CN106204508A - WorldView 2 remote sensing PAN and multi-spectral image interfusion method based on non-negative sparse matrix - Google Patents

Abstract

nullThe invention provides a kind of WorldView based on non-negative sparse matrix 2 remote sensing PAN and multi-spectral image interfusion method，Relate to image co-registration field，By using a kind of Algorithms of Non-Negative Matrix Factorization that multispectral image is carried out luminance component extraction，Then use HCS conversion that image is merged，Obtain fusion image，Inject in details and be obtained for certain lifting in terms of spectrum holding，Finally give high-quality fusion results，Owing to the extraction of I component be have employed NMF method，Improve the extraction accuracy of luminance component，Relatively contrast algorithm is more reasonable，Make WV 2 satellite fusion image total quality higher，Incorporate in detailed information and all improve in terms of spectrum holding，Subjective assessment can reach consistent with objective analysis results，The fusion image obtained is visual more preferably，Picture is apparent.More traditional remote sensing PAN and multi-spectral image interfusion method is advantageously.

Description

WorldView-2 remote sensing PAN and multi-spectral image based on non-negative sparse matrix melts Conjunction method

Technical field

The present invention relates to a kind of method in image co-registration field, especially remote sensing PAN and multi-spectral image co-registration.

Background technology

In recent years, the wave band number of New Satellite remote sensing multispectral image is being continuously increased, and the resolution of image is also quickly Ground improves.As a example by WorldView-2 (WV-2) satellite, WV-2 satellite launches in 2009, using the teaching of the invention it is possible to provide 8 wave bands 1.84 meters The multispectral image of resolution and the full-colour image of 0.46 meter of resolution of single band.There is compared with conventional satellite image following spy Point: wave band increases, spectrum divides thinner；The spectral coverage of full-colour image narrows, and is allowed to the spectrum with multi light spectrum hands Allot raw bigger change.In remote sensing application, generally require the image with high spatial and high spectral resolution.Image Integration technology is exactly the spatial resolution utilizing the full-colour image of high spatial resolution to remove to improve multispectral image, protects simultaneously as far as possible The spectral characteristic holding multispectral image is constant.WV-2 satellite image represents the development trend of ultra high resolution remote sensing images, with Time also have higher requirement for the fusion of remote sensing images.Change just because of these, make existing fusion method effect not Good.

The most a lot of Remote sensing image fusion algorithms have the process of extract light intensity level, and such as, Brovey conversion, HCS become Changing and MSFIM method, the quality that luminance component extracts will directly affect fusion results, have a great impact the performance of algorithm. Wherein, MSFIM algorithm is a kind of innovatory algorithm based on brightness smothing filtering modulation (SFIM) algorithm, although but this innovatory algorithm Improve incorporating of detailed information, but create compared with the worse spectrum distortion of SFIM algorithm.For reducing the spectrum in MSFIM method Distortion, needs the ratio changing luminance component with full-colour image, makes ratio closer to 1, i.e. should make luminance component and full-colour picture The spectral response characteristic of picture is more like.The extracting method of luminance component is had averaging method, weighted mean approach and each wave band to calculate by other Art averaging method.These methods are all the differentiation of traditional method in fact, and the luminance component of extraction is the most accurate.

Summary of the invention

Existing fusion method is directed to conventional satellite remote sensing multispectral image, for ultra high resolution remote sensing images Speech, is not optimum image interfusion method, and existing luminance component extracting method can not preferably solve details and incorporate With spectrum distortion problem so that the luminance component of extraction is the most accurate.

In order to overcome the deficiencies in the prior art, by the present invention in that with a kind of Non-negative Matrix Factorization (Non-negative Matrix Factorization, NMF) algorithm multispectral image is carried out luminance component extraction, then use HCS conversion right Image merges, and obtains fusion image, and the fusion image of the present invention is injected in details and is obtained for one in terms of spectrum holding Fixed lifting, has finally given high-quality fusion results.

The technical solution adopted for the present invention to solve the technical problems comprises the steps:

Step 1. uses Non-negative Matrix Factorization method to extract I component

First, by the multispectral image X of full-colour image Pan and eight wave bands₁,X₂,…,X₈Stretch by row, obtain P, M₁, M₂,…,M₈Vector, then form matrix V to be decomposed by formula (1), i.e.

V=[P, M₁,M₂,...,M₈] (1)

Wherein, P, M₁,M₂,…,M₈It is respectively full-colour image Pan and eight wave band multispectral image X₁,X₂,…,X₈During computing Image array is pulled into corresponding column vector；

Next, then make

[P,M₁,M₂,...,M₈]=WH (2)

Wherein W is n*r matrix, and n is the line number of matrix W, and r is the columns of matrix W, and H is r*9 matrix, and after decomposition, W is one Column vector, reverts to the I component that image array is the most obtained by W；

Step 2. uses Pan component to mate I component

Order

P^'2=(Pan)² (3)

Wherein Pan is full-colour image, i.e. replaces Pan variable with P' variable, and then, below equation i.e. uses P^'2Component mates The I of step 1 gained²Component:

$P^{\′\′2}=\frac{{\mathrm{\σ}}_0}{{\mathrm{\σ}}_1}(P^{\′2}-{\mathrm{\μ}}_1+{\mathrm{\σ}}_1)+{\mathrm{\μ}}_0-{\mathrm{\σ}}_0---\left(4\right)$

Wherein μ₀、σ₀It is respectively I²Average and standard variance, μ₁、σ₁It is respectively P^'2Average and standard variance, P^”2For I after joining²Component；

$I_{adj}=\sqrt{P^{\′\′2}}---\left(5\right)$

Use I_adjComponent replaces P " component represents the I component after coupling；

Step 3. uses hypersphere color space resolution to merge HCS conversion and finally gives eight new wave band components

First, the multispectral image X to eight wave bands₁,X₂,...,X₈Carry out HCS direct transform and obtain corresponding angle component φ 1, φ 2 ..., φ 7, HCS direct transform is as follows:

${\mathrm{\φ}}_1=\arctan \left(\frac{\sqrt{{X_8}^2+{X_7}^2+\mathrm{...}+{X_2}^2}}{X_1}\right)$

${\mathrm{\φ}}_6=\arctan \left(\frac{\sqrt{{X_8}^2+{X_7}^2}}{X_6}\right)$

${\mathrm{\φ}}_7=\arctan \left(\frac{X_8}{X_7}\right)---\left(6\right)$

Secondly, the I component I after the coupling that step 2 is tried to achieve_adjWith angle component φ 1, φ 2 ..., φ 7 does HCS contravariant Get eight new wave band component X' in return₁,X'₂,…,X'₈, HCS inverse transformation is as follows:

X₁'=I_adjcosφ₁

X₂'=I_adjsinφ₁cosφ₂

X₇'=I_adjsinφ₁sinφ₂...sinφ₆cosφ₇

X₈'=I_adjsinφ₁sinφ₂...sinφ₆sinφ₇ (7)

The each Band fusion of step 4.

X' in selecting step 3₁,X'₂,…,X'₈In any three wave bands merge, will be directly placed into by three band images In RGB triple channel, i.e. can get fusion image.

The invention has the beneficial effects as follows owing to the extraction of I component be have employed NMF method, improve the extraction of luminance component In precision, relatively the innovatory algorithm MSFIM algorithm of contrast algorithm SFIM, the extraction algorithm of I component is more reasonable so that WV-2 satellite melts Close overall picture quality higher, incorporate in detailed information and all improve in terms of spectrum holding, subjective assessment and objective Analysis result can reach consistent, and additionally the inventive method relatively contrasts fusion image visuality that algorithm MSFIM algorithm obtains more Good, picture is apparent.More traditional remote sensing PAN and multi-spectral image interfusion method is advantageously.

Accompanying drawing explanation

Fig. 1 is the technology path block diagram of the present invention.

Fig. 2 is multispectral image X of the present invention₂,X₃,X₅Fusion results.

Detailed description of the invention

The present invention is further described with embodiment below in conjunction with the accompanying drawings.

Step 1. uses Non-negative Matrix Factorization (NMF) method to extract I component

Non-negative Matrix Factorization, as the means of numerical analysis of a relative maturity, digs in graphical analysis, text cluster, data The aspects such as pick, speech processes are widely applied.In view of from multispectral image extract I component will as far as possible with entirely The spectral response of color image is consistent, therefore extracts at NMF in the method for I component with full-colour image and the multispectral figure of eight wave bands As forming matrix V to be decomposed.

First, by the multispectral image X of full-colour image Pan and eight wave bands₁,X₂,…,X₈Stretch by row, obtain P, M₁, M₂,…,M₈Vector, then form matrix V to be decomposed by formula (1), i.e.

V=[P, M₁,M₂,...,M₈] (1)

Wherein, P, M₁,M₂,…,M₈It is respectively full-colour image Pan and eight wave band multispectral image X₁,X₂,…,X₈During computing Image array is pulled into corresponding column vector；

Next, then make

[P,M₁,M₂,...,M₈]=WH (2)

Wherein W is n*r matrix, and n is the line number of matrix W, and r is the columns of matrix W, and H is r*9 matrix.W after decomposition is one Individual column vector, reverts to the I component that image array is the most obtained by W.

Wherein, W is n*r matrix, and n is the line number of matrix W, and r is the columns of matrix W, and H is r*9 matrix, it is contemplated that extraction I component to match with full-colour image, now should extract the global characteristics W of full-colour image and multispectral image, the W after decomposing Revert to the I component that image array is the most obtained；

Step 2. uses Pan component to mate I component

Order

P^'2=(Pan)² (3)

Wherein Pan is full-colour image, i.e. replaces Pan variable with P' variable, and then, below equation i.e. uses P^'2Component mates The I of step 1 gained²Component:

$P^{\′\′2}=\frac{{\mathrm{\σ}}_0}{{\mathrm{\σ}}_1}(P^{\′2}-{\mathrm{\μ}}_1+{\mathrm{\σ}}_1)+{\mathrm{\μ}}_0-{\mathrm{\σ}}_0---\left(4\right)$

Wherein μ₀、σ₀It is respectively I²Average and standard variance, μ₁、σ₁It is respectively P^'2Average and standard variance, P^”2For I after joining²Component；

$I_{adj}=\sqrt{P^{\′\′2}}---\left(5\right)$

Use I_adjComponent replaces P " component represents the I component after coupling；

Step 3. uses hypersphere color space resolution to merge HCS conversion and finally gives eight new wave band components

Fusion method HCS (Hyperspherical Color Space towards WorldView-2 satellite image Resolution Merge) conversion wave band number is not limited, be therefore suitable for Multi-Band Remote Sensing Images merge.In HCS converts, Angle variables determines the spectral information of image, and I determines the monochrome information of image, the change of the I variable spectral information to image Not impact, this is the key point of HCS conversion.

First, the multispectral image X to eight wave bands₁,X₂,...,X₈Carry out HCS direct transform and obtain corresponding angle component φ 1, φ 2 ..., φ 7, HCS direct transform is as follows:

${\mathrm{\φ}}_1=\arctan \left(\frac{\sqrt{{X_8}^2+{X_7}^2+\mathrm{...}+{X_2}^2}}{X_1}\right)$

${\mathrm{\φ}}_6=\arctan \left(\frac{\sqrt{{X_8}^2+{X_7}^2}}{X_6}\right)$

${\mathrm{\φ}}_7=\arctan \left(\frac{X_8}{X_7}\right)---\left(6\right)$

Secondly, the I component I after the coupling that step 2 is tried to achieve_adjWith angle component φ 1, φ 2 ..., φ 7 does HCS contravariant Get eight new wave band component X' in return₁,X'₂,…,X'₈, HCS inverse transformation is as follows:

X₁'=I_adjcosφ₁

X₂'=I_adjsinφ₁cosφ₂

X₇'=I_adjsinφ₁sinφ₂...sinφ₆cosφ₇

X₈'=I_adjsinφ₁sinφ₂...sinφ₆sinφ₇ (7)

The each Band fusion of step 4.

X' in selecting step 3₁,X'₂,…,X'₈In any three wave bands merge, will be directly placed into by three band images In RGB triple channel, i.e. can get fusion image.

For just verifying WorldView-2 remote sensing PAN and multi-spectral New Image Fusion based on non-negative sparse matrix Really property and Optimality, carry out following experiment.Fusion method used in experiment is HCS conversion.The I component of two kinds of fusion methods carries Access method is respectively adopted MSFIM method and NMF method.Experimental data is one group of real WV-2 image, is shooting on April 3rd, 2011 Sydney Australia, for verification algorithm and can show that image is used for subjective assessment clearly, experimental data uses image A part, size is 300*300pixels, gray level 256.For obtaining reference picture during evaluation result, the first light how Spectrogram picture is down sampled to original 1/4th, then is upsampled to original size, and the most original multispectral image can be as ginseng Examine image.

The full-colour image that in Fig. 2, (a) is original is down sampled to the image after 1/4th of original image；Fig. 2 (b) is former The multispectral image begun, contrasts for experimental result.Fig. 2 (c) is original extraction I component method in HCS conversion fusion method Fusion results；Fig. 2 (d) is the fusion results extracting I component herein by NMF method.Observe Fig. 2, Fig. 2 (d) spectral signature with Fig. 2 (b) is the most close, i.e. context of methods fusion results is better than original method on spectrum keeps, it can be seen that, institute of the present invention The fusion results obtained in the constant situation of spectral preservation characteristic, the incorporating more traditional fusion method and to get well of detailed information, and The fusion results of the present invention also improves on spectrum keeps.

The present invention chooses space correlation coefficient (spatial correlate coefficient, sCC), correlation coefficient (correlate coeffcient, CC), average gradient (average gradient, AG), comentropy (information Entropy, IE) etc. conventional objective evaluation index fusion results is carried out objective evaluation, these desired values the most greatly show merge knot Fruit is the best.

Table 1 objective evaluation result

Band1 ... Band8 represents eight wave bands of multispectral image.Table 1 can be seen that, the every of the inventive method comments Valency index is the biggest, and this explanation present invention shows relatively traditional method better performance in remote sensing image fusion.

Claims (1)

1. a WorldView-2 remote sensing PAN and multi-spectral image interfusion method based on non-negative sparse matrix, its feature exists In comprising the steps:

Step 1. uses Non-negative Matrix Factorization method to extract I component

First, by the multispectral image X of full-colour image Pan and eight wave bands₁, X₂..., X₈Stretch by row, obtain P, M₁, M₂..., M₈ Vector, then form matrix V to be decomposed by formula (1), i.e.

V=[P, M₁, M₂..., M₈] (1)

Wherein, P, M₁, M₂..., M₈It is respectively full-colour image Pan and eight wave band multispectral image X₁, X₂..., X₈Will figure during computing As matrix pulls into corresponding column vector；

Next, then make

[P, M₁, M₂..., M₈]=WH (2)

Wherein W is n*r matrix, and n is the line number of matrix W, and r is the columns of matrix W, and H is r*9 matrix, after decomposition W be one row to Amount, reverts to the I component that image array is the most obtained by W；

Step 2. uses Pan component to mate I component

Order

P′²=(Pan)² (3)

Wherein Pan is full-colour image, i.e. replaces Pan variable with P ' variable, and then, below equation i.e. uses P '²Component coupling step The I of 1 gained²Component:

$P^{\′\′2}=\frac{{\mathrm{\σ}}_0}{{\mathrm{\σ}}_1}(P^{\′2}-{\mathrm{\μ}}_1+{\mathrm{\σ}}_1)+{\mathrm{\μ}}_0-{\mathrm{\σ}}_0---\left(4\right)$

Wherein μ₀、σ₀It is respectively I²Average and standard variance, μ₁、σ₁It is respectively P '²Average and standard variance, P "²After coupling I²Component；

$I_{adj}=\sqrt{P^{\′\′2}}---\left(5\right)$

Use I_adjComponent replaces P, and " component represents the I component after coupling；

Step 3. uses hypersphere color space resolution to merge HCS conversion and finally gives eight new wave band components

First, the multispectral image X to eight wave bands₁, X₂..., X₈Carry out HCS direct transform and obtain corresponding angle component φ 1, φ 2 ..., φ 7, HCS direct transform is as follows:

${\mathrm{\φ}}_1=\arctan \left(\frac{\sqrt{{X_8}^2+{X_7}^2+\mathrm{...}+{X_2}^2}}{X_1}\right)$

${\mathrm{\φ}}_6=\arctan \left(\frac{\sqrt{{X_8}^2+{X_7}^2}}{X_6}\right)$

${\mathrm{\φ}}_7=\arctan \left(\frac{X_8}{X_7}\right)---\left(6\right)$

Secondly, the I component I after the coupling that step 2 is tried to achieve_adjWith angle component φ 1, φ 2 ..., φ 7 does HCS inverse transformation and obtains To eight new wave band component X '₁, X '₂..., X '₈, HCS inverse transformation is as follows:

X′₁=I_adjcosφ₁

X₂'=I_adjsinφ₁cosφ₂

X₇'=I_adj sinφ₁sinφ₂...sinφ₆cosφ₇

X₈'=I_adjsinφ₁sinφ₂...sinφ₆sinφ₇ (7)

The each Band fusion of step 4.

X ' in selecting step 3₁, X '₂..., X '₈In any three wave bands merge, RGB will be directly placed into by three band images In triple channel, i.e. can get fusion image.