of ID-CNN noisy +αx HF , where α is the weight;

[0047] Step 3.2: The training set is predicted through residual network training to obtain a noise image of the same dimension, and the input and output logarithmic subtraction is used to obtain the final predicted image;

[0048] Step 3.3: Calculate the loss function and reversely derive to update the convolution kernel parameters to complete the training.

[0049] As a specific implementation, the training set in step 3.2 is predicted through residual network training to obtain a noise image of the same dimension, and the input and output logarithms are subtracted to obtain the final predicted image, as follows:

[0050] The training labels of ID-CNN are set to x clean +αx′ HF , where x clean For x noisy The corresponding original image without noise, x′ HF For x clean High frequency feature images.

[0051] As a specific implementation, the calculation of the loss function and the reverse derivation to update the convolution kernel parameters described in step 3.3 are as follows:

[0052] The loss function uses the mean square error (MSE). For the input image Y and the output image X, the elements are denoted as y i and x i , calculate the mean square error between the input image Y and the output image X And reverse the derivation to reduce the loss value.

[0053] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0054] Example

[0055] Combination Figure 1 The present invention provides a method for suppressing coherent speckle noise in SAR images based on machine learning, comprising the following steps:

[0056] Step 1: Artificially add noise to the noise-free optical grayscale image, and use the SAR-BM3D (three-dimensional block matching filter method for SAR image denoising) method to perform preliminary spot suppression to obtain a preliminary spot suppression image set, as follows:

[0057] Step 1.1: Use the multiplicative noise model to artificially add noise to the noise-free optical grayscale image dataset to obtain a noisy simulated SAR image dataset. The multiplicative noise model is as follows:

[0058] Y=XN

[0059] In the formula, Y represents the simulated SAR image, X represents the noise-free optical grayscale image, N represents the multiplicative noise, N obeys the Gamma distribution, and the distribution function is as follows:

[0060]

[0061] Where, L represents the visibility coefficient;

[0062] Step 1.2: Randomly flip the images in the noisy dataset horizontally or vertically and scale them by 1, 0.9, 0.8, and 0.7, and then extract 40×40 image blocks with a step size of 10 pixels for training;

[0063] Step 1.3: Use the SAR-BM3D method to denoise the extracted noisy image blocks to obtain preliminary despeckle image blocks, as follows:

[0064] Step 1.3.1, starting from the first K×K image block in the extracted noisy image block (K is a configurable parameter, the larger K is, the more complicated the calculation is, and K is set to 5 in this algorithm), use the raster scanning method, shift 3 columns each time until the entire image is traversed, and finally select N K×K reference blocks (the size of N is related to the size of the noisy image block and the raster scanning step length). And slide and search for the difference less than the threshold τ in an area of ​​n×n (n is set to 8 in the algorithm) centered on the reference block with a step length of 1. hard The comparison blocks of the same size as the reference block are integrated into a three-dimensional matrix. The first two dimensions are the two-dimensional pixel values ​​of the comparison blocks, and the third dimension is the number of comparison blocks. The process of finding comparison blocks G(P) is as follows:

[0065] G(P)={Q:d(P,Q)≤τ hard}

[0066] In the formula, τ hard represents the threshold, d(P,Q) represents the Euclidean distance between the comparison block and the reference block;

[0067] Step 1.3.2, after forming N (the number of three-dimensional matrices is equal to the number of reference blocks) three-dimensional matrices, perform a two-dimensional DCT (discrete cosine) transform (perform a two-dimensional DCT transform on the two-dimensional pixel values ​​of the comparison block), and then perform an Hadamard transform on the third dimension of the transformed matrix (the third dimension refers to the number of comparison blocks); perform hard threshold processing on the three-dimensional matrix after the Hadamard transform, and set the pixels less than the threshold value τ γ(The threshold here is the threshold of hard threshold processing, which is different from the previous step) The coefficient is set to 0; the three-dimensional matrix after hard threshold processing is subjected to one-dimensional inverse transformation and two-dimensional inverse transformation in the third dimension to obtain the processed image block. The process is as follows:

[0068]

[0069] Where Q(P) hard represents the processed image block, represents the normalized 3D linear transformation, represents its inverse transform, and γ represents the hard threshold operation.

[0070] At this point, each 2D image patch is an estimate of the denoised image;

[0071] Step 1.4: fuse the preliminary despeckle image blocks to their original positions respectively, and then make a final estimate to obtain a preliminary despeckle image.

[0072] Step 2: Perform discrete Haar wavelet transform on the preliminary anti-spot image, and perform wavelet reconstruction on the extracted high-frequency part to obtain the high-frequency information of the preliminary anti-spot image, as follows:

[0073] Step 2.1, using discrete Haar wavelet transform on the data after preliminary spot suppression, and obtaining 4 wavelet components for each image;

[0074] Step 2.1.1: For the two-dimensional image y after preliminary spot suppression, 1 ; Y 2 ;...;Y n ), Y n Represents a single row matrix, using the discrete Haar wavelet as a low-pass or high-pass filter on Y n Transform as follows:

[0075] Y nL =Y n *L,Y nH =Y n *H

[0076] Where * represents a two-dimensional convolution operation, L is a low-pass filter [1,1], H is a high-pass filter [-1,1], and Y nL Y n The low-frequency component, Y nH Y n high frequency components.

[0077] Step 2.1.2, for y=(Y 1 ; Y 2 ;...;Y n ) in each single-row matrix Y nBoth are subjected to discrete wavelet transform to obtain the low-frequency component y of y L =(Y 1L ; Y 2L ;...;Y nL ) and high frequency component y H =(Y 1H ; Y 2H ;...;Y nH );

[0078] Step 2.1.3: For the separated low-frequency component y L and high frequency component y H Further discrete wavelet transform is performed along the columns, and the low-frequency component y L The low-frequency component y is obtained by discrete transformation along the column LL and high frequency component y LH , high frequency component y H Get the low frequency component y HL and high frequency component y HH .

[0079] Step 2.2: Remove the high frequency component y HH The other three components y LL ,y LH ,y HL Set all to 0, and then use inverse wavelet transform to fuse all components into a high-frequency feature image.

[0080] Step 3: Use the high-frequency feature image to guide the training and testing of the convolutional neural network to obtain the final anti-spot image, as follows:

[0081] Step 3.1: First, the noisy image x noisy With high-frequency feature image x HF Perform element-by-element superposition to obtain the training input x of ID-CNN noisy +αx HF , where α is the weight;

[0082] Step 3.2: The training set is predicted through residual network training to obtain a noise image N of the same dimension, and the input and output logarithmic subtraction is used to obtain the final predicted image.

[0083] Furthermore, the training label of ID-CNN is set to x clean +αx′ HF , where x clean For x noisy The corresponding original image without noise, x′ HF For x clean High frequency feature images.

[0084] Step 3.3, calculate the loss function and reversely derive to update the convolution kernel parameters, and finally complete the training. Furthermore, the loss function in the algorithm uses MSE (mean squared error). For the input image Y and the output image X, the elements are denoted as y i and x i , calculate the mean square error between the input image Y and the output image X And reverse derivative to reduce the loss value.

[0085] Combination Figure 1 to Figure 3 This embodiment provides a SAR image coherent speckle suppression method based on machine learning. The training data used in the experiment is 400 randomly selected optical images from the NWPU-RESISC45 dataset, and the image size is 256 pixels × 256 pixels. The real SAR image used for testing in the experiment comes from ALOS PALSAR, and the image size is 1000 pixels × 1000 pixels.

[0086] Firstly, the SAR images are artificially denoised, and the data set is enhanced by randomly scaling the images and flipping them horizontally or vertically. Then, the enhanced data set is sent to the SAR-BM3D method for preliminary speckle suppression.

[0087] Then, Haar wavelet transform is used to extract the high-frequency feature image of the image after preliminary spot suppression, and the original noisy image and high-frequency features with different weights are used as the input of the ID-CNN network for training. Figure 2 It can be seen that adding high-frequency features to guide training can effectively improve the spot suppression effect, and the best effect is achieved when the weight α is 1.25.

[0088] Finally, when the weight α is 1.25, the real SAR image is tested. Figure 3 This is a diagram showing the effect of the present invention on SAR image despeckle. Figure 3 It can be seen that the present invention can effectively suppress the coherent speckle noise of SAR images.

Claims

1. A SAR image speckle suppression method based on machine learning, characterized in that: The following steps are involved: Step 1, artificially add noise to the noise-free optical grayscale image, use the SAR-BM3D method to perform preliminary spot suppression, and obtain a preliminary spot suppression image; Among them, SAR-BM3D is a three-dimensional block matching filtering method for SAR image denoising; Step 2, performing discrete Haar wavelet transform on the preliminary anti-spot image, and performing wavelet reconstruction on the extracted high-frequency part to obtain a high-frequency feature image of the preliminary anti-spot image; Step 3: Use the high-frequency feature image to guide the training and testing of ID-CNN to obtain the final despeckle image, where ID-CNN is a residual neural network used for SAR image denoising; In step 3, the high-frequency feature image is used to guide the training and testing of ID-CNN to obtain the final anti-spot image, as follows: Step 3.1: The noisy image x noisy With high-frequency feature image x HF Perform element-by-element superposition to obtain the training input x of ID-CNN noisy +αx HF , where α is the weight; Step 3.2: The training set is predicted through residual network training to obtain a noise image of the same dimension, and the input and output logarithmic subtraction is used to obtain the final predicted image; Step 3.3, calculate the loss function and reversely derive to update the convolution kernel parameters, and finally complete the training; The training set described in step 3.2 is predicted through residual network training to obtain a noise image of the same dimension, and the input and output logarithmic subtraction is used to obtain the final predicted image, as follows: The training labels of ID-CNN are set to x clean +αx′ HF , where x clean For x noisy The corresponding original image without noise, x′ HF For x clean High-frequency feature images; The calculation loss function described in step 3.3 and the reverse derivative update convolution kernel parameters are as follows: The loss function uses the mean square error (MSE). For the input image Y and the output image X, the elements are denoted as y i and x i , calculate the mean square error between the input image Y and the output image X And reverse the derivation to reduce the loss value.

2. The SAR image speckle suppression method based on machine learning according to claim 1, characterized in that: The noise-free optical grayscale image described in step 1 is artificially noised, and the SAR-BM3D method is used to perform preliminary spot suppression to obtain a preliminary spot suppression image, which is specifically as follows: Step 1.1, using a multiplicative noise model to artificially add noise to the noise-free optical grayscale image to obtain a noisy simulated SAR image data set; Step 1.2: The images in the noisy simulated SAR image dataset are randomly flipped horizontally or vertically and randomly scaled by multiples of 1, 0.9, 0.8, and 0.7, and then 40×40 image blocks are extracted using a 10-pixel step size for training; Step 1.3, using the SAR-BM3D method to denoise the extracted noisy image block to obtain a preliminary despeckle image block; Step 1.4: fuse the preliminary anti-spot image blocks to their original positions respectively, and then make a final estimate to obtain a preliminary anti-spot image.

3. The SAR image speckle suppression method based on machine learning according to claim 2, characterized in that: The multiplicative noise model described in step 1.1 is as follows: Y=XN (1) Where Y represents the simulated SAR image, X represents the noise-free optical grayscale image, N represents the multiplicative noise, and N obeys the Gamma distribution. The distribution function is as shown in formula (2): Where L represents the visibility coefficient.

4. The SAR image speckle suppression method based on machine learning according to claim 2, characterized in that: The SAR-BM3D method described in step 1.3 is used to denoise the extracted noisy image block to obtain a preliminary despeckle image block, as follows: Step 1.3.1, starting from the first K×K image block in the extracted noisy image block, K is a set natural number, using the raster scanning method, shifting 3 columns each time until the entire image is traversed, and finally selecting N K×K reference blocks, the value of N is determined according to the size of the noisy image block and the raster scanning step; and sliding search with a step size of 1 in an n×n area centered on the reference block for a difference less than the first threshold τ hard The comparison blocks of the same size as the reference block are integrated into a three-dimensional matrix. The first two dimensions are the two-dimensional pixel values ​​of the comparison blocks, and the third dimension is the number of comparison blocks. The process of finding the comparison blocks G(P) is as shown in formula (3): G(P)={Q:d(P,Q)≤τ hard } (3) In the formula, τ hard represents the first threshold, d(P,Q) represents the Euclidean distance between the comparison block P and the reference block Q; Step 1.3.2, after forming N three-dimensional matrices, perform a two-dimensional DCT transformation on the two-dimensional pixel values ​​of the comparison block, and then perform a Hadamard transform on the third dimension of the transformed matrix; Perform hard threshold processing on the three-dimensional matrix after Hadamard transformation, and set the value less than the second threshold τ γ The coefficients are set to 0; the three-dimensional matrix after hard threshold processing is subjected to one-dimensional inverse transformation and two-dimensional inverse transformation in the third dimension to obtain the processed image block. The process is as shown in formula (4): Where, Q(P) hard represents the processed image block, represents the normalized 3D linear transformation, represents its inverse transform, and γ represents the hard threshold operation.

5. The SAR image speckle suppression method based on machine learning according to claim 1, characterized in that: In step 2, the preliminary anti-spot image is subjected to discrete Haar wavelet transform, and the extracted high-frequency part is subjected to wavelet reconstruction to obtain a high-frequency feature image of the preliminary anti-spot image, as follows: Step 2.1, using discrete Haar wavelet transform on the data after preliminary spot suppression, and obtaining 4 wavelet components for each image; Step 2.2: Set all components except the high-frequency components to 0, and then use the discrete Haar wavelet inverse transform to fuse all components into a high-frequency feature image.

6. The SAR image speckle suppression method based on machine learning according to claim 5, characterized in that: The data after preliminary spot suppression described in step 2.1 is transformed using discrete Haar wavelet transform, and 4 wavelet components are obtained for each image, as follows: Step 2.1.1: for the two-dimensional image y after preliminary spot suppression, n ), Y n Represents a single row matrix, using the discrete Haar wavelet as a low-pass or high-pass filter on Y n Transform, as shown in formula (5): AND nL =And n *L,Y nH =And n *H (5) Where * represents a two-dimensional convolution operation, L is a low-pass filter [1,1], H is a high-pass filter [-1,1], and Y nL Y n The low-frequency component, Y nH Y n The high frequency component of Step 2.1.2, for y=(Y1; Y2; ...; Y n ) in each single-row matrix Y b Both are subjected to discrete wavelet transform to obtain the low-frequency component y of y L =(Y 1L ; Y 2L ;...;Y nL ) and high frequency component y H =(Y 1H ; Y 2H ;...;Y nH ); Step 2.1.3: For the separated low-frequency component y L and high frequency component y H Further perform discrete Haar wavelet transform along the columns to obtain the low-frequency component y L The low frequency component y LL and high frequency component y LH , high frequency component y H The low frequency component y HL and high frequency component y HH .

7. The SAR image speckle suppression method based on machine learning according to claim 6, characterized in that: In step 2.2, all components except the high-frequency component are set to 0, and then all components are fused into a high-frequency feature image using the discrete Haar wavelet inverse transform, as follows: Remove the high frequency component y HH The other three components y LL ,y LH ,y HL Set all to 0, and then use the discrete Haar wavelet inverse transform to fuse all components into a high-frequency feature image.

Images