CN1472544A - Remote sensing image picture element and characteristic combination optimizing mixing method - Google Patents

Abstract

The method carries on the optimum mixture in picture element grade with two characteristic indexes of space detail information and spectral information at IHS space by utilizing low-frequency baseband coefficient obtained from wavelet multilayer decomposition of multispectral image intensity component and low-frequency baseband coefficient obtained from corres ponding multilayer wavelet decomposition of high resolution image. The new intensity component is obtained by carrying on characteristic grade mixture of high-frequency coefficient decomposed with wavelet and then carrying on wavelet inverse transform for wavelet coefficient. Finally, the mixed image is obtained after IHS inverse transform is done.

Description

Remote sensing image pixel and characteristic binding optimum fusion method

Technical field:

The present invention relates to a kind of remote sensing image pixel and characteristic binding optimum fusion method, decompose time-frequency characteristic and IHS (intensity I ntensity-colourity Hue-saturation degree Saturation) the conversion fusion method that has in conjunction with the wavelet multiresolution rate, carry out Pixel-level and feature level associating optimum fusion, be the core technology that remote sensing image merges, in fields such as all kinds of military or civilian Remote Sensing Information Processing System, digital city space information system, all can be widely used.

Background technology:

The multispectral remote sensing image of panchromatic remote sensing image of effective fusion high resolving power and low resolution how, two characteristic indexs of space detailed information and spectral information in the balanced fusion results are one of multi-source remote sensing visual fusion Study on Technology focuses.

The IHS fusion method that people such as Haydn at first propose is one of classical practical algorithm.This method transforms to IHS space by the IHS conversion from rgb space with multispectral image, simultaneously high-resolution panchromatic image is carried out linear stretch, the feasible average of back image and the strength component I in variance and the IHS space of stretching ₀Consistent.Then, the high resolution image after stretching as new strength component, is arrived former rgb space with H and S component according to IHS inverse transformation fortran.Like this, make that the image after merging had both had higher spatial resolution, kept former low resolution multispectral image identical colourity and saturation degree simultaneously again.Yet, the IHS fusion method of this classics exists certain defective, because the data of different-waveband have different spectral patterns, the IHS fusion method has been twisted original spectral characteristic, produced spectrum degradation phenomena in various degree, thereby be unfavorable for the correct identification and the classification of image, particularly for the visual fusion of the multisensor remote sensing image of phase simultaneously not, the IHS fusion method can't make the tone of the former multispectral image of color harmony that merges image be consistent, it is this because the conversion of spectral information has caused image can not be used for atural object identification and inverting.People such as Te-Ming have carried out mathematical proof in the IHS space, have discussed the defective of IHS fusion method, and the conclusion that obtains is: although be used to replace strength component I ₀High-resolution panchromatic image I _NewBefore replacement, carried out the coupling of the statistical property of image, but matching error δ=I _New-I has caused colored distortion.

Summary of the invention:

The objective of the invention is to deficiency at above-mentioned IHS conversion integration technology, a kind of remote sensing image pixel and characteristic binding optimum fusion method are provided, can improve the spatial resolution that merges the back image, can reduce colored distortion again, effectively improve the spectral information index that merges image.

For realizing such purpose, the present invention is in the IHS space, the strength component of multispectral image is decomposed the low frequency base band coefficient that obtains through the small echo multilayer, obtaining low frequency base band coefficient with high resolution image through corresponding multilayer wavelet decomposition carries out merging with the optimum Pixel-level of spatial detail information and two characteristic indexs of spectral information, high frequency coefficient after wavelet decomposition is carried out the feature level to be merged, then wavelet coefficient is carried out corresponding wavelet inverse transformation, obtain new strength component, the image after carrying out again obtaining merging after the IHS inverse transformation.

Because wavelet transformation has good frequency division characteristic at transform domain, the statistical property of wavelet coefficient has reflected the notable features such as edge, line and zone of remote sensing image, and the present invention introduces multiresolution analysis (Multi-resolution Analysis) method of wavelet transformation in the fusion of multispectral remote sensing image of panchromatic remote sensing image of high resolving power and low resolution.

Method of the present invention comprises following concrete steps:

(1) multispectral image B is carried out the IHS conversion, obtain colourity H, saturation degree S and strength component I respectively, then I component is carried out 3 layers of wavelet decomposition in the IHS color space;

(2) high resolution image A is carried out linear stretch and histogram coupling, carry out wavelet decomposition then, decompose the number of plies and also be 3 layers;

(3) determine one 3 * 3 spatial domain window, respectively the I component of image A and the subband wavelet coefficient of image B are added up, try to achieve average μ (2 ^j) and variance D (2 ^j);

(4) the high-frequency sub-band coefficient of corresponding layers of resolution carries out the fusion of feature level according to (1) formula:

(1) in the formula, 2 ^jBe the wavelet decomposition number of plies, W ^k(2 ^j, x is 2 y) ^jThe high-frequency sub-band coefficient fusion results that obtains under the resolution; W _A ^k(2 ^j, x, y) and W _B ^k(2 ^j, x y) is the corresponding high-frequency sub-band coefficient of image A and I ', D _A ^k, D _B ^kBe respectively so that (x y) is the variance of 3 * 3 spatial domain window of center pixel;

(5) to the low frequency base band coefficient of the subband wavelet coefficient of the I component of image A and image B, carry out optimum Pixel-level according to (2) formula and merge the k of weight coefficient _OptAsk excellent according to spectral information evaluation index and spatial resolution evaluation index:

A(2 ^j，x，y)＝k ₁A _A(2 ^j，x，y)+k ₂A _B(2 ^j，x，y)??????????????(2)

(2) in the formula, A _A(2 ^j, x, y), A _B(2 ^j, x y) is respectively the correspondence 2 of image A and I ' ^jThe low frequency base band data of resolution, k ₁, k ₂For needs are asked excellent weight coefficient.According to normalization requirement, k ₁, k ₂Satisfy k ₁+ k ₂=1, promptly ask excellent Determination of Weight Coefficient can be summed up as the k that satisfies objective function _Opt=k ₁=1-k ₂

(6) to the low frequency base band coefficient of the wavelet coefficient of the Pixel-level optimum fusion that obtains, and each high frequency wavelet coefficient that the feature level of carrying out merges carries out corresponding wavelet inverse transformation, obtains new strength component I ';

(7) I ', H, S are carried out the IHS inverse transformation, the image C after obtaining merging.

The present invention combines the characteristics of IHS fusion method and small echo fusion method, ask the feature level of excellent fusion and high-frequency sub-band to merge by respectively the weight coefficient of small echo base band coefficient having been carried out Pixel-level, its beneficial effect is presented as: make the image after merging both reach the highest spatial resolution, reduced colored distortion simultaneously again to greatest extent.Balanced space detailed information and two characteristic indexs of spectral information in the fusion results have effectively been improved the spectral information index that merges image.

Description of drawings:

Fig. 1 is the present invention-based on the remote sensing images pixel and the characteristic binding optimal characteristics fusion method process flow diagram of wavelet statistic characteristics.

Fig. 2 is the contrast of remote sensing image fusion results of the present invention and IHS method.Wherein, Fig. 2 (a) is multispectral remote sensing image (512 * 512); Fig. 2 (b) is the panchromatic remote sensing image of high resolving power (512 * 512); Fig. 2 (c) is the fusion results (512 * 512) of IHS method, and Fig. 2 (d) is fusion results of the present invention (512 * 512).

Fig. 3 is that the feature level of each sub-band coefficients after the wavelet decomposition merges.Wherein, Fig. 3 (a)～(c) is respectively the horizontal high frequency imaging of panchromatic high resolution image (P-Panchromatic) and multispectral image (M-Multispectral) and feature level fusion results image (F-Fused); Fig. 3 (d)～(f) is respectively panchromatic, multispectral vertical high frequency image and feature level fusion results image thereof; Fig. 3 (g)～(i) is respectively panchromatic, multispectral oblique line high frequency imaging and feature level fusion results image thereof.

Fig. 4 is a performance evaluation index curve of the present invention.

Embodiment:

In order to understand technical scheme of the present invention better, be further described below in conjunction with the embodiment of accompanying drawing.

Detailed process of the present invention such as Fig. 1.The present invention chooses a multispectral remote sensing image A (512 * 512) as Fig. 2 (a), and the panchromatic remote sensing image B of high resolving power (512 * 512) behind the strict registration of A, B, implements following steps as Fig. 2 (b):

1, multispectral image B is carried out the IHS conversion, obtain colourity H, saturation degree S and strength component I respectively, then I component is carried out 3 layers of wavelet decomposition in the IHS color space;

2, high resolution image A is carried out linear stretch and histogram coupling, carry out wavelet decomposition then, decompose the number of plies and also be 3 layers;

3, determine one 3 * 3 spatial domain window, respectively the I of image A and the subband wavelet coefficient of image B are added up, try to achieve average μ (2 ^j) and variance D (2 ^j);

4, determine the high-frequency sub-band coefficient of corresponding layers of resolution according to (1) formula, carry out the feature level and merge; Can obtain the feature level fusion results of corresponding layers of resolution, as shown in Figure 3; Fig. 3 is that the feature level of each sub-band coefficients after the wavelet decomposition merges.Fig. 3 (a)～(c) is respectively the horizontal high frequency imaging of panchromatic high resolution image (P-Panchromatic) and multispectral image (M-Multispectral) and feature level fusion results image (F-Fused); Fig. 3 (d)～(f) is respectively panchromatic, multispectral vertical high frequency image and feature level fusion results image thereof; Fig. 3 (g)～(i) is respectively panchromatic, multispectral oblique line high frequency imaging and feature level fusion results image thereof;

5, to the low frequency base band coefficient of the subband wavelet coefficient of the I of image A and image B, carry out Pixel-level according to (2) formula and merge weight coefficient k _OptAsk excellent according to the objective function shown in (3) formula:

F(k _opt)＝F(k ₁，k ₂)

＝Max(E _SP(k ₁，k ₂)；?E _HF(k ₁，k ₂))OR?Max(E _SP(k ₁，k ₂)；AG(k ₁，k ₂))??(3)

＝{E _SP(k _opt)；E _HF(k _opt)}OR{E _SP(k _opt)；AG(k _opt)}

D:g(k _opt):k _opt＝k ₁＝1-k ₂(0≤k _opt≤1)

k _opt∈D∈R

Wherein, fused images optimizing evaluation index is defined as follows respectively:

(i) spectral information evaluation index

Utilize the degree of correlation of fused images and multispectral image to define the evaluation index of spectral information.

Make that f is a fused image, f ₀Be multispectral image.The definition spectral information is estimated E _SPIndex is as (4) formula. $E_{\mathrm{SP}}=\mathrm{Corr}(f,f_0)=\frac{{\mathrm{\Σ}}_{j=1}^{\mathrm{npix}}(f_i-\stackrel{\‾}{f})(f_{0j}-{\stackrel{\‾}{f}}_0)}{\sqrt{{\mathrm{\Σ}}_{j=1}^{\mathrm{npix}}(f_i-\stackrel{\‾}{f}{)}^2{(f_{0j}-{\stackrel{\‾}{f}}_0)}^2}}........\left(4\right)$

Wherein, npix is the number of pixel in the image, f and f ₀The gray average of presentation video, degree of correlation Corr (f, f ₀) reflected image f and f ₀Similarity degree.

(ii) spatial resolution evaluation index

Utilize the high fdrequency component and the degree of correlation between the high-definition picture high fdrequency component of the gray scale of fused images correspondence to come definition space resolution index.Make f _HBe the high resolving power panchromatic image.At first the image that merges is converted into gray level image, carries out wavelet decomposition then, obtain four component (f of fused images ^a, f ^h, f ^v, f ^d), represent the low frequency component of fused images, the high fdrequency component of horizontal direction, the high fdrequency component of vertical direction and the high fdrequency component of diagonal respectively.Equally also can obtain four component (f of high-definition picture wavelet decomposition _H ^a, f _H ^h, f _H ^v, f _H ^d).Definition space resolution evaluation index is shown in (5) formula: $E_{\mathrm{HF}}=\frac{\mathrm{Corr}(f^h,f_H^h)+\mathrm{Corr}(f^v,f_H^v)+\mathrm{Corr}(f^d,f_H^d)}{3}......\left(5\right)$

In addition, the most direct index of improvement of image definition quality is the average gradient of image, it has reflected the readability of image, also reflects minor detail contrast and texture transformation feature in the image simultaneously, and definition average gradient (Average grads) evaluation index is shown in (6) formula: $\mathrm{AG}=\frac{1}{M*N}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{[\mathrm{\Δxf}(i,j)}^2+\mathrm{\Δyf}{(i,j)}^2{]}^{1/2}......\left(6\right)$

(6) in the formula, and Δ xf (i, j), (i j) is respectively pixel (i, j) first order difference on the x/y direction to Δ yf;

Weight coefficient k according to the fusion of the base band data in the optimizing evaluation index correction fusion criterion of fused images ₁, k ₂Fusion weights k along with high resolution image low frequency base band coefficient ₁Increase, spatial resolution evaluation index E _HFIncrease spectral information evaluation index E with AG thereupon _SPReduce.Therefore, the k in (3) formula _OptBe to make objective function reach maximum weight coefficient, promptly make E _SP, E _HFPerhaps E _SP, AG reaches simultaneously maximum weight coefficient.k _OptSpan is [0,1].

According to weight coefficient k _OptAsk excellent objective function, in interval [0,1], along with weight coefficient merges weights k ₁Increase, obtain spatial resolution evaluation index E by (7) formula _HFWith AG and spectral information evaluation index E _SPCurve, intersections of complex curve is the k of optimum weight coefficient _Opt, as shown in Figure 4. $E\left(i\right)=\frac{E\left(i\right)-\mathrm{MinE}\left(i\right)}{\mathrm{Max}[E\left(i\right)-\mathrm{Min}(E\left(i\right)]}......\left(7\right)$

(7) in the formula, E (i) is an evaluation index, and i is weight coefficient k ₁The number of times of optimizing.Figure 4 shows that the E that obtains with optimizing step-length 0.001 in [0,1] _SP, E _HF, the AG curve.Can see the E after the normalization _SP, E _HF, AG is non-linear, E _HFWith merging weights k ₁Increase and the trend that increases greater than the trend of the increase of AG.Satisfy the weight coefficient k that asks excellent objective function _OptBe E _SPAnd E _HF, E _SPIntersection point k with AG _Opt1=0.474 or k _Opt2=0.586 (convergence precision δ≤0.001).Therefore, k _Opt1And k _Opt2Balanced space detailed information and two characteristic indexs of spectral information in the fusion results reach optimum.In other words, according to (4)～(6) formula, carry out objective qualitative evaluation calculating and can obtain optimum weight coefficient k _Opt1Or k _Opt2, it makes the image after merging both reach the highest spatial resolution, has reduced colored distortion simultaneously again to greatest extent;

6, to the low frequency base band coefficient of the wavelet coefficient of the Pixel-level optimum fusion that obtains, and each high frequency wavelet coefficient that the feature level of carrying out merges carries out corresponding wavelet inverse transformation, obtains new strength component I ';

7, I ', H, S are carried out the IHS inverse transformation, the image C after obtaining merging;

Fusion results compares: Fig. 2 (c) is the fusion results (512 * 512) of IHS method, and Fig. 2 (d) is fusion results of the present invention (512 * 512).

Performance evaluation: table 1 when choosing optimum weight coefficient, two index quantification values of the related coefficient of fusion results and average gradient, and contrast with the IHS method.

The performance evaluation index of table 1 fusion results relatively

	Wave band	Related coefficient		Average gradient
						????IHS	????R	????0.5624		????10.6146
????G	????0.4644		????11.3663
					????B		????0.5060		????9.5875
The inventive method	????k i	k opt1＝k 1＝0.474；k 2＝0.526		k opt1＝k 1＝0.586；k 2＝0.414
						Related coefficient	Average gradient	Related coefficient	Average gradient
		????R	??0.8488	??8.9448	??0.7950					??9.0954
	????G					??0.7691	??9.4102	??0.6992	??9.6055
		????B	??0.8247	??7.9353	??0.7626					??8.1007

Claims (3)

1, a kind of remote sensing image pixel and characteristic binding optimum fusion method is characterized in that comprising following concrete steps:

(1) multispectral image B is carried out the IHS conversion, obtain colourity H, saturation degree S and strength component I respectively, then I component is carried out 3 layers of wavelet decomposition in the IHS color space;

(2) high resolution image A is carried out linear stretch and histogram coupling, carry out wavelet decomposition then, decompose the number of plies and also be 3 layers;

(3) determine one 3 * 3 spatial domain window, respectively the I component of image A and the subband wavelet coefficient of image B are added up, try to achieve average μ (2 ^j) and variance D (2 ^j);

(4) the high-frequency sub-band coefficient of corresponding layers of resolution is according to formula Carry out the feature level and merge, in the formula, 2 ^jBe the wavelet decomposition number of plies, W ^k(2 ^j, x is 2 y) ^jThe high-frequency sub-band coefficient fusion results that obtains under the resolution, W _A ^k(2 ^j, x, y) and W _B ^k(2 ^j, x y) is the corresponding high-frequency sub-band coefficient of image A and I ', D _A ^k, D _B ^kBe respectively so that (x y) is the variance of 3 * 3 spatial domain window of center pixel;

(5) to the low frequency base band coefficient of the subband wavelet coefficient of the I component of image A and image B, according to formula

A (2 ^j, x, y)=k ₁A _A(2 ^j, x, y)+k ₂A _B(2 ^j, x y) carries out optimum Pixel-level and merges the k of weight coefficient _OptAsk excellent according to spectral information evaluation index and spatial resolution evaluation index, in the formula, A _A(2 ^j, x, y), A _B(2 ^j, x y) is respectively the correspondence 2 of image A and I ' ^jThe low frequency base band data of resolution, k ₁, k ₂For needs are asked excellent weight coefficient, k ₁, k ₂Satisfy k ₁+ k ₂=1, promptly ask excellent Determination of Weight Coefficient to be summed up as the k that satisfies objective function _Opt=k ₁=1-k ₂

(6) to the low frequency base band coefficient of the wavelet coefficient of the Pixel-level optimum fusion that obtains, and each high frequency wavelet coefficient that the feature level of carrying out merges carries out corresponding wavelet inverse transformation, obtains new strength component I ';

(7) I ', H, S are carried out the IHS inverse transformation, the image C after obtaining merging.

2, as said remote sensing image pixel of claim 1 and characteristic binding optimum fusion method, it is characterized in that weight coefficient k _OptWhen asking excellent, ask excellent objective function according to formula

F(k _opt)＝F(k ₁，k ₂)

＝Max(E _SP(k ₁，k ₂)；E _HF(k ₁，k ₂))OR?Max(E _SP(k ₁，k ₂)；AG(k ₁，k ₂))

＝{E _SP(k _opt)；E _HF(k _opt)}OR{E _SP(k _opt)；AG(k _opt)}

D:g(k _opt):k _opt＝k ₁＝1-k ₂(0≤k _opt≤1)

k _opt∈D∈R

Determine.

3, as said remote sensing image pixel of claim 1 and characteristic binding optimum fusion method, it is characterized in that weight coefficient k _OptWhen asking excellent, utilize formula according to spectral information evaluation and spatial resolution evaluation index $E\left(i\right)=\frac{E\left(i\right)-\mathrm{MinE}\left(i\right)}{\mathrm{Max}[E\left(i\right)-\mathrm{Min}(E\left(i\right)]}$

Obtain the curve of spectral information evaluation and spatial resolution evaluation index, intersections of complex curve is the k of optimum weight coefficient _Opt, in the formula, E (i) is an evaluation index, i is weight coefficient k ₁The number of times of optimizing.