Summary of the invention
For the deficiencies in the prior art, the invention provides a kind of Hyperspectral imagery processing method that feature based is strengthened,
Can effectively catch and describe trickle texture.
A kind of Hyperspectral imagery processing method strengthened based on textural characteristics that the present invention provides, described high spectrum image bag
Containing some two dimensional images corresponding with different wave length, described Hyperspectral imagery processing method includes:
1) arbitrary two dimensional image is filtered, obtains the local direction response vector of all pixels in two dimensional image,
The local direction response vector combination repeatedly filtering the multiple directions obtained forms the local direction response vector of all pixels;
2) successively described local direction response vector is normalized and N-state encodes, obtain the direction of numeralization
Vector, obtains the eigenvalue of all pixels, and builds textural characteristics matrix according to described direction vector;
3) circulation step 1)~2) obtain the textural characteristics matrix of all two dimensional images, according to the ripple of each textural characteristics matrix
Long range dependent, carries out textural characteristics reinforcement to each textural characteristics matrix, obtains corresponding textural characteristics and strengthens matrix;
4) strengthen matrix extracts main textural characteristics from all of textural characteristics, formed for representing high spectrum image
Main texture feature vector.
The Hyperspectral imagery processing method based on textural characteristics reinforcement of the present invention uses the eigenvalue of all pixels to retouch
State the textural characteristics of corresponding two dimensional image, it is ensured that the rotational invariance of image, the textural characteristics of all two dimensional images is entered
Row is strengthened and goes out main textural characteristics according to strengthened texture feature extraction and form main texture feature vector for table
Show high spectrum image.By reasonably utilizing the multi-wavelength information of high spectrum image, it is possible to effective seizure high spectrum image is rich
Rich textural characteristics.
Described step 1) according to formula:
It is filtered, wherein:
I represents and works as forward two-dimensional viewing;
LθRepresent as all pixels local direction response vector on θ direction ,-π≤θ≤π in forward two-dimensional viewing;
For second order class Gaussian function:
G2a, G2b, G2cRespectively by second order Gauss function G, (x, y) along counterclockwise rotation 0, pi/2 and the knot of 3 π/4
Really, (x y) is the pixel value in two dimensional image;
For second order class Gaussian function Hilbert transform:
It is and the basic function of x and y independence.Described H2a, H2b, H2cAnd H2dSuch as document Freeman, W.T. , &Adelson,
E.H. (1991), The design and use of steerable filters, IEEE Transactions on
Pattern analysis and machine intelligence,13,891-906。
By change value θ, obtaining each pixel of acquisition has the local direction on direction corresponding in difference, and obtain is multiple
The local direction response vector combination in direction forms the local direction response vector of all pixels.
Described step 2) in pass through formulaTo in described local direction response vector
Each local direction response vector be added the normalization of standard deviation, obtain normalization directional response vector,It is two
In dimension image, pixel is at θpLocal direction response vector on direction,For rightLocal direction after normalization rings
Should be vectorial, P is the direction number of the local direction response vector of two dimensional image,It is right to representSeek standard deviation.
By using the method for normalizing adding standard deviation, it is to avoid during normalization molecule denominator equal proportion occurs
Problem.
Described step 2) N-state coding carry out according to probabilistic model, including:
2-1) according to probabilistic model, the interval of all elements in the local direction response vector after normalization is drawn
It is divided into N number of region;
2-2) described N number of region is used 0 the most successively, 1 ..., N-1 is numbered;
2-3) number table of element affiliated area is shown as the N-state coding result of respective element;
The N-state coding result of all elements in described local direction response vector i.e. constitute the direction of numeralization to
Amount.
Encoded by probabilistic model, improve the accuracy of N-state coding.
Described step 2) the middle eigenvalue using LRP method to obtain all pixels:LOL(LRPP,N, i) represent for showing for P position
LRPP,NPerform the shift left operation of i position, wherein:DpFor pth the element in described direction vector,
N is the state number of N-state coding.
Use LRP method, the pixel represented with direction vector employing eigenvalue is represented, owing to each pixel has P
The local direction of individual different directions is corresponding, therefore, for guaranteeing the rotational invariance of high spectrum image, it is ensured that each pixel is only
One textural characteristics value, the LRP to P positionP,NCarry out shifting function (step-length of movement is i), it may be assumed that
LOL(LRPP,N, i)=DiN0+Di+1N1+...DPNP-i+D0NP-i+1+...+Di-1NP, by P shifting function, take every
The minima of secondary shift result, makes the corresponding eigenvalue of each pixel.The rotation behaviour of the most corresponding two dimensional image of shifting function
Making, the step-length corresponding rotation angle of displacement is operated by multi-shift, it is ensured that the rotational invariance of image.
In step 3) in, when each textural characteristics matrix is carried out textural characteristics reinforcement, first determine whether current texture feature
The dependency of matrix and other each textural characteristics matrixes, by current texture eigenmatrix and with current texture eigenmatrix ripple
Long other relevant all textural characteristics matrixes carry out point-to-point fusion and obtain textural characteristics reinforcement matrix.
By point-to-point fusion method, by and all textural characteristics matrixes relevant to its wavelength carry out Matrix Calculating and put down
All, the new matrix obtained is the reinforcement textural characteristics matrix that this textural characteristics matrix is corresponding, and method is simple, it is easy to accomplish.
In step 3) in, any two textural characteristics matrix whether wavelength is relevant is to obtain according to correlation coefficient, described phase
Close coefficient according to formula:Calculate, CijRepresent the i-th stricture of vagina corresponding with j wavelength
The correlation coefficient of reason eigenmatrix, aiAnd ajIt is respectively the row vector that the i-th textural characteristics matrix tensile corresponding with j wavelength becomes,
If Cij> 0, then judge that two textural characteristics matrix wavelength are correlated with, otherwise, it is determined that wavelength is uncorrelated.
The present invention merges by the textural characteristics of two dimensional image corresponding for all wavelengths, and therefrom obtains three-dimensional Gao Guang
The main textural characteristics of spectrogram picture, compare with the processing method of prior art, the multi-wavelength letter of Appropriate application high spectrum image
Breath, it is possible to capture abundant textural characteristics accurately, it is simple to carry out the differentiation of grain details, be particularly suited for close grain image
(such as the texture image of fillet) is analyzed.By the dimensionality reduction that the big measure feature extracted is carried out, EO-1 hyperion can either be represented accurately
Image, reduces again data volume, is conducive to improving the speed of subsequent applications.
Detailed description of the invention
The Hyperspectral imagery processing the method below in conjunction with specific embodiments feature based of the present invention strengthened do into
One step describes, but protection scope of the present invention is not limited thereto embodiment, and any those familiar with the art exists
In the technical scope that the invention discloses, the change that can readily occur in or replacement, all should contain within protection scope of the present invention.
In the present embodiment as a example by distinguishing the flesh of fish experiment under varying environment.
Instrument prepares
Experimental facilities is by electronic computer, EO-1 hyperion instrument, Halogen light, rectification black-white board.EO-1 hyperion instrument uses U.S. ASD
The Handheld Field Spec spectrogrph of (Analytical Spectral Device) company, spectrum sample is spaced apart
1.5nm, sample range is 380nm~1030nm, uses diffuse-reflectance mode to carry out sample spectrum sampling;Use and join with spectral instrument
14.5 Halogen lights of set, must use rectification black-white board that EO-1 hyperion instrument is carried out conventional corrective before carrying out spectra collection.
Material prepares
Prepare 54 fresh and alive flatfish (turbot), butcher, blood-letting, remove internal organs, clean up, freeze stand-by.Fish
Body weight is between 372g to 580g value (average 512g), and length is at 27.5cm to 32cm (average 30.5cm).Subsequently on chopping block
It is cut into 240 section samples from right to left.Front 96 samples are used as fresh not freezing sample (Fresh), remaining sample
144 samples, take respectively in the environment of 72 samples are placed in-70 DEG C and-20 DEG C and generate quick freezing sample and slow freezing sample
This.Room temperature is set in constant 20 DEG C.Through 9 days, all of freezing sample solved through the time at night in the environment of 4 DEG C
Freeze, form quick freezing defrosting sample set (FFT) and slow freezing defrosting sample set (SFT).For Fresh sample, 64 with
Press proof this be training set, 32 immediately sample be test set.For FFT and SFT sample, employ 48 sample conducts the most respectively
Training set, 24 samples are as test set.
High spectrum image pretreatment
To all of high spectrum image, use the area-of-interest that size is by Hyperspectral imagery processing software ENVI5.0
Region high spectrum image choose.In order to ensure the accuracy of experiment, eliminate by before instrumental effects and illumination effect
Spectrum picture corresponding to 100 wavelength, only chooses the 101-512 clear high spectrum image that 412 light waves are corresponding altogether.In order to
Remove effect of noise, all of high spectrum image uses minimal noise separate conversion (Minimum Noise Fraction
Rotation, MNF Rotation) carry out high s/n ratio correction.
