CN103679210B - The Objects recognition method mixed based on high spectrum image solution - Google Patents
The Objects recognition method mixed based on high spectrum image solution Download PDFInfo
- Publication number
- CN103679210B CN103679210B CN201310647509.0A CN201310647509A CN103679210B CN 103679210 B CN103679210 B CN 103679210B CN 201310647509 A CN201310647509 A CN 201310647509A CN 103679210 B CN103679210 B CN 103679210B
- Authority
- CN
- China
- Prior art keywords
- matrix
- mrow
- spectrum image
- high spectrum
- msub
- 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.)
- Active
Links
Landscapes
- Investigating Or Analysing Biological Materials (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention discloses a kind of Objects recognition method mixed based on high spectrum image solution, mainly solves the problems, such as that existing method judges that the affiliated atural object classification of mixed pixel point is inaccurate.Implementation step is：A panel height spectrum picture is inputted, the mixed pixel in the high spectrum image is pressed into row and lines up a matrix, composition data matrix；Constrained with the manifold of data matrix, the bound term that the sparse constraint of abundance matrix and the smoothness constraint of end member matrix are formed, is added in the object function of NMF algorithms, forms new object function；It is mixed that solution is optimized to new object function, obtains the end member matrix and abundance matrix after the high spectrum image solution is mixed；End member matrix and abundance matrix after being mixed according to solution judge the atural object classification of all mixed pixel points in the high spectrum image.The present invention can improve the mixed obtained end member value of solution and the precision of Abundances, so that the precision of high spectrum image Objects recognition is improved, available for target following.
Description
Technical field
The invention belongs to technical field of remote sensing image processing, is a kind of Objects recognition side mixed based on high spectrum image solution
Method, this method can be used for the analysis of high spectrum image, a mixed pixel point are decomposed into end member and corresponding Abundances.
Background technology
High-spectrum seems that the tens of same earth surface area or even hundreds of wave bands are imaged at the same time using imaging spectrometer
And the 3-D view obtained, it is made of two-dimensional space information and one-dimensional spectral information.Using these abundant spectral informations over the ground
Thing is finely divided and differentiates, is widely applied multi-field.Because the spectrum sensor of high spectrum image is spectrally resolved
Rate is high, so the spectral coverage formed is more, but the energy that each spectral coverage receives is small, so the ground face for receiving spectrum can only be improved
Product, that is, reduce spatial resolution.So while the not high complexity for also having nature atural object of the spatial resolution of high spectrum image
So as to form mixed pixel point.The generally existing of mixed pixel point not only influences identification and the nicety of grading of atural object, but also is distant
The significant obstacle that sense technology develops to quantification.Therefore it is high-spectrum remote sensing that mixed pixel point, which how is effectively performed, and decomposes
One of key issue of application.
The model of mixed pixel point generally uses line style mixed model in high spectrum image, it the advantages of be that algorithm is simple,
Explicit physical meaning.The mathematical procedure of the model is briefly described below：One pixel with L spectral coverage is expressed as Xij∈RL ×1, have P end member end member matrix be expressed as M ∈ RL×P, the corresponding abundance matrix of M is expressed as Sij∈RP×1.Then have：Xij=
MSij+ n, wherein n are noise.The model is limited by two conditions in the actual environment：
① Mup≥0,(1≤u≤L,1≤p≤P)
②
Two formulas above represent respectively the energy of spectrum there is no the size of negative value and the energy of mixing be it is certain, can not
Can be infinitely great.Above-mentioned model and restrictive condition all meets Non-negative Matrix Factorization（Nonnegative Matrix
Factorization, NMF）Mathematical model, it is possible to it is mixed to carry out solution with NMF algorithms.
It is proposed at present based on Non-negative Matrix Factorization（Nonnegative Matrix Factorization, NMF）Height
The solution mixing method of spectrum picture is all that regular terms is added on the object function of NMF algorithms.Because this method does not take into full account
The characteristic of high spectrum image so that the mixed effect of solution is poor, so as to cause the precision of Objects recognition not high.
