A method for automatically removing noise of SDGSAT-1 low-light image considering waveband correlation

Abstract

The present application belongs to the technical field of satellite remote sensing data processing, and particularly relates to a method for automatically removing noise of SDGSAT-1 faint light image considering band correlation. The background regions of the panchromatic image and the multispectral image are respectively sampled to the spatial resolution of the other as a mask image, and the background regions of the two types of images are unified. The region containing dark current noise and salt and pepper noise in the panchromatic image is identified, and the noise in the panchromatic and multispectral images is removed by using the region. Further, the isolated low-value pixels in the multispectral image are identified by band, and the dark current noise and salt and pepper noise still existing in the multispectral image are removed. The potential stripe noise of the panchromatic image is identified, the stripe noise direction is obtained by matching the collinear line segments of adjacent blocks, and the optimal noise direction is determined. The noise is repaired by using the noise neighborhood pixel value. Higher precision faint light data denoising can be realized, the effect and automation level are improved, and high-quality data basis is provided for city spatial structure analysis, social economic index calculation and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite remote sensing data processing, and particularly relates to a method for automatically removing noise of micro-light images of a sustainable development science satellite No. 1 (SDGSAT-1), and is particularly suitable for a high-precision denoising scenario that takes into account the correlation of panchromatic and multispectral bands and automatically determines the direction of stripe noise. BACKGROUND

[0002] The sustainable development science satellite No. 1 (SDGSAT-1) is the first scientific satellite in the world that is dedicated to serving the United Nations 2030 Agenda for Sustainable Development (2030 Agenda), and carries a high-performance micro-light imager. A stable night light source is an important indicator for quantifying the intensity of human activities, and the micro-light data of SDGSAT-1 has achieved effective application in the fields of urban planning, ecological monitoring, disaster assessment, energy management, etc. Currently, the micro-light remote sensing data of SDGSAT-1 has noise interference, which reduces the authenticity of the spectral characteristics of night light and affects the analysis results in actual application. Effective removal of the noise of SDGSAT-1 micro-light images helps to improve the signal-to-noise ratio of micro-light images and enhance the reliability of monitoring the status of human activities on the ground, thereby providing a scientific basis for the evaluation of sustainable development goals.

[0003] In previous studies on the denoising method of SDGSAT-1 micro-light data, the existing methods can remove salt and pepper noise and stripe noise, but there are two main problems: first, the correlation of pixel values in the same area of panchromatic and multispectral images is not considered, resulting in the residual of part of the salt and pepper noise; second, the determination of the direction of stripe noise requires human intervention, which hinders the automated batch production of micro-light data products. These problems affect the denoising effect of micro-light data and limit the automation degree and efficiency of micro-light data processing. SUMMARY

[0004] The present application aims to provide a method for automatically removing noise of SDGSAT-1 micro-light images that takes into account the correlation of bands, thereby solving the aforementioned problems in the prior art.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0006] A method for automatically removing noise of SDGSAT-1 micro-light images that takes into account the correlation of bands, comprising the following steps:

[0007] The background regions of the panchromatic image and the multispectral image are respectively sampled to the resolution of the other as a mask image, and the background regions of the two types of images are unified by using the mask image; the regions containing dark current noise and salt and pepper noise in the panchromatic image are identified, and the noise in the panchromatic and multispectral images is removed by using the regions; further, isolated low-value pixels in the multispectral image are identified by waveband, and the dark current noise and salt and pepper noise still existing in the multispectral image are removed; potential stripe noise in the panchromatic image is preliminarily identified according to the characteristics that the pixel values of the stripe noise are low, the width is limited, and the direction is consistent; straight line segments existing in the potential stripe noise are detected by block, and the direction of the stripe noise is obtained by matching collinear line segments in adjacent blocks; a line segment is constructed in the neighborhood interval of the noise direction, and the proportion of the noise is calculated, so as to gradually search for the optimal direction of the stripe noise by reducing the interval and step; the edge pixels of the panchromatic image are taken as the starting point, and the stripe noise is detected along the optimal noise direction, and the noise is repaired by using the neighborhood pixel values of the noise.

[0008] Preferably, the method comprises the following steps:

[0009] S1, a background region in which the pixel values of each waveband of an SDGSAT-1 multispectral image containing red, green and blue wavebands are all 1 is detected, the background region is resampled to 10 m resolution to generate a binary mask, which is used to update the background regions of PL, PH and HDR wavebands of an SDGSAT-1 panchromatic image, and the background region of the panchromatic image is further detected, and the background region is resampled to 40 m resolution to generate another binary mask, which is used to update the background regions of the multispectral image;

[0010] S2, dark current noise and salt and pepper noise in PL and PH wavebands of the panchromatic image are respectively detected, common noise regions of the two wavebands are extracted, the common noise regions are subjected to erosion processing, noise pixels in the corresponding regions of each waveband of the panchromatic image are removed, and the common noise regions are resampled to 40 m resolution to remove noise in the corresponding regions of each waveband of the multispectral image;

[0011] S3, threshold segmentation is performed on the red waveband of the multispectral image to remove dark current noise, eight-neighborhood analysis is adopted to remove salt and pepper noise, and the same method is respectively adopted to remove noise in the green waveband and the blue waveband of the multispectral image;

[0012] S4, the background pixels of the HDR waveband of the panchromatic image are scanned one by one, and when the eight-neighborhood of the background pixels meets the following conditions, the background pixels are marked as potential stripe noise: the number of pixels with a value of 1 in the eight-neighborhood is in the range of 3-5, the background pixels in each row and each column are adjacent, and the maximum and minimum column coordinate values of the background pixels in each row are monotonically increasing or decreasing;

[0013] S5, dividing the image map marked with potential stripe noise into 1024*1024 non-overlapping image blocks, performing Hough transform on each image block to detect line segments, screening line segment pairs with absolute value of slope difference less than 0.01 and absolute value of coordinate difference less than 15 between adjacent image blocks, calculating potential noise coverage of each line segment pair, selecting 5 groups of line segment pairs with highest coverage for slope similarity analysis, and taking the slope with highest frequency as the preliminary stripe noise direction;

[0014] S6, in the interval of the preliminary stripe noise direction as center and 0.2 offset in positive and negative directions and the corresponding start point interval, traversing parameter combinations with 0.1 and 1 as steps respectively, calculating noise coverage of each combination and determining the preliminary optimal slope; in the interval of the preliminary optimal slope as center and 0.1 offset in positive and negative directions and the adjusted start point interval, performing secondary traversal with 0.01 and 1 as steps respectively, and determining the optimal stripe noise direction;

[0015] S7, taking the edge pixel of the panchromatic image as the start point, drawing line segments along the optimal stripe noise direction and screening noise line segments; detecting stripe noise based on the spatial relationship, aggregation degree and size relationship of low-value pixels on the line segments for the HDR band of the panchromatic image, replacing the pixel value of the stripe noise with the 75% quantile value of non-stripe noise pixels in a 3*4 neighborhood; repeating the above stripe noise detection and repair steps for the PL and PH bands of the panchromatic image.

[0016] Preferably, the specific steps of step S1 include:

[0017] S11, generating a multispectral binary map L according to the pixel values of the red band M R , the green band M G and the blue band M B of the multispectral image M M , and the calculation formula is:

[0018]

[0019] wherein, the pixel value L M (i,j) represents the value of the i-th row and j-th column of the binary map L M , and the images M R , M G and M B are the red, green and blue bands of the multispectral image M respectively;

[0020] S12, resampling the binary map L M into a binary mask L P with a resolution of 10m, and the resampling formula is:

[0021]

[0022] wherein x M and y M are the starting coordinates of the multispectral image space, the scaling ratio h M = 4;

[0023] S13, in the region where the value of the binary mask L P is 1, the full-color image PL, PH, and the pixel corresponding to the HDR band are assigned a value of 1;

[0024] S14, according to the pixel values of the PL band P PL , the PH band P PH , and the HDR band P HDR of the full-color image P, a full-color binary graph K P is generated according to the formula, and the calculation formula is:

[0025]

[0026] wherein the images P PL , P PH , and P HDR are the PL, PH, and HDR bands of the full-color image P in sequence;

[0027] S15, the full-color binary graph K P is resampled into a binary mask K M with a resolution of 40m, and the resampling formula is:

[0028]

[0029] wherein y P and y M represent the starting vertical coordinates of the full-color image and the multispectral image space range, respectively, and x P and x M represent the starting horizontal coordinates of the full-color image and the multispectral image space range, respectively;

[0030] S16, in the region where the value of the binary mask K M is 1, the pixels corresponding to the red, green, and blue bands of the multispectral image are assigned a value of 1.

[0031] Preferably, the specific steps of the step S2 include:

[0032] S21, detecting the PL band noise region V PL and the PH band noise region V PH of the full-color image;

[0033] S22, performing intersection processing on the V PL and the V PH , and performing a 3x3 full 1 matrix twice to obtain the common noise region VP ;

[0034] S23, re-sampling the V P value of 1 in the full-color image to 1;

[0035] S24, re-sampling the V P to 40m resolution to obtain a multi-spectral noise region V M ;

[0036] S25, re-sampling the V M value of 1 in the multi-spectral image to 1.

[0037] Preferably, the specific steps of the step S21 include:

[0038] S211, detecting a PL band noise region V PL according to the formula, and the calculation formula is:

[0039]

[0040] wherein, PL min is the minimum pixel value of the PL band, and the dark current threshold g = 3;

[0041] S212, for the pixel p with a value of 0 in the V PL , calculating the number s of pixels with a value of 1 in the eight-neighbor region of the pixel p, and if the pixel number s is greater than 5, resetting the value of the pixel p to 1;

[0042] S213, repeating the steps S211-S212 to detect a PH band noise region V PH .

[0043] Preferably, the specific steps of the step S3 include:

[0044] S31, detecting a multi-spectral red band M R noise region N R according to the formula, and the calculation formula is:

[0045]

[0046] wherein, R min is the minimum pixel value of the red band M R , and the dark current noise threshold q = 5;

[0047] S32, for each pixel u with a value of 0 in the N R , calculating the number w of pixels with a value of 1 in the eight-neighbor region of the pixel u, and if the pixel number w is 8, resetting the value of u to 1;

[0048] S33, re-sampling the multi-spectral red band MR N R The region pixel with value 1 is assigned value 1;

[0049] S34, repeating steps S31 to S33, removing dark current noise and salt and pepper noise of the multi-spectrum green band M G , and blue band M B .

[0050] Preferably, the specific operation of "calculating the potential noise coverage of each line segment pair" in step S5 includes:

[0051] S561, calculating 8-connected regions of the potential stripe noise image N, each connected region corresponding to a connected image C i , the connected image C i is the same size as the noise image N, the connected region pixel value is 1, and the other region pixel value is 0;

[0052] S562, constructing a line segment image graph L with the same size as the potential stripe noise image N, connecting the start point coordinates and end point coordinates of the line segment l n , constructing a line segment with a width of 1, and assigning the pixel value as 1 if the line segment passes through it, and assigning the pixel value as 0 if the line segment does not pass through it;

[0053] S563, in the connected region of the stripe noise image N, selecting the connected region intersected with the line segment l n , denoted as set I, and then calculating the potential noise coverage r n , the calculation formula is:

[0054]

[0055] S564, for the other line segment l m,n of the line segment pair l m , calculating the potential noise coverage r m according to the steps of S562 to S563;

[0056] S565, calculating the potential noise coverage r of the line segment pair l m,n , the calculation formula is r = r m + r n .

[0057] Preferably, the specific operation of "taking the full-color image edge pixel as the starting point" in step S7 includes:

[0058] S711, if the optimal slope k e is greater than zero, selecting the pixels one by one as the starting point from bottom to top along the first column of the potential stripe noise image N, and from left to right along the first row, if the optimal slope k eIf the value is less than zero, select pixels one by one as starting points along the first row of image N from left to right, and then along the last column from top to bottom. Then, use each starting point and the optimal slope k as the starting point. e Construct a line segment l with a width of 1. i ;

[0059] S712, Statistical analysis of each line segment l i The number of potentially noisy pixels n i The total number of pixels n i The sequence is constructed by following the order of line segments. If the sequence contains n pixels... i , satisfying n i >n i-1 And n i >n i+1 If the value is greater than the threshold a, then the number of pixels n is recorded. i Corresponding line segment l i The starting coordinates form the starting set D.

[0060] Preferably, the specific steps for detecting stripe noise in step S7 include:

[0061] S721, for a single starting point (x) in the starting point set D. i ,y i ), in turn (x i -1,y i ), (x i ,y i ), (x i +1,y i Starting from the optimal slope k, e Construct a line segment l with a width of 1 in the direction. i-1 l i l i+1 ;

[0062] S722, Set threshold Each pixel in the HDR band of the panchromatic image is evaluated individually; if it is less than a threshold... In binary image B, assign a value of 1; otherwise, assign a value of 0.

[0063] S723, take line segment l i-1 l i l i+1 In binary image B, what is the total number of pixels with a value of 1 in the t-th row (3 pixels)? like If the value is 1, then pixels with a value of 1 are marked as noise. If the value of t is 2 and the two pixels with a value of 1 are adjacent, they are both marked as noise. Change the value of t and judge the noise for each row.

[0064] S724, threshold Setting 2 to 9 in turn, repeating the steps of S722-S723 mark new noise pixels, merge different threshold Detect all noise pixels and generate a binary graph Q, noise pixels are assigned a value of 1 in the binary graph Q, otherwise a value of 0;

[0065] S725, for each detected noise pixel, count the number of pixels with a value of 1 in its 3x3 neighborhood in the binary graph Q, and count the number of pixels less than the center point in its 3x3 neighborhood in the full-color image HDR band, if the sum of c1 and c2 is greater than 9, it is confirmed that the pixel is a stripe noise, otherwise it is restored to a non-noise pixel.

[0066] The beneficial effects of the present application are:

[0067] The present application overcomes the problem that the previous method ignores the correlation of the pixel values of the same region of the full-color and multispectral images, and the direction of the stripe noise needs to be determined manually, and can realize higher precision SDGSAT-1 low-light data denoising, which helps to improve the effect and automation level of SDGSAT-1 low-light data noise processing, and can provide high-quality data basis for urban spatial structure analysis, social and economic indicator measurement, urban environment parameter inversion, and provide strong support for monitoring and evaluation of the United Nations Sustainable Development Goals. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 is the method flow chart of the SDGSAT-1 low-light image noise automatic removal of the present application considering the band correlation;

[0069] Figure 2 is the full-color noise binary graph in the embodiment of the present application;

[0070] Figure 3 is the full-color image contrast graph before and after dark current and salt and pepper noise removal in the embodiment of the present application;

[0071] Figure 4 is the multispectral image contrast graph before and after dark current and salt and pepper noise removal in the embodiment of the present application;

[0072] Figure 5 is the full-color image graph with stripe noise in the embodiment of the present application;

[0073] Figure 6 is the potential stripe noise result graph in the embodiment of the present application;

[0074] Figure 7 is the stripe noise pixel extraction result graph in the embodiment of the present application. DETAILED DESCRIPTION

[0075] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings. It should be understood that the specific embodiments described herein are merely intended to explain the present application and not to limit the present application.

[0076] As shown in the accompanying Figures 1-7 The present application discloses a method for automatically removing SDGSAT-1 low-light image noise considering band correlation, comprising the following steps:

[0077] The background regions of the panchromatic image and the multispectral image are respectively sampled to the other's spatial resolution as a mask image, and the mask image is used to unify the background regions of the two types of images; the regions containing dark current noise and salt and pepper noise in the panchromatic image are identified, and the noise in the panchromatic and multispectral images is removed by using the regions; further, the isolated low-value pixels in the multispectral image are identified by band, and the dark current noise and salt and pepper noise still existing in the multispectral image are removed; the potential stripe noise in the panchromatic image is preliminarily identified according to the characteristics of low pixel value, limited width and consistent direction of the stripe noise; the straight line segments existing in the potential stripe noise are detected by block, and the direction of the stripe noise is obtained by matching the collinear line segments of adjacent blocks; the line segments are constructed in the neighborhood interval of the noise direction, and the noise proportion is calculated, so as to gradually search for the optimal direction of the stripe noise by reducing the interval and step length; the edge pixels of the panchromatic image are taken as the starting point, and the stripe noise is detected along the optimal noise direction, and the noise is repaired by using the noise neighborhood pixel value.

[0078] Specifically, the method comprises the following steps:

[0079] S1, the background regions of the multispectral image in which the pixel values of each band are all 1 are detected, and resampled to 10m resolution, which is used to update the background region of the panchromatic image; further, the background region of the panchromatic image is detected, and resampled to 40m resolution, which is used to update the background region of the multispectral image;

[0080] S2, the dark current noise and salt and pepper noise in the PL and PH bands of the panchromatic image are respectively detected, the noise regions common to the two bands are extracted, the noise pixels in the noise regions of each band of the panchromatic image are removed, and the common noise regions are resampled to 40m resolution to remove the noise existing in each band of the multispectral image;

[0081] S3, for the red band of the multispectral image, threshold segmentation is implemented to remove the dark current noise, eight-neighborhood analysis is adopted to remove the salt and pepper noise, and further, the same method is respectively used for the green band and the blue band to remove noise;

[0082] S4, the background pixels of the HDR band of the panchromatic image are scanned one by one, and when the eight-neighborhood pixels of the background pixels meet the following conditions, the background pixels are marked as potential stripe noise: the number of the background pixels is within a certain range, the background pixels of each row and each column are adjacent, and the maximum and minimum column coordinate values of the background pixels of each row change monotonously.

[0083] S5, the potential stripe noise image is divided into multiple image blocks, Hough transform is performed in the blocks to detect line segments, line segment pairs that can be connected across blocks and have similar slopes between adjacent blocks are screened, 5 groups of line segment pairs with the largest noise proportion are further selected for slope similarity analysis, and the slope with the highest frequency is taken as the stripe noise direction;

[0084] S6, in the preset stripe noise direction and starting point range, all combinations of slopes and starting points are traversed at a fixed step, line segments are drawn for each parameter combination, and the noise pixel proportion is calculated, the slope of the combination with the largest noise proportion is taken, and then the slope search range and step are reduced for secondary traversal to determine the optimal noise direction;

[0085] S7, line segments are drawn along the image boundary and in the optimal noise direction, noise line segments are screened according to the number of noise pixels on the line segments, and stripe noise is detected based on the spatial relationship, aggregation degree and size relationship of low-value pixels on the noise line segments, and the pixel value of the stripe noise pixel is replaced by 75% quantile value of the effective pixels in the neighborhood of the stripe noise pixel.

[0086] In some embodiments, the specific steps of step S1 include:

[0087] S11, generating a multispectral image binary image L M according to the pixel value of the multispectral image M

[0088]

[0089] wherein the pixel value L M (i,j) represents the value of the binary image L M at the i-th row and j-th column, and the image M R , M G , M B are the red, green and blue bands of the multispectral image M respectively;

[0090] S12, resampling the binary image L M to generate a binary mask L P with a spatial resolution of 10m, and the calculation formula is:

[0091]

[0092] wherein x M and y M are the starting horizontal and vertical coordinates of the spatial range of the multispectral image, and the scaling ratio h M is set to 4;

[0093] S13, in the binary mask L PThe region with value 1 assigns the corresponding pixels of the three bands PL, PH and HDR of the panchromatic image 3 to the background pixel value 1;

[0094] S14, generating a panchromatic image binary graph K according to the pixel value of the panchromatic image P P , and the calculation formula is:

[0095]

[0096] wherein, the image P PL , P PH , P HDR are the PL, PH and HDR bands of the panchromatic image P in sequence;

[0097] S15, resampling the panchromatic image binary graph K P to generate a binary mask K M with a spatial resolution of 40m, and the calculation formula is:

[0098]

[0099] wherein, y P , y M represent the starting vertical coordinates of the spatial range of the panchromatic image and the multi-spectral image respectively, and x P , x M represent the starting horizontal coordinates of the spatial range of the panchromatic image and the multi-spectral image respectively;

[0100] S16, in the region with value 1 of the binary mask K M , assigning the corresponding pixels of the red, green and blue bands of the multi-spectral image M to 1.

[0101] In some specific embodiments, the specific steps of step S2 include:

[0102] S21, detecting the noise regions V PL , V PH of the panchromatic image PL and PH;

[0103] S22, performing intersection processing on the noise regions V PL , V PH , and then performing twice erosion operation using a 3x3 all-1 matrix to obtain the noise region V P ;

[0104] S23, obtaining the region with pixel value 1 of the noise region V P , and setting the pixel value of the corresponding region in the three bands of the panchromatic image to 1;

[0105] S24, resampling the noise region V P to 40m resolution to obtain the multi-spectral noise region V M, the resampling calculation formula is same as S15;

[0106] S25, obtaining the noise region V M The region with pixel value of 1 is set as 1 in the corresponding region of the three bands of the multi-spectral image.

[0107] In the embodiment, the attached Figure 2 is the noise image V common to the two bands of the panchromatic image PL and PH P The white region represents the noise region (pixel value of 1), and the black region represents the effective urban region (pixel value of 0). The removal of the dark current and salt and pepper noise of the panchromatic and multi-spectral images is based on the mask processing of the image.

[0108] In some specific embodiments, the specific steps of step S21 include:

[0109] S211, detecting the noise region V of the panchromatic image band PL PL The calculation formula is:

[0110]

[0111] Wherein, PL min is the minimum pixel value of the PL band, the threshold g is the dark current threshold of the panchromatic band, and the threshold g is set to 3;

[0112] S212, for each pixel p with a value of 0 in the noise region V PL , calculating the number s of pixels with a value of 1 in the eight-neighbor domain of the pixel p, and if the number s of pixels is greater than 5, resetting the value of the pixel p to 1;

[0113] S213, same as steps S211-S212, identifying the noise region V of the band PH PH .

[0114] In some specific embodiments, the specific steps of step S3 include:

[0115] S31, calculating the noise region N R of the red band M R of the multi-spectral image, the calculation formula is:

[0116]

[0117] Wherein, R min is the minimum pixel value of the red band M R of the multi-spectral image, the threshold q is the dark current noise threshold of the multi-spectral band, and the threshold q is generally set to 5;

[0118] S32, for the noise region N RIf the value of the pixel u is 0, the number of pixels with value 1 in the eight-neighborhood of the pixel u is calculated, and if the value of the pixel number w is 8, the value of the pixel u is reset to 1.

[0119] S33, obtaining the noise region N R The region with pixel value 1 is set to 1 in the red wave band M R of the multispectral image.

[0120] S34, repeating the steps of S31 to S33 to remove the dark current noise and salt and pepper noise in the blue wave band M B and the green wave band M G of the multispectral image.

[0121] In this embodiment, the multispectral image results before and after denoising are attached Figure 4 under the condition that the background pixels are set to be colorless, and the multispectral image results before and after denoising are attached Figure 4 (a) is the original SDGSAT-1 multispectral image, and (b) is the multispectral image result after removing the dark current and salt and pepper noise, which shows that the dark current and salt and pepper noise are effectively removed while the urban effective boundary is preserved as much as possible. Figure 4

[0122] In some specific embodiments, the specific steps of step S4 include:

[0123] S41, for the panchromatic image P, the pixel p is determined to be a potential noise pixel if the following conditions are met simultaneously:

[0124] (1) the value of the pixel p is 1;

[0125] (2) the number of pixels with value 1 in the eight-neighborhood of the pixel p is in the range of 3 to 5;

[0126] (3) in the eight-neighborhood of the pixel p, the column coordinates of the pixels with value 1 in each row are continuous, and the row coordinates of the pixels with value 1 in each column are continuous;

[0127] (4) in the eight-neighborhood of the pixel p, the maximum value of the column coordinates of the pixels with value 1 in each row satisfies the monotone increasing or monotone decreasing, and the minimum value of the row coordinates of the pixels with value 1 in each column satisfies the monotone increasing or monotone decreasing;

[0128] S42, according to the judgment conditions of S41, each pixel of the panchromatic image HDR wave band is judged one by one, and the pixels meeting the conditions are assigned a value of 1 in the potential stripe noise image N.

[0129] In this example, Figure 6 ​The result image of the preliminary screening of potential stripe noise is shown in Fig. 1, in which white pixels represent potential noise and black pixels are non-noise pixels. A discontinuous line segment composed of potential noise pixels in the image is the stripe noise possibly existing in the full-color image.

[0130] In some embodiments, the specific step of step S5 comprises:

[0131] S51, starting from the upper left corner of the potential stripe noise image N, the image is cut into image blocks N of 1024x1024 pixels without overlapping i,j , the range of image block N i,j in the noise image N is: the ithx1024 row to the (i+1)thx1024-1 row, the jthx1024 column to the (j+1)thx1024-1 column, and image blocks with insufficient size are discarded;

[0132] S52, all image blocks N i,j are processed respectively, the image coordinate system is established starting from the upper left corner of image block N i,j , and for each pixel coordinate (x i,j , y n ) with a value of 1 in image block N n , a corresponding curve ρ=x n cosθ+y n sinθ is constructed in the Hough space H, the range of θ is -90°≤θ≤89°, the frequency of all curves passing through each parameter point is counted, the highest frequency is H max , the threshold value z=H max 0.75, and the parameter points (ρ m , θ m ) not less than the threshold value z are outputted;

[0133] S53, each parameter point (ρ m , θ m ) represents a line segment l m , the slope k m and the starting horizontal coordinate x ms and the ending horizontal coordinate x me of the line segment l m in the image block are calculated, and the calculation formula is:

[0134]

[0135] The starting horizontal coordinate x ms is obtained by substituting y=0, and the ending horizontal coordinate x me is obtained by substituting y=1023;

[0136] S54, for any two image blocks N i,j and N i+1,j, if there are line segments l m and l n , the absolute value difference between the slope k m of line segment l m and the slope k n of line segment l n is less than 0.01, and the absolute value difference between the starting horizontal coordinate x m of line segment l ms and the ending horizontal coordinate x n of line segment l ne is less than 15, then the line segments l m and l n are recorded as a line segment pair l m,n , and the slope k m is the slope of the line segment pair l m,n ;

[0137] S55, if there is no line segment pair approximately on a straight line in any two image blocks, then the stripe noise detection is terminated, and the processed panchromatic image is outputted;

[0138] S56, the potential noise coverage r of each line segment pair l m,n is calculated, the five line segment pairs with the highest potential noise coverage r values are taken, other line segment pairs similar to each line segment pair are obtained, the similarity condition is that the absolute value difference between the slopes of two line segment pairs is less than 0.02, the line segment pair l r with the largest number of similar line segment pairs is selected, and the slope thereof is recorded as the preliminary stripe noise direction k r .

[0139] In some specific embodiments, the specific steps of calculating the potential noise coverage r of the line segment pair l m,n in step S56 include:

[0140] S561, the 8-connected regions of the potential stripe noise image N are calculated, each connected region corresponds to a connected image C i , the connected image C i is the same size as the noise image N, the connected region pixel value is 1, and the other region pixel value is 0;

[0141] S562, a line segment image graph L with the same size as the potential stripe noise image N is constructed, the starting point coordinates and the ending point coordinates of the line segment l n are connected to construct a line segment with a width of 1, and the pixels passed by the line segment are assigned a value of 1, and the pixels not passed by the line segment are assigned a value of 0;

[0142] S563, in the connected region of the stripe noise image N, the connected region intersecting with the line segment l n is selected and recorded as a set I, and then the potential noise coverage r n is calculated, and the calculation formula is:

[0143]

[0144] S564, for the line segment pair l m,n , the potential noise coverage r m is calculated according to the steps of S562 to S563; m ;

[0145] S565, the potential noise coverage r of the line segment pair l m,n is calculated, and the calculation formula is r = r m + r n .

[0146] In some embodiments, the specific steps of step S6 include:

[0147] S61, the line segment pair l r with the largest number of similar line segment pairs is selected, and the corresponding line segments l1 and l2 are in the upper and lower adjacent image blocks N1 and N2, respectively, and the line segment l2 is extended to the longitudinal coordinate 0 of the image block N1 according to the preliminary stripe noise direction k r , and then the horizontal coordinate x 2s ’ at this position is calculated, and the calculation formula is:

[0148]

[0149] wherein x 2e is the horizontal coordinate of the end point of the line segment l2;

[0150] S62, in the interval [k r -0.2, k r +0.2], [min(x 1s , x 2s ’), max(x 1s , x 2s ’)], the values are taken at a step of 0.1 and 1 respectively and combined in pairs to form an ordered pair (k w , x w ), and all ordered pairs form a set F, and x 1s is the horizontal coordinate of the starting point of the line segment l1;

[0151] S63, the image blocks N1 and N2 are spliced along the vertical direction to form an image block N 12 , for any ordered pair (k w , x w ) in the set F, a line segment l 12 with a width of 1 is drawn in the image block N w with the coordinate (x w , 0) as the starting point and according to the slope k w , and the line segment l wIn image block N 12 Given the potential noise coverage r, select the line segment with the highest noise coverage, and use its slope as the initial optimal slope k. f ;

[0152] S64, in [k] f -0.1,k f +0.1]、[min(x 1s ,x 2s ')-10,max(x 1s ,x 2s Within the interval [k)+10], values ​​are taken with step sizes of 0.01 and 1 respectively, and then combined in pairs to form ordered pairs (k) t ,x t All ordered pairs constitute the set E;

[0153] S65, for any ordered pair (k) in set E t ,x t Following the same steps as in S63, draw the line segment and calculate the potential noise coverage r of the line segment. Select the line segment with the highest noise coverage and use its slope as the optimal slope k. e .

[0154] In some specific embodiments, step S7 includes the following steps:

[0155] S71, take all pixels in the first row and first column or the first row and last column of the latent stripe noise image N, and use each pixel as a starting point to detect the optimal slope k. e Number of noise pixels in the direction n i According to the quantity n i Select the starting pixels that meet the criteria, and denote them as the pixel set D;

[0156] S72, in the panchromatic image HDR band, using each pixel in pixel set D as a starting point, detects the optimal slope k. e Stripe noise pixels in the direction are merged, and all stripe noise detected at the starting point is combined.

[0157] S73, for stripe noise pixels in the HDR band of panchromatic image, replace their pixel values ​​with the 75th percentile value of non-stripe noise pixel values ​​in a 3×4 neighborhood.

[0158] S74. For the PL and PH bands of the panchromatic image P, starting from each pixel in the pixel set D, determine and remove the stripe noise of each band according to steps S72-S73.

[0159] In some specific embodiments, step S71 includes the following steps:

[0160] S711, if the optimal slope ke If the slope is greater than zero, select pixels one by one as the starting point along the first column of the latent stripe noise image N from bottom to top, and then along the first row from left to right. If the optimal slope k e If the value is less than zero, select pixels one by one as starting points along the first row of image N from left to right, and then along the last column from top to bottom. Then, use each starting point and the optimal slope k as the starting point. e Construct a line segment l with a width of 1. i ;

[0161] S712, Statistical analysis of each line segment l i The number of potentially noisy pixels n i The total number of pixels n i The sequence is constructed by following the order of line segments. If the sequence contains n pixels... i , satisfying n i >n i-1 And n i >n i+1 If the value is greater than the threshold a (threshold a is set to 100), then the number of pixels n is recorded. i Corresponding line segment l i The starting coordinates form the starting set D.

[0162] In some specific embodiments, step S72 includes the following steps:

[0163] S721, for a single starting point (x) in the starting point set D. i ,y i ), in turn (x i -1,y i ), (x i ,y i ), (x i +1,y i Starting from the optimal slope k, e Construct a line segment l with a width of 1 in the direction. i-1 l i l i+1 ;

[0164] S722, Set threshold Each pixel in the HDR band of the panchromatic image is evaluated individually; if it is less than a threshold... In binary image B, assign a value of 1; otherwise, assign a value of 0.

[0165] S723, take line segment l i-1 l i l i+1 In binary image B, what is the total number of pixels with a value of 1 in the t-th row (3 pixels)? like If the value is 1, then pixels with a value of 1 are marked as noise. If the value of t is 2 and the two pixels with a value of 1 are adjacent, they are both marked as noise. Change the value of t and judge the noise for each row.

[0166] S724, threshold Set the thresholds sequentially from 2 to 9, repeat steps S722-S723 to mark new noisy pixels, and merge pixels with different thresholds. All detected noise pixels are processed and a binary image Q is generated. Noise pixels are assigned a value of 1 in binary image Q, otherwise they are assigned a value of 0.

[0167] S725: For each detected noise pixel, count the number of pixels with a value of 1 in its 3×3 neighborhood in the binary image Q (c1), and count the number of pixels smaller than the center point in the 3×3 neighborhood of the pixel in the panchromatic image HDR band (c2). If the sum of c1 and c2 is greater than 9, the pixel is confirmed as stripe noise; otherwise, it is restored to a non-noise pixel.

[0168] In this embodiment, it should be noted that the Hough Transform is a feature extraction technique widely used in image processing and computer vision, primarily for detecting geometric shapes. This method detects shapes in an image by finding the local maxima of the accumulator corresponding to a specific curve in the parameter space.

[0169] Basic principle: When investigating whether a set of pixels in an image satisfies specific geometric features (such as a straight line), we can typically map each edge point (x, y) in the image space to a parameter space (ρ, θ). The general polar coordinate form of a straight line is ρ = x·cosθ + y·sinθ (where ρ is the distance from the point to the origin, and θ is the angle between the normal and the x-axis). By accumulating and voting on all edge points in the parameter space, we obtain an accumulator array. The straight lines in the image correspond to the peak positions in the parameter space; by finding these peaks, we can deduce the straight lines in the image space.

[0170] By adopting the above-disclosed technical solution of this invention, the following beneficial effects are obtained:

[0171] This invention overcomes the problems of previous methods that ignore the correlation of pixel values ​​in the same area of ​​panchromatic and multispectral images and require manual determination of the direction of stripe noise. It can achieve higher precision denoising of SDGSAT-1 low-light data, which helps to improve the effect and automation level of SDGSAT-1 low-light data noise processing. It can provide a high-quality data foundation for urban spatial structure analysis, socio-economic indicator calculation, and urban environmental parameter inversion, and provide strong support for the monitoring and assessment of the United Nations Sustainable Development Goals.

[0172] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatic noise removal of SDGSAT-1 low-light images considering band correlation, characterized in that, Includes the following steps: The background regions of panchromatic and multispectral images are sampled to the spatial resolution of the other to serve as mask images, which are then used to unify the background regions of the two types of images. Regions containing dark current noise and salt-and-pepper noise in the panchromatic image are identified, and these regions are used to remove noise from both the panchromatic and multispectral images. Furthermore, isolated low-value pixels in the multispectral image are identified by band, and the remaining dark current noise and salt-and-pepper noise in the multispectral image are removed. Based on the characteristics of low pixel values, limited width, and consistent direction of stripe noise, potential stripe noise in the panchromatic image is initially identified. Straight line segments in potential stripe noise are detected in blocks, and the direction of stripe noise is obtained by matching collinear line segment pairs in adjacent blocks. Line segments are constructed within the neighborhood of the noise direction, and the noise percentage is calculated. The interval and step size are narrowed to gradually search for the optimal direction of the stripe noise. Stripe noise is detected along the optimal noise direction, starting from the edge pixels of the panchromatic image, and the noise is repaired using the pixel values ​​of the noise neighborhood. Specifically, the following steps are included: S1, detect the background region where the pixel value of each band of the SDGSAT-1 multispectral image containing red, green and blue bands is 1, resample the background region to 10m resolution to generate a binary mask, which is used to update the background region of the PL, PH and HDR bands of the SDGSAT-1 panchromatic image, further detect the background region of the panchromatic image, and resample it to 40m resolution to generate another binary mask, which is used to update the background region of the multispectral image; S2, detect the dark current noise and salt-and-pepper noise in the PL and PH bands of the panchromatic image respectively, extract the common noise region of the two bands, perform erosion processing on the common noise region, remove the noise pixels in the corresponding regions of each band of the panchromatic image, and then resample the common noise region to 40m resolution to remove the noise in the corresponding regions of each band of the multispectral image. S3, threshold segmentation is performed on the red band of the multispectral image to remove dark current noise, eight-neighborhood analysis is used to remove salt-and-pepper noise, and the same method is used to denoise the green and blue bands of the multispectral image respectively. S4. Scan the background pixels of the panchromatic image HDR band one by one. When the eight neighborhoods of the background pixel meet the following conditions, it is marked as potential stripe noise: the number of pixels with a value of 1 in the eight neighborhoods is in the range of 3-5, the background pixels in each row and column are adjacent, and the maximum and minimum column coordinate values ​​of the background pixels in each row are monotonically increasing or decreasing. S5. Divide the image map marked with potential stripe noise into non-overlapping image blocks of 1024×1024 pixels. Perform Hough transform on each image block to detect line segments. Filter line segment pairs with an absolute slope difference of less than 0.01 and an absolute coordinate difference of less than 15 between adjacent image blocks. Calculate the potential noise coverage of each line segment pair. Select the 5 line segment pairs with the highest coverage for slope similarity analysis. The slope with the highest frequency is taken as the initial stripe noise direction. S6, within the interval centered on the initial stripe noise direction and offset by 0.2 in both positive and negative directions, and within the corresponding starting interval, the parameter combinations are traversed with step sizes of 0.1 and 1 respectively, the noise coverage of each combination is calculated, and the initial optimal slope is determined; within the interval centered on the initial optimal slope and offset by 0.1 in both positive and negative directions, and within the adjusted starting interval, a second traversal is performed with step sizes of 0.01 and 1 respectively to determine the optimal direction of the stripe noise; S7, starting from the edge pixels of the panchromatic image, draw line segments along the optimal direction of the stripe noise and filter out the noise line segments; for the HDR band of the panchromatic image, detect stripe noise based on the spatial relationship, aggregation degree and size relationship of low-value pixels on the line segments, and replace the stripe noise pixel value with the 75th percentile value of the non-stripe noise pixels in the 3×4 neighborhood; repeat the above stripe noise detection and repair steps for the PL and PH bands of the panchromatic image.

2. The method for automatic noise removal of SDGSAT-1 low-light images considering band correlation according to claim 1, characterized in that, The specific steps of step S1 include: S11, based on the red band M of the multispectral image M R Green band M G Blue band M B The pixel values ​​are used to generate a multispectral binary image L according to the formula. M The calculation formula is: ； Wherein, pixel value L M (i,j) represents the binary graph L. M The value in the i-th row and j-th column, image M R M G M B The images are, in order, the red, green, and blue bands of the multispectral image M; S12, the binary image L M Resampling to a 10m resolution binary mask L P The resampling formula is: ； Where, x M and y M The starting coordinates of the multispectral image space, with a scaling ratio of h. M =4; S13, in the binary mask L P For regions with a value of 1, the corresponding pixels of the panchromatic image's PL, PH, and HDR bands will be assigned a value of 1. S14, based on the PL band P of the panchromatic image P PL PH band P PH HDR band P HDR The pixel values ​​are used to generate a panchromatic binary image K according to the formula. P The calculation formula is: ； Among them, image P PL P PH P HDR The bands are, in order, the PL, PH, and HDR bands of the panchromatic image P; S15, the panchromatic binary image K P Resampling to a 40m resolution binary mask K M The resampling formula is: ； Among them, y P y M The x-coordinates represent the starting ordinates of the spatial extent of the panchromatic and multispectral images, respectively. P x M These represent the starting x-coordinates of the spatial range of the panchromatic image and the multispectral image, respectively. S16, in the binary mask K M For regions with a value of 1, the corresponding pixels in the red, green, and blue bands of the multispectral image will be assigned a value of 1.

3. The method for automatic noise removal of SDGSAT-1 low-light images considering band correlation according to claim 1, characterized in that, The specific steps of step S2 include: S21, Detect noise region V in panchromatic image PL band. PL PH band noise region V PH ; S22, for the V PL With V PH Intersection processing is performed, and two erosion operations are carried out using a 3×3 all-1 matrix to obtain the common noise region V. P ; S23, V in each band of the panchromatic image P The pixels in the region with a value of 1 are assigned a value of 1. S24, according to the sampling formula in step S15, the V P Resampling to 40m resolution yields the multispectral noise region V M ; S25, V in each band of the multispectral image M The pixels in the region with a value of 1 are assigned a value of 1.

4. The method for automatic noise removal of SDGSAT-1 low-light images considering band correlation according to claim 3, characterized in that, The specific steps of step S21 include: S211, Detect the noise region V in the PL band according to the formula. PL The calculation formula is: ； Among them, PL min The minimum pixel value for the PL band is given by dark current threshold g=3. S212, for the V PL For a pixel p with a value of 0, calculate the number of pixels s with a value of 1 in its eight neighborhoods. If the number of pixels s is greater than 5, then reset the value of pixel p to 1. S213, Repeat steps S211-S212 to detect the noise region V in the PH band. PH .

5. The method for automatic noise removal of SDGSAT-1 low-light images considering band correlation according to claim 1, characterized in that, The specific steps of step S3 include: S31, Detect the multispectral red band M according to the formula. R Noise region N R The calculation formula is: ； Among them, R min Red band M R The minimum pixel value, dark current noise threshold q=5; S32, for the N R For each pixel u with a median value of 0, calculate the number w of pixels with a value of 1 in its eight neighborhoods. If the value of the number of pixels w is 8, then reset the value of u to 1. S33, the multispectral red band M R N R The pixels in the region with a value of 1 are assigned a value of 1. S34, Repeat steps S31 to S33 to remove the multispectral green band M. G Blue band M B Dark current noise and salt-and-pepper noise.

6. The method for automatic noise removal of SDGSAT-1 low-light images considering band correlation according to claim 1, characterized in that, The specific operations for calculating the potential noise coverage of each line segment pair in step S5 include: S561, Calculate the 8-connected regions of the latent stripe noise image N, where each connected region corresponds to a connected image C. i Connecting Image C i The image is of the same size as the noisy image N, with connected regions having a pixel value of 1 and other regions having a pixel value of 0. S562, Construct a line segment image L of the same size as the potential stripe noise image N, and connect the line segments l. n Given the starting and ending coordinates, construct a line segment with a width of 1. Assign a value of 1 to the pixels the line segment passes through and a value of 0 to the pixels it does not pass through. S563, in the connected region of the stripe noise image N, select the line segment l n The intersecting connected regions are denoted as set I, and then the potential noise coverage r is calculated. n The calculation formula is: ； S564, for line segment pairs l m,n The other line segment l m Calculate the potential noise coverage r according to steps S562 to S563. m ; S565, Calculate line segment pairs l m,n The potential noise coverage r is calculated using the formula r = r m +r n .

7. The method for automatic noise removal of SDGSAT-1 low-light images considering band correlation according to claim 1, characterized in that, The specific operations in step S7, starting from the edge pixels of the panchromatic image, include: S711, if the optimal slope k e If the slope is greater than zero, select pixels one by one as the starting point along the first column of the latent stripe noise image N from bottom to top, and then along the first row from left to right. If the optimal slope k e If the value is less than zero, select pixels one by one as starting points along the first row of image N from left to right, and then along the last column from top to bottom. Then, use each starting point and the optimal slope k as the starting point. e Construct a line segment l with a width of 1. i ; S712, Statistical analysis of each line segment l i The number of potentially noisy pixels n i The total number of pixels n i The sequence is constructed by following the order of line segments. If the sequence contains n pixels... i , satisfying n i >n i-1 And n i >n i+1 If the value is greater than the threshold a, then the number of pixels n is recorded. i Corresponding line segment l i The starting coordinates form the starting set D.

8. The method for automatic noise removal of SDGSAT-1 low-light images considering band correlation according to claim 7, characterized in that, The specific steps for detecting stripe noise in step S7 include: S721, for a single starting point (x) in the starting point set D. i ,y i ), in turn (x i -1,y i ), (x i ,y i ), (x i +1,y i Starting from the optimal slope k, e Construct a line segment l with a width of 1 in the direction. i-1 l i l i+1 ; S722, Set threshold Each pixel in the HDR band of the panchromatic image is evaluated individually; if it is less than a threshold... If the value is 1, then it is assigned a value of 1 in the binary image B; otherwise, it is assigned a value of 0. S723, take line segment l i-1 l i l i+1 In the binary image B, for the 3 pixels in the t-th row, determine the total number of pixels with a value of 1, ∂. If ∂ is 1, then the pixels with a value of 1 are marked as noise. If ∂ is 2 and the two pixels with a value of 1 are adjacent, then both are marked as noise. Change the value of t and perform noise judgment on each row. S724, threshold Set the thresholds sequentially from 2 to 9, repeat steps S722-S723 to mark new noisy pixels, and merge pixels with different thresholds. All detected noise pixels are processed and a binary image Q is generated. Noise pixels are assigned a value of 1 in binary image Q, otherwise they are assigned a value of 0. S725: For each detected noise pixel, count the number of pixels with a value of 1 in its 3×3 neighborhood in the binary image Q (c1), and count the number of pixels smaller than the center point in the 3×3 neighborhood of the pixel in the panchromatic image HDR band (c2). If the sum of c1 and c2 is greater than 9, the pixel is confirmed as stripe noise; otherwise, it is restored to a non-noise pixel.

Images