High spectrum image pending in the present embodiment is the high spectrum image that n wavelength size is, each wavelength is corresponding
Two-dimentional high spectrum image M × M rank matrix I represent, two dimension high spectrum image in each pixel have P directional response.
The Hyperspectral imagery processing method that the present embodiment is strengthened based on textural characteristics, as it is shown in figure 1, include:
1) filtering, obtains local direction response vector
Second order class Gaussian functionAnd second order class Gaussian function Hilbert transformMatrix I is carried out convolution, passes through
Quadrature filtering pairWithRespectively the two-dimentional high spectrum image that all wavelengths is corresponding is filtered, obtains tieing up high-spectrum
A local direction response vector of single pixel in Xiang:
P is to be formed to represent pth direction, 0≤p≤P-1, θpThe angle responded mutually for pth direction ,-π≤θp≤π。
Change θpValue, calculate P time, obtain the local direction response vector on P direction of all pixels, and
Form the local direction response vector of all pixels:
WithAccording to second order Gauss filter function G (x, y) and Hilbert transform H2Obtain,
σ2For the variance of class Gaussian function, (x y) is the pixel value in two dimensional image.
Definition:
WhereinG2a, G2b, G2cRespectively
It is corresponding that by G, (x, y) along counterclockwise rotation 0, pi/2 and the result of 3 π/4.The multinomial using three rank goes to approach two
The Hilbert transform H of rank Gaussian function2, obtain:
WhereinWith
H2a, H2b, H2cAnd H2dIt is and the basic function of x and y independence.
2) quantify, obtain textural characteristics matrix
2-1) normalization: use the normalization operation adding standard deviation to local directional response vectorIt is normalized
Pretreatment.
Order:
It is right to representAsk for standard deviation computing, obtain normalized local direction response vector further:
2-2) N-state coding: the dynamic N state encoding based on probabilistic model.
Probabilistic model is defined by following process:
If F (t) is normalized local direction response vector VNormThe value of middle all elements, its span is
[minVNorm, maxVNorm], and 0≤minVNorm< maxVNorm< 1, minVNormAnd maxVNormIt is respectively normalized local side
The minima of element and maximum in response vector, t is certain normalized local direction response vector, i.e. LNormθp, value
Scope is identical with F (t), minVNormAnd maxVNormIt is respectively minima and the maximum of element in normalization directional response vector.
The method using compounded trapezoidal to approach carries out the Integration Solving of F (t), uses parzen window sound directive to normalized local
Probability density f (ω) answered carries out printenv Multilayer networks.
Probability density f (ω) according to normalization directional response sets up probabilistic model:
It is [minV by the codomain of F (t)Norm, maxVNorm] carry out N decile, obtain N number of region and (the present embodiment uses N
Decile divides, and is divided into the region that N number of size is identical), t is divided into N number of respective regions by the region then divided according to F (t), uses
0,1,2 ... N-1 is numbered, and represent that the N-state of the t in respective regions encodes result by numbering.As in figure 2 it is shown, normalizing
Local direction after change is correspondingBe divided in the region of numbered N-2, then the local direction after normalization is correspondingN-state coding result be N-2.So, according to probabilistic model by normalization local direction response vector VNormConvert
Direction vector V for numeralizationD:
VD=(D0,D1,...,DP-1)。 (9)
2-3) calculate textural characteristics: use LRP method to calculate the textural characteristics value of each pixel, it is ensured that the rotation of picture
Invariance.
According to formula:
Calculate textural characteristics value LRP of each pixelP,N, DpThe direction vector V of numeralizationDIn element, to P position
LRPP,NCarry out P shifting function, then to P the LRP obtainedP,NMinimizing, this minima is then the spy of respective pixel point
Value indicative, it may be assumed that
Wherein, LOL (LRPP,N, i) represent the LRP for P positionP,NPerforming shift left operation, the step-length of movement is i position, and i takes
Value is followed successively by 0, and 1 ... P-1.
If VD=(0,1,1,3), it is believed that i=0, represents that two dimensional image does not rotates, and direction vector isRepresenting that two dimensional image hour hands rotate pi/2, direction vector isThe rest may be inferred
The direction vector of different rotation angle can be obtained, then:
LOL(LRPP,N, 0)=0 × 40+1×41+1×42+3×43=192
LOL(LRPP,N, 1)=1 × 40+1×41+3×42+0×43=53
LOL(LRPP,N, 2)=1 × 40+3×41+0×42+1×43=77
LOL(LRPP,N, 3)=3 × 40+0×41+1×42+1×43=83
Take minima, obtain the eigenvalue of this pixel:
The textural characteristics value of all pixels in two dimensional image corresponding for each wavelength is combined the texture forming M × M
Eigenmatrix.
3) textural characteristics is strengthened, and obtains textural characteristics and strengthens matrix
3-1) judge wavelength dependence
The correlation matrix C that definition is symmetrical:
Wherein CijFor the element in correlation matrix C, the phase of the textural characteristics matrix that expression i-th is corresponding with jth wavelength
Pass coefficient:
aiRepresent wavelength i correspondence textural characteristics matrix GiThe row vector being drawn into, ajRepresent textural characteristics corresponding to wavelength j
Matrix GjThe row vector being drawn into, wherein cov (ai,aj) it is ajAnd ajBetween covariance.If CijIn more than 0, then judge GiWith
GjWavelength is correlated with.
3-2) point-to-point fusion
All and G is found out according to correlation coefficientiThe textural characteristics matrix that wavelength is relevant, by special for texture relevant for all wavelengths
Levy matrix and (include Gi) be added be averaging, carry out point-to-point fusion, formed new textural characteristics matrix be textural characteristics strengthen
Matrix.For the high spectrum image of n wavelength, carry out n point-to-point fusion process successively, obtain n textural characteristics and strengthen matrix.
4) extract main textural characteristics, form the main texture feature vector for representing high spectrum image
4-1) respectively n the textural characteristics re-formed after fused is strengthened n the M that matrix is stretched into2Dimensional vector, and
Form a M2× n rank matrix R:
This matrix i.e. represents high spectrum image matrix after textural characteristics reinforcement.
4-2) matrix R is used PCA algorithm, choose three main constituent P1,P2,P3, and it is drawn into row vector, form texture
Feature vector, Xk, P1, P2, P3M for mutual pairwise orthogonal2Dimensional vector.
Repeatedly carry out step 1)~4) obtain the texture feature vector X of the high spectrum image of all samplesk, XkRepresent kth
The texture feature vector of the high spectrum image that sample is corresponding.
5) texture feature vector of all EO-1 hyperion is combined into new matrix and carries out dimension-reduction treatment, to main textural characteristics
Vector carries out dimension-reduction treatment, greatly reduces data volume, improve subsequent applications (predominantly close grain high spectrum image point
Class) the execution time, improve subsequent applications efficiency.
It is l (l=k × 3M that the texture feature vector of all EO-1 hyperion is combined into new matrix X, matrix X2) rank matrix, k is
Number of samples, is 240 in the present embodiment.Using in the present embodiment has the epidemiology learning method DLPP of supervision to be obtained by matrix X dimensionality reduction
To a d dimension data set Y=(y1,y2,...,ym),yi∈Rd(d < < l) wherein:
YT=ATX, (16)
A is that d ties up row vector, tries to achieve by the following method:
At ATXLWXTEquation is solved under the restrictive condition of A=1:
Minimize.
B and W represents the weight matrix that size is l × k, BijRepresent that some i and some j belongs to inhomogeneity, WijRepresent some i and some j
Belong to same class.DBAnd DWFor diagonal matrix, representing matrix B and W is by row or by arranging the sum being added, L respectivelyB=DB-B and LW=
DW-W is Laplacian Matrix.
By (17) algebraic transformation is obtained:
XLBXTA=λ XLWXTA, (18)
λ is characterized value.
By solving eigenvalue characteristic of correspondence vector big for d before in (18), i.e. can get A, A=(a0,a1,...,ad)
It is to have by according to eigenvalue λ1> λ2> ... λdThe characteristic vector composition of arrangement.
Again the A obtained substitution (16) is tried to achieve YT, to YTCarry out the d after dimensionality reduction i.e. tried to achieve by transposition and tie up main textural characteristics square
Battle array Y.
The Y becoming d to tie up main texture spy's vector X dimensionality reduction that l ties up by this method, it is ensured that distance in higher dimensional space
Point farther out remains on larger distance, and some distance close together remains in that nearer.
To result Y after dimensionality reduction in the present embodiment, a young waiter in a wineshop or an inn is used to take advantage of support vector machine storehouse lib-LSSVM (library-
LeastSquare Support Vector Machine, lib-LSSVM build the support vector machine of least square
Different fillet samples is classified by (LeastSquare Support Vector Machine, LSSVM), and has obtained standard
True classification results.