CN103927551B - Polarimetric SAR semi-supervised classification method based on superpixel correlation matrix - Google Patents
Polarimetric SAR semi-supervised classification method based on superpixel correlation matrix Download PDFInfo
- Publication number
- CN103927551B CN103927551B CN201410162755.1A CN201410162755A CN103927551B CN 103927551 B CN103927551 B CN 103927551B CN 201410162755 A CN201410162755 A CN 201410162755A CN 103927551 B CN103927551 B CN 103927551B
- Authority
- CN
- China
- Prior art keywords
- pixel
- super
- matrix
- correlation matrix
- training
- 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
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a polarimetric SAR semi-supervised classification method based on a superpixel correlation matrix. The polarimetric SAR semi-supervised classification method mainly solves the problem that an existing technology needs large quantity of samples. The achieving steps are as follows: (1) reading a polarimetric SAR image, conducting related preprocessing on speckle noise on the SAR image, and synchronizing a pseudo-color image; (2) calculating and testing a superpixel, training the area center of the superpixel, and constructing a data array and a physical feature correlation matrix; (3) calculating a sparsity structure feature correlation matrix through a data matrix; (4) conducting weight fusion on the physical feature correlation matrix and the sparsity structure feature correlation matrix; (5) classifying the related fused arrays through a semi-supervised method, and outputting a final classification result. The polarimetric SAR semi-supervised classification method reduces the influence of the speckle noise on the classification result, effectively reduces the requirements on the quantity of training samples, improves accuracy rate of classification, and can be used for classifying and identifying surface features.
Description
Technical field
The invention belongs to technical field of image processing, more particularly to Polarimetric SAR Image semisupervised classification method, can be used for mesh
Mark is other.
Background technology
Synthetic aperture radar SAR is a kind of effective means from earth observation from space, is worked with round-the-clock, round-the-clock
Feature, has application widely at aspects such as geological prospecting, urban planning, military detections.Polarimetric synthetic aperture radar
PolSAR is a kind of coherent multichannel microwave imaging system that SAR technologies constantly develop and produce, and it is each by measurement ground
Scattering propertiess of the resolution cell under different polarization combinations obtaining the polarization information of target, compared with single polarization SAR, PolSAR
More complete have recorded the back scattering information of target, thus greatly strengthen the ability that radar obtains target information.
At present, the airborne and borne polarization SAR system of external many mechanisms can provide the full-polarization SAR number of different-waveband
According to.Than AIRSAR, the EMISAR of Denmark that more typical airborne polarization SAR system has the U.S., the ESAR of Germany is Japanese
PISAR, the RAMSES of France, Canadian CV580 etc.；Typical borne polarization SAR system includes the SIR-C in the U.S., and Europe is empty
The Envisat ASAR of office, ALOS, the Canadian RADARSAT2 of Japan, TerraSAR-X of Germany etc..At home, it is many
Colleges and universities to institute also in the research work that development is related, and polarization SAR theory and application aspect achieve it is huge
Progress.
Classification of Polarimetric SAR Image is an important step during Polarimetric SAR Image interpretation, is also polarization SAR information
The important research direction for processing.According to the Land use systems of polarization information, the method for polarization SAR classification can be divided into direct profit
With the method for Polarization scattering feature, using the method for polarization statistical nature and with reference to scattering signatures and the method for statistical nature；Root
According to the difference of algorithm used, it is divided into the methods such as statistics, knowledge, neutral net, fuzzy logic, wavelet analysises, support vector machine；Though
So these methods have been achieved for preferable effect, but these traditional methods based on single pixel point are inevitably
Affected by coherent speckle noise in Polarimetric SAR Image.According to whether artificial instruct, can be divided into has supervision and non-supervisory two
Class, the sorting technique for having supervision can fast and accurately realize the classification of image, but be because needing model in processing procedure
Foundation and study, time complexity is higher；The realization of unsupervised approaches is more quick, but the accuracy of unsupervised segmentation is but
Than relatively low.
So the classification that Polarimetric SAR Image how is accurately realized in the case where number of training is less is urgently to solve
Problem certainly.
The content of the invention
Present invention aims to the deficiency of above-mentioned technology, proposes a kind of polarization based on super-pixel correlation matrix
SAR image semisupervised classification method, to reduce impact of the coherent speckle noise to image procossing, and in the less feelings of number of training
The accuracy of classification is improved under condition.
The technical scheme is that what is be achieved in that：
Over-segmentation is carried out to the corresponding pseudo color image of polarization SAR, several super-pixel are obtained；The super picture of test is calculated respectively
The regional center of element and training super-pixel；The dependency between any two super-pixel is calculated using the Wishart distances of amendment, is obtained
To physical features correlation matrix；Sparsity structure correlation matrix is obtained using the method for carrying out minimizing constraint to rank of matrix；Will
Above-mentioned two correlation matrix is weighted fusion, and the classification of Polarimetric SAR Image is finally realized using the method for semi-supervised figure, its skill
Art step includes as follows：
(1) Polarimetric SAR Image is input into, it is carried out to drop the pretreatment of coherent speckle noise, and synthesize pseudo color image；
1.1) a width polarization SAR data are read in, it is carried out using exquisite Lee wave filter drop the pre- place of coherent speckle noise
Reason, obtains corresponding covariance matrix, and the window size of wave filter is set as 7 × 7；
1.2) Pauli energy feature decomposition is carried out to covariance matrix, and synthesizes the corresponding pseudocolour picture of polarization SAR data
Picture；
(2) to Polarimetric SAR Image over-segmentation, regional center Θ of test super-pixel is calculatediWith the region of training super-pixel
Center Σi：
2.1) over-segmentation is carried out to pseudo color image, obtains several super-pixel, S1,S2,…Si,…Sk, it is super as test
Pixel, wherein SiI-th super-pixel, i=1 ... k are represented, k is the number of the super-pixel for dividing；
2.2) to each super-pixel SiThe corresponding covariance matrix of middle single pixel point is sued for peace and calculates average, obtains each
Regional center Θ of individual test super-pixeli；
2.3) in the corresponding standard drawing of the Polarimetric SAR Image per class N number of pixel is randomly selected as training sample point,
And m set is divided into, each set is used as a training super-pixel；To single pixel point pair in each training super-pixel
The covariance matrix answered is sued for peace and calculates average, obtains training corresponding regional center Σ of super-pixeliIf standard drawing includes c classes
Target, the then sum for training super-pixel is c × m.
(3) by regional center Σ of training super-pixeliWith regional center Θ of test super-pixeliDischarge successively, counted
According to matrix：X=[x1,…xi,…xnum]=[Ψ1,Ψ2,…,Ψc,Θ1,Θ2,…,Θn], wherein xiIn representing data matrix
The regional center of i-th super-pixel, i=1 ... num, num for super-pixel total number, ΨiRepresent the i-th class training super-pixel
Set, Ψi=[Σ1Σ2…Σm]；
(4) Wishart of the amendment between calculating any two super-pixel is apart from disi,j, obtain dependency simi,j, structure
Build physical features correlation matrix Sim:
Wherein,I=1 ... num, j=1 ... num,It is respectively x,
xi,xjMaximal possibility estimation, Tr [] for matrix mark, parameter q represent polarization SAR transmitting and reception antenna whether be same
Instruction, if so, then q=3, otherwise q=4；
(5) sparsity structure feature correlation matrix Z is built*：
5a) data matrix X itself is minimized about as the dictionary of rarefaction representation by the order to coefficient matrix Z
Beam, tries to achieve coefficient matrix Z：
Wherein, matrix A is the base of rarefaction representation, i.e. dictionary, and E is noise, and λ is the parameter for controlling influence of noise, and β is constraint
Matrix Z openness parameter, | | | |*For nuclear norm,For l2,1Norm, EijFor matrix
The element of the i-th rows of E jth row, | | | |0For l0Norm, is not the number of 0 value in representing matrix, because l0Norm Solve problems
It is a NP-hard problem, therefore this constraint is relaxed as l1Norm is solved, and is expressed asZij
For the element of the i-th rows of matrix Z jth row；
5b) sparsity structure feature correlation matrix is calculated by coefficient matrix ZWherein T representing matrixs turn
Put；
(6) to physical features correlation matrix Sim and sparsity structure feature correlation matrix Z*Fusion is weighted, obtains new
Correlation matrix W=Sim+ α Z*, wherein α is the weight for controlling matrix fusion, and span is 0 to 50；
(7) semisupervised classification is carried out using the correlation matrix W after fusion, obtains the classification of each test super-pixel；
7.1) Laplacian Matrix C=M-W is calculated, wherein, M is be made up of the summation of correlation matrix W each row elements right
Angular moment battle array；
7.2) the class probability matrix P of input training super-pixeltrain=[P1,…,Pi,…,Pt], calculate test super-pixel
Class probability matrix
Wherein, Pi=[p1,…pj,…pc]T,K=1 ... c, j=1 ... c, i=1 ...
T, c for target classification number, t be train super-pixel number, PiThe class probability vector of j-th training super-pixel is represented,It is the submatrix of Laplacian Matrix C, represents the Laplacian Matrix of training super-pixel,
ce,fRepresenting matrix C e rows f row element, e=1 ... t, f=t+1 ... num, num for super-pixel total number, Cul=
Clu T, represent the Laplacian Matrix of test super-pixel, the transposition of T representing matrixs；
7.3) class probability matrix P is taken outtestH-th column vector Ph=[p1,h,…ps,h,…pc,h], find PhMiddle maximum
The position s that value is locatedh, the shValue be classification belonging to h-th test super-pixel, wherein h=1 ... n, s=1 ... c, ps,hTable
Show that h-th test super-pixel belongs to the probability of s classes；
(8) classification belonging to each test super-pixel is corresponded to into original polarization SAR image, exports final classification knot
Really.
There is advantages below compared with prior art in the present invention：
1. the present invention carries out the sorted of Polarimetric SAR Image from based on the method for super-pixel rather than single pixel point
Journey, reduces image of the coherent speckle noise to classification results, substantially increases the accuracy rate of classification.
2. the present invention merged physics spy's correlation levy matrix and sparsity structure it is special it is related levy matrix, make use of physical message and
Global structure information, becomes apparent from the dependency between super-pixel.
3. the present invention carries out final Classification of Polarimetric SAR Image using semi-supervised method, make use of training sample and survey
Sample this dependency, reduces the requirement to training samples number, compares other supervised classification methods, and the present invention can
The classification results being more satisfied with using less training sample.
Description of the drawings
Fig. 1 is the flowchart of the present invention；
Fig. 2 is the sub-process figure of coefficient matrix solution procedure in the present invention；
Fig. 3 is the sub-process figure of semisupervised classification process in the present invention；
Fig. 4 is present invention emulation Polarimetric SAR Image used；
Fig. 5 is the result figure that present invention emulation Polarimetric SAR Image used is too segmented into super-pixel.
Fig. 6 is the classification results figure of present invention emulation.
Specific embodiment
Below in conjunction with the accompanying drawings the step of 1 couple of present invention is described in further detail.
Step one：Polarimetric SAR Image is carried out to drop the pretreatment of coherent speckle noise, and synthesizes pseudo color image：
1.1) a width polarization SAR data are read in, it is carried out using exquisite Lee wave filter drop the pre- place of coherent speckle noise
Reason, obtains corresponding covariance matrix, and the window size of wave filter is set as 7 × 7.
1.2) Pauli energy feature decomposition, the corresponding pseudocolour picture of synthesis polarization SAR data are carried out to covariance matrix
Picture.
Step 2：Over-segmentation is carried out to pseudo color image, during the region of test super-pixel and training super-pixel is calculated respectively
The heart.
2.1) over-segmentation is carried out using the method for Normalized cut to pseudo color image, obtains several super-pixel,
S1,S2,…Si,…Sk, using these super-pixel as test super-pixel, wherein SiI-th super-pixel, i=1 ... k are represented, k is to draw
The number of the super-pixel divided；
The Normalized Cut methods：It is that image is mapped as into weighted-graph G, G=(V, E), wherein V are undirected
The set on all summits in figure, E is the set on side in non-directed graph；By removing some of non-directed graph side, can be figure segmentation
Into two disjoint vertex sets.For cutting minimum and avoiding the occurrence of the feelings for splitting single summit between K subgraph for ensureing segmentation
Condition, a kind of new undirected subgraph A of measurement two degree uncorrelated between B is defined in Normalized Cut methods：
Wherein,W () represents the weights between two nodes,
It is the first non-directed graph subgraph A interior joints and all remaining nodes in non-directed graph link summation, assoc (B, V) is second undirected
Figure subgraph B interior joints link summation with all remaining nodes in non-directed graph.To can be calculated most with Normalized Cut methods
Excellent minimal cut problem is transformed in the solution to the eigen vector of matrix, simplifies calculating process；
2.2) to each super-pixel SiThe corresponding covariance matrix of middle single pixel point is sued for peace and calculates average, obtains each
Regional center Θ of individual test super-pixeli；
2.3) in the corresponding standard drawing of Polarimetric SAR Image, N number of pixel is randomly selected as instruction to every class ground object target
Practice sample point, and be divided into m set, each set is used as a training super-pixel；To single in each training super-pixel
The corresponding covariance matrix of individual pixel is sued for peace and calculates its average, obtains training corresponding regional center Σ of super-pixeliIf, mark
Quasi- figure includes c class atural objects, then the sum for training super-pixel is c × m.
Polarimetric SAR Image used includes 9 class atural objects in the present invention, for every class atural object randomly selects respectively 900 pictures
Vegetarian refreshments makees training sample point, and these sample points are equally divided into into 30 groups, therefore common property gives birth to 270 training super-pixel.
Step 3：Data matrix X is built using the regional center of super-pixel.
By regional center Σ of training super-pixeliWith regional center Θ of test super-pixeliDischarge successively, obtain data square
Battle array：X=[x1,…xi,…xnum]=[Ψ1,Ψ2,…,Ψc,Θ1,Θ2,…,Θn], wherein xiRepresent i-th in data matrix
The regional center of super-pixel, i=1 ... num, num for super-pixel total number, ΨiThe set that the i-th class trains super-pixel is represented,
Ψi=[Σ1Σ2…Σm]。
Step 4：Physical features correlation matrix Sim is calculated using the regional center of super-pixel.
4.1) Wishart of the amendment between calculating any two super-pixel is apart from disi,j：
Wherein, i=1 ... num, j=1 ... num,It is respectively x, xi,xjMaximal possibility estimation, Tr [] is
The mark of matrix, parameter q represents whether polarization SAR transmitting and reception antenna are same instruction, if so, then q=3, otherwise q=4；
4.2) according to the Wishart for correcting apart from disi,j, obtain dependency simi,j：
4.2) using dependency simi,jBuild physical features correlation matrix Sim：
Step 5：Sparsity structure feature correlation matrix Z is calculated using data matrix X*。
5.1) data matrix X itself is minimized about as the dictionary of rarefaction representation by the order to coefficient matrix Z
Beam, tries to achieve coefficient matrix Z：
Wherein, matrix A is the base of rarefaction representation, i.e. dictionary, and E is noise, and λ is the parameter for controlling influence of noise, and β is control
Matrix Z openness parameter, | | | |*For nuclear norm,For l2,1Norm, EijFor matrix
The element of the i-th rows of E jth row, | | | |0For l0Norm, is not the number of 0 value in representing matrix, because l0Norm Solve problems
It is a NP-hard problem, therefore this constraint is relaxed as l1Norm is solved, and is expressed asZij
For the element of the i-th rows of matrix Z jth row；In solution procedure, Constrained problem is converted into into unconfined non-precision augmentation glug
Bright day multiplier problem：
Wherein, J1, J2It is to solve for the pilot process variable of coefficient matrix Z, J1=Z, J2=Z, Y1,Y2,Y3It is three differences
Lagrange take advantage of the factor, u is penalty factor, | | | |FFor F norms, Tr [] is the mark of matrix；
With reference to Fig. 2, this step is as follows to the solution procedure of non-precision augmentation Lagrange multiplier：
5.1a) input data matrix X, the parameter lambda of impact noise and influence coefficient matrix Z openness parameter beta；
5.1b) initialize matrix coefficient matrix Z=J1=J2=0, noise matrix E=0, Lagrange takes advantage of factor Y1=Y2
=Y3=0, penalty factor u=10-6, max (u)=1010, parameter ρ=1.1 for affecting penalty factor to update, loop termination parameter ε
=10-6；
The cyclic process for solving coefficient matrix Z 5.1c) is entered, its dependent variable is fixed, first pilot process variable is updated
J1：
Its dependent variable 5.1d) is fixed, second pilot process variable J is updated2：
Its dependent variable 5.1e) is fixed, noise matrix E is updated：
Its dependent variable 5.1f) is fixed, coefficient matrix Z is updated：
5.1g) update three different Lagranges and take advantage of the factor：
Y1=Y1+u(X-XZ-E)
Y2=Y2+u(Z-J1)
Y3=Y3+u(Z-J2)；
5.1h) update penalty factor u：U=min (ρ umax (u))；
5.1i) judge whether the condition of convergence meets, if meeting, terminate solving the cyclic process of coefficient matrix Z, otherwise return
Return step 5.1c) process is continued cycling through, the condition of convergence is：
||X-XZ-E||∞＜ ε and | | Z-J1||∞＜ ε and | | Z-J2||∞＜ ε；
5.2) sparsity structure feature correlation matrix is calculated by coefficient matrix ZWherein T representing matrixs
Transposition.