The content of the invention
It is an object of the invention to the deficiency for existing method, proposes that a kind of atural object mixed based on high spectrum image solution is known
Other method, this method is by the flatness of the end member matrix of high spectrum image, the manifold of the openness and data matrix of abundance matrix
Assuming that structure is combined together so that high spectrum image solution mixes that effect is more preferable, so that the precision higher of Objects recognition.
Realizing the technical solution of this method is：A panel height spectrum picture is inputted, by the mixed pixel in the high spectrum image
Press row and line up a matrix, composition data matrix, is constrained, the sparse constraint and end member of abundance matrix with the manifold of data matrix
The bound term that the smoothness constraint of matrix is formed, is added in the object function of NMF algorithms, forms new object function, then right
This new object function optimizes solution and mixes, and obtains the end member matrix and abundance matrix of the high spectrum image, and then basis should
The end member matrix and abundance matrix that solution is mixed out judge the atural object classification of all mixed pixel points in the high spectrum image.Specific steps
Including as follows：
（1）Input a panel height spectrum picture X ∈ RM×N×L, and by mixed pixel point X in the high spectrum imageij∈R1×LPress
Row arrangement, forms data matrix Z ∈ RL×B, wherein M and N are the row and column of two dimensional image, i and be the abscissa that j is two dimensional image
And ordinate, L are spectral coverage number, B is mixed pixel point sum in high spectrum image, and B=M × N, R represent real number set；
（2）Theoretical, the manifold bound term of construction data matrix Z is assumed according to manifold：
Wherein ziBe Z i-th row, zjIt is the jth row of Z, and ziIt is zjK neighbour in one, S is abundance matrix, si
It is the i-th row of S, W is the weight matrix of Z, WijIt is an element of W,For ziAnd zjWeights, Tr ()
The mark of representing matrix, the transposition of T representing matrixes, D are the diagonal weight matrixs of Z, DiiIt is an element on D diagonal, Dii
=ΣjWij, Y is flow shape factor matrix, Y=D-W；
（3）According to high spectrum image imaging theory, L is added in abundance matrix S1/2Norm, obtains sparse constraint expression
Formula ‖ S ‖1/2, using the sparse constraint item as abundance matrix S；
（4）According to high spectrum image imaging theory, Frobenius norms are added in end member matrix M, obtain smoothness constraint
Expression formulaUsing the smoothness constraint term as end member matrix M；
（5）Three bound terms that step (2)-(4) are obtained are added to the object function of NMF algorithms
In, to form new object function：
Wherein, α be abundance matrix S sparse constraint regular parameter, β be end member matrix M smoothness constraint regular parameter, γ
Regular parameter is constrained for the manifold of data matrix Z；
（6）To step（5）Obtained object function f '（M, S）Solution is optimized with iteration multiplication, obtains high-spectrum
As X ∈ RM×N×LEnd member matrix M and abundance matrix S；
（7）By above-mentioned high spectrum image X ∈ RM×N×LMiddle mixed pixel point XijUse step（6）Solve obtained end member matrix M
With abundance vector siRepresent, i.e. mixed pixel point Xij=Msi；
（8）It is theoretical according to high spectrum image statistical distribution, by step（7）In abundance vector siTo mixed pixel point XijInto
Row atural object classification judges, i.e., as max (si)=saiWhen, then sentence mixed pixel point XijBelong to a classes, obtain the mixed pixel point
Class label is vij=a, wherein max () represent the maximum in amount of orientation, and a=1,2 ..., P are represented in the high spectrum image
Corresponding atural object class number, P represent atural object classification sum, s in the high spectrum imageaiIt is siA-th of element；
（9）To above-mentioned high spectrum image X ∈ RM×N×LIn all mixed pixel point steps（8）Operation carry out ground species
Do not judge, obtain high spectrum image X ∈ RM×N×LAtural object classification matrix V ∈ RM×N。
The present invention has the following advantages compared with prior art：
1. the present invention more fully considers the architectural characteristic of high spectrum image compared to existing method, by data matrix
Manifold constraint is added to the NMF algorithms that solution is mixed, is constrained with not adding manifold, and CNMF algorithms are compared with piecewise smooth NMF algorithms, obtained
The precision higher of the abundance matrix arrived so that judge that the affiliated atural object classification of mixed pixel point is more accurate, so as to reach higher ground
Thing accuracy of identification.
It is and existing 2. it is mixed that the sparse constraint of the smoothness constraint of end member matrix and abundance matrix is added to solution by the present invention at the same time
GLNMF algorithms compare, more taken into full account high spectrum image characteristic, substantially increased the precision of end member matrix so that calculate
The atural object classification gone out is other closer to truly species, is more in line with the requirement of practical application.
Brief description of the drawings
Fig. 1 be the present invention realize flow chart；
Fig. 2 is the high spectrum image schematic diagram that the present invention uses；
Fig. 3 is the contrast of the curve of spectrum and actual spectrum curve of the end member of the invention for mixing out with existing several algorithm solutions
Schematic diagram；
Fig. 4 is that the spectral modeling distance SAD of the end member of the invention mixed out with existing several algorithm solutions conciliates the abundance mixed out
Average 30 error bar comparison diagrams of root-mean-square error RMSE.
Embodiment
It is as follows with reference to Fig. 1, specific implementation step of the invention：
Step 1：High spectrum image is inputted, builds data matrix, obtains the real atural object classification square of the high spectrum image
Battle array, real end member matrix and real abundance matrix.
1.1) high spectrum image as shown in Figure 2 is inputted, which is 145 × 145, shares 16 class atural objects, the figure
Each mixed pixel point as in can regard the spectral vector being made of the spectral information of 200 spectral coverages as；
1.2) by high spectrum image X ∈ RM×N×LMiddle mixed pixel point Xij∈R1×LArranged by row, form data matrix Z
∈RL×B, wherein, M and N are the row and column of two dimensional image, and i and j are respectively the abscissa and ordinate of two dimensional image, and L is spectral coverage
Number, P are atural object classification number, and B is mixed pixel point sum in high spectrum image, and B=M × N, R represent real number set, in this example
Middle L values are that 200, B values are that 21025, P values are 16；
1.3) the true classification matrix for obtaining the high spectrum image is V ∈ RM×N, true end member is M ∈ RL×P, it is true rich
Spend for S ∈ RP×B。
Step 2：Structure constraint item：
1.1）Theoretical, the manifold bound term of construction data matrix Z is assumed according to manifold：
1.1a）The i-th row z of data matrix Z is calculated with heat kernel functioniWith the jth row z of data matrix ZjWeights Wij：
Wherein, σ is thermonuclear parameter, value 1；
1.1b）Use L2The i-th row s of norm calculation abundance matrix SiWith the jth row s of abundance matrix SjDistance：
‖si-sj‖2；
1.1c）Theory is assumed according to manifold, by step 1.1a）Formula and step 1.1b）Formula combine composition
Manifold bound term：
Wherein ziIt is zjK neighbour in one, W is the weight matrix of Z, the mark of Tr () representing matrix, and T represents square
The transposition of battle array, D are the diagonal weight matrixs of Z, DiiIt is an element on D diagonal, Dii=ΣjWij, Y is flow shape factor
Matrix, Y=D-W；
1.2）According to high spectrum image imaging theory, L is added in abundance matrix S1/2Norm, obtains sparse constraint expression
Formula ‖ S ‖1/2, using the sparse constraint item as abundance matrix S；
1.3）According to high spectrum image imaging theory, Frobenius norms are added in end member matrix M, are obtained smoothly about
Beam expression formulaUsing the smoothness constraint term as end member matrix M.
Step 3：Construct new object function
1.1）Three bound terms that step 2 is obtained are added into the object function of NMF algorithms
In, to form new object function：
Wherein, α be abundance matrix S sparse constraint regular parameter, value 2, β be end member matrix M smoothness constraint just
Then parameter, value 1, the manifold that γ is data matrix Z constrain regular parameter, value 0.6.
Step 4：The new object function f ' obtained to step 3（M, S）Solution is optimized with iteration multiplication, obtains end member
Matrix M and abundance matrix S.
1.1）End member matrix M and abundance matrix S is initialized with random number between [0,1], place is normalized in each column of S
Reason；
1.2）Input terminal variable matrix M and abundance matrix S；
1.3）By to new object function f '（M, S）In end member matrix M and abundance matrix S derivations, respectively obtain end
The calculation formula of the iteration multiplication of variable matrix M and abundance matrix S：
M’=M.*(ZST-βM)./MSST,
Wherein,（.*）The element multiplication of representing matrix,（./）The element division of representing matrix；
1.4）With step 1.3）The formula new abundance matrix S ' that calculates and new end member matrix M ', by new abundance
Matrix S ' and new end member matrix M ' are used as step 1.2）Input；
1.5）Repeat step 1.2）~1.4）N times common, output finally needs end member matrix M and the abundance matrix obtained
S, calculating terminate, wherein, n is to perform number, value 3000.
Step 5：The high spectrum image X ∈ R that judgment step 1 inputsM×N×LIn all mixed pixel points atural object classification.
1.1）The high spectrum image X ∈ R inputted to step 1M×N×LIn mixed pixel point XijMix what is obtained with step 3 solution
End member matrix M and abundance vector siRepresent, i.e. Xij=Msi；
1.2）It is theoretical according to high spectrum image statistical distribution, by step 1.1）In obtained abundance vector siTo mixing picture
Vegetarian refreshments XijAtural object classification judgement is carried out, i.e., as max (si)=saiWhen, then sentence mixed pixel point XijBelong to a classes, obtain the mixing
The class label of pixel is vij=a, wherein max () represent the maximum in amount of orientation, and a=1,2 ..., P represent the bloom
Corresponding atural object class number in spectrogram picture, P represent atural object classification sum, s in the high spectrum imageaiIt is siA-th of element；
1.3）To high spectrum image X ∈ RM×N×LIn all mixed pixel point steps 1.2）Operation carry out ground species
Do not judge, obtain high spectrum image X ∈ RM×N×LSolution it is mixed after atural object classification matrix V ∈ RM×N。
The effect of the present invention is further illustrated by following emulation experiment：
（1）Experiment simulation condition：
The high-spectrum that this experiment uses seems typical AVIRIS high spectrum images：It is derived from U.S. of in June, 1992 shooting
The Indian remote sensing trial zone in the state state of Indiana northwestward, atural object classification amount to 16 classes, and the size of image is 145 × 145.It is original
Data share 220 spectral coverages, remove 20 spectral coverages by noise pollution and water pollution, only retain remaining 200 spectral coverages.This
Experiment is Intel (R) Core (TM) i5-2450, dominant frequency 2.5GHz in CPU, is inside saved as in the WINDOWS7 systems of 4G using soft
Part MATLAB2009a is emulated.
（2）Evaluation index：
1.1）For end member, with spectral modeling distance SAD come the true end member M of comparisontWith estimation end member MtSimilitude.Spectrum
The curve of spectrum of smaller then two end members of value of angular distance SAD is closer.For abundance, with root-mean-square error RMSE come than serious
Real abundance StWith estimation abundance StBetween difference.Smaller then two Abundances of value of root-mean-square error RMSE are closer.Above-mentioned two
The formula of a evaluation criterion is respectively：
Wherein, MtArranged for the t of real end variable matrix M, MtTo estimate the t of end member matrix M row, StFor true abundance square
The t row of battle array S, StFor estimate abundance matrix S t row, t represent atural object classification, t=1,2 ..., P, P be atural object classification
Sum；
（3）Experiment simulation content：
Experiment one
Using the present invention to step（1）Described in high spectrum image to optimize solution mixed, obtain end member matrix M and abundance square
Battle array S, then use NMF algorithms, CNMF algorithms, GLNMF algorithms and piecewise smooth NMF algorithms excellent to the progress of above-mentioned high spectrum image respectively
Neutralizing is mixed, is contrasted with the mixed obtained result of present invention solution, experimental result is as shown in figure 3, wherein:
Fig. 3（a）It is bent with actual atural object zunyite spectrum with a kind of atural object zunyite curve of spectrum that NMF algorithm solutions are mixed out
The contrast schematic diagram of line；
Fig. 3（b）It is with a kind of atural object zunyite curve of spectrum that CNMF algorithm solutions are mixed out and actual atural object zunyite spectrum
The contrast schematic diagram of curve；
Fig. 3（c）It is with a kind of atural object zunyite curve of spectrum that GLNMF algorithm solutions are mixed out and actual atural object zunyite spectrum
The contrast schematic diagram of curve；
Fig. 3（d）It is yellow with actual atural object chlorine with a kind of atural object zunyite curve of spectrum that piecewise smooth NMF algorithm solutions are mixed out
The contrast schematic diagram of the brilliant curve of spectrum；
Fig. 3（e）It is with a kind of atural object zunyite curve of spectrum that the algorithm solution of the present invention is mixed out and actual atural object zunyite
The contrast schematic diagram of the curve of spectrum；
As seen from Figure 3, the present invention is compared to other existing methods, and the curve of spectrum of obtained atural object is closer to true atural object
The curve of spectrum of atural object.
Experiment two
Utilize step（2）The end member matrix M and abundance matrix S that are obtained to experiment one calculate the spectrum of end member matrix M respectively
The root-mean-square error RMSE value of angular distance sad value and abundance matrix S, experimental result is as shown in figure 4, wherein:
Fig. 4（a）It is the spectral modeling of the end member of the invention mixed out with above-mentioned 4 kinds existing algorithm solutions apart from average the 30 of sad value
Secondary resultant error rod comparison diagram；
Fig. 4（b）It is average 30 knots of the root-mean-square error RMSE value of the abundance of the invention mixed out with above-mentioned 4 kinds of algorithm solutions
Fruit error bar comparison diagram；
From fig. 4, it can be seen that the present invention, compared to other existing methods, the precision of the mixed obtained end member value of solution and Abundances is more
It is high.
Experiment three
The end member matrix M and abundance matrix S obtained using experiment one clicks through all mixed pixels in the high spectrum image
Row classification judges, obtains the classification matrix V after solution is mixed, and recycles supporting vector machine SVM to solving the mixed classification square obtained afterwards
Battle array V and true classification matrix V carry out Objects recognition precision measurement, by the present invention measurement gained Objects recognition precision with it is above-mentioned
The Objects recognition precision of 4 kinds of existing method measurement gained is contrasted, as shown in table 1.
1 Objects recognition accurate values index of table contrasts
As seen from Table 1, the Objects recognition precision obtained using present invention progress Objects recognition is significantly better than above-mentioned 4
The Objects recognition precision of the existing algorithm of kind.
In conclusion the present invention can increase substantially end member value and the precision of Abundances in high spectrum image, so that more
The precision of Objects recognition is improved well, so the present invention is existed as a kind of Objects recognition method mixed based on high spectrum image solution
High spectrum image identification field has broad application prospects.
Claims (3)
1. a kind of Objects recognition method mixed based on high spectrum image solution, is included the following steps：
(1) a panel height spectrum picture X ∈ R are inputtedM×N×L, and by mixed pixel point X in the high spectrum imageij∈R1×LArranged by row
Row, form data matrix Z ∈ RL×B, wherein M and the row and column that N is two dimensional image, i and the abscissa that j is two dimensional image are sat with vertical
Mark, L are spectral coverage number, and B is mixed pixel point sum in high spectrum image, and B=M × N, R represent real number set；
(2) theoretical, the manifold bound term of construction data matrix Z is assumed according to manifold：
<mfenced open = "" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>B</mi>
</munderover>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>s</mi>
<mi>j</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<msub>
<mi>W</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>B</mi>
</munderover>
<msubsup>
<mi>s</mi>
<mi>i</mi>
<mi>T</mi>
</msubsup>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<msub>
<mi>D</mi>
<mrow>
<mi>i</mi>
<mi>i</mi>
</mrow>
</msub>
<mo>-</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>B</mi>
</munderover>
<msubsup>
<mi>s</mi>
<mi>i</mi>
<mi>T</mi>
</msubsup>
<msub>
<mi>s</mi>
<mi>j</mi>
</msub>
<msub>
<mi>W</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mo>=</mo>
<mi>T</mi>
<mi>r</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>SYS</mi>
<mi>T</mi>
</msup>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
Wherein ziBe Z i-th row, zjIt is the jth row of Z, and ziIt is zjK neighbour in one, S is abundance matrix, siIt is S
I-th row, sjIt is the jth row of S, W is the weight matrix of Z, WijIt is an element of W,For ziAnd zjPower
Value, σ is thermonuclear parameter, and value 1, the mark of Tr () representing matrix, the transposition of T representing matrixes, D is the diagonal weights square of Z
Battle array, DiiIt is an element on D diagonal, Dii=∑jWij, Y is flow shape factor matrix, Y=D-W；
(3) according to high spectrum image imaging theory, L is added in abundance matrix S1/2Norm, obtains sparse constraint expression formula | | S |
|1/2, using the sparse constraint item as abundance matrix S；
(4) according to high spectrum image imaging theory, Frobenius norms are added in end member matrix M, obtain smoothness constraint expression
FormulaUsing the smoothness constraint term as end member matrix M；
(5) three bound terms for obtaining step (2)-(4) are added to the object function of NMF algorithms
In, to form new object function：
<mrow>
<msup>
<mi>f</mi>
<mo>,</mo>
</msup>
<mrow>
<mo>(</mo>
<mi>M</mi>
<mo>,</mo>
<mi>S</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<mo>|</mo>
<mo>|</mo>
<mi>Z</mi>
<mo>-</mo>
<mi>M</mi>
<mi>S</mi>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mi>&alpha;</mi>
<mo>|</mo>
<mo>|</mo>
<mi>S</mi>
<mo>|</mo>
<msub>
<mo>|</mo>
<mrow>
<mn>1</mn>
<mo>/</mo>
<mn>2</mn>
</mrow>
</msub>
<mo>+</mo>
<mi>&beta;</mi>
<mo>|</mo>
<mo>|</mo>
<mi>M</mi>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mi>&gamma;</mi>
<mi>T</mi>
<mi>r</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>SYS</mi>
<mi>T</mi>
</msup>
<mo>)</mo>
</mrow>
</mrow>
Wherein, α is the sparse constraint regular parameter of abundance matrix S, and β is the smoothness constraint regular parameter of end member matrix M, and γ is number
Regular parameter is constrained according to the manifold of matrix Z；
(6) solution is optimized with iteration multiplication to the object function f ' (M, S) that step (5) obtains, obtains high spectrum image X ∈
RM×N×LEnd member matrix M and abundance matrix S；
(7) by above-mentioned high spectrum image X ∈ RM×N×LMiddle mixed pixel point XijObtained end member matrix M and rich is solved with step (6)
Spend vector siRepresent, i.e. mixed pixel point Xij=Msi；
(8) it is theoretical according to high spectrum image statistical distribution, by the abundance vector s in step (7)iTo mixed pixel point XijCarry out ground
Species do not judge, i.e., as max (si)=saiWhen, then sentence mixed pixel point XijBelong to a classes, obtain the classification of the mixed pixel point
Label is vij=a, wherein max () represent the maximum in amount of orientation, and a=1,2 ..., P represent phase in the high spectrum image
The atural object class number answered, P represent atural object classification sum, s in the high spectrum imageaiIt is siA-th of element；
(9) to above-mentioned high spectrum image X ∈ RM×N×LIn all mixed pixel points carry out atural object classification with the operation of step (8) and sentence
It is disconnected, obtain high spectrum image X ∈ RM×N×LAtural object classification matrix V ∈ RM×N。
2. described in the Objects recognition method mixed based on high spectrum image solution according to claims 1, wherein step (2)
Theory is assumed according to manifold, the manifold bound term of construction data matrix Z, carries out as follows：
The i-th row z of data matrix Z 2a) is calculated with heat kernel functioniWith the jth row z of data matrix ZjWeights Wij：
<mrow>
<msub>
<mi>W</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
<mo>=</mo>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>z</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>z</mi>
<mi>j</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mi>&sigma;</mi>
</mfrac>
</mrow>
</msup>
<mo>,</mo>
</mrow>
Wherein, σ is thermonuclear parameter, value 1；
2b) use L2The i-th row s of norm calculation abundance matrix SiWith the jth row s of abundance matrix SjDistance：
||si-sj||2；
2c) assumed according to manifold it is theoretical, by step 2a) formula and step 2b) formula combine and form manifold bound term：Wherein, B is the sum of mixed pixel point in high spectrum image.
3. the Objects recognition method mixed based on high spectrum image solution according to claims 1, wherein in the step (6)
With iteration multiplication Optimization Solution object function f ' (M, S), carry out as follows：
3a) with random number initialization end member matrix M and abundance matrix S between (0,1), each column of abundance matrix S is subjected to normalizing
Change is handled；Input terminal variable matrix M and abundance matrix S；
Derivation 3b) is carried out to the abundance matrix S in object function f ' (M, S)=0, the iteration multiplication for obtaining abundance matrix S calculates
Formula：
<mrow>
<mi>S</mi>
<mo>.</mo>
<mo>*</mo>
<mrow>
<mo>(</mo>
<msup>
<mi>M</mi>
<mi>T</mi>
</msup>
<mi>Z</mi>
<mo>+</mo>
<mn>2</mn>
<mi>&gamma;</mi>
<mi>S</mi>
<mi>W</mi>
<mo>)</mo>
</mrow>
<mo>.</mo>
<mo>/</mo>
<mrow>
<mo>(</mo>
<msup>
<mi>M</mi>
<mi>T</mi>
</msup>
<mi>M</mi>
<mi>S</mi>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<msup>
<mi>&alpha;S</mi>
<mrow>
<mo>-</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
</mrow>
</msup>
<mo>+</mo>
<mn>2</mn>
<mi>&gamma;</mi>
<mi>S</mi>
<mi>D</mi>
<mo>)</mo>
</mrow>
</mrow>
The wherein element multiplication of (.*) representing matrix, the element division of (/) representing matrix, W are the weight matrixs of Z, and D is pair of Z
Linea angulata weight matrix, Z are data matrix, and the transposition of T representing matrixes, α is the sparse constraint regular parameter of abundance matrix S, and γ is
The manifold constraint regular parameter of data matrix Z；
Derivation 3c) is carried out to the end member matrix M in object function f ' (M, S)=0, the iteration multiplication for obtaining end member matrix M calculates
Formula：
M.*(ZST-βM)./MSST
Wherein, β is the smoothness constraint regular parameter of end member matrix M；
3d) with step 3b) and step 3c) the formula new abundance matrix S ' that calculates and new end member matrix M ', by abundance
Matrix S ' and new end member matrix M ' are used as step 3a) input；
3e) repeat step 3a)~3d) n times, output finally needs the end member matrix M and abundance matrix S obtained, then calculates
Terminate, wherein, n is to perform number, value 3000.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310647509.0A CN103679210B (en) | 2013-12-03 | 2013-12-03 | The Objects recognition method mixed based on high spectrum image solution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310647509.0A CN103679210B (en) | 2013-12-03 | 2013-12-03 | The Objects recognition method mixed based on high spectrum image solution |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103679210A CN103679210A (en) | 2014-03-26 |
| CN103679210B true CN103679210B (en) | 2018-04-17 |
Family
ID=50316698
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310647509.0A Active CN103679210B (en) | 2013-12-03 | 2013-12-03 | The Objects recognition method mixed based on high spectrum image solution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103679210B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104462196B (en) * | 2014-10-30 | 2019-01-22 | 南京信息工程大学 | Multiple features combining Hash information search method |
| CN104931504B (en) * | 2015-05-06 | 2017-12-05 | 浙江工业大学 | A kind of space junk material analysis method based on Non-negative Matrix Factorization |
| CN104933425B (en) * | 2015-07-10 | 2018-04-27 | 中国地质大学（武汉） | A kind of hyperspectral data processing method |
| CN105320959B (en) * | 2015-09-30 | 2018-11-16 | 西安电子科技大学 | High spectrum image sparse solution mixing method based on end member study |
| CN106056524A (en) * | 2016-05-25 | 2016-10-26 | 天津商业大学 | Hyper-spectral image nonlinear de-mixing method based on differential search |
| CN106788714B (en) * | 2016-12-05 | 2019-01-18 | 重庆工商大学 | A kind of sparse solution mixing method based on optical computing |
| CN106650811B (en) * | 2016-12-26 | 2019-08-13 | 大连海事大学 | A kind of EO-1 hyperion mixed pixel classification method cooperateing with enhancing based on neighbour |
| CN110132866A (en) * | 2019-05-27 | 2019-08-16 | 东北大学 | A kind of soil EO-1 hyperion solution mixing method and system |
| CN110458760B (en) * | 2019-06-20 | 2021-11-05 | 中国地质大学（武汉） | HNMF remote sensing image unmixing method based on information entropy |
| CN112504975B (en) * | 2020-12-14 | 2022-12-30 | 杭州电子科技大学 | Hyperspectral unmixing method based on constrained nonnegative matrix factorization |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101770584A (en) * | 2009-12-30 | 2010-07-07 | 重庆大学 | Extraction method for identification characteristic of high spectrum remote sensing data |
| CN102324047A (en) * | 2011-09-05 | 2012-01-18 | 西安电子科技大学 | High spectrum image atural object recognition methods based on sparse nuclear coding SKR |
-
2013
- 2013-12-03 CN CN201310647509.0A patent/CN103679210B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101770584A (en) * | 2009-12-30 | 2010-07-07 | 重庆大学 | Extraction method for identification characteristic of high spectrum remote sensing data |
| CN102324047A (en) * | 2011-09-05 | 2012-01-18 | 西安电子科技大学 | High spectrum image atural object recognition methods based on sparse nuclear coding SKR |
Non-Patent Citations (3)
| Title |
|---|
| Manifold Regularized Sparse NMF for Hyperspectral Unmixing;Xiaoqiang Lu 等;《IEEE Transactions on Geoscience and Remote Sensing》;20130531;第51卷(第5期);2815-2826 * |
| Nonnegative matrix factorization for spectral data analysis;V.Paul Pauca 等;《Linear Algebra and its Applications》;20060701;第416卷(第1期);29-47 * |
| 基于权重与混合模型的遥感图像分类方法研究;何海清 等;《国土资源遥感》;20080615(第2期);18-21 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103679210A (en) | 2014-03-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103679210B (en) | The Objects recognition method mixed based on high spectrum image solution | |
| CN103824050B (en) | A kind of face key independent positioning method returned based on cascade | |
| CN103413151B (en) | Hyperspectral image classification method based on figure canonical low-rank representation Dimensionality Reduction | |
| CN110659718A (en) | Small convolution nuclear cell counting method and system based on deep convolution neural network | |
| CN105528589A (en) | Single image crowd counting algorithm based on multi-column convolutional neural network | |
| CN104732566B (en) | Compression of hyperspectral images cognitive method based on non-separation sparse prior | |
| CN105261004A (en) | Mean shift and neighborhood information based fuzzy C-mean image segmentation method | |
| CN105678757A (en) | Object displacement measurement method | |
| CN105975912A (en) | Hyperspectral image nonlinearity solution blending method based on neural network | |
| CN112560967B (en) | Multi-source remote sensing image classification method, storage medium and computing device | |
| CN103413292B (en) | Based on the hyperspectral image nonlinear abundance estimation method of constraint least square | |
| CN107680116A (en) | A kind of method for monitoring moving object in video sequences | |
| CN106443627A (en) | Bayesian based K distribution sea clutter shape parameter estimation method | |
| CN103942774A (en) | Multi-target collaborative salient-region detection method based on similarity propagation | |
| CN106405533A (en) | Radar target combined synchronization and positioning method based on constraint weighted least square | |
| CN103325120A (en) | Rapid self-adaption binocular vision stereo matching method capable of supporting weight | |
| CN102609944B (en) | Hyper-spectral remote sensing image mixed pixel decomposition method based on distance geometry theory | |
| CN105930846A (en) | Neighborhood information and SVGDL (support vector guide dictionary learning)-based polarimetric SAR image classification method | |
| CN110399820B (en) | Visual recognition analysis method for roadside scene of highway | |
| CN105069774A (en) | Object segmentation method based on multiple-instance learning and graph cuts optimization | |
| CN104734724B (en) | Based on the Compression of hyperspectral images cognitive method for weighting Laplce's sparse prior again | |
| CN106096660A (en) | Convolutional neural networks based on independent composition analysis algorithm | |
| CN103679719A (en) | Image segmentation method | |
| CN104732546A (en) | Non-rigid SAR image registration method based on region similarity and local spatial constraint | |
| CN104778718B (en) | Single image truck volume measuring method based on 3D models |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |