An incremental time series SAR image denoising method based on sparse expression

Abstract

The present invention belongs to the technical field of SAR image denoising, and in particular, relates to an incremental time series SAR image denoising method based on sparse expression, which uses the time series images of the area where the target image is located, selects the images of the adjacent dates of the target image as a window; then calculates the super-image of the window, namely the temporary super-image; then, introduces the experience image, calculates the weighted average of the temporary super-image and the experience image by similarity weighted average, and uses this as the super-image corresponding to the target image; then, filters the super-image to obtain a low-noise super-image; then, calculates the ratio image of the target image and the low-noise super-image; filters the ratio image by a RuLoG filter to obtain a low-noise ratio image; finally, multiplies the low-noise ratio image and the low-noise super-image to obtain the denoised target image. The present invention can ensure the denoising effect of the target image while taking into account the processing efficiency.

Description

Technical Field

[0001] The invention belongs to the technical field of SAR image denoising, and in particular relates to an incremental time series SAR image denoising method based on sparse expression. Background Art

[0002] Synthetic Aperture Radar (SAR) images have been widely used in disaster monitoring, environmental monitoring, ocean monitoring, resource exploration, crop yield estimation, mapping and military, and are increasingly valued by countries around the world. However, SAR is a coherent imaging system, and its inherent speckle noise makes subsequent SAR image analysis and interpretation difficult. Therefore, SAR image noise reduction is usually a necessary step in the use of SAR images.

[0003] The main problems of SAR image denoising are the maintenance of spatial resolution, the restoration of edges and textures, and the maintenance of point targets. After decades of development, SAR image speckle suppression methods have been continuously proposed and improved. Compared with traditional filtering methods (Lee filter, Frost filter, etc.), spatial filters (such as SAR-BM3D or NL-SAR, etc.) have significantly better performance in terms of maintaining spatial resolution and restoring edges and textures. Among them, sparse expression, as an efficient signal processing method, has also been widely used in related fields. Under fully developed coherent noise conditions, the signal model of SAR images can be expressed as I = R × S, where I is the radar observation value, R is the true reflection coefficient of the radar, and S is the noise; taking the logarithm of I can obtain the additive model of SAR data Z = log(I) = log(R) + log(S) = R z +S z At this time, sparse expression can be used to z and S z The solution is obtained to recover the real reflection coefficient of the radar.

[0004] With the launch of a new generation of satellite radar systems (Cosmo-SkyMed, TerraSAR-X, ALOS-2, Sentinel series, etc.), the revisit time of SAR satellites is shorter, making it easier to obtain earth observation data for the same area at different time periods. How to use the spatial and temporal information of multi-phase SAR images to improve the performance of the filter is the idea followed by current time series denoising methods. The main time series denoising methods include spatiotemporal weighting methods (M-TSF and MSAR-BM3D), change detection-aware multi-spatiotemporal averaging (2SPPB), and the use of three-dimensional adaptive neighborhood filtering. The disadvantage of this type of method is that the spatiotemporal complexity of the calculation will increase as the time series is extended. To address this problem, some researchers have proposed a time series image denoising method based on ratio images (RABSAR), which is to calculate a "super image (x) based on the time series images. s )", and then compare the target image with it to get the ratio image And denoise this ratio image, and finally multiply the denoised ratio image and the "super image" to get the final denoised image, that is: RABASAR obtains a super-image through time-series arithmetic averaging. The advantages of this method are fast speed and simple implementation, but there are two disadvantages: 1) the obtained super-image may contain a large amount of spot information; 2) while the noise reduction effect increases with the increase of the time-series image length, the algorithm efficiency will drop significantly, and SAR time-series image is an "incremental" data source, that is, for a certain area, the SAR satellite will regularly obtain its remote sensing data.

[0005] In summary, researchers in this field have found that the existing technology has the following problems: on the one hand, denoising only a single image can easily lead to loss of spatial resolution of the resulting image; on the other hand, the efficiency of existing time-series image denoising methods will decrease rapidly with the increase in data volume and cannot adapt to incremental data sources.

[0006] Therefore, how to ensure the noise reduction effect of the target image while taking into account the processing efficiency has become a problem that needs to be solved urgently. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the present invention provides an incremental time series SAR image denoising method based on sparse expression, which can ensure the denoising effect of the target image while taking into account the processing efficiency.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] An incremental time series SAR image denoising method based on sparse expression comprises the following steps:

[0010] Step 1, obtaining time series data of SAR images of a target area, wherein the time series data includes the target image and a preset number of historical images adjacent to the target image;

[0011] Step 2: According to a window of preset size, the window where the target image is located in the data time series data is obtained and recorded as the current window; the temporary super image of the current window is calculated by iterative sparse and averaging; then the corresponding experience image is obtained according to the preset processing method, and the experience image and the temporary super image are matched for similar blocks to obtain the super image of the current window;

[0012] Step 3, performing noise reduction processing on the super image obtained in step 2 by using a SAR image filter to obtain a low-noise super image;

[0013] Step 4, dividing the low-noise super image and the target image to obtain a corresponding ratio image;

[0014] Step 5, denoising the ratio image by using a SAR image filter to obtain a low-noise ratio image;

[0015] Step 6, calculating the denoised target image using the low noise ratio image and the low noise super image obtained in step 3.

[0016] Preferably, in step 1, the time series data is also preprocessed so that the sizes of all images in the time series data are the same.

[0017] Preferably, in step 2, the calculating of the temporary super-image of the current window by iterative sparse and averaging comprises:

[0018] First, through the sparse expression F s Calculate the low-rank matrix L1 of the first image z1 in the current window, then add the obtained L1 to the second image in the current window to get z′2, and then use z′2 as F s The input is used to calculate the low-rank matrix L2 of z′2, and the above process is repeated until L m-1 Accumulate to the last image target image z in the current window m We can get z′ m , and through F s Calculate z′ m The low-rank matrix L m , and finally get the temporary hyperimage of the current window Where m is the size of the window.

[0019] Preferably, 3≤m≤10.

[0020] Preferably, in step 2, performing similar block matching on the experience image and the temporary super image to obtain the super image of the current window includes:

[0021] Get the set of blocks of the temporary hyperimage and a collection of blocks of empirical images Where n is the amount of data in the block set;

[0022] Calculate the similarity A=[α1,α2,…,α n ];

[0023] Then solve the hyperimage x of the current window s :

[0024] Where J is a matrix of all 1s, is a temporary super-image, For the experience image.

[0025] Preferably, in step 2, the process of acquiring the empirical image includes: determining whether the method has been used to reduce noise for the historical images adjacent to the target image in the current window; if not, randomly sampling a preset proportion of images from the time series data for iterative sparseness and then averaging to obtain the corresponding empirical image; if so, using the super image in the noise reduction process of the historical image adjacent to the target image as the corresponding empirical image.

[0026] Preferably, in step 2, the preset ratio is greater than or equal to 20%.

[0027] Preferably, in step 6, the calculation formula for calculating the target image after noise reduction is:

[0028] In the formula, is the target image after denoising, is a low noise ratio image, is a low-noise super-image.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. Unlike the prior art that directly performs denoising on the target image or the RABSAR method, the present invention uses the idea of ​​"incremental learning" to deal with the "incremental" characteristics of SAR time series images. Incremental learning is a new learning model, which refers to a learning model in which a learning system can continuously "learn" new knowledge from new samples and can preserve most of the knowledge that has been "learned" before. In actual perception data, the amount of data often increases gradually. Therefore, when faced with new data, the use of incremental learning methods can use existing knowledge to speed up the model's learning of new data, while also better optimizing the results.

[0031] Specifically, the present invention uses the time series images of the area where the target image is located, selects the images of the adjacent dates of the target image as windows; then calculates the super image of the window, namely the temporary super image; then, introduces the experience image, calculates the weighted average of the temporary super image and the experience image by similarity weighted average, and uses this as the super image corresponding to the target image; then, filters the super image to obtain a low-noise super image; then, calculates the ratio image of the target image and the low-noise super image; filters the ratio image by the RuLoG filter to obtain a low-noise ratio image; finally, multiplies the low-noise ratio image and the low-noise super image to obtain the denoised target image. The target image processed in this way retains more shape structure information, improves spatial resolution, and has the advantages of simple operation and low calculation amount.

[0032] Compared with directly reducing the noise of the target image, the present invention can ensure the noise reduction effect; compared with the RABSAR method, the present method can ensure the noise reduction effect while taking into account the processing efficiency.

[0033] In summary, the present invention can ensure the noise reduction effect of the target image while taking into account the processing efficiency.

[0034] 2. When calculating the super-image of the current window, the present invention first calculates the low-rank matrix of the first image in the window, then accumulates the low-rank matrix to the next image and uses the obtained image as the new current image, thereby iteratively accumulating and calculating the low-rank matrices of all images in the window, and then taking the arithmetic mean of the low-rank matrix of the last image as a temporary super-image; then, the empirical image is introduced, and the weighted average of the temporary super-image and the empirical image is calculated by similarity weighted averaging, and this is used as the super-image corresponding to the target image. In this way, the spatial and temporal information of multi-phase SAR images can be fully utilized to improve the performance of the filter.

[0035] 3. When introducing the empirical image, the present invention will select the corresponding empirical image acquisition method according to whether the method has just been used. If the method has just been used, that is, the historical image adjacent to the target image in the current window has not been denoised using the method, then a preset proportion of images are randomly sampled from the time series data for iterative sparse processing and then averaged to obtain the corresponding empirical image; if it has not just been used, the super image in the denoising process of the historical image adjacent to the target image is used as the corresponding empirical image. In this way, the overall efficiency of the method can be further guaranteed while ensuring the effectiveness of the empirical image.

[0036] 4. In this method, the window size is greater than or equal to 3 and less than or equal to 10, which can take into account both the efficiency and effect of iterative sparsity. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0038] Figure 1 is a flow chart in an embodiment;

[0039] Figure 2 A schematic diagram of the process in the embodiment;

[0040] Figure 3 is a schematic diagram of a SAR image time series data set in an embodiment;

[0041] Figure 4 It is a schematic diagram of the effects of each step in the embodiment. DETAILED DESCRIPTION

[0042] The following is a further detailed description through specific implementation methods:

[0043] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation method, structure, features and effects of an incremental time series SAR image denoising method based on sparse expression proposed by the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0044] Unless otherwise defined, all techniques and terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0045] Example:

[0046] like Figure 1 , Figure 2 As shown, this embodiment discloses an incremental time series SAR image denoising method based on sparse expression. It should be noted that: Figure 2 The subscript of the parameter in (such as x t-m ) is only for illustration, and the specific contents in the embodiments are for the convenience of those skilled in the art to implement and Figure 2 The subscripts of some parameters in the method are not completely unified, but the principles and processes are exactly the same and will not affect the understanding of those skilled in the art.

[0047] The method comprises the following steps:

[0048] Step S1, obtaining time series data of SAR images of a target area, wherein the time series data includes the target image and a preset number of historical images adjacent to the target image; and preprocessing the time series data so that the sizes of all images in the time series data are the same.

[0049] In this specific implementation, image data of different dates in the area where the target image is located are prepared as time series data, and preprocessed (data clipping, thermal noise removal, orbit correction, removal of spatial correlation, etc.), and then all images are aligned in coordinates, such as Figure 3 For ease of processing, all images can also be saved as a three-dimensional matrix. The dataset in this embodiment has a dimension of [500×500×35], i.e., 35 images of 500×500. The i-th image is x i , then the historical image of the embodiment can be expressed as [x1,x2,…,x 35 ]. The target image is denoted as x 36 , also a 500×500 image.

[0050] Step S2, according to a window of preset size, obtain the window where the target image is located in the data time series data, and record it as the current window; calculate the temporary super image of the current window by iterative sparse and averaging; then obtain the corresponding empirical image according to the preset processing method, and perform similar block matching between the empirical image and the temporary super image to obtain the super image of the current window.

[0051] The calculation of the temporary super-image of the current window by iterative sparse and averaging includes:

[0052] First, through the sparse expression F s Calculate the low-rank matrix L1 of the first image z1 in the current window, then add the obtained L1 to the second image in the current window to get z′2, and then use z′2 as F s The input is used to calculate the low-rank matrix L2 of z′2, and the above process is repeated until L m-1 Accumulate to the last image target image z in the current window m We can get z′ m , and through F s Calculate z′ m The low-rank matrix L m , and finally get the temporary hyperimage of the current window Where m is the size of the window. The temporary hyperimage is Figure 4 (d) as shown.

[0053] In this embodiment, the window size is 5, and the window time series slice is D=[x 32 ,x 33 ,…,x36 ]. For the convenience of description, let the i-th image in the window be z i , then the images in the window can be renumbered as D=[z1,z2,…,z5], where the target image is z5. In this embodiment, RobustPCA is used to calculate the low-rank matrix of each image, and other sparse expression methods can also be selected in other embodiments.

[0054] The process of acquiring the experience image includes: determining whether the method has been used to reduce noise for the historical image adjacent to the target image in the current window; if not, randomly sampling a preset ratio of images from the time series data for iterative sparse and averaging to obtain the corresponding experience image (the same as the solution process of the temporary super image); if yes, taking the super image in the noise reduction process of the historical image adjacent to the target image as the corresponding experience image. The preset ratio is greater than or equal to 20%, and is 25% in this embodiment.

[0055] The experience image is matched with the temporary super-image by similar blocks, and the super-image of the current window is obtained including:

[0056] Get the set of blocks of the temporary hyperimage and a collection of blocks of empirical images Calculate the similarity A=[α1,α2,…,α n ]; then solve the hyperimage x of the current window s :

[0057] Where J is a 500×500 all-1 matrix, that is, all its elements are 1. is a temporary super-image, For the experience image.

[0058] In the specific implementation, the block size is 5×5, and the amount of data in the block set is 10000. Then the block set of the temporary super image is The block set of the empirical image is The similarity of the corresponding blocks is A = [α1, α2, ..., α 10000 ]. For the convenience of description, each pair of similar blocks and their corresponding similarities are represented by a triplet, namely At this point, the final hyperimage of the current window is Figure 4 (e) as shown.

[0059] Step S3, performing noise reduction processing on the super image obtained in step 2 by using a SAR image filter to obtain a low-noise super image In the specific implementation, the super image obtained by S2 is input into the MuLoG-BM3D filter, and the output result is the low-noise super image Reference Figure 4 (f) In other embodiments, the MuLoG-BM3D filter may also be replaced by other filters.

[0060] Step S4: Divide the low-noise super-image by the target image to obtain the corresponding ratio image Ratio Image R 36 like Figure 4 (g) as shown.

[0061] Step S5, denoising the ratio image by using a SAR image filter to obtain a low-noise ratio image. In this embodiment, the ratio image R is denoised by using a RuLoG (Ratio extension of MuLoG) filter. 36 Perform noise reduction to obtain a low noise ratio image Reference Figure 4 (h) In other embodiments, the RuLoG filter may also be replaced by other SAR image filters.

[0062] Step S6, using the low noise ratio image and the low noise super image obtained in step 3 Target image after denoising like Figure 4 (b) as shown.

[0063] Unlike the prior art that directly performs denoising on the target image or the RABSAR method, the present invention uses the idea of ​​"incremental learning" to deal with the "incremental" characteristics of SAR time series images. Incremental learning is a new learning model, which refers to a learning model in which a learning system can continuously "learn" new knowledge from new samples and can preserve most of the knowledge that has been "learned" before. In actual perception data, the amount of data often increases gradually. Therefore, when faced with new data, the use of incremental learning methods can use existing knowledge to speed up the model's learning speed for new data, and at the same time better optimize the results obtained. Specifically, the present invention uses the time series images of the area where the target image is located, selects the images of the adjacent dates of the target image as windows; then calculates the super-image of the window, namely the temporary super-image; then, introduces the empirical image, calculates the weighted average of the temporary super-image and the empirical image by similarity weighted average, and uses this as the super-image corresponding to the target image; then, filters the super-image to obtain a low-noise super-image; then, calculates the ratio image of the target image and the low-noise super-image; filters the ratio image by the RuLoG filter to obtain a low-noise ratio image; finally, multiplies the low-noise ratio image and the low-noise super-image to obtain the denoised target image. The target image processed in this way retains more shape structure information, improves the spatial resolution, and has the advantages of simple operation and less calculation. Compared with directly denoising the target image, the present invention can ensure the denoising effect; compared with the RABSAR method, the present method can take into account the processing efficiency while ensuring the denoising effect.

[0064] In addition, when calculating the super-image of the current window, the present invention first calculates the low-rank matrix of the first image in the window, then accumulates the low-rank matrix to the next image and uses the obtained image as the new current image, thereby iteratively accumulating and calculating the low-rank matrices of all images in the window, and then the arithmetic mean of the low-rank matrix of the last image is used as a temporary super-image; then, the empirical image is introduced, and the weighted average of the temporary super-image and the empirical image is calculated by similarity weighted average, and this is used as the super-image corresponding to the target image. In this way, the spatial and temporal information of the multi-phase SAR image can be fully utilized to improve the performance of the filter. In addition, when introducing the empirical image, the present invention will select the corresponding empirical image acquisition method according to whether the method has just been used. If the method has just been used, that is, the historical image adjacent to the target image in the current window has not been denoised using the method, a preset proportion of images are randomly sampled from the time series data for iterative sparse and then averaged to obtain the corresponding empirical image; if it has not just been used, the super-image in the denoising process of the historical image adjacent to the target image is used as the corresponding empirical image. In this way, the validity of the empirical image can be guaranteed while further ensuring the overall efficiency of the method.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. An incremental time series SAR image denoising method based on sparse expression, characterized in that: The following steps are involved: Step 1, obtaining time series data of SAR images of a target area, wherein the time series data includes the target image and a preset number of historical images adjacent to the target image; Step 2: According to a window of preset size, the window where the target image is located in the data time series data is obtained and recorded as the current window; the temporary super image of the current window is calculated by iterative sparse and averaging; then the corresponding experience image is obtained according to the preset processing method, and the experience image and the temporary super image are matched for similar blocks to obtain the super image of the current window; Step 3, performing noise reduction processing on the super image obtained in step 2 by using a SAR image filter to obtain a low-noise super image; Step 4, dividing the low-noise super image and the target image to obtain a corresponding ratio image; Step 5, denoising the ratio image by using a SAR image filter to obtain a low-noise ratio image; Step 6, calculating the denoised target image by using the low noise ratio image and the low noise super image obtained in step 3; In step 2, the calculation of the temporary super-image of the current window by iterative sparse and averaging includes: First, through the sparse expression F s Calculate the low-rank matrix L1 of the first image z1 in the current window, and then add the obtained L1 to the second image in the current window to get z2 ′ , and then z2 ′ As F s The input calculation z2 ′ The low-rank matrix L2 is repeated until L m-1 Accumulate to the last image target image z in the current window m Get z ′ m , and through F s Calculate z ′ m The low-rank matrix L m , and finally get the temporary hyperimage of the current window Where m is the size of the window.

2. The incremental time series SAR image denoising method based on sparse expression according to claim 1, characterized in that: In step 1, the time series data is also preprocessed so that the sizes of all images in the time series data are the same.

3. The incremental time series SAR image denoising method based on sparse expression according to claim 1, characterized in that: 3≤m≤10。 4. The incremental time series SAR image denoising method based on sparse expression according to claim 1, characterized in that: In step 2, the similarity block matching between the experience image and the temporary super image to obtain the super image of the current window includes: Get the set of blocks of the temporary hyperimage and a collection of blocks of empirical images Where n is the amount of data in the block set; Calculate the similarity A=[α1,α2,…,α n ]; Then solve the hyperimage x of the current window s : Where J is a matrix of all 1s. is a temporary super-image, For the experience image.

5. The incremental time series SAR image denoising method based on sparse expression according to claim 4, characterized in that: In step 2, the process of acquiring the empirical image includes: determining whether the method has been used to reduce the noise of the historical images adjacent to the target image in the current window; if not, randomly sampling a preset proportion of images from the time series data for iterative sparseness and then averaging to obtain the corresponding empirical image; if so, using the super image in the noise reduction process of the historical image adjacent to the target image as the corresponding empirical image.

6. The incremental time series SAR image denoising method based on sparse expression according to claim 5, characterized in that: In step 2, the preset ratio is greater than or equal to 20%.

7. The incremental time series SAR image denoising method based on sparse expression according to claim 1, characterized in that: In step 6, the calculation formula for calculating the denoised target image is: In the formula, is the target image after denoising, is a low noise ratio image, is a low-noise super-image.

Images