CN1540586A - Method for picking up and comparing spectral features in remote images - Google Patents

Abstract

The method includes following steps: after completing decorrelation among multispectral image spectrum from principal component transformation, determining center vectors of each cluster in image; according to 'nearby principle' assigning each vector element of image into cluster represented by its center; carrying out mergering and splitting operation for result of cluster to make tend result be balanced so as to obtain of final result of picking up features. Secularization is carried out for picking up and comparing operation so as to make degree of similarity or dissimilarity be expressed in qualitative and quantitative. Advantages of the invention are: reducing amount of calculation easy of parameters control since iteration self-organizing data analytic technique algorithm 'ISODATA' is adopted.

Description

The spectral signature of remote sensing images is extracted and comparative approach

Technical field:

The spectral signature that the present invention relates to a kind of remote sensing images is extracted and comparative approach, relate in particular to a kind of based on improving ISODATA (Iterative Self-Organizing Data Analysis Technique, iteration self-organization data analysis technique A) spectral signature of the remote sensing images of algorithm is extracted and comparative approach, all extensive utilization can be arranged in content-based view data library searching, image classification and pattern-recognition.

Background technology:

CBIR is exactly the technology that the content (spectrum, texture, shape etc.) by analysis image is retrieved image.Be characterized in: 1, retrieving has interactivity, and the user can participate in retrieving; 2, introduced the auxiliary notion of feature database and knowledge; What 3, it was more paid attention to is information fast query.

Remote sensing images are carried out spectral signature is extracted and feature relatively is the important component part that the remote sensing image data storehouse is retrieved, by Chen Hua etc. propose based on principal component transformation (Karhunen-Loeve transformation, Karhunen Loeve conversion) and the multi-spectral remote sensing image feature extracting method (Chen Hua of ISODATA clustering algorithm, pacify refined, Chen Shuhai, the utilization of Liu Yongchang .KL conversion in the multispectral image cluster. infrared and laser engineering, 2002,30 (2): 79-82.) obtained good effect, but, because the data volume of multispectral image is very big, makes that the calculated amount of this algorithm is very big, on the other hand, parameter in this algorithm in the ISODATA cluster is too much, and being difficult to control, the feature that this algorithm extracted does not compare simultaneously, can't apply to the view data library searching.

Summary of the invention:

The objective of the invention is to deficiency at the prior art existence, provide a kind of based on remote-sensing image spectrum feature extraction that improves the ISODATA algorithm and method relatively, can be under the situation that guarantees former algorithm effect, (especially at the remote sensing images wave band more for a long time) significantly reduces calculated amount, reduce parameters needed, make parameter control easier, and feature method relatively is provided, make algorithm can be used for the view data library searching.

For realizing such purpose, innovative point of the present invention is to keep the achievement that former algorithm obtains in initial clustering, the scope that iterates is dwindled, and allow and iterate nature and finish; The preset parameter of the absolute property of former algorithm is changed into the dynamic parameter of ratio character; The design feature comparison algorithm makes that the different degree of phase Sihe between image obtains qualitative and quantitative embodiment.Finish between multispectral image spectrum after the decorrelation with principal component transformation, determine each cluster centre vector of image, according to " nearby principle " each vector element of image is assigned in the class of each cluster centre representative and gone.Clustering result is merged and splitting operation, make clustering result tend to be balanced, obtain the final result of feature extraction.After two width of cloth images are finished feature extraction, their result is carried out feature relatively, obtain their difference standard.

Implementation method of the present invention comprises that mainly the initial clustering of image vector element, merging and division, the feature of class compare three basic steps:

1, the initial clustering of image vector element

Earlier former multi-spectral remote sensing image is carried out principal component transformation, the number of the definite principal component that will keep of precision according to actual needs, the order descending according to principal component characteristic of correspondence value keeps principal component, obtains the principal component matrix.Make the initial cluster center number equal the principal component number square, the average and the standard deviation of computed image principal component matrix are the center with the average, about five equilibrium between 2 times of standard deviations, obtain end points as initial cluster center.Each vector element in the principal component matrix is carried out initial clustering with " nearby principle ", promptly ask each vector element of image and two norm distances of each cluster centre, if certain vector element and certain distances of clustering centers minimum just are divided into such with this vector element.Obtain the element set and the element number of each class at last.

2, the merging of class and division

Calculate the number percent of element number, number percent is removed less than the class of a certain given threshold value.Calculate all kinds of averages and standard deviation and overall standard deviation.If the ratio of the standard deviation of a certain class and population standard deviation is greater than a certain given threshold value, with such decomposition, promptly with such cluster centre and standard deviation and form a pair of new cluster centre with difference, then this dvielement is divided in the new class according to " nearby principle ".If the ratio of the standard deviation of a certain class and population standard deviation is less than a certain given threshold value, with it and next adjacent class merging with it, calculate the average of the later class of merging and the cluster centre and the standard deviation of the new class of standard deviation conduct, but if this kind situation appears in last class, just not as any processing.All kinds of averages of double counting and standard deviation and overall standard deviation are also carried out above judgement, till merging that does not have class and division take place.The cluster centre of Xing Chenging, all kinds of element number, all kinds of standard deviation are the feature extraction result at last.

3, feature relatively

With the cluster centre and the standard deviation scalarization respectively of two width of cloth images, all kinds of element number are carried out normalization, compare by following four standards then.Standard 1: with the cluster centre is horizontal ordinate, and all kinds of element percentage are that ordinate is made curve under same coordinate system, their similarity degree of qualitative comparison; Standard 2: the absolute value of the integral area of calculated curve and their differences obtains the ratio between area difference and the area; Standard 3: the cluster centre of two width of cloth images with " nearby principle " pairing, and is calculated the mean value of two norms between the former vector of their correspondences; Standard 4: calculate by two norms between the former vector of the standard deviation correspondence of the class of standard 3 pairings, calculate the ratio of distance and their geometrical mean between the two cluster centre standard deviations.

Spectral signature of the present invention is extracted with comparative approach has following beneficial effect:

By keep the achievement that the ISODATA clustering algorithm obtains in initial clustering, the scope that iterates is reduced into the merging and the division of class, and allows and iterate nature and finish, make under the situation that guarantees former algorithm effect, especially at the remote sensing images wave band more for a long time, significantly reduced calculated amount; Change the dynamic parameter of ratio character into by the preset parameter with the absolute property of former algorithm, make that under the situation that guarantees former algorithm effect, parameters needed reduces, parameter control is easier, and the parameter meaning is clearer and more definite; By designing four feature standards of comparison, will relatively carry out scalarization between the vector array, make that the different degree of phase Sihe between image obtains qualitative and quantitative embodiment, and make the feature extraction result of the inventive method can be used for the view data library searching.

Description of drawings:

Fig. 1 the present invention is based on the remote-sensing image spectrum feature extraction of improvement ISODATA algorithm and the process flow diagram of comparative approach.

As shown in the figure: original image is carried out determining each cluster centre vector after principal component transformation finishes; According to " nearby principle " each vector element of image is assigned to and to be gone in all kinds of; Clustering result is merged and division, cluster result is tended to be balanced; Cluster result is carried out feature relatively, obtain four difference standards; Provide the qualitative and quantitative explanation of two width of cloth image similarity degree.

The remote sensing images that Fig. 2 adopts for the embodiment of the invention.

Figure 2 shows that in " first ride of digital Shanghai " remote sensing image data storehouse three 200 * 200 * 128 image, show all are the 120th wave bands among the figure, background all is land along the river, Huangpu.

Fig. 3 is three width of cloth figure comparative standards, 1 curve among Fig. 2, and Fig. 3 (a) is Fig. 2 (a) and (b) comparison curves, and Fig. 3 (b) be Fig. 2 (a) and (c) comparison curves, and Fig. 3 (c) is Fig. 2 (b) and (c) comparison curves.

Embodiment:

In order to understand technical scheme of the present invention better, embodiments of the present invention are further described below in conjunction with drawings and Examples.

The original remote sensing images that one embodiment of the present of invention adopt be three 200 * 200 * 128 image in " first ride of digital Shanghai " remote sensing image data storehouse, as shown in Figure 2, show all are the 120th wave bands among the figure, and background is land, riverside, Huangpu.The spectral signature of remote sensing images of the present invention extract with the comparative approach flow process as shown in Figure 1, specific implementation method is carried out as follows:

1, the initial clustering of image vector element

Earlier original image is carried out principal component transformation, remove correlativity between spectrum.Each wave band of multispectral image is designated as gray matrix

Vector form: f _i=[α ₁₁... α _1nα ₂₁... α _2nα ₃₁... α _M1... α _Mn], B=[f then ₁f ₂... f _p] ¹Express piece image, the covariance matrix of B

The proper vector ν of C _kBy eigenvalue _kCorresponding from big to small the arrangement formed conversion battle array A, C _D=A ' CA, C _DBe diagonal matrix.Linear transformation D=A ' B, the row of D is called principal component, λ _kIt is the principal component variance.If principal component is more, remove the principal component of less relatively eigenwert correspondence, the result is not had much affect, but can reduce calculated amount.Kept preceding 8 principal components in the present embodiment, promptly principal component is 200 * 200 * 83 dimension matrixes.

The principal component matrix is carried out initial clustering.The initial cluster center number K of ISODATA algorithm is what rule of thumb to determine, generally get principal component count p square.The initial value of K is also definite like this among the present invention, and still, the iteration of step 2 is that nature finishes in the inventive method, and iteration result is an equilibrium point, and then K reduces greatly to result's influence, and the robustness of algorithm itself is strengthened, and promptly K is at p ²Value in the certain limit is as p ²=9, getting K=8,9,10 o'clock, results change was very little.Calculate mean vector u and the standard deviation vector σ of D:

The mean vector of matrix D:

$u=\frac{1}{\mathrm{mn}}\underset{l=1}{\overset{\mathrm{mn}}{\mathrm{\Σ}}}{x}_l---\left(1\right)$

The standard deviation vector of matrix D:

$\mathrm{\σ}=\sqrt{\frac{1}{\mathrm{mn}-1}\underset{l=1}{\overset{\mathrm{mn}}{\mathrm{\Σ}}}{(x_l-u)}^2}---\left(2\right)$

Five equilibrium (K-1) part gets K initial cluster center center (i) between (u-2 σ) and (u+2 σ), and the element that matrix D is all is with " nearby principle " cluster.If x is a vector element, min minimizes, and norm2 asks two norms, and Index asks sequence number.As follows, x is divided into the r class:

$\mathrm{center}\left(i\right)=(u-2\mathrm{\σ})+\frac{4\mathrm{\σ}}{K-1}(i-1),i=\mathrm{1,2},......K---\left(3\right)$

r＝Index(min(norm2(x-center(1)，norm2(x-center(2)，......，norm2(x-center(K))))????(4)

Each element is repeated above operation, obtain the element set M of each class _iAnd number N _iPresent embodiment has formed 64 classes, the vector element that has number not wait in each class.

2, the merging of class and division

If certain class satisfies $N_i<\frac{1}{\mathrm{\&alpha;}}\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{N}_i,$ Remove such (5) because element number does not just have the qualification that forms class very little, do the precision that helps improving cluster like this.Calculate all kinds of parameters: (max is a maximizing)

Average:

$\mathrm{aver}\left(i\right)=\frac{1}{N_i}\underset{l=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{M}_i\left(l\right)---\left(6\right)$

Standard deviation maximum value:

$\mathrm{stda}\left(i\right)=\max \left(\sqrt{\frac{1}{N_i-1}\underset{l=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{(M_i\left(l\right)-\mathrm{aver}\left(i\right))}^2}\right)---\left(7\right)$

Standard deviation maximum value average:

$\mathrm{av}\_\mathrm{stda}=\frac{1}{W}\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}\mathrm{stda}\left(i\right),i=\mathrm{1,2},......W---\left(8\right)$

If std is the standard deviation computing, the merging of class is as follows:

If stda (i)＜γ _c(av_stda) but i=W does not process;

If stda (i)＜γ _c(av_stda) and i ≠ W, then M _I+1=M _i+ M _I+1, N _I+1=N _i+ N _I+1,

$\mathrm{aver}(i+1)=\frac{1}{N_i+N_{i+1}}(\underset{l=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{M}_i\left(l\right)+\underset{l=1}{\overset{N_{i+1}}{\mathrm{\&Sigma;}}}{M}_{i+1}\left(l\right)),\mathrm{stda}\left(M_{i+1}\right)=\mathrm{std}\left(M_{i+1}\right)---\left(9\right)$

Abandon aver (i), stda (i), M _iAnd N _i, γ _c＜1, γ _cMore near 1 just tendency merging more.

The splitting of class is as follows:

If stda (i)＞γ _s(av_stda)

Aver (W+1)=aver (i)-stda (i) then, aver (W+2)=aver (i)+stda (i) (10)

With " nearby principle " with M _iIn element assign to M _W+1With M _W+2In, and calculate N _W+1, N _W+2And stda (W+1), stda (W+2), abandon aver (i), stda (i), M ₁And N ₁γ _s＞1, γ _sMore near 1 just tendency division more.New cluster centre is inserted in the former cluster centre array with " nearby principle ".

According to new cluster centre, repeating step 2 till merging that does not have class and division, finally forms three characteristic quantities: cluster centre set of vectors aver (i), all kinds of standard deviation set of vectors stda (i), all kinds of element number N ₁Three width of cloth figure of present embodiment respectively form 34,36 and 34 classes, and the cluster centre of all kinds of correspondences and standard deviation.

3, feature relatively

With aver1 (i), the big back apteryx of dimension removes among the aver2 (j), makes dimension identical, N1 _i, N2 _jNormalization.

Make r=norm2 (aver1 (1)-aver2 (1)), s=norm2 (aver1 (1)-aver2 (2)), t=norm2 (aver2 (1)-aver2 (2)),

u＝norm2(aver1(m)-aver2(n))，v＝norm2(aver1(m)-aver2(n-1))，

w＝norm2(aver2(n)-aver2(n-1))

If r＜s and t＜s, min=aver1 (1) otherwise min=aver2 (1),

If u＜v and w＜v, max=aver1 (m) otherwise max=aver2 (n).

aver1(i)＝norm2(aver1(i)-min)/norm2(max-min)，

aver2(i)＝norm2(aver2(i)-min)/norm2(max-min)

Stda1 (i) and stda2 (j) in like manner handle.

Standard 1: with aver1 (i), aver2 (j) is a horizontal ordinate, N _i, N2 _jBe ordinate mapping g ₁, g ₂Qualitative comparison;

Standard 2:

$q=\mathrm{abs}\left(S\right)/\sqrt{S_1{S}_2}---\left(11\right)$

S ₁, S ₂With S be respectively g ₁, g ₂And g ₁-g ₂Integrated value, q is more little, image is similar more;

Standard 3:Index (i, j), the aver1 of two width of cloth images (i), aver2 (j) be with " nearby principle " pairing, and calculate the mean value of distance between the former vector of their correspondences, and the more little image of mean value is similar more;

Standard 4:Qstda (i, j)=mean (stda1 (i)-stda2 (j)) (12)

I wherein, j from Index (i, j), mean represents to ask on average, (i, j) more little image is similar more for Qstda.

It is more as follows that 3 width of cloth remote sensing images shown in Figure 2 are carried out feature:

(1) as parameter alpha=0.01, γ _s=0.5, γ _c=2 o'clock, Fig. 2 feature standard of comparison 1 was seen Fig. 3.(the solid line presentation graphs 2 (a) among Fig. 3 (a), dotted line presentation graphs 2 (b); Solid line presentation graphs 2 (a) among Fig. 3 (b), dotted line presentation graphs 2 (c); Solid line presentation graphs 2 (b) among Fig. 3 (c), dotted line presentation graphs 2 (c)).

Table 1: as parameter alpha=0.01, γ _s=0.5, γ _c=2 o'clock, standard 2,3,4 results

	Fig. 2 (a) and Fig. 2 (b)	Fig. 2 (a) and Fig. 2 (c)	Fig. 2 (b) and Fig. 2 (c)
				The area ratio	0.66048	0.85134	0.87367
The cluster centre mean distance	5.2222	9.7727	7.6111
				The standard deviation mean distance	9.1215	10.909	9.0091

According to the definition of 4 standards, draw judgement from Fig. 3 with table 1: the intensity profile scope of Fig. 2 (a) and Fig. 2 (b) is all more similar with the regularity of distribution, Fig. 2 (a) and Fig. 2's (c) and Fig. 2 (b) and Fig. 2's (c) differ bigger.(2) adopt former ISODATA algorithm and above used standard of comparison, operation parameter: dvielement number minimum value parameter θ _N=400, class splitting parameter θ _S=56.9, class merges parameter θ _C=151.4, L=3, θ _NBe 1% of the total pixel of original image, θ _SBe between the original image initial cluster center ultimate range 5%, θ _CIt is the maximal value in the standard deviation mean value behind the initial clustering first.

Table 2: former algorithm and improvement algorithm of the present invention are finished cluster required time (CPU frequency: 866MHZ, unit: second) under Matlab6.5

	Fig. 2 (a) and Fig. 2 (b)	Fig. 2 (a) and Fig. 2 (c)	Fig. 2 (b) and Fig. 2 (c)
				Improve algorithm simulating (1) the cluster time	65.334	66.375	66.967
Former algorithm simulating (2) the cluster time	254.33	337.69	245.34

Improvement algorithm of the present invention as can be found from Table 2 and used time of former algorithm are far from each other, and the used time of former algorithm is about 3 times of improvement used time of algorithm, and visible method of the present invention is more superior than former algorithm really on calculated amount.

Claims (1)

1, a kind of spectral signature of remote sensing images is extracted and comparative approach, it is characterized in that comprising following three basic steps:

1) initial clustering of image vector element: earlier former multi-spectral remote sensing image is carried out principal component transformation, the number of the definite principal component that will keep of precision according to actual needs, the order descending according to principal component characteristic of correspondence value keeps principal component, obtain the principal component matrix, make the initial cluster center number equal the principal component number square, the average and the standard deviation of computed image principal component matrix, with the average is the center, about five equilibrium between 2 times of standard deviations, obtain end points as initial cluster center, each vector element in the principal component matrix is carried out initial clustering with " nearby principle ", promptly ask each vector element of image and two norm distances of each cluster centre, if certain vector element and certain distances of clustering centers minimum just are divided into such with this vector element, obtain the element set and the element number of each class at last;

2) merging of class and division: the number percent that calculates element number, number percent is removed less than the class of a certain given threshold value, calculate all kinds of averages and standard deviation and overall standard deviation, if the ratio of the standard deviation of a certain class and population standard deviation is greater than a certain given threshold value, with such decomposition, promptly with such cluster centre and standard deviation and form a pair of new cluster centre with difference, then this dvielement is divided in the new class according to " nearby principle ", if the ratio of the standard deviation of a certain class and population standard deviation is less than a certain given threshold value, with it and next adjacent class merging with it, calculate the average of the later class of merging and the cluster centre and the standard deviation of the new class of standard deviation conduct, if this kind situation appears in last class, just not as any processing; All kinds of averages of double counting and standard deviation and overall standard deviation are also carried out above judgement, and till merging that does not have class and division took place, the cluster centre of Xing Chenging, all kinds of element number, all kinds of standard deviation were the feature extraction result at last;

3) feature relatively: with the cluster centre and the standard deviation scalarization respectively of two width of cloth images, all kinds of element number are carried out normalization, compare by following four standards then, standard 1: with the cluster centre is horizontal ordinate, all kinds of element percentage are that ordinate is made curve under same coordinate system, their similarity degree of qualitative comparison; Standard 2: the absolute value of the integral area of calculated curve and their differences obtains the ratio between area difference and the area; Standard 3: the cluster centre of two width of cloth images with " nearby principle " pairing, and is calculated the mean value of two norms between the former vector of their correspondences; Standard 4: calculate by two norms between the former vector of the standard deviation correspondence of the class of standard 3 pairings, calculate the ratio of distance and their geometrical mean between the two cluster centre standard deviations.

Images