Step 6：To physical features correlation matrix Sim and sparsity structure feature correlation matrix Z*Fusion is weighted, is obtained
New correlation matrix W=Sim+ α Z*, wherein α is the weight for controlling matrix fusion, and span is 0 to 50.
Step 7：Semisupervised classification is carried out using correlation matrix W, the classification of each test super-pixel is obtained；
With reference to Fig. 3, this step is implemented as follows：
7.1) Laplacian Matrix C=M-W is calculated, wherein, M is be made up of the summation of correlation matrix W each row elements right
Angular moment battle array；
7.2) the class probability matrix P of input training super-pixeltrain=[P1,…,Pi,…,Pt], calculate test super-pixel
Class probability matrix
Wherein, Pi=[p1,…pj,…pc]T,K=1 ... c, j=1 ... c, i=1 ...
T, c for target classification number, t be train super-pixel number, PiThe class probability vector of j-th training super-pixel is represented,It is the submatrix of Laplacian Matrix C, represents the Laplacian Matrix of training super-pixel,
ce,fRepresenting matrix C e rows f row element, e=1 ... t, f=t+1 ... num, num for super-pixel total number, Cul=
Clu T, represent the Laplacian Matrix of test super-pixel, the transposition of T representing matrixs；
7.3) class probability matrix P is taken outtestH-th column vector Ph=[p1,h,…ps,h,…pc,h], find PhMiddle maximum
The position s that value is locatedh, the shValue be classification belonging to h-th test super-pixel, wherein h=1 ... n, s=1 ... c, ps,hTable
Show that h-th test super-pixel belongs to the probability of s classes.
Step 8：Classification belonging to each test super-pixel is corresponded to into original polarization SAR image, final classification is exported
As a result.
The effect of the present invention can be further illustrated by following emulation：
1. simulated conditions
The emulation of the present invention is Pentium Dual Core CPU E5200, the hardware of internal memory 2GB in dominant frequency 2.5GHZ
Carry out under the software environment of environment and MATLAB R2010.
2. emulation content
2.1) a width polarization SAR data are input into, enter to be about to the pretreatment of coherent speckle noise, obtain corresponding covariance square
Battle array, and synthesize pseudo color image, as a result such as Fig. 4；
(a) is the pseudo color image of present invention emulation polarization SAR data used in wherein Fig. 4, and visual angle is 4, and image is big
Little is 388 × 430, is the full-polarization SAR figure of the 1m resolution in Flevoland regional NASA/JPL AIRSAR L-bands
The part intercepted as in；(b) is the corresponding atural object distribution standard figure of Polarimetric SAR Image in Fig. 4, wherein 9 class atural objects are had, point
It is not：Semen Brassicae campestriss, soil, Rhizoma Solani tuber osi, Radix Betae, alfalfa, meadow, Semen Tritici aestivi, Semen Pisi sativi, Fructus Hordei Vulgaris, in legend such as Fig. 4 shown in (c)；
2.2) over-segmentation is carried out using the method for Normalized Cut to the pseudo color image shown in Fig. 4, is emulated
Test super-pixel used, as a result such as Fig. 5, the number for testing super-pixel is 334；
2.3) with reference to correlation matrix W, classification is realized using semi-supervised method to the Polarimetric SAR Image shown in Fig. 4, as a result
Such as Fig. 6；
Wherein, (a) is the classification results figure of Polarimetric SAR Image in Fig. 6, and (b) is table of the classification results on standard drawing in 6
Diagram, (c) is the cartogram of classification results in 6, is indicated with confusion matrix.
3. simulated effect analysis
(a) and (b) in Fig. 6 can be seen that the inventive method and carry out polarization SAR figure by elementary cell of super-pixel from 6
The classification of picture, the region of segmentation preferably remains the integrity of zone boundary than more complete, it will be apparent that reduce coherent spot
Impact of the noise to classification results；
(c) is as can be seen that present invention incorporates the physical features and sparsity structure feature of polarization SAR data, profit from Fig. 6
With semi-supervised sorting technique, make in the case where training sample is less, can still obtain preferable classification results, effectively
Improve the accuracy of Classification of Polarimetric SAR Image.
Claims (4)
1. a kind of Polarimetric SAR Image semisupervised classification method based on super-pixel correlation matrix, comprises the steps：
(1) Polarimetric SAR Image is input into, it is carried out to drop the pretreatment of coherent speckle noise, and synthesize pseudo color image；
(2) to pseudo color image over-segmentation, regional center Θ of test super-pixel is calculatediWith the regional center of training super-pixel
Σi；
(3) by regional center Σ of training super-pixeliWith regional center Θ of test super-pixeliDischarge successively, obtain data square
Battle array：X=[x1,…xi,…xnum]=[Ψ1,Ψ2,…,Ψc,Θ1,Θ2,…,Θn], wherein xiRepresent i-th in data matrix
The regional center of super-pixel, i=1 ... num, num for super-pixel total number, ΨiThe set that the i-th class trains super-pixel is represented,
Ψi=[Σ1 Σ2 … Σm]；
(4) Wishart of the amendment between calculating any two super-pixel is apart from disi,j, obtain dependency simi,j, construction
Reason feature correlation matrix Sim:
Wherein,I=1 ... num, j=1 ... num,It is respectively x, xi,xj's
Maximal possibility estimation, Tr [] is the mark of matrix, and parameter q represents whether polarization SAR transmitting and reception antenna are same instruction, if
It is, then q=3, otherwise q=4；
(5) sparsity structure feature correlation matrix Z is built*：
5.1) data matrix X itself is carried out into minimum constraint as the dictionary of rarefaction representation by the order to coefficient matrix Z,
Try to achieve coefficient matrix Z：
Wherein, matrix A is the base of rarefaction representation, i.e. dictionary, and E is noise, and λ is the parameter for controlling influence of noise, and β is constraint square
Battle array Z openness parameter, | | | |*For nuclear norm,For l2,1Norm, EijFor matrix E
The element of the i-th row jth row, | | | |0For l0Norm, is not the number of 0 value in representing matrix, because l0Norm Solve problems
It is a NP-hard problem, therefore this constraint is relaxed as l1Norm is solved, and is expressed asZij
For the element of the i-th rows of matrix Z jth row；
5.2) sparsity structure feature correlation matrix is calculated by coefficient matrix ZThe wherein transposition of T representing matrixs；
(6) to physical features correlation matrix Sim and sparsity structure feature correlation matrix Z*Fusion is weighted, new correlation is obtained
Matrix W=Sim+ α Z*, wherein α is the weight for controlling matrix fusion, and span is 0 to 50；
(7) semisupervised classification is carried out using the correlation matrix W after fusion, obtains the classification of each test super-pixel；
(8) classification belonging to each test super-pixel is corresponded to into original polarization SAR image, exports final classification results.
2. the Polarimetric SAR Image semisupervised classification method based on super-pixel correlation matrix according to claim 1, its feature
It is：Input Polarimetric SAR Image described in step (1), carries out dropping the pretreatment of coherent speckle noise to it, and synthesizes pseudocolour picture
Picture, is carried out in accordance with the following steps：
1.1) a width polarization SAR data are read in, it is carried out using exquisite Lee wave filter drop the pretreatment of coherent speckle noise, obtained
To corresponding covariance matrix, the window size of wave filter is set as 7 × 7；
1.2) Pauli energy feature decomposition is carried out to covariance matrix, and synthesizes the corresponding pseudo color image of polarization SAR data.
3. the Polarimetric SAR Image semisupervised classification method based on super-pixel correlation matrix according to claim 1, its feature
It is：Described in step (2) to pseudo color image over-segmentation, calculate regional center Θ of test super-pixeliWith training super-pixel
Regional center Σi, carry out in accordance with the following steps：
2.1) over-segmentation is carried out to pseudo color image, obtains several super-pixel, S1,S2,…Si,…Sk, as test super-pixel,
Wherein SiI-th super-pixel, i=1 ... k are represented, k is the number of the super-pixel for dividing；
2.2) to each super-pixel SiThe corresponding covariance matrix of middle single pixel point is sued for peace and calculates average, obtains each survey
Regional center Θ of examination super-pixeli；
2.3) in the corresponding standard drawing of the Polarimetric SAR Image per class N number of pixel is randomly selected as training sample point, and will
It is divided into m set, and each set is used as a training super-pixel；It is corresponding to single pixel point in each training super-pixel
Covariance matrix is sued for peace and calculates average, obtains training corresponding regional center Σ of super-pixeliIf standard drawing includes c class targets,
The sum for then training super-pixel is c × m.
4. the Polarimetric SAR Image semisupervised classification method based on super-pixel correlation matrix according to claim 1, its feature
It is：Semisupervised classification is carried out using the correlation matrix W after fusion described in step (7), is carried out as follows：
7.1) Laplacian Matrix C=M-W is calculated, wherein, M is made up of each row element summations of correlation matrix W to angular moment
Battle array；
7.2) the class probability matrix P of input training super-pixeltrain=[P1,…,Pi,…,Pt], the class for calculating test super-pixel is general
Rate matrix
Wherein, Pi=[p1,…pj,…pc]T,K=1 ... c, j=1 ... c, i=1 ... t, c is mesh
Target classification number, t be train super-pixel number, PiThe class probability vector of i-th training super-pixel is represented,It is the submatrix of Laplacian Matrix C, represents the Laplacian Matrix of training super-pixel, ce,f
Representing matrix C e rows f row element, e=1 ... t, f=t+1 ... num, num for super-pixel total number, Cul=Clu T,
Represent the Laplacian Matrix of test super-pixel, the transposition of T representing matrixs；
7.3) class probability matrix P is taken outtestH-th column vector Ph=[p1,h,…ps,h,…pc,h], find PhMiddle maximum institute
Position sh, the shValue be classification belonging to h-th test super-pixel, wherein h=1 ... n, s=1 ... c, ps,hRepresent the
H test super-pixel belongs to the probability of s classes.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410162755.1A CN103927551B (en) | 2014-04-21 | 2014-04-21 | Polarimetric SAR semi-supervised classification method based on superpixel correlation matrix |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410162755.1A CN103927551B (en) | 2014-04-21 | 2014-04-21 | Polarimetric SAR semi-supervised classification method based on superpixel correlation matrix |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103927551A CN103927551A (en) | 2014-07-16 |
| CN103927551B true CN103927551B (en) | 2017-04-12 |
Family
ID=51145767
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410162755.1A Active CN103927551B (en) | 2014-04-21 | 2014-04-21 | Polarimetric SAR semi-supervised classification method based on superpixel correlation matrix |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103927551B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104156918B (en) * | 2014-08-01 | 2017-02-15 | 西安电子科技大学 | SAR image noise suppression method based on joint sparse representation and residual fusion |
| CN104392251B (en) * | 2014-11-28 | 2017-05-24 | 西安电子科技大学 | Hyperspectral image classification method based on semi-supervised dictionary learning |
| CN104408481B (en) * | 2014-12-05 | 2017-07-28 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on depth wavelet neural network |
| CN104463805B (en) * | 2014-12-16 | 2017-03-15 | 西安电子科技大学 | A kind of polarization SAR Approach for Coherent Speckle Reduction based on homogeney significance and set direction |
| CN105989614B (en) * | 2015-02-13 | 2020-09-01 | 中国科学院西安光学精密机械研究所 | Dangerous object detection method fusing multi-source visual information |
| CN104751184B (en) * | 2015-04-08 | 2017-11-21 | 西安电子科技大学 | Based on the full polarimetric SAR sorting technique that intensity statistics is sparse |
| CN105549009B (en) * | 2015-12-09 | 2017-12-15 | 电子科技大学 | A kind of SAR image CFAR object detection methods based on super-pixel |
| CN105787517B (en) * | 2016-03-11 | 2018-12-14 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on the sparse self-encoding encoder of small echo |
| CN106127224A (en) * | 2016-06-13 | 2016-11-16 | 西安电子科技大学 | Based on the semi-supervised polarization SAR sorting technique quickly auctioning figure |
| CN106355625B (en) * | 2016-08-31 | 2019-12-24 | 甘宗平 | Single-polarization SAR false color image generation method |
| CN107239757B (en) * | 2017-05-23 | 2020-12-08 | 西安电子科技大学 | Polarized SAR image target detection method based on deep ladder network |
| CN108197638B (en) * | 2017-12-12 | 2020-03-20 | 阿里巴巴集团控股有限公司 | Method and device for classifying sample to be evaluated |
| CN113344013B (en) * | 2021-07-15 | 2023-06-30 | 西安电子科技大学 | Polarized SAR image feature mining method based on graph low-rank sparse decomposition |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2665018A1 (en) * | 2012-05-14 | 2013-11-20 | Thomson Licensing | Object identification in images or image sequences |
| CN103700091A (en) * | 2013-12-01 | 2014-04-02 | 北京航空航天大学 | Image significance object detection method based on multiscale low-rank decomposition and with sensitive structural information |
-
2014
- 2014-04-21 CN CN201410162755.1A patent/CN103927551B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2665018A1 (en) * | 2012-05-14 | 2013-11-20 | Thomson Licensing | Object identification in images or image sequences |
| CN103700091A (en) * | 2013-12-01 | 2014-04-02 | 北京航空航天大学 | Image significance object detection method based on multiscale low-rank decomposition and with sensitive structural information |
Non-Patent Citations (2)
| Title |
|---|
| "Superpixel-wise semi-supervised structural sparse coding classifier for image segmentation";shuyuan yang等;《Engineering Applications of Artificial Intelligence》;20130820;第2608-2612页 * |
| "非凸低秩稀疏约束的图像超像素分割方法";张文娟，冯象初;《西安电子科技大学学报（自然科学版）》;20131031;第40卷(第5期);第86-91页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103927551A (en) | 2014-07-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103927551B (en) | Polarimetric SAR semi-supervised classification method based on superpixel correlation matrix | |
| Liu et al. | Application of deep convolutional neural networks for detecting extreme weather in climate datasets | |
| CN106355151B (en) | A kind of three-dimensional S AR images steganalysis method based on depth confidence network | |
| Li et al. | Deep learning-based classification methods for remote sensing images in urban built-up areas | |
| Chen et al. | PolSAR image classification using polarimetric-feature-driven deep convolutional neural network | |
| CN102096825B (en) | Graph-based semi-supervised high-spectral remote sensing image classification method | |
| CN108038445B (en) | SAR automatic target identification method based on multi-view deep learning framework | |
| CN103413151B (en) | Hyperspectral image classification method based on figure canonical low-rank representation Dimensionality Reduction | |
| CN104166859B (en) | Based on SSAE and FSALS SVM Classification of Polarimetric SAR Image | |
| CN104077599B (en) | Polarization SAR image classification method based on deep neural network | |
| CN110363151A (en) | Based on the controllable radar target detection method of binary channels convolutional neural networks false-alarm | |
| CN103258324B (en) | Based on the method for detecting change of remote sensing image that controlled kernel regression and super-pixel are split | |
| CN103440505B (en) | The Classification of hyperspectral remote sensing image method of space neighborhood information weighting | |
| CN112052754B (en) | Polarization SAR image ground object classification method based on self-supervision characterization learning | |
| CN108846426A (en) | Polarization SAR classification method based on the twin network of the two-way LSTM of depth | |
| CN107292317A (en) | Polarization SAR sorting technique based on shallow-layer feature Yu T matrix deep learnings | |
| CN105608465B (en) | Full-polarization SAR coastline terrain classification method | |
| CN105741309B (en) | A kind of remote sensing image variation detection method based on the conversion of card side and samples selection | |
| CN107992891A (en) | Based on spectrum vector analysis multi-spectral remote sensing image change detecting method | |
| CN109446894A (en) | The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture | |
| CN109784401A (en) | A kind of Classification of Polarimetric SAR Image method based on ACGAN | |
| CN110276746A (en) | A kind of robustness method for detecting change of remote sensing image | |
| Wang et al. | Multi-view CNN-LSTM neural network for SAR automatic target recognition | |
| CN104751183A (en) | Polarimetric SAR image classification method based on tensor MPCA | |
| Jin et al. | Deep learning for seasonal precipitation prediction over China |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |