CN113870159A - Hyperspectral fusion method based on dynamic gradient group sparsity and low-rank regularization - Google Patents
Hyperspectral fusion method based on dynamic gradient group sparsity and low-rank regularization Download PDFInfo
- Publication number
- CN113870159A CN113870159A CN202111019135.9A CN202111019135A CN113870159A CN 113870159 A CN113870159 A CN 113870159A CN 202111019135 A CN202111019135 A CN 202111019135A CN 113870159 A CN113870159 A CN 113870159A
- Authority
- CN
- China
- Prior art keywords
- resolution
- image
- matrix
- low
- hyperspectral
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000007500 overflow downdraw method Methods 0.000 title claims abstract description 20
- 239000011159 matrix material Substances 0.000 claims abstract description 199
- 230000004927 fusion Effects 0.000 claims abstract description 52
- 238000000034 method Methods 0.000 claims abstract description 33
- 230000003595 spectral effect Effects 0.000 claims description 40
- 238000005070 sampling Methods 0.000 claims description 34
- 230000004044 response Effects 0.000 claims description 20
- 238000001228 spectrum Methods 0.000 claims description 18
- 239000013598 vector Substances 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 9
- 230000010076 replication Effects 0.000 claims description 9
- 239000000126 substance Substances 0.000 claims description 8
- 238000000354 decomposition reaction Methods 0.000 claims description 7
- 230000009977 dual effect Effects 0.000 claims description 7
- 230000003190 augmentative effect Effects 0.000 claims description 6
- 230000015556 catabolic process Effects 0.000 claims description 5
- 238000006731 degradation reaction Methods 0.000 claims description 5
- 238000004458 analytical method Methods 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 4
- 238000011161 development Methods 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 description 6
- 230000008602 contraction Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000013441 quality evaluation Methods 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10036—Multispectral image; Hyperspectral image
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
The invention provides a hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization. The method comprises the steps of obtaining a low-resolution hyperspectral image through a hyperspectral sensor, collecting a low-resolution multispectral image and a high-resolution panchromatic image of the same picture through the multispectral sensor, constructing a fusion model among the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image through dynamic gradient group sparse regularization and low-rank regularization, solving the fusion model through an alternating direction multiplier method, obtaining a coefficient matrix, and further multiplying the coefficient matrix and a subspace matrix to obtain the high-resolution hyperspectral image. The fusion method combines the subspace regularization of the image with the image fusion, so that the solution of the fusion target is converted into the solution of the low-dimensional coefficient matrix, the calculation efficiency is improved, and the high-resolution hyperspectral image which is superior to the contrast method in qualitative and quantitative aspects is obtained.
Description
Technical Field
The invention belongs to the field of remote sensing image fusion, and particularly relates to a hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization.
Background
In the field of remote sensing, Panchromatic (PAN), Multispectral (MS) and Hyperspectral (HS) data have wide applications, including land classification, change detection, and the like. Due to the cost limitations of optical sensors and the limitations of data storage and transmission bandwidth, the design of optical sensors requires a trade-off between spatial and spectral resolution. For example, hyperspectral data with bands up to several hundred provides higher spectral resolution, but its spatial picture is generally blurred. Full-color images with a single band have relatively single spectral information, but are rich in spatial information. Due to the application requirements for remote sensing data with both high spatial resolution and spectral resolution, various image fusion techniques have been developed.
The Multispectral fusion technology fuses a Low-Resolution Multispectral Image (LR-MSI) and a High-Resolution Panchromatic Image (HR-PAN) to improve the spatial Resolution of the Multispectral Image, and the fusion technology is developed more mature at present and mainly comprises a component replacement method, a multiresolution analysis method and a variation method and a machine learning-based method. The Hyperspectral fusion technology generally fuses a Low-Resolution Hyperspectral Image (LR-HSI) and a High-Resolution Multispectral Image (HR-MSI) to improve the spatial Resolution of the Hyperspectral Image, and compared with the Hyperspectral Image fusion technology, the Hyperspectral Image fusion technology has higher data dimension and more variables to be estimated. The hyperspectral image fusion technology mainly comprises the expansion of a multispectral image fusion method, a Bayes-based method and a linear spectral decomposition-based method.
In consideration of high spectrum, information contained in the multispectral image and information contained in the panchromatic image are complementary, so that the spatial and spectral information contained in the three kinds of data of the same picture can be better mined by fusing the three kinds of data of the same picture. The research of the data fusion method is less, and in order to better utilize the total information of a hyperspectral image, a multispectral image and a panchromatic image, the invention provides a novel model-based method for combining the sparseness and low-rank regularization of a dynamic gradient group with the fusion of hyperspectral, multispectral and panchromatic data.
Disclosure of Invention
Aiming at the defects of the existing remote sensing image fusion method, the invention provides a hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization.
Step 1: acquiring a high-spectrum image with low resolution by a high-spectrum sensor, and acquiring a low-resolution multi-spectrum image and a high-resolution full-color image of the same picture by the multi-spectrum sensor;
step 2: constructing a fusion model among the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image according to the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image in the step 1;
and step 3: solving a coefficient matrix X by an alternative direction multiplier method based on the fusion model in the step 2(3)And further multiplying the coefficient matrix and the subspace matrix to obtain a high-resolution hyperspectral image F(3);
Preferably, the low-resolution hyperspectral image in step 1 is recorded as a tensorWhose 3-mode expansion matrix isIndicates that the image has LhIndividual wave band, Wh×HhA plurality of pixels;
step 1, the multispectral image with low resolution is recorded as tensorWhose 3-mode expansion matrix isIndicates that the image has LmIndividual wave band, Wm×HmA plurality of pixels;
step 1 the high resolution panchromatic image is recorded as a tensorWhose 3-mode expansion matrix isIndicating that the image has 1 band, Wp×HpA plurality of pixels;
the ratio of the resolution between the high-resolution panchromatic image and the low-resolution hyperspectral image is shThe ratio of the resolution between the high-resolution panchromatic image and the low-resolution multispectral image is smI.e. by
Preferably, the fusion model among the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image in the step 2 is composed of a hyperspectral image fitting term, a multispectral image fitting term, a dynamic gradient group sparse regularization term and a low-rank regularization term;
the hyperspectral image fitting item and the multispectral image fitting item are low-resolution hyperspectral images H(3)Low resolution multispectral image M(3)The low-resolution hyperspectral image H is constructed by the space degradation relation and the spectrum degradation relation between the high-resolution hyperspectral image H and the high-resolution hyperspectral image(3)Spatial down-sampling version of high-resolution equivalent high-spectral image, low-resolution multi-spectral image M(3)The high-spectrum image equivalent to high resolution is a version after spatial down-sampling and spectral down-sampling;
the hyperspectral image fitting term, the multispectral image fitting term, the dynamic gradient group sparse regularization term and the low-rank regularization term form a target energy function related to a coefficient matrix;
preferably, the step 2 of constructing the fusion model among the hyperspectral image, the multispectral image and the panchromatic image specifically comprises the following sub-steps:
step 2.1: the low-resolution hyperspectral image H (3) can be considered as a spatially down-sampled version of the high-resolution hyperspectral image F (3):
H(3)=F(3)BhSh
wherein, the matrixIs a spatial blur matrix.Is a spatial down-sampling operation for reducing the spatial resolution of the hyperspectral image.
Step 2.2: multispectral image M of low resolution(3)Hyperspectral image F which can be regarded as high resolution(3)Spatial down-sampling and spectral down-sampling versions:
M(3)=RmF(3)BmSm
wherein the content of the first and second substances,is the spectral response matrix of the multispectral instrument. Matrix arrayIs a spatial blur matrix.Is a spatial down-sampling operation for reducing the spatial resolution of the hyperspectral image.
Step 2.3: the high-resolution panchromatic image P and the high-resolution hyperspectral image F have the same spatial edge information, the gradient difference of the high-resolution panchromatic image P and the high-resolution hyperspectral image F meets the group sparsity characteristic, and a regularization item for dynamic gradient group sparsity can be established:
the high-resolution hyperspectral image is recorded as a tensorWhose 3-mode expansion matrix isIndicates that the image has LhIndividual wave band, Wp×HpA plurality of pixels;
wherein the content of the first and second substances,indicating the replication of high resolution panchromatic images to LmAnd (4) a plurality of wave bands.
B in step 2.1hMatrix, R in step 2.2m,BmThe matrix can be composed of a low-resolution hyperspectral image H(3)Low resolution multispectral image M(3)High resolution full color image P(3)The estimation is obtained, and the estimation method provided by the hyperspectral image fusion method named HySure is used in the invention.
||X||*
step 2.5: converting the high-resolution hyperspectral image into a coefficient matrix model of the high-resolution hyperspectral image in a low-dimensional subspace by combining a product model of the subspace matrix and the coefficient matrix;
step 2.5 the product model of the subspace matrix and the coefficient matrix is:
wherein the extract is3The mode product of the 3 rd mode is shown,is a subspace matrix representing a vector composed of D pure spectral features, the subspace momentsThe array is directly obtained by the low-resolution hyperspectral image through image decomposition, and the image decomposition method selected by the invention is vertex component analysis;
Zhang Liangis a coefficient matrix, representingIs represented by a linear combination of vector members in a subspace matrix,representing a pixelAnd corresponding coefficients, i is the spatial transverse coordinate of the pixel point, and j is the spatial longitudinal coordinate of the pixel point.
Step 2.5, the coefficient matrix model of the high-resolution hyperspectral image converted into the high-resolution hyperspectral image in the low-dimensional subspace is as follows:
F(3)=EX(3)
step 2.6: based on the steps 2.1 to 2.5, establishing a fusion model of the three components:
whereinRepresenting Frobenius norm, | | · |. luminance2,1Representing L2,1 norm, | | · |. non-volatile memory*Representing a kernelNumber, lambdam,λφAnd λlParameters for balancing the respective items are respectively. X(3)In the form of a two-dimensional matrix representation of the coefficients.
Preferably, the specific implementation of step 3 comprises the following sub-steps:
step 3.1: introducing an auxiliary variable O, wherein O is equal to X(3)Bh(ii) a Introducing an auxiliary variable U, wherein U is equal to X(3)Bm(ii) a Introducing an auxiliary variable V, wherein V is equal to X(3)(ii) a Introducing an auxiliary variable W, and satisfying that W is equal to RmEV; introducing an auxiliary variable Q, wherein Q is equal to X(3). The three fusion model is expressed as:
s.t.O=X(3)Bh
U=X(3)Bm
V=X(3)
W=RmEV
Q=X(3)
the augmented Lagrange function of the three fusion models is expressed as:
wherein, Y1,Y2,Y3,Y4And Y5For scale Dual variables (Scaled Dual Variable),representing Frobenius norm, | | · |. luminance2,1Representing L2,1 norm, | | · |. non-volatile memory*Denotes the nuclear norm, λm,λφAnd λlParameters for balancing the respective items are respectively. μ is a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients. H(3)Is a two-dimensional matrix representation of a low-resolution hyperspectral image. M(3)Multiple lights of low resolutionTwo-dimensional matrix representation of the spectral image.Indicating the replication of high resolution panchromatic images to LmAnd (4) a plurality of wave bands.The sign of the gradient is indicated.In the form of a spatial blur matrix, the matrix is,in the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,in order to perform a spatial down-sampling operation,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix. The optimization problem of the augmented Lagrangian function of the three-component fusion model can be decomposed into an X subproblem, an O subproblem, a U subproblem, a V subproblem, a W subproblem, a Q subproblem and a Y subproblem.
Step 3.2: solving the X subproblem:
wherein t in the upper right corner represents the number of iterations and Y1,Y2,Y3And Y5In the form of a dual-scale variable,represents Frobenius norm, mu is penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients.In the form of a spatial blur matrix, the matrix is,is a spatial fuzzy matrix;
the solution to the above problem is:
step 3.3: solving an O subproblem:
wherein, Y1In the form of a dual-scale variable,represents Frobenius norm, mu is penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients. H(3)Is a two-dimensional matrix representation of a low-resolution hyperspectral image.In the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,is a subspace matrix.
Divide O into OShAndwhereinRepresentation is not represented by matrix ShAnd (3) solving the O subproblem by the selected pixel point:
step 3.4: solving the U sub-problem:
wherein, Y2In the form of a dual-scale variable,denotes the Frobenius norm, λmTo balance the parameters of the various items. μ is a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients. M(3)Is a two-dimensional matrix representation of the low resolution multispectral image.In the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix.
Divide U into USmAndwhereinRepresentation is not represented by matrix SmThe solution of the selected pixel point and the U subproblem is as follows:
step 3.5: solving the V subproblem:
wherein, Y3,Y4In the form of a dual-scale variable,represents Frobenius norm, mu is penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients.Is a spectral response matrix of a multi-spectral instrument,is a subspace matrix.
The solution to the above problem is:
step 3.6: solving the W sub-problem:
wherein, Y4In the form of a dual-scale variable,representing Frobenius norm, | | · |. luminance2,1Denotes L2,1 norm,. lambda.φTo balance the parameters of the various items. μ is a penalty parameter.Indicating the replication of high resolution panchromatic images to LmAnd (4) a plurality of wave bands.The sign of the gradient is indicated.Is a spectral response matrix of a multi-spectral instrument,is a subspace matrix.
Let G be W-P, then the W sub-problem becomes
The W subproblem) can be solved directly by a vector Total Variation (vector Total Variation) algorithm.
Step 3.7: solving the Q sub-problem:
wherein, Y5In the form of a dual-scale variable,representing Frobenius norm, | | · |. luminance*Denotes the nuclear norm, λlTo balance the parameters of the various items. μ is a penalty parameter. X(3)Is a two-dimensional matrix representation of coefficientsForm (a).
The Q sub-problem can be solved directly by a singular value contraction algorithm.
Step 3.8: and solving the Y subproblem. This subproblem can be directly updated by the following equation:
wherein, Y1,Y2,Y3,Y4And Y5For scale-duality variables, μ is a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients.In the form of a spatial blur matrix, the matrix is,in the form of a spatial blur matrix, the matrix is,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix. O, U, V, W and Q as auxiliary variablesAmount of the compound (A).
Step 3.9: obtaining the coefficient tensor by the iterative solution of the subproblemAccording to step 2E, obtaining a high-resolution hyperspectral image
The method mainly meets the application requirement of obtaining the high-resolution hyperspectral image by fusing the low-resolution hyperspectral image with the low-resolution multispectral image and the high-resolution panchromatic image. The fusion method combines the subspace regularization of the image with the image fusion, thereby converting the solution of the fusion target into the solution of the low-dimensional coefficient matrix. The model provided by the invention comprises two secondary data fitting terms, a dynamic gradient group sparse regularization term and a low-rank regularization term, and is solved by an ADMM algorithm, so that a hyperspectral fusion image which is superior to a comparison method in qualitative and quantitative aspects is obtained.
Drawings
FIG. 1: is an experimental flow chart of the invention.
FIG. 2: is a fusion result graph of CNMF.
FIG. 3: is a fusion result graph of the GSA of the comparison method of the invention.
FIG. 4: is a fusion result graph of the method provided by the invention.
FIG. 5: is a reference image of the present invention.
Detailed Description
In order to facilitate the understanding and implementation of the present invention for those of ordinary skill in the art, the present invention will be described in further detail with reference to the accompanying drawings and examples, it is to be understood that the examples described herein are only for the purpose of illustrating the present invention and are not to be construed as limiting the present invention.
The data used in the present invention was taken from the department of university of Pavia taken by a rosss aerial imaging spectrometer in germany. The invention selects 93 wave bands from 115 wave bands of the data to obtain HR-HSI (320 multiplied by 93), and the HR-HSI is used as a reference image Ref for quality evaluation after fusion. To generate LR-HSI (20 × 20 × 93), the present invention first blurs HR-HSI using a 7 × 7 Gaussian blur (mean zero, standard deviation of 2), and then downsamples every sixteen pixels in the width and height modes of HR-HSI. To simulate the LR-MSI (80 × 80 × 4) of the same scene, HR-HSI was sampled along the spectral mode using a reflective spectral response filter similar to IKONOS, then blurred with 7 × 7 Gaussian blur (mean zero, standard deviation of 2), and finally sampled every four pixels in the width and height modes of HR-HSI. The HR-PAN (320 x 320) for the same scene is generated by first sampling the HR-HSI into four band images along the spectral mode using a reflected spectral response filter similar to IKONOS, and then linearly combining the first three bands with [0.114, 0.587, 0.299] weights, respectively. The flow chart of the fusion experiment is shown in FIG. 1.
The embodiment provides a hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization, which comprises the following steps of:
step 1: acquiring a high-spectrum image with low resolution by a high-spectrum sensor, and acquiring a low-resolution multi-spectrum image and a high-resolution full-color image of the same picture by the multi-spectrum sensor;
step 1, recording the low-resolution hyperspectral image as tensorWhose 3-mode expansion matrix isIndicates that the image has Lh93 bands, Wh×Hh20 × 20-400 pixels;
step 1, the multispectral image with low resolution is recorded as tensorWhose 3-mode expansion matrix isIndicates that the image has Lm4 bands, Wm×Hm6400 pixels 80 × 80;
step 1 the high resolution panchromatic image is recorded as a tensorWhose 3-mode expansion matrix isIndicating that the image has 1 band, Wp×Hp320 × 320 — 102400 pixels;
the ratio of the resolution between the high-resolution panchromatic image and the low-resolution hyperspectral image is shThe ratio of the resolution between the high-resolution panchromatic image and the low-resolution multispectral image is smI.e. by
Step 2: constructing a fusion model among the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image according to the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image in the step 1;
preferably, the fusion model among the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image in the step 2 is composed of a hyperspectral image fitting term, a multispectral image fitting term, a dynamic gradient group sparse regularization term and a low-rank regularization term;
the hyperspectral image fitting item and the multispectral image fitting item are low-resolution hyperspectral images H(3)Low resolution multispectral image M(3)The low-resolution hyperspectral image H is constructed by the space degradation relation and the spectrum degradation relation between the high-resolution hyperspectral image H and the high-resolution hyperspectral image(3)Spatial down-sampling version of high-resolution equivalent high-spectral image, low-resolution multi-spectral image M(3)The high-spectrum image equivalent to high resolution is a version after spatial down-sampling and spectral down-sampling;
the hyperspectral image fitting term, the multispectral image fitting term, the dynamic gradient group sparse regularization term and the low-rank regularization term form a target energy function related to the coefficient matrix.
The specific implementation of the construction of the fusion model among the hyperspectral image, the multispectral image and the panchromatic image in the step 2 comprises the following substeps:
step 2.1: low-resolution hyperspectral image H(3)Hyperspectral image F which can be regarded as high resolution(3)Spatially down-sampled version of (a):
H(3)=F(3)BhSh
wherein, the matrixIs a spatial blur matrix.Is a spatial down-sampling operation for reducing the spatial resolution of the hyperspectral image.
Step 2.2: multispectral image M of low resolution(3)Hyperspectral image F which can be regarded as high resolution(3)Spatial down-sampling and spectral down-sampling versions:
M(3)=RmF(3)BmSm
whereinIs the spectral response matrix of the multispectral instrument. Matrix arrayIs a spatial blur matrix.Is a spatial down-sampling operation for reducing the spatial resolution of the hyperspectral image.
Step 2.3: the high-resolution panchromatic image P and the high-resolution hyperspectral image F have the same spatial edge information, the gradient difference of the high-resolution panchromatic image P and the high-resolution hyperspectral image F meets the group sparsity characteristic, and a regularization item for dynamic gradient group sparsity can be established:
the high-resolution hyperspectral image is recorded as a tensorWhose 3-mode expansion matrix isIndicates that the image has Lh93 bands, Wp×Hp320 × 320 — 102400 pixels;
wherein the content of the first and second substances,indicating the replication of high resolution panchromatic images to Lm4 bands.
B in step 2.1hMatrix, R in step 2.2m,BmThe matrix can be composed of a low-resolution hyperspectral image H(3)Low resolution multispectral image M(3)High resolution full color image P(3)The estimation is obtained by using an estimation method provided in a hyperspectral image fusion method named HySure.
||X||*
step 2.5: converting the high-resolution hyperspectral image into a coefficient matrix model of the high-resolution hyperspectral image in a low-dimensional subspace by combining a product model of the subspace matrix and the coefficient matrix;
step 2.5 the product model of the subspace matrix and the coefficient matrix is:
wherein the extract is3The mode product of the 3 rd mode is shown,the method comprises the following steps that a subspace matrix is used, wherein the subspace matrix is composed of vectors with D being 10 pure spectral features, the subspace matrix is directly obtained by low-resolution hyperspectral images through image decomposition, and the selected image decomposition method is vertex component analysis;
Zhang Liangis a coefficient matrix, representingIs represented by a linear combination of vector members in a subspace matrix,representing a pixelAnd corresponding coefficients, i is the spatial transverse coordinate of the pixel point, and j is the spatial longitudinal coordinate of the pixel point.
Step 2.5, the coefficient matrix model of the high-resolution hyperspectral image converted into the high-resolution hyperspectral image in the low-dimensional subspace is as follows:
F(3)=EX(3)
step 2.6: based on the steps 2.1 to 2.5, establishing a fusion model of the three components:
whereinRepresenting Frobenius norm, | | · |. luminance2,1Representing L2,1 norm, | | · |. non-volatile memory*Denotes the nuclear norm, λm=0.01,λφ0.01 and λlEach parameter 0.001 is a parameter for balancing each item. X(3)In the form of a two-dimensional matrix representation of the coefficients.
And step 3: solving a coefficient matrix X by an alternative direction multiplier method based on the fusion model in the step 2(3)And further multiplying the coefficient matrix and the subspace matrix to obtain a high-resolution hyperspectral image F(3);
The specific implementation of the step 3 comprises the following substeps:
step 3.1: introducing an auxiliary variable O, wherein O is equal to X(3)Bh(ii) a Introducing an auxiliary variable U, wherein U is equal to X(3)Bm(ii) a Introducing an auxiliary variable V, wherein V is equal to X(3)(ii) a Introducing an auxiliary variable W, and satisfying that W is equal to RmEV; introducing an auxiliary variable Q, wherein Q is equal to X(3). The three fusion model is expressed as:
s.t.O=X(3)Bh
U=X(3)Bm
V=X(3)
W=RmEV
Q=X(3)
the augmented Lagrange function of the three fusion models is expressed as:
wherein, Y1,Y2,Y3,Y4And Y5For scale Dual variables (Scaled Dual Variable),representing Frobenius norm, | | · |. luminance2,1Representing L2,1 norm, | | · |. non-volatile memory*Denotes the nuclear norm, λm=0.01,λφ0.01 and λlEach parameter 0.001 is a parameter for balancing each item. μ ═ 0.001 is a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients. H(3)Is a two-dimensional matrix representation of a low-resolution hyperspectral image. M(3)Is a two-dimensional matrix representation of the low resolution multispectral image.Indicating the replication of high resolution panchromatic images to Lm4 bands.The sign of the gradient is indicated.In the form of a spatial blur matrix, the matrix is,in the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,in order to perform a spatial down-sampling operation,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix. The optimization problem of the augmented Lagrangian function of the three-component fusion model can be decomposed into an X subproblem, an O subproblem, a U subproblem, a V subproblem, a W subproblem, a Q subproblem and a Y subproblem.
Step 3.2: solving the X subproblem:
wherein t in the upper right corner represents the number of iterations and Y1,Y2,Y3And Y5In the form of a dual-scale variable,denotes Frobenius norm, with μ ═ 0.001 as a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients.In the form of a spatial blur matrix, the matrix is,is a spatial fuzzy matrix;
the solution to the above problem is:
step 3.3: solving an O subproblem:
wherein, Y1Is a scale pairThe amount of the even variable is,denotes Frobenius norm, with μ ═ 0.001 as a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients. H(3)Is a two-dimensional matrix representation of a low-resolution hyperspectral image.In the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,is a subspace matrix.
Divide O into OShAndwhereinRepresentation is not represented by matrix ShAnd (3) solving the O subproblem by the selected pixel point:
step 3.4: solving the U sub-problem:
wherein, Y2In the form of a dual-scale variable,denotes the Frobenius norm, λm0.01 is a parameter for balancing the respective items. μ ═ 0.001 is a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients. M(3)Is a two-dimensional matrix representation of the low resolution multispectral image.In the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix.
Divide U into USmAndwhereinRepresentation is not represented by matrix SmThe solution of the selected pixel point and the U subproblem is as follows:
step 3.5: solving the V subproblem:
wherein, Y3,Y4In the form of a dual-scale variable,denotes Frobenius norm, with μ ═ 0.001 as a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients.Is a spectral response matrix of a multi-spectral instrument,is a subspace matrix.
The solution to the above problem is:
step 3.6: solving the W sub-problem:
wherein, Y4In the form of a dual-scale variable,representing Frobenius norm, | | · |. luminance2,1Denotes L2,1 norm,. lambda.φ0.01 is a parameter for balancing the respective items. μ ═ 0.001 is a penalty parameter.Indicating the replication of high resolution panchromatic images to Lm4 bands.The sign of the gradient is indicated.Is a spectral response matrix of a multi-spectral instrument,is a subspace matrix.
Let G be W-P, then the W sub-problem becomes
The W subproblem) can be solved directly by a vector Total Variation (vector Total Variation) algorithm.
Step 3.7: solving the Q sub-problem:
wherein, Y5In the form of a dual-scale variable,representing Frobenius norm, | | · |. luminance*Denotes the nuclear norm, λl0.001 is the parameter balancing each term. μ ═ 0.001 is a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients.
The Q sub-problem can be solved directly by a singular value contraction algorithm.
Step 3.8: and solving the Y subproblem. This subproblem can be directly updated by the following equation:
wherein, Y1,Y2,Y3,Y4And Y5For scale-duality variables, μ ═ 0.001 is a penalty parameter. X(3)In the form of a two-dimensional matrix representation of the coefficients.In the form of a spatial blur matrix, the matrix is,in the form of a spatial blur matrix, the matrix is,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix. O, U, V, W and Q are auxiliary variables.
Step 3.9: obtaining the coefficient tensor by the iterative solution of the subproblemAccording to step 2Further obtaining a high-resolution hyperspectral image
The flow of the optimization solution is shown in table 1.
TABLE 1
And 4, step 4: based on the stepsThe invention will be based on hyperspectral imagesMultispectral imagesFull color imageAnd fusing the three data to obtain a high-resolution hyperspectral image. In order to quantitatively and qualitatively evaluate the fused image, the invention selects CNMF and GSA as comparison methods, and the invention uses the comparison methods to fuse the hyperspectral image and the multispectral image and compare the fusion result of the invention. The invention uses four quality evaluation indexes of ERGAS, SAM, RMSE and PSNR, the quantitative analysis result is shown in Table 2, the bold represents the best result, and the visual comparison result is shown in figures 2-5. Fig. 2 is a graph of the fusion result of the comparison method CNMF, fig. 3 is a graph of the fusion result of the comparison method GSA, fig. 4 is a graph of the fusion result of the present invention, and fig. 5 is a reference image. It can be seen that fig. 2 and 3 are clearly obscured, with much detail obscured, whereas fig. 4 is much clearer than fig. 2 and 3, taking the upper right hand small house as an example, the grid division can be seen in fig. 4, while fig. 2 and 3 are not. Fig. 4 is closer to the reference image fig. 5 than to fig. 2 and 3.
TABLE 2
It can be seen that, because the fusion method of the present invention utilizes the information in the full-color image, compared with the HSI/MSI fusion algorithm and the HSI/PAN fusion algorithm, the fusion result of the present invention has higher definition, each index is also closest to the ideal value, and the present invention is superior to the comparison method in both qualitative and quantitative aspects.
It should be understood that parts of the description not set forth in detail are of prior art.
It should be understood that the above-mentioned embodiments are described in some detail, and not intended to limit the scope of the invention, and those skilled in the art will be able to make alterations and modifications without departing from the scope of the invention as defined by the appended claims.
Claims (5)
1. A hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization is characterized in that,
step 1: acquiring a high-spectrum image with low resolution by a high-spectrum sensor, and acquiring a low-resolution multi-spectrum image and a high-resolution full-color image of the same picture by the multi-spectrum sensor;
step 2: constructing a fusion model among the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image according to the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image in the step 1;
and step 3: solving a coefficient matrix X by an alternative direction multiplier method based on the fusion model in the step 2(3)And further multiplying the coefficient matrix and the subspace matrix to obtain a high-resolution hyperspectral image F(3)。
2. The hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization according to claim 1, wherein the low-resolution hyperspectral image in the step 1 is recorded as tensorWhose 3-mode expansion matrix isIndicates that the image has LhIndividual wave band, Wh×HhA plurality of pixels;
step 1 the low resolution ofThe spectral image is noted as tensorWhose 3-mode expansion matrix isIndicates that the image has LmIndividual wave band, Wm×HmA plurality of pixels;
step 1 the high resolution panchromatic image is recorded as a tensorWhose 3-mode expansion matrix isIndicating that the image has 1 band, Wp×HpA plurality of pixels;
3. The hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization according to claim 1, wherein the fusion model among the low-resolution hyperspectral image, the low-resolution multispectral image and the high-resolution panchromatic image in the step 2 is composed of a hyperspectral image fitting term, a multispectral image fitting term, a dynamic gradient group sparseness regularization term and a low-rank regularization term;
the hyperspectral image fitting item and the multispectral image fitting item are low-resolution hyperspectral images H(3)Low resolution multispectral image M(3)Low resolution highlight constructed based on spatial and spectral degradation relationships with high resolution hyperspectral imagesSpectral image H(3)Spatial down-sampling version of high-resolution equivalent high-spectral image, low-resolution multi-spectral image M(3)The high-spectrum image equivalent to high resolution is a version after spatial down-sampling and spectral down-sampling;
the hyperspectral image fitting term, the multispectral image fitting term, the dynamic gradient group sparse regularization term and the low-rank regularization term form a target energy function related to the coefficient matrix.
4. The hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization according to claim 1, wherein the step 2 of constructing the fusion model among the hyperspectral image, the multispectral image and the panchromatic image specifically comprises the following sub-steps:
step 2.1: low-resolution hyperspectral image H(3)Hyperspectral image F which can be regarded as high resolution(3)Spatially down-sampled version of (a):
H(3)=F(3)BhSh
wherein, the matrixIs a spatial fuzzy matrix;the method is a spatial down-sampling operation and is used for reducing the spatial resolution of the hyperspectral image;
step 2.2: multispectral image M of low resolution(3)Hyperspectral image F which can be regarded as high resolution(3)Spatial down-sampling and spectral down-sampling versions:
M(3)=RmF(3)BmSm
wherein the content of the first and second substances,is a spectral response matrix of a multispectral instrument; matrix arrayIs a spatial fuzzy matrix;the method is a spatial down-sampling operation and is used for reducing the spatial resolution of the hyperspectral image;
step 2.3: the high-resolution panchromatic image P and the high-resolution hyperspectral image F have the same spatial edge information, the gradient difference of the high-resolution panchromatic image P and the high-resolution hyperspectral image F meets the group sparsity characteristic, and a regularization item for dynamic gradient group sparsity can be established:
the high-resolution hyperspectral image is recorded as a tensorWhose 3-mode expansion matrix isIndicates that the image has LhIndividual wave band, Wp×HpA plurality of pixels;
wherein the content of the first and second substances,indicating the replication of high resolution panchromatic images to LmA plurality of wave bands;
b in step 2.1hMatrix, R in step 2.2m,BmThe matrix can be composed of a low-resolution hyperspectral image H(3)Low resolution multispectral image M(3)High resolution full color image P(3)The estimation is obtained, and the estimation method provided by the hyperspectral image fusion method named HySure is used in the invention;
||X||*
step 2.5: converting the high-resolution hyperspectral image into a coefficient matrix model of the high-resolution hyperspectral image in a low-dimensional subspace by combining a product model of the subspace matrix and the coefficient matrix;
step 2.5 the product model of the subspace matrix and the coefficient matrix is:
wherein the extract is3The mode product of the 3 rd mode is shown,the method is characterized in that the method is a subspace matrix, the subspace matrix is represented by vectors of D pure spectral features, the subspace matrix is directly obtained by low-resolution hyperspectral images through image decomposition, and the image decomposition method selected by the invention is vertex component analysis;
Zhang Liangis a coefficient matrix, representingIs represented by a linear combination of vector members in a subspace matrix,representing a pixelCorresponding coefficients, i is the spatial transverse coordinate of the pixel point, and j is the spatial longitudinal coordinate of the pixel point;
step 2.5, the coefficient matrix model of the high-resolution hyperspectral image converted into the high-resolution hyperspectral image in the low-dimensional subspace is as follows:
F(3)=EX(3)
step 2.6: based on the steps 2.1 to 2.5, establishing a fusion model of the three components:
5. The hyperspectral fusion method based on dynamic gradient group sparseness and low-rank regularization according to claim 1 is characterized in that the specific implementation of step 3 comprises the following substeps:
step 3.1: introducing an auxiliary variable O, wherein O is equal to X(3)Bh(ii) a Introducing an auxiliary variable U, wherein U is equal to X(3)Bm(ii) a Introducing an auxiliary variable V, wherein V is equal to X(3)(ii) a Introducing an auxiliary variable W, and satisfying that W is equal to RmEV; introducing an auxiliary variable Q, wherein Q is equal to X(3)(ii) a The three fusion model is expressed as:
s.t.O=X(3)Bh
U=X(3)Bm
V=X(3)
W=RmEV
Q=X(3)
the augmented Lagrange function of the three fusion models is expressed as:
wherein, Y1,Y2,Y3,Y4And Y5For scale Dual variables (Scaled Dual Variable),representing Frobenius norm, | | · |. luminance2,1Representing L2,1 norm, | | · |. non-volatile memory*Denotes the nuclear norm, λm,λφAnd λlParameters of each balance item are respectively; mu is a penalty parameter; x(3)Is a two-dimensional matrix representation of the coefficients; h(3)A two-dimensional matrix representation of the hyperspectral image at low resolution; m(3)A two-dimensional matrix representation of the low-resolution multispectral image;indicating the replication of high resolution panchromatic images to LmA plurality of wave bands;represents a gradient sign;in the form of a spatial blur matrix, the matrix is,for spatially fuzzy matrices,In order to perform a spatial down-sampling operation,in order to perform a spatial down-sampling operation,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix; the optimization problem of the augmented Lagrangian function of the three-part fusion model can be decomposed into an X subproblem, an O subproblem, a U subproblem, a V subproblem, a W subproblem, a Q subproblem and a Y subproblem;
step 3.2: solving the X subproblem:
wherein t in the upper right corner represents the number of iterations and Y1,Y2,Y3And Y5In the form of a dual-scale variable,representing Frobenius norm, mu is a penalty parameter; x(3)Is a two-dimensional matrix representation of the coefficients;in the form of a spatial blur matrix, the matrix is,is a spatial fuzzy matrix;
the solution to the above problem is:
step 3.3: solving an O subproblem:
wherein, Y1In the form of a dual-scale variable,representing Frobenius norm, mu is a penalty parameter; x(3)Is a two-dimensional matrix representation of the coefficients; h(3)A two-dimensional matrix representation of the hyperspectral image at low resolution;in the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,is a subspace matrix;
divide O into OShAndwhereinRepresentation is not represented by matrix ShAnd (3) solving the O subproblem by the selected pixel point:
step 3.4: solving the U sub-problem:
wherein, Y2In the form of a dual-scale variable,denotes the Frobenius norm, λmTo balance the parameters of the various items; mu is a penalty parameter; x(3)Is a two-dimensional matrix representation of the coefficients; m(3)A two-dimensional matrix representation of the low-resolution multispectral image;in the form of a spatial blur matrix, the matrix is,in order to perform a spatial down-sampling operation,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix;
divide U into USmAndwhereinRepresentation is not represented by matrix SmThe solution of the selected pixel point and the U subproblem is as follows:
step 3.5: solving the V subproblem:
wherein, Y3,Y4In the form of a dual-scale variable,representing Frobenius norm, mu is a penalty parameter; x(3)Is a two-dimensional matrix representation of the coefficients;is a spectral response matrix of a multi-spectral instrument,is a subspace matrix;
the solution to the above problem is:
step 3.6: solving the W sub-problem:
wherein, Y4In the form of a dual-scale variable,represents FrobeniusNorm, | · | luminance2,1Denotes L2,1 norm,. lambda.φTo balance the parameters of the various items; mu is a penalty parameter;indicating the replication of high resolution panchromatic images to LmA plurality of wave bands;represents a gradient sign;is a spectral response matrix of a multi-spectral instrument,is a subspace matrix;
let G be W-P, then the W sub-problem becomes
The W subproblem) can be solved directly by a vector Total Variation (vector Total Variation) algorithm;
step 3.7: solving the Q sub-problem:
wherein, Y5In the form of a dual-scale variable,representing Frobenius norm, | | · |. luminance*Denotes the nuclear norm, λlTo balance the parameters of the various items; mu is a penalty parameter; x(3)Is a two-dimensional matrix representation of the coefficients;
the Q sub-problem can be directly solved through a singular value shrinkage algorithm;
step 3.8: solving the Y subproblem; this subproblem can be directly updated by the following equation:
wherein, Y1,Y2,Y3,Y4And Y5Is a scale dual variable, mu is a penalty parameter; x(3)Is a two-dimensional matrix representation of the coefficients;in the form of a spatial blur matrix, the matrix is,in the form of a spatial blur matrix, the matrix is,is a spectral response matrix of a multi-spectral instrument,is a subspace matrix; o, U, V, W and Q are auxiliary variables;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111019135.9A CN113870159A (en) | 2021-09-01 | 2021-09-01 | Hyperspectral fusion method based on dynamic gradient group sparsity and low-rank regularization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111019135.9A CN113870159A (en) | 2021-09-01 | 2021-09-01 | Hyperspectral fusion method based on dynamic gradient group sparsity and low-rank regularization |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113870159A true CN113870159A (en) | 2021-12-31 |
Family
ID=78989033
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111019135.9A Pending CN113870159A (en) | 2021-09-01 | 2021-09-01 | Hyperspectral fusion method based on dynamic gradient group sparsity and low-rank regularization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113870159A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114998167A (en) * | 2022-05-16 | 2022-09-02 | 电子科技大学 | Hyperspectral and multispectral image fusion method based on space-spectrum combined low rank |
CN115880199A (en) * | 2023-03-03 | 2023-03-31 | 湖南大学 | Long-wave infrared hyperspectral and multispectral image fusion method, system and medium |
CN116402726A (en) * | 2023-06-08 | 2023-07-07 | 四川工程职业技术学院 | Denoising fusion method of hyperspectral-multispectral image |
-
2021
- 2021-09-01 CN CN202111019135.9A patent/CN113870159A/en active Pending
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114998167A (en) * | 2022-05-16 | 2022-09-02 | 电子科技大学 | Hyperspectral and multispectral image fusion method based on space-spectrum combined low rank |
CN114998167B (en) * | 2022-05-16 | 2024-04-05 | 电子科技大学 | High-spectrum and multi-spectrum image fusion method based on space-spectrum combined low rank |
CN115880199A (en) * | 2023-03-03 | 2023-03-31 | 湖南大学 | Long-wave infrared hyperspectral and multispectral image fusion method, system and medium |
CN116402726A (en) * | 2023-06-08 | 2023-07-07 | 四川工程职业技术学院 | Denoising fusion method of hyperspectral-multispectral image |
CN116402726B (en) * | 2023-06-08 | 2023-08-22 | 四川工程职业技术学院 | Denoising fusion method of hyperspectral-multispectral image |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113870159A (en) | Hyperspectral fusion method based on dynamic gradient group sparsity and low-rank regularization | |
Xu et al. | Nonlocal patch tensor sparse representation for hyperspectral image super-resolution | |
Wang et al. | Deep blind hyperspectral image fusion | |
Xie et al. | Hyperspectral pansharpening with deep priors | |
Loncan et al. | Hyperspectral pansharpening: A review | |
Xie et al. | Hyperspectral image super-resolution using deep feature matrix factorization | |
Zhang et al. | A super-resolution reconstruction algorithm for hyperspectral images | |
CN110070518B (en) | Hyperspectral image super-resolution mapping method based on dual-path support | |
Fu et al. | Joint camera spectral sensitivity selection and hyperspectral image recovery | |
Li et al. | Hyperspectral pansharpening via improved PCA approach and optimal weighted fusion strategy | |
Dian et al. | Hyperspectral image super-resolution via local low-rank and sparse representations | |
CN107316309B (en) | Hyperspectral image saliency target detection method based on matrix decomposition | |
CN109360147B (en) | Multispectral image super-resolution reconstruction method based on color image fusion | |
CN111223049B (en) | Remote sensing image variation fusion method based on structure-texture decomposition | |
CN108520495B (en) | Hyperspectral image super-resolution reconstruction method based on clustering manifold prior | |
CN113205453B (en) | Hyperspectral fusion method based on space-spectrum total variation regularization | |
Paris et al. | A novel sharpening approach for superresolving multiresolution optical images | |
Lanaras et al. | Advances in hyperspectral and multispectral image fusion and spectral unmixing | |
CN109859153B (en) | Multispectral image fusion method based on adaptive spectrum-spatial gradient sparse regularization | |
Rangzan et al. | Supervised cross-fusion method: a new triplet approach to fuse thermal, radar, and optical satellite data for land use classification | |
Teodoro et al. | Sharpening hyperspectral images using plug-and-play priors | |
Licciardi et al. | Super-resolution of hyperspectral images using local spectral unmixing | |
CN113744134A (en) | Hyperspectral image super-resolution method based on spectrum unmixing convolution neural network | |
Dai et al. | Spatial-spectral representation for x-ray fluorescence image super-resolution | |
Cao et al. | Universal high spatial resolution hyperspectral imaging using hybrid-resolution image fusion |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |