Remote sensing image cloud removal method
By applying the connectivity domain technology in remote sensing images, the problem of insufficient accuracy and generalization capabilities of cloud detection and removal methods is solved, high-precision image quality evaluation and automated reference image selection are achieved, and the efficiency of de-cloud processing is improved and stability.
Patent Information
- Application Number
- CN202111200522.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-14
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-10-14
AI Technical Summary
The cloud detection and removal methods in existing remote sensing images have problems such as low accuracy and weak generalization capabilities. Especially when dealing with thick clouds, it is difficult to effectively extract edge data of cloud and cloud shadows, and it is impossible to reasonably evaluate the same cloud proportion. Image quality of different cloud shapes and distributions leads to inaccurate image quality evaluation and reliance on manual intervention for reference image selection.
In de-cloud processing, the connection domain technology is used to analyze the shape and distribution of the cloud area, evaluate the image quality, and automatically select and filter reference images to achieve high-precision quality judgment and de-cloud processing of the image.
It improves the accuracy of image quality evaluation and the stability of automated selection, solves the problem of image quality evaluation of different cloud shapes and distributions, and improves the efficiency and effectiveness of cloud removal.
Smart Images

Figure CN114092343B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of remote sensing images, and in particular to a remote sensing image cloud removal method. Background Art
[0002] In remote sensing images, clouds are a common element that not only causes spectrum distortion and seriously affects the quality of remote sensing data, but also has an adverse effect on further analysis. As the main source of invalid pixels in remote sensing images, clouds should be detected and excluded as much as possible before further processing of remote sensing images. It is particularly important to eliminate the occlusion of clouds without losing the original effective information as much as possible for the restoration of remote sensing images. Cloud detection is regarded as the key to subsequent identification, classification and analysis of remote sensing images, and is an important basis for remote sensing image restoration. After obtaining the cloud location area information through cloud detection, the occluded area is restored according to the information of neighboring objects or reference images, which can improve the visualization effect of the image and increase the usability of the image; for the reconstruction of remote sensing object information, a large amount of satellite resources can be saved and repeated shooting can be avoided. Although cloud detection and cloud removal / cloud restoration have been widely studied in the field of remote sensing images, there are still problems of low accuracy and weak generalization ability. With the rise of deep learning, people have tried to apply this method to remote sensing images. Although some research results have been achieved, there are still many key technologies that need to be broken through. For example, when using deep learning for cloud detection, it is difficult to effectively extract the edge data of clouds and cloud shadows when extracting clouds and cloud shadows.
[0003] In cloud removal / cloud restoration, clouds can be divided into thin clouds and thick clouds according to their types. There is a small amount of surface reflection information that can be used under thin clouds, so the processing of thin clouds is relatively simple and the method is relatively mature. Thin cloud removal is similar to dehazing of natural images, and it can also be considered to be implemented based on the spectral differences of different bands. In common thick cloud images, thick cloud areas and cloud shadow areas often appear in pairs. However, the surface occlusion information under thick clouds is completely invalid, so the thin cloud removal method cannot be applied to remove thick clouds. At present, thick cloud removal methods generally require multi-temporal cloud-free images as reference images to restore the spectral information blocked by clouds.
[0004] In traditional cloud processing methods, the threshold method is usually used to judge the image quality and then decide whether to process the image. This method can effectively evaluate the size of the cloud ratio in the image, but it cannot evaluate the quality of images containing clouds of different shapes and distributions at the same cloud ratio, such as an image of a large area of thick clouds with a cloud ratio of 20% and an image of multiple discrete point clouds with a cloud ratio of 20%. This image of a large area of thick clouds has a low overall cloud coverage, especially after expansion processing, the image quality is significantly better than that of images with discrete point clouds, and should be used first in cloud removal processing. There is no point in processing images with multiple discrete point clouds. At the same time, the existing multi-temporal image thick cloud removal methods lack effective and stable standards for selecting reference images when selecting reference images, and require a lot of manual intervention.
[0005] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention
[0006] The present invention proposes a remote sensing image declouding method that can solve the above-mentioned problems of the prior art. The method of the present invention uses the connected domain technology to achieve high-precision judgment of the quality of useful images in declouding, achieves quality assessment of images based on cloud shape and distribution, and achieves automatic selection and screening of reference images that do not rely on human intervention.
[0007] To achieve the above technical objectives, the technical solution of the present invention is implemented as follows:
[0008] S1 searches for images intersecting with the image I to be processed at different times according to the spatial position of the image I to be processed, and obtains a first image;
[0009] S2 selects the first image to obtain the second image according to the spatial position relationship between the image to be processed I and the cloud-free area of the first image;
[0010] S3 uses the connected domain technology to analyze the cloud area of the second image, evaluate the image quality, and filter the second image to obtain the third image;
[0011] S4 uses the third image to perform cloud removal processing on the image to be processed I to obtain a cloud-removed image.
[0012] Optionally, the step of filtering the first image to obtain the second image according to the spatial position relationship between the image to be processed I and the cloud-free area of the first image includes:
[0013] Perform cloud region detection on the image to be processed I to obtain a cloud region mask mask1 of the image to be processed I;
[0014] Perform cloud area detection on the first image to obtain a cloud-free area mask mask3;
[0015] Screen the images where the cloud-free area in the first image intersects with the image to be processed I, determine the spatial position relationship, and obtain the effective intersecting image according to the spatial position relationship. The cloud-free area in the first image is represented by mask mask3;
[0016] The images whose cloud-free areas in the valid intersecting images intersect with the cloud area of the image to be processed I are screened to determine the spatial position relationship, and the second image is obtained according to the spatial position relationship. The cloud area of the image to be processed I is represented by mask mask1.
[0017] Optionally, the using of the connected domain technology to analyze the cloud area of the second image, evaluate the image quality, screen the second image, and obtain a third image, wherein the third image constitutes a screened image set S3, including:
[0018] (1) performing cloud region detection on the first image to obtain a cloud region mask mask2, performing binarization processing on the second image, setting cloud and cloud shadow pixels as foreground pixels with a pixel value of 1 according to the cloud region mask mask2 of the second image, and setting cloud-free region pixels as background pixels with a pixel value of 0 according to the cloud-free region mask mask3;
[0019] (2) Scanning pixels of the second image until the scanned pixel value is 1, taking the pixel as a seed, assigning a new label, and pushing all foreground pixels adjacent to the seed into a stack;
[0020] (3) Pop the top pixel of the stack, assign the same label to all pixels in the stack, and push all foreground pixels adjacent to the new top pixel of the stack into the stack;
[0021] (4) Repeat the above step (3) until the stack is empty;
[0022] (5) Repeat steps (2) to (4) until the scanning is completed and all connected domains of the cloud and cloud shadow are obtained. The number of connected domains is obtained based on the number of connected domain labels.
[0023] (6) determining whether the second image is qualified according to a connected domain parameter and the cloud amount of the second image, wherein the connected domain parameter includes the number of connected domains and the size of connected domains; if the connected domain parameter is less than a preset connected domain parameter threshold and the image cloud amount is less than a preset cloud amount threshold, determining that the second image is qualified, and obtaining a third image;
[0024] The cloud amount of the second image is the ratio of the cloud area to the entire image.
[0025] Optionally, using the third image to perform cloud removal processing on the image to be processed I to obtain a cloud-removed image includes:
[0026] Prioritizing the third image to obtain a prioritized third image;
[0027] Using the sorted third image in sequence to perform cloud removal processing on the image to be processed I;
[0028] When the cloud area of the image to be processed I is completely filled, the processing is stopped and a cloud-free image is obtained.
[0029] Optionally, the prioritizing the third images to obtain the prioritized third images includes:
[0030] Sort the third image according to the time phase to obtain a first priority sorted image;
[0031] The first priority ranked images are further ranked according to the size of the overlapped area between the cloud-free area of the first priority ranked image and the cloud area of the image to be processed I, with the larger overlapped area being prioritized, to obtain the second priority ranked images;
[0032] The second priority ranked images are sorted according to their image resolutions, and are sorted in the order of image resolutions being equal to, higher than, and lower than the resolution of the image to be processed I, to obtain a sorted third image.
[0033] Optionally, the step of sorting the third image relative to the time to obtain the first priority sorted image includes:
[0034] Select the images of the third image that are in the same period of the same year, the same period of the same year, and the same period of the previous year as the image I to be processed according to the time phase;
[0035] Prioritize the images in the third image that are of the same year as the image to be processed I, then prioritize the images in the third image that are of the same period as the same period last year as the image to be processed I, and finally prioritize the images in the third image that are of the same year before the image to be processed I, to obtain images ranked with the first priority.
[0036] The beneficial effects of the present invention are as follows: the present invention provides a remote sensing image declouding method. Through the connected domain technology, high-precision judgment of the quality of useful images is achieved in declouding, quality assessment of images based on cloud shape and distribution is achieved, and automatic selection and screening of reference images that do not rely on human intervention is achieved. The efficiency and stability of screening useful images and reference images are improved, and the problem that the quality of images with different cloud shapes and distributions at the same cloud proportion cannot be reasonably evaluated in traditional declouding methods is solved. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0038] Figure 1 A flowchart of a remote sensing image cloud removal method provided by the present invention;
[0039] Figure 2 It is a flow chart of sorting the third image according to the time relative to one embodiment of the present invention;
[0040] Figure 3 (a)-(b) shows the comparison of multiple point clouds and a large area of thick cloud image under the same cloud area ratio;
[0041] Figure 4 This is a flowchart of a deep learning network model processing based on Deeplab V3+ according to an embodiment of the present invention;
[0042] Figure 5 The remote sensing image to be processed after preprocessing according to an embodiment of the present invention;
[0043] Figure 6 This is an image after cloud detection according to an embodiment of the present invention;
[0044] Figure 7 The cloud and cloud shadow images extracted by an embodiment of the present invention;
[0045] Figure 8 1 and 2 are cloud-removed images and padded areas obtained in one embodiment of the present invention.
[0046] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0047] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.
[0048] Figure 1 is a flow chart of the remote sensing image cloud removal method of the present invention, such as Figure 1 As shown, the method comprises the following steps:
[0049] A remote sensing image cloud removal method, the method comprising the following steps:
[0050] S1 searches for images that intersect with the image I to be processed at different times according to the spatial position of the image I to be processed, and obtains a first image.
[0051] The remote sensing image to be processed is selected by the user. The user can select the image by drawing a rectangular frame or by geographical location according to the application scenario. The image status determines whether to perform preprocessing and the preprocessing process. The preprocessing process in this embodiment includes orthorectification, geometric correction, radiation correction and atmospheric correction. The preprocessing process can be adjusted according to the actual application requirements, using only part of the preprocessing process or all of the processes. Figure 5 The figure shows the remote sensing image to be processed after preprocessing according to an embodiment of the present invention. Figure 6-8 The image is obtained by cutting out the rectangular frame at the lower left corner of the image.
[0052] S2 selects the first image according to the spatial position relationship between the image to be processed I and the cloud-free area of the first image to obtain the second image.
[0053] In this embodiment, cloud area detection is performed on the image to be processed I to obtain a cloud area mask mask1 of the image to be processed I; cloud area detection is performed on the first image to obtain a cloud area mask mask2 and a cloud-free area mask mask3; the image in which the cloud-free area in the first image intersects with the image to be processed I is screened to determine the spatial position relationship, and a valid intersecting image is obtained based on the spatial position relationship, and the cloud-free area of the first image is represented by mask mask3; the image in which the cloud-free area in the valid intersecting image intersects with the cloud area of the image to be processed I is screened to determine the spatial position relationship, and a second image is obtained based on the spatial position relationship, and the cloud area of the image to be processed I is represented by mask mask1.
[0054] The cloud area detection method is: constructing a cloud detection deep learning network model and training it, and inputting the image into the trained cloud detection deep learning network model to obtain the detection result. Figure 6 shown.
[0055] In this embodiment, DeepLab V3+ is used to build a cloud detection deep learning network model. Figure 4 As shown in Figure 2, DeepLab V3 is used as the encoder. The cross entropy loss is used as the loss function to fit the semantic classification score map and the true label, and the stochastic gradient descent method is used as the optimization method to back-propagate the entire semantic segmentation network.
[0056] In this embodiment, the deep learning network model structure consists of two parts: an encoder and a decoder. Figure 4As shown in the figure, the data first enters the encoding or feature extraction stage. DeepLab V3 is used as the encoder, which is composed of Atrous Conv and 1*1Conv in sequence. Conv represents the convolution layer, Atrous Conv is the dilated convolution layer, and the dilated convolution layer is composed of 1*1Conv, 3*3Conv, 3*3Conv, 3*3Conv, and Image Pooling in sequence. The dilation factors of the three 3*3 dilated convolutions in the dilated convolution layer are 6, 12, and 18 respectively. Image Pooling is the image pooling layer. The output of each encoder part is retained as the underlying feature, and the output underlying feature is downsampled by 2 times. The decoder is composed of 1*1Conv, Upsample, Concat, 3*3Conv, and Upsample in sequence. Upsample is the upsampling layer with an upsampling ratio of 4 times, and Concat is the cascade layer. Before fusing the low-level information, a 1*1 convolution is performed first.
[0057] S3 uses the connected domain technology to analyze the cloud area of the second image, evaluate the image quality, filter the second image, and obtain the third image:
[0058] (1) Perform cloud region detection on the first image to obtain a cloud region mask mask2, perform binarization on the second image, set cloud and cloud shadow pixels as foreground pixels according to the cloud region mask mask2 of the second image, with a pixel value of 1, and set cloud-free region pixels as background pixels according to the cloud-free region mask mask3, with a pixel value of 0;
[0059] (2) Scan the pixels of the second image until the value of the current pixel is 1, use the pixel as a seed, assign a new label, and push all foreground pixels adjacent to the seed into a stack;
[0060] (3) Pop the top pixel of the stack, assign the same label to all pixels in the stack, and push all foreground pixels adjacent to the new top pixel of the stack into the stack;
[0061] (4) Repeat the above step (3) until the stack is empty;
[0062] (5) Repeat steps (2) to (4) until the scanning is completed and all connected domains of the cloud and cloud shadow are obtained. The number of connected domains is obtained based on the number of connected domain labels.
[0063] (6) determining whether the second image is qualified according to a connected domain parameter and the cloud cover of the second image, wherein the connected domain parameter includes the number of connected domains and the size of the connected domains, and presetting a connected domain parameter threshold and a cloud cover threshold. If the connected domain parameter is less than the connected domain parameter threshold and the image cloud cover is less than the preset threshold, the second image is determined to be qualified, and a third image is obtained according to the determination result.
[0064] As shown in Figure 3, a comparison of multiple point clouds 3(a) and a large area of thick cloud 3(b) is shown under the same cloud area ratio. Figure 3(a) and 3(b) These are sample images of different areas of the same size captured at the same resolution from the same remote sensing image. Figure 3(a) and 3(b) The threshold method used in the traditional method cannot further evaluate the image quality of the two images in Figure 3. The connected domain technology used in the present invention can evaluate Figure 3 (a) as an image with better quality than Figure 3 (b).
[0065] S4 uses the third image to perform cloud removal processing on the image to be processed I to obtain a cloud-removed image.
[0066] Among them, the method for de-clouding is: prioritize the third image to obtain the sorted third image; use the sorted third image in sequence to perform de-clouding on the image to be processed I; when the cloud area of the image to be processed I is completely filled, stop processing to obtain a de-clouded image.
[0067] Among them, the priority sorting method is: sort the third image according to the time relative to obtain the first priority sorted image; further sort the first priority sorted image according to the size of the overlapping area between the cloud-free area of the first priority sorted image and the cloud area of the image I to be processed, and prioritize the image with the larger overlapping area to obtain the second priority sorted image; sort the second priority sorted image according to its image resolution, and sort it in the order of its image resolution being equal to, higher than, and lower than the resolution of the image I to be processed, to obtain the sorted third image.
[0068] Among them, the priority of sorting the third image according to the time is as follows: Figure 2 As shown: images of the same year and the same period as the image to be processed I are prioritized, followed by images of the same period last year as the image to be processed I, and finally images of the same year before the image to be processed I are prioritized.
[0069] Among them, after the priority sorting is completed and before the cloud removal processing is performed, if the resolution of the third image is inconsistent with the resolution of the image to be processed I, the third image is downsampled or upsampled to keep the resolution consistent.
[0070] The method of downsampling from high to low is preferred. Upsampling can be used optionally under special requirements. Upsampling uses the traditional bilinear interpolation algorithm. Four pixel points (x, y), (x, y+1), (x+1, y), (x+1, y+1) constitute an interpolation area. f(x, y) is the pixel value at (x, y) in the original image. Set the floating point coordinates obtained by inverse transformation to (x+a, y+b), where a and b are floating point numbers in the interval [0, 1]. The target pixel value f(x+a, y+b) can be constructed by the interpolation area formed by the above four pixel points:
[0071] f(x+a,y+b)=(1-a)(1-b)f(x,y)+(1-a)bf(x,y+1)+a(1-b)f(x+1,y)+abf(x+1,y+1)
[0072] This embodiment uses the Poisson equation method to remove clouds and obtain a cloud-removed image.
[0073] Poisson information reconstruction:
[0074]
[0075]
[0076] ΔI(x,y)=divv,
[0077] I(x,y) is the image to be processed, I ref (x, y) is the third image used for cloud removal, Ω represents the cloud area, represents the cloud boundary, represents the gradient operator, Δ represents the Laplace operator, divv represents the derivative of the gradient field, and v represents the gradient field of the reference image in Ω. Make the I(x,y) gradient field as close as possible to the guided gradient field v. As a boundary condition, boundary consistency is guaranteed in the reconstruction optimization.
[0078] Use spectral features to select similar pixels for replacement:
[0079]
[0080] RMSD is the root mean square error, L(x i ,y i ,t i ,b) is at t i The time reference image is located at (x i ,y i) is the spectral value of the i-th similar pixel in band b, L(x,y,t1,b) is the spectral value of the corresponding missing pixel under the same definition, and n is the number of bands. Pixels below the preset RMSD threshold are selected as similar pixels.
[0081] Set similar pixel weights and calculate the spectral values of the pixels to be filled.
[0082] CD i =RMSD i *D i ,
[0083] D i Represents geometric distance, CD i represents the composite index, W j Represents the weight of similar pixels, 0≤W j ≤1, L(x,y,t2,b) is the spectral value of the pixel to be filled. Fill the pixels covered by mask1 in the image to be processed I to obtain a cloud-free image, such as Figure 8 Shown are the declouded image and the filled area obtained in the embodiment of the present invention.
[0084] The beneficial effects of the present invention are as follows: the present invention provides a remote sensing image declouding method. Through the connected domain technology, high-precision judgment of the quality of useful images is achieved in declouding, quality assessment of images based on cloud shape and distribution is achieved, and automatic selection and screening of reference images that do not rely on human intervention is achieved. The efficiency and stability of screening useful images and reference images are improved, and the problem that the quality of images with different cloud shapes and distributions at the same cloud proportion cannot be reasonably evaluated in traditional declouding methods is solved.
[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A remote sensing image cloud removal method, characterized in that: include: S1 searches for images intersecting with the image I to be processed at different times according to the spatial position of the image I to be processed, and obtains a first image; S2 selects the first image to obtain the second image according to the spatial position relationship between the image to be processed I and the cloud-free area of the first image; S3 uses the connected domain technology to analyze the cloud area of the second image, evaluate the image quality, and filter the second image to obtain the third image; S4 uses the third image to perform cloud removal processing on the image to be processed I to obtain a cloud-removed image; The method of using the connected domain technology to analyze the cloud area of the second image, evaluate the image quality, and screen the second image to obtain the third image includes: (1) performing cloud region detection on the first image to obtain a cloud region mask mask2, performing binarization processing on the second image, setting cloud and cloud shadow pixels as foreground pixels with a pixel value of 1 according to the cloud region mask mask2 of the second image, and setting cloud-free region pixels as background pixels with a pixel value of 0 according to the cloud-free region mask mask3; (2) Scanning pixels of the second image until the scanned pixel value is 1, taking the pixel as a seed, assigning a new label, and pushing all foreground pixels adjacent to the seed into a stack; (3) Pop the top pixel of the stack, assign the same label to all pixels in the stack, and push all foreground pixels adjacent to the new top pixel of the stack into the stack; (4) Repeat the above step (3) until the stack is empty; (5) Repeat steps (2) to (4) until the scanning is completed and all connected domains of the cloud and cloud shadow are obtained. The number of connected domains is obtained based on the number of connected domain labels. (6) determining whether the second image is qualified according to a connected domain parameter and a cloud amount of the second image, wherein the connected domain parameter includes the number of connected domains and the size of connected domains; if the connected domain parameter is less than a preset connected domain parameter threshold and the image cloud amount is less than a preset cloud amount threshold, determining that the second image is qualified, and obtaining a third image; The method of screening the first image to obtain the second image according to the spatial position relationship between the image to be processed I and the cloud-free area of the first image includes: Perform cloud region detection on the image to be processed I to obtain a cloud region mask mask1 of the image to be processed I; Perform cloud area detection on the first image to obtain a cloud-free area mask mask3; Screen the images where the cloud-free area in the first image intersects with the image to be processed I, determine the spatial position relationship, and obtain the effective intersecting image according to the spatial position relationship. The cloud-free area in the first image is represented by mask mask3; The images whose cloud-free areas in the valid intersecting images intersect with the cloud area of the image to be processed I are screened to determine the spatial position relationship, and the second image is obtained according to the spatial position relationship. The cloud area of the image to be processed I is represented by mask mask1.
2. A remote sensing image cloud removal method according to claim 1, characterized in that: The cloud amount of the second image is the ratio of the cloud area to the entire image.
3. The remote sensing image cloud removal method according to claim 1, characterized in that: The method of using the third image to perform cloud removal processing on the image to be processed I to obtain a cloud-removed image includes: Prioritizing the third image to obtain a prioritized third image; Using the sorted third image in sequence to perform cloud removal processing on the image to be processed I; When the cloud area of the image to be processed I is completely filled, the processing is stopped and a cloud-free image is obtained.
4. A remote sensing image cloud removal method according to claim 3, characterized in that: The step of prioritizing the third image to obtain the prioritized third image includes: Sort the third image according to the time phase to obtain a first priority sorted image; The first priority ranked images are further ranked according to the size of the overlapped area between the cloud-free area of the first priority ranked image and the cloud area of the image to be processed I, with the larger overlapped area being prioritized, to obtain the second priority ranked images; The second priority ranked images are sorted according to their image resolutions, and are sorted in the order of image resolutions being equal to, higher than, and lower than the resolution of the image to be processed I, to obtain a sorted third image.
5. A remote sensing image cloud removal method according to claim 4, characterized in that: The step of sorting the third image according to the time relative time to obtain the first priority sorted image comprises: Select the images of the third image that are in the same period of the same year, the same period of the same year, and the same period of the previous year as the image I to be processed according to the time phase; Prioritize the images in the third image that are of the same year as the image to be processed I, then prioritize the images in the third image that are of the same period as the same period last year as the image to be processed I, and finally prioritize the images in the third image that are of the same year before the image to be processed I, to obtain images ranked with the first priority.
Citation Information
Patent Citations
Cloud distribution characteristic based natural object integrity analysis method of remote sensing image
CN108416308A
Method for removing cloud and cloud shadows in remote sensing image
CN111899194A
Cloud and fog automatic identification method and system based on static meteorological satellite image sequence
CN112287838A
Cloud detection on remote sensing imagery
US20170161584A1
