CN105869161A - Method for selecting wave bands of hyperspectral image based on image quality assessment - Google Patents

Abstract

The invention provides a method for selecting wave bands of a hyperspectral image based on image quality assessment, aiming at addressing poor property of selected wave band subsets in determining that is present in non-monitored wave band selection methods. The method includes the following steps: inputting a hyperspectral image that is subject to wave band selection; normalizing the hyperspectral image; averaging wave band images of the normalized hyperspectral image to obtain an average image; quantifying the wave band images and the average image; with the average image as a reference, obtaining the image quality assessment value of each wave band image; replacing mutual information between wave bands and sample labels in the MRMR wave band selection methods with the image quality assessment values; based on the improved MRMR, conducting wave band selection. According to the invention, the method, through introducing image quality assessment, can select high quality wave bands, and has small correlation among wave bands, has excellent determining property of the selected wave band subsets, and is applicable to dimensionality reduction on hyperspectral images.

Description

Hyperspectral image band selection method based on image quality evaluation

Technical field

The invention belongs to technical field of image processing, relate to a kind of hyperspectral image band selection method, be specifically related to one Plant hyperspectral image band selection method based on image quality evaluation, can be used for high spectrum image is carried out dimensionality reduction.

Background technology

Remote sensing is a kind of technology detecting nature object from afar, and all objects in nature all have reflection or launch The characteristic of electromagnetic wave, and object is different, and present environment is different, and the electromagnetic wave of object emission or reflection also differs, quilt The electromagnetic wave of object emission or reflection comprises the characteristic information of object, it is possible to the perceptually instrument of different objects.Based on This principle, the electromagnetic wave that remote sensing technology collects object emission or reflection by physical equipment carrys out perception and differentiates object.By distant The image sensed is exactly remote sensing images.

Multispectral image is that we contact more remote sensing image, its several visible light wave ranges in electromagnetic spectrum usually Upper obtain information simultaneously.Through development for a long time, remote sensing technology has been achieved for the biggest progress, and high-spectrum remote-sensing is remote sensing Technology develops into the product after certain phase.High-spectrum remote sensing the ultraviolet of electromagnetic spectrum, visible ray, near-infrared and in red Obtain information in tens of to hundreds of individual the narrowest and continuous print spectrum segment in exterior domain simultaneously.Compared to traditional remote sensing figure As multispectral image, high spectrum image can provide abundant information and have the highest spectral resolution.EO-1 hyperion Image also there will be the phenomenon of information redundancy while providing bulk information.The information of a large amount of height correlations is high spectrum image Process further and bring computation complexity height, expend the difficult problems such as memory space.Therefore at utmost retain EO-1 hyperion enrich Reducing dimension on the basis of information is an important step in Hyperspectral imagery processing.

Method for high-spectrum image dimensionality reduction has feature extraction and waveband selection.Feature extraction generally by linear or High-dimensional data space is mapped to low-dimensional data space to realize dimensionality reduction by person's non-linear method.Feature extracting method is by mapping The data of the lower dimensional space obtained lose the physical significance representated by initial data.Relative to the method for feature extraction, wave band Selection is based on unmapped dimension reduction method, and this method is the band subset selecting performance performance good from original wave band, Relative to the method for feature extraction, waveband selection will not change original data, preferably remains the physics in initial data Implication and the information of atural object.

High-spectrum image dimensionality reduction has supervision waveband selection and without supervision ripple according to whether the participation of marker samples is divided into Section selects.Having supervision waveband selection to need the participation of marker samples, existing have supervision band selection method to have based on can between class The method of point property, method etc. based on maximal correlation minimal redundancy (MRMR), wherein MRMR is the most classical to have supervision wave band System of selection, its principle is simple and has a higher computational efficiency, and can select high-quality band subset and wave band it Between redundancy relatively low.There is supervision waveband selection because there being the participation of marker samples, it is possible to select the ripple of better performances Cross-talk collection, but obtain marker samples and need to expend substantial amounts of manpower and materials, and also some region is not marked.This Largely limit the range of application of supervision waveband selection.The participation of marker samples is not had without supervision waveband selection, it Be according to wave band between immanent structure select high-quality band subset.Because need not the participation of marker samples, institute With applied widely without supervision band selection method.

Existing have band selection method based on quantity of information without supervision band selection method, and wave band based on similarity selects Selection method and band selection method based on cluster, wherein band selection method based on quantity of information specifically has based on maximum variance The method (MVPCA) of principal component analysis, original wave band is ranked up according to load factor matrix, thus picks out by MVPCA The wave band that variance is big, although the method can select high-quality wave band, but the redundancy between wave band is the highest, wave band it Between information to a great extent be repeat, so the performance performance of the Band Set selected is poor.The second be based on The band selection method of similarity, its thinking is in the residue wave band never selected, and selects to select Band Set the most similar Wave band, the shortcoming of the method be only ensure that selected wave band and with select wave band between redundancy minimum, but Its useful information comprised may seldom, and the quality of Band Set performance that the method is selected is initial with two of setting Wave band has much relations.The third band selection method based on cluster specifically has based on K central point band selection method (k- Medoids), its implementing though, for cluster wave band, then selects centrally located wave band, and what the method selected lacks Point is that the Band Set selected easily is affected by initial center point, so the performance of the wave band selected also has much room for improvement.

Image is human perception and the important information source of Machinery model identification, however image collection, compress, locate Reason, the problem transmitted and create image quality decrease during display etc..How to set up effective image quality evaluation mechanism Widely studied by scholars.Image quality evaluation can be divided into subjective evaluation method and method for objectively evaluating.Subjectivity is commented Valency, objective evaluation has and easily operates, spends the advantages such as low, so objective evaluation has become as the emphasis of research at present, objective comments Whether valency participates in image quality evaluation according to original reference image is divided into full reference, half reference and without with reference to 3 class methods, full reference Image quality evaluating method needs to use original reference image.Image quality evaluation (SSIM) based on structural similarity is a kind of Classical full reference image quality appraisement criterion, SSIM be a kind of weigh original image based on structural information and process after image it Between the method for similarity degree, it calculates the strongest with subjective quality assessment relatedness.Quality evaluation mechanism in recent years exists The aspects such as transmission of video, character recognition, security monitoring, medical science are used widely, but image quality evaluation is not also at height Being applied in terms of spectrum picture dimensionality reduction, and each wave band of high spectrum image is exactly piece image in fact, to each wave band Image carries out quality evaluation, and the value of the quality evaluation obtained may have directive function to EO-1 hyperion dimensionality reduction dimensionality reduction.

Summary of the invention

It is an object of the invention to the defect overcoming above-mentioned prior art to exist, it is proposed that a kind of based on image quality evaluation Hyperspectral image band selection method, by introduce image quality evaluation, MRMR band selection method is improved, propose A kind of new without supervision band selection method, for solving the existing wave band selected existed without supervision band selection method The problem of identification ability difference.

The technical thought of the present invention is: obtained a kind of new nothing supervision waveband selection side by introducing evaluating criterion of quality Method, the method can select the wave band that quality evaluation value is big, the redundancy between the wave band that this wave band selects again and simultaneously Low, such that it is able to obtain comprising the band subset of more discriminant information, reach the purpose of high-spectrum image dimensionality reduction.

According to above-mentioned technical thought, it is achieved the technical scheme that the present invention takes, comprise the steps:

Step 1, inputs high spectrum image A；

Step 2, normalizes between 0-1 by the spectral value of high spectrum image A, obtains the high spectrum image after normalization A′；

Step 3, is designated as x by each band image in the high spectrum image A ' after normalization_i, wherein x_iRepresent i-th ripple Section image, i=1,2...N, each band image in high spectrum image A ' is averaging, obtains the average image, be designated as y '；

Step 4, to band image x each in the high spectrum image A ' after normalization_iQuantify with the average image y ', Band image x after quantization_i' and quantify after the average image y '；

Step 5, utilizes evaluating criterion of quality, with the average band image after quantization as reference, after obtaining described quantization Band image x_i' quality evaluation value, i-th quantify after band image x_i' quality evaluation value, be designated as Q (x_i′,y′)；

Step 6, the quality evaluation value Q (x described in utilization_i', y '), replace formula medium wave used by MRMR band selection method Mutual information between section and sample label, the formula after being replaced:

$\underset{{x_i}^{\′}\∈X-S_{m-1}}{max}\[Q({x_i}^{\′},y^{\′})-\mathrm{\α}\frac{1}{m-1}\underset{{x_q}^{\′}\∈S_{m-1}}{\Σ}I({x_i}^{\′},{x_q}^{\′})\]$

Wherein m refers to the m-th wave band selected, and this wave band is X-S_m-1A wave band in wave band, X is wave band complete or collected works, S_m-1The set of m-1 the wave band for having selected, X-S_m-1For the set of the wave band of pending selection, x_i' it is pending selection One wave band, I (x_i′,x_q') refer to wave band x_i' and wave band x_q' mutual information, x_q' it is the wave band elected, Q (x_i′, Y ') refer to wave band x_i' and wave band x_q' image quality evaluation values, parameter alpha be used for adjusting image quality evaluation values and wave band it Between importance shared by redundancy；

Step 7, according to the formula after described replacement, obtains the MRMR band selection method improved, and utilizes the method to enter Row waveband selection.

The present invention compared with prior art, has the advantage that

First: due to the fact that employing evaluating criterion of quality, by the wave band in MRMR band selection method formula and sample Mutual information between label, replaces with image quality evaluation values, then carries out waveband selection according to the MRMR after replacing it, with existing There is technology to compare, the wave band of informative can be selected, in turn ensure that the dependency between wave band and wave band is less simultaneously, Improve the discriminating power of band subset.

Second: due to the fact that employing evaluating criterion of quality, by the wave band in MRMR band selection method formula and sample Mutual information between label, replaces with quality evaluation value, need not the participation of sample label during waveband selection, with existing MRMR band selection method compare, the high spectrum image not having marker samples can be processed.

Accompanying drawing explanation

Fig. 1 be the present invention realize FB(flow block)；

Fig. 2 is the inventive method and other band selection methods performance comparison on indian pines high spectrum image Curve chart；

Fig. 3 is the present invention and other band selection methods performance comparison curve chart on paviaU high spectrum image；

Detailed description of the invention:

Below in conjunction with the drawings and specific embodiments, the present invention is described in further detail:

Referring to the drawings 1, the high spectrum image employed in the present embodiment is state of Indiana high spectrum image.

Step 1, inputs high spectrum image A；

Step 2, normalizes between 0-1 by the spectral value of high spectrum image A, obtains the high spectrum image after normalization A′；

Step 3, is designated as x by each band image in the high spectrum image A ' after normalization_i, wherein x_iRepresent i-th ripple Section image, i=1,2...220, each band image in high spectrum image A ' is averaging, obtains the average image, be designated as y '；

Step 4, to band image x each in the high spectrum image A ' after normalization_iQuantify with the average image y ', Band image x after quantization_i' and quantify after the average image y '；

Step 5, utilizes evaluating criterion of quality, with the average image after quantization as reference, obtains the wave band after described quantization Image x_i' quality evaluation value, i-th quantify after band image x_i' quality evaluation value, be designated as Q (x_i′,y′)；

The present invention uses the structural similarity evaluating criterion of quality SSIM (structural similarity) matter to image Amount is evaluated, and SSIM is a kind of based on the method for similarity degree between structural information measurement image, and it calculates simple and subjective Quality evaluation relatedness is stronger.Shown in it asks for comprising the following steps that:

(5a) the band image x after quantifying is calculated_i', the average image y's ' after i=1,2...220 and quantization is average strong Degree

${\mathrm{\μ}}_{{x_i}^{\′}}=\frac{1}{M}\underset{j=1}{\overset{M}{\Σ}}{x_{i_j}}^{\′}$

${\mathrm{\μ}}_{y^{\′}}=\frac{1}{M}\underset{j=1}{\overset{M}{\Σ}}{y_j}^{\′}$

Represent the band image x after quantifying_i' mean intensity, μ_y′Represent that the average image y's ' after quantifying is average strong Degree, M refers to band image x_i' and the pixel number of the average image y ',Represent band image x_i' jth pixel, y_jThe jth pixel of ' expression the average image y ', j=1,2...21025；

(5b) the band image x after quantifying is calculated_i' and the luminance contrast function I (x of the average image y ' after quantifying_i′,y′)

$I({x_i}^{\′},y^{\′})=\frac{2{\mathrm{\μ}}_{{x_i}^{\′}}{\mathrm{\μ}}_{y^{\′}}+C_1}{{{\mathrm{\μ}}_{{x_i}^{\′}}}^2+{{\mathrm{\μ}}_{y^{\′}}}^2+C_1}$

Wherein C₁It is in order to avoid denominator is the zero little non-zero constant added, the present invention makes C₁=(K₁L)², ginseng Number L can take the arbitrary value in pixel value excursion, and the excursion of pixel value is 0-255 in this example, so L can take Arbitrary value in 0-255, L value is 255 in the present invention, parameter K₁It is a constant less than 1, K in the present invention₁Value is 0.01。

(5c) the band image x after quantifying is calculated_i' and the standard deviation of the average image y ' after quantifying

${\mathrm{\δ}}_{{x_i}^{\′}}={\left(\frac{1}{M-1}\underset{j=1}{\overset{M}{\Σ}}{\left({x_{i_j}}^{\′}-{\mathrm{\μ}}_{{x_i}^{\′}}\right)}^2\right)}^{\frac{1}{2}}$

${\mathrm{\δ}}_{y^{\′}}={\left(\frac{1}{M-1}\underset{j=1}{\overset{M}{\Σ}}{\left({y_j}^{\′}-{\mathrm{\μ}}_{y^{\′}}\right)}^2\right)}^{\frac{1}{2}}$

WhereinFor the band image x after quantifying_i' standard deviation, δ_y′Standard deviation for the average image y ' after quantifying；

(5d) the band image x after quantifying is calculated_i' and the contrast contrast function C (x of the average image y ' after quantifying_i′, y′)

$C({x_i}^{\′},y^{\′})=\frac{2{\mathrm{\δ}}_{{x_i}^{\′}}{\mathrm{\δ}}_{y^{\′}}+C_2}{{{\mathrm{\δ}}_{{x_i}^{\′}}}^2+{{\mathrm{\δ}}_{y^{\′}}}^2+C_2}$

Wherein C₂It is in order to avoid denominator is the zero little non-zero constant added, the present invention makes C₂=(K₂L)², ginseng Number K₂It is a constant less than 1, K in the present invention₂Value is 0.03.

(5e) the band image x after quantifying is calculated_i' and the Structure Comparison function S (x of the average image y ' after quantifying_i′,y′)

$S({x_i}^{\′},y^{\′})=\frac{{\mathrm{\δ}}_{{x_i}^{\′}{y}^{\′}}+C_3}{{\mathrm{\δ}}_{{x_i}^{\′}}{\mathrm{\δ}}_{y^{\′}}+C_3}$

Wherein C₃It is in order to avoid denominator is the zero little non-zero constant added, order Definition such as Shown in lower

${\mathrm{\δ}}_{{x_i}^{\′}{y}^{\′}}=\frac{1}{M-1}\underset{j=1}{\overset{M}{\Σ}}({x_{i_j}}^{\′}-{\mathrm{\μ}}_{{x_i}^{\′}})({y_j}^{\′}-{\mathrm{\μ}}_{y^{\′}})$

(5f) value of SSIM is calculated

SSIM(x_i', y ')=[I (x_i′,y′)]^η[C(x_i′,y′)]^β[S(x_i′,y′)]^γ

η, β, γ are the constant more than zero, control I (x_i′,y′)、C(x_i′,y′)、S(x_i', y ') shared when seeking SSIM value Importance, SSIM (x_i', y ') i.e. step 5 is designated as Q (x_i', y ') image quality evaluation values.Make in the present invention η=β= γ, by above-mentioned required I (x_i′,y′)、C(x_i′,y′)、S(x_i', y ') and the value of parameter η, β, γ be brought into SSIM (x_i′,y′) Formula arranges and obtains:

$SSIM({x_i}^{\′},y^{\′})=\frac{(2{\mathrm{\μ}}_{{x_i}^{\′}}{\mathrm{\μ}}_{y^{\′}}+C_1)(2{\mathrm{\δ}}_{{x_i}^{\′}{y}^{\′}}+C_2)}{({{\mathrm{\μ}}_{{x_i}^{\′}}}^2+{{\mathrm{\μ}}_{y^{\′}}}^2+C_1)({{\mathrm{\δ}}_{{x_i}^{\′}}}^2+{{\mathrm{\δ}}_{y^{\′}}}^2+C_2)}$

Step 6, the quality evaluation value SSIM (x described in utilization_i', y '), replace in formula used by MRMR band selection method Mutual information between wave band and sample label, the formula after being replaced:

$\underset{{x_i}^{\′}\∈X-S_{m-1}}{max}\[SSIM({x_i}^{\′},y^{\′})-\mathrm{\α}\frac{1}{m-1}\underset{{x_q}^{\′}\∈S_{m-1}}{\Σ}I({x_i}^{\′},{x_q}^{\′})\]$

Wherein m refers to the m-th wave band selected, and this wave band is X-S_m-1A wave band in wave band, X is wave band complete or collected works, S_m-1The set of m-1 the wave band for having selected, X-S_m-1For the set of the wave band of pending selection, x_i' it is pending selection One wave band, I (x_i′,x_q') refer to wave band x_i' and wave band x_q' mutual information, x_q' it is the wave band elected, SSIM (x_i', y ') refer to wave band x_i' and the image quality evaluation values of y ', parameter alpha be used for adjusting image quality evaluation values and wave band it Between importance shared by redundancy；

Step 7, according to the formula after replacing, obtains the MRMR band selection method improved, and utilizes the method to carry out wave band Select, comprise the following steps that shown:

(7a) Band Set of 220 wave band compositions in high spectrum image is designated as X, null set S=[] is set；

(7b) according to the SSIM value tried to achieve, select a wave band from Band Set X, be added into null set S, selected Band Set S₁, original wave band collection X removes first wave band selected, obtains Band Set X-S to be selected₁；

The SSIM value of maximum is selected, using the band image of its correspondence as first elected from 220 SSIM values Wave band, and this wave band is designated as p₁, this wave band is added null set S, has been selected Band Set S₁=[p₁], at original wave band Collection X gets rid of wave band p₁, obtaining Band Set to be selected is X-S₁, its wave band number is 219；

(7c) according to the formula improved from Band Set X-S to be selected₁Second wave band of middle selection, is added into S₁Obtain Select Band Set S₂, at set X-S to be selected₁Second wave band that middle removal is selected, obtains new Band Set X-S to be selected₂

MRMR band selection method, the mutual information between formula medium wave band used and sample label replaces with picture quality Evaluation of estimate SSIM, the formula of improvement is as follows:

$\underset{{x_k}^{\′}\∈X-S_1}{max}\[SSIM({x_k}^{\′},y^{\′})-\mathrm{\α}\underset{{x_q}^{\′}\∈S_1}{\Σ}I({x_k}^{\′},{x_q}^{\′})\]$

$I({x_k}^{\′},{x_q}^{\′})=\∫\∫p({x_k}^{\′},{x_q}^{\′})\log \frac{p({x_k}^{\′},{x_q}^{\′})}{p\left({x_k}^{\′}\right)p\left({x_q}^{\′}\right)}{{\mathrm{dx}}_k}^{\′}{{\mathrm{dx}}_q}^{\′}$

Second wave band, wherein x is selected according to above formula_k' it is Band Set X-S to be selected₁In a wave band, k=1, 2...219, x_q' it is to have selected Band Set S₁In a wave band, now select only one of which wave band in Band Set, i.e. (7b) the wave band p elected in₁, the purpose of this formula be select the big wave band of SSIM value this wave band simultaneously again with select wave band collection The redundancy closing medium wave band is minimum.Second wave band elected is designated as p₂, this wave band is added Band Set S₁, obtain Select Band Set S₂=[p₁p₂], S₂Represent to have selected in Band Set have two wave bands, at X-S₁Set is removed selected Two wave bands, obtain new Band Set X-S to be selected₂, its wave band number is 218, and parameter alpha is used for adjusting image quality evaluation values Importance shared by redundancy between SSIM and wave band, on Indian pines image, α is set to 1.573；

(7d) according to the formula improved from described Band Set X-S to be selected₂The 3rd wave band of middle selection, is added into S₂? To selecting Band Set S₃, at Band Set X-S to be selected₂The 3rd wave band that middle removal is selected, obtains new wave band collection to be selected Close X-S₃, the formula of improvement is as follows:

$\underset{{x_i}^{\′}\∈X-S_2}{max}\[SSIM({x_k}^{\′},y^{\′})-\mathrm{\α}\frac{1}{2}\underset{{x_q}^{\′}\∈S_2}{\Σ}I({x_k}^{\′},{x_q}^{\′})\]$

The 3rd wave band, wherein x is selected according to above formula_k' for Band Set X-S to be selected₂In a wave band, k=1, 2...218, x_q' it is to have selected Band Set S₂In a wave band, selecting now has two wave bands, i.e. (7b) in Band Set In the wave band p that elects₁(7c) the wave band p elected in₂, the purpose of this formula be select the big wave band of SSIM value simultaneously this Wave band is minimum with the redundancy selecting Band Set medium wave band again.The 3rd wave band elected is designated as p₃, this wave band is added Enter S₂, selected Band Set S₃=[p₁p₂p₃], S₃Represent to have selected in Band Set have three wave bands, at X-S₂In set Except the 3rd wave band selected, obtain new Band Set X-S to be selected₃, its wave band number is 217；

(7e) according to the formula improved from Band Set X-S to be selected_m-1Middle selection m-th wave band, is added into S_m-1Obtain Select Band Set S_m, at Band Set X-S to be selected_m-1The m-th wave band that middle removal is selected, obtains new wave band collection to be selected Close X-S_m

$\underset{{x_i}^{\′}\∈X-S_{m-1}}{max}\[SSIM({x_k}^{\′},y^{\′})-\mathrm{\α}\frac{1}{m-1}\underset{{x_q}^{\′}\∈S_{m-1}}{\Σ}I({x_k}^{\′},{x_q}^{\′})\]$

M-th wave band, wherein x is selected according to above formula_k' for Band Set X-S to be selected_m-1In a wave band, k=1, 2...N-(m-1), x_q' it is to have selected Band Set S_m-1In a wave band, selecting now has m-1 wave band in Band Set, The wave band p i.e. elected by above-mentioned steps₁p₂...p_m-1, the purpose of this formula is to select the big wave band of SSIM value this ripple simultaneously Section is minimum with the redundancy of the wave band selecting Band Set again.The m-th wave band elected is designated as p_m, this wave band has been added Select Band Set S_m-1In selected Band Set S_m=[p₁p₂p₃...p_m-1p_m],S_mRepresent to have selected in Band Set have m ripple Section, at X-S_m-1Set is removed the m-th wave band selected, obtains new Band Set X-S to be selected_m；

(7f): repeated execution of steps (7e), wave band is progressively selected, until reaching required wave band number p, selected works the most Close S_pIn be required wave band.

Below in conjunction with analogous diagram, the effect of the present invention is described further.

Simulation process includes:

1. simulated conditions:

The hardware test platform that the emulation experiment of the present invention uses is: processor is Inter Core i7-4720HQ, main Frequency is 2.60GHz, internal memory 8GB；Software platform is: Windows 10 Ultimate 64 bit manipulation system, Matlab R2015a are carried out Emulation testing.

2. emulation data:

The present invention uses the image of the university of Pavia that ROSIS-03 optical pickocff is recorded, and passes through machine The state of Indiana high spectrum image that load visible ray and Infrared Imaging Spectrometer AVIRIS obtain carries out sorting algorithm emulation and tests The card present invention selects the effectiveness of wave band；The high spectrum image of university of Pavia is designated as paviaU, state of Indiana EO-1 hyperion Image is designated as Indian pines, and the two data set is the hyperspectral data processing experimental data that industry is generally acknowledged.

The ground truth figure of above-mentioned high-spectral data and correspondence both be from:

http://www.ehu.es/ccwintco/index.php？Title=Hyperspectral_Remote_ Sensing_Scenes

3. emulation content:

In order to verify the effectiveness of the Band Set of selection, typically after carrying out waveband selection, selected wave band pair can be used The high-spectral data answered, using the classification in ground truth figure as label, classifies to high spectrum image, with standard of classifying Exactness is as the criterion of waveband selection performance.

(3.1) emulation one:

Existing representative the doing with the method that the present invention mentions without supervision band selection method of experimental selection three Relatively.These three method is SBBS (band selection method based on similarity) respectively, MVPCA (divide by maximum variance principal component Analysis), k-medoids (K central point band selection method)；

For Indian pines image, the method for the present invention is used all to have chosen with SBBS, MVPCA and k-medoids 100 wave bands, select 10% known mark sample as training sample, remaining known mark sample as test sample, SVM (svm) grader is used to do classification experiments.In the experiment of this high spectrum image, all of svm grader all uses Libsvm-3.20 realizes, and parameter is set to c=1024, g=2^-7, remaining parameter uses default value；

Effectiveness of the invention is illustrated below by analysis of simulation result；Result such as accompanying drawing 2；

Accompanying drawing 2 is the inventive method and existing three kinds of waveband selection sides obtained on Indian pines high spectrum image The Contrast on effect curve chart of method, the discriminant information that the band subset that band selection method selects comprises is the most, and the classification obtained is accurate Exactness is the highest.Figure presents the accuracy obtained for classification after four kinds of methods pick out 10 to 100 wave bands respectively.From figure In it can be seen that the inventive method (SSIM) except being slightly below k-medoids when 10 wave band, but effect than SBBS and MVPCA is good, and on its all band, effect will be better than three control methods, the especially present invention and picks out acquired by 40 wave bands Effect performance when be better than 100 wave bands that existing three kinds of methods obtain, the waveband selection of the present invention is described Performance is the most outstanding, the abundant discriminant information comprised, dimensionality reduction effective.

(3.2) emulation two:

For paviaU image, the method for the present invention and SBBS, MVPCA and k-medoids is used all to have chosen 50 ripples Section, the known mark sample of selection 1% is as training sample, and remaining known mark sample, as test sample, uses and supports Vector machine (svm) grader does classification experiments.In the experiment of this high spectrum image, svm grader used all uses libsvm- 3.20 realize, and parameter is set to c=1024, g=2, and remaining parameter uses default value；Result such as accompanying drawing 3；

Accompanying drawing 3 is the effect of the inventive method and the existing three kinds of band selection methods obtained on paviaU high spectrum image Really correlation curve figure.Figure presents the accuracy obtained for classification after four kinds of methods pick out 5 to 50 wave bands respectively.From In figure it can be seen that in whole wave band excursion the effect of the inventive method be substantially better than existing three kinds of methods, Er Qieben Invention is when selecting 10 wave bands, and recognition performance has tended towards stability level, and classifying quality is fine, has reached additive method Select level during 50 wave bands, illustrate that the inventive method just can reach optimal effectiveness when picking out less wave band.

To sum up, the effect of the present invention is substantially better than other several method, in the case of different-waveband number, except individually Wave band number, relative additive method recognition performance has had and has been obviously improved, and shows to comprise in the band subset that the present invention selects rich Rich discriminant information, has preferable recognition performance.

Claims (3)

1. a hyperspectral image band selection method based on image quality evaluation, it is characterised in that comprise the following steps that

(1) input high spectrum image Α；

(2) spectral value of described high spectrum image Α is normalized between 0-1, obtain the high spectrum image Α ' after normalization；

(3) each band image in the high spectrum image Α ' after described normalization is designated as x_i, wherein x_iRepresent i-th wave band Image, i=1,2...N, each band image in the high spectrum image Α ' after normalization is averaging, obtains the average image, It is designated as y '；

(4) to band image x each in the high spectrum image Α ' after described normalization_iQuantify with the average image y ', obtain Band image x after quantization_i' and quantify after the average image y '；

(5) utilize evaluating criterion of quality, with the average image y ' after quantization as reference, obtain the band image x after described quantization_i′ Quality evaluation value, i-th quantify after band image x_i' quality evaluation value, be designated as Q (x_i′,y′)；

(6) the quality evaluation value Q (x described in utilization_i', y '), replace formula medium wave band and sample used by MRMR band selection method Mutual information between label, the formula after being replaced:

$\underset{{x_i}^{\′}\∈X-S_{m-1}}{max}\[Q({x_i}^{\′},y^{\′})-\mathrm{\α}\frac{1}{m-1}\underset{{x_q}^{\′}\∈S_{m-1}}{\Σ}I({x_i}^{\′},{x_q}^{\′})\]$

Wherein m refers to select m-th wave band, and this wave band is X-S_m-1A wave band in Band Set, X is wave band complete or collected works, S_m-1The set of m-1 the wave band for having selected, X-S_m-1For the set of the wave band of pending selection, x_i' it is pending selection One wave band, I (x_i′,x_q') refer to wave band x_i' and wave band x_q' mutual information, x_q' it is the wave band elected, Q (x_i′, Y ') refer to wave band x_i' and the image quality evaluation values of wave band y ', parameter alpha be used for adjusting image quality evaluation values and wave band it Between importance shared by redundancy；

(7) according to the formula after described replacement, obtain the MRMR band selection method improved, and utilize the method to carry out wave band choosing Select.

Hyperspectral image band selection method based on image quality evaluation the most according to claim 1, it is characterised in that Quality evaluation value described in step (5), uses evaluating criterion of quality SSIM based on structural similarity to ask for, according to following step Rapid realization:

(5a) the band image x after quantifying is calculated_i', the mean intensity of the average image y ' after i=1,2...N and quantization

${\mathrm{\μ}}_{{x_i}^{\′}}=\frac{1}{M}\underset{j=1}{\overset{M}{\Σ}}{x_{i_j}}^{\′}$

${\mathrm{\μ}}_{y^{\′}}=\frac{1}{M}\underset{j=1}{\overset{M}{\Σ}}{y_j}^{\′}$

Represent the band image x after quantifying_i' mean intensity, μ_yThe mean intensity of the average image y ' after ' expression quantization, M Refer to band image x_i' and the pixel number of the average image y ',Represent band image x_i' jth pixel, y_j' table Show the jth pixel of the average image y ', j=1,2...M；

(5b) the band image x after quantifying is calculated_i' and the luminance contrast function I (x of the average image y ' after quantifying_i′,y′)

$I({x_i}^{\′},y^{\′})=\frac{2{\mathrm{\μ}}_{{x_i}^{\′}}{\mathrm{\μ}}_{y^{\′}}+C_1}{{{\mathrm{\μ}}_{{x_i}^{\′}}}^2+{{\mathrm{\μ}}_{y^{\′}}}^2+C_1}$

Wherein C₁=(K₁L)², parameter L can take the arbitrary value in pixel value excursion, parameter K₁It it is a constant less than 1；

(5c) the band image x after quantifying is calculated_i' and the standard deviation of the average image y ' after quantifying

${\mathrm{\δ}}_{{x_i}^{\′}}={\left(\frac{1}{M-1}\underset{j=1}{\overset{M}{\Σ}}{\left({x_{i_j}}^{\′}-{\mathrm{\μ}}_{{x_i}^{\′}}\right)}^2\right)}^{\frac{1}{2}}$

${\mathrm{\δ}}_{y^{\′}}={\left(\frac{1}{M-1}\underset{j=1}{\overset{M}{\Σ}}{\left({y_j}^{\′}-{\mathrm{\μ}}_{y^{\′}}\right)}^2\right)}^{\frac{1}{2}}$

WhereinFor the band image x after quantifying_i' standard deviation, δ_y' for the standard deviation of the average image y ' after quantifying；

(5d) the band image x after quantifying is calculated_i' and the contrast contrast function C (x of the average image y ' after quantifying_i′,y′)

$C({x_i}^{\′},y^{\′})=\frac{2{\mathrm{\δ}}_{{x_i}^{\′}}{\mathrm{\δ}}_{y^{\′}}+C_2}{{{\mathrm{\δ}}_{{x_i}^{\′}}}^2{{\mathrm{\δ}}_{y^{\′}}}^2+C_2}$

Wherein C₂=(K₂L)², parameter K₂It it is a constant less than 1；

(5e) the band image x after quantifying is calculated_i' and the Structure Comparison function S (x of the average image y ' after quantifying_i′,y′)

$S({x_i}^{\′},y^{\′})=\frac{{\mathrm{\δ}}_{{x_i}^{\′}{y}^{\′}}+C_3}{{\mathrm{\δ}}_{{x_i}^{\′}}{\mathrm{\δ}}_{y^{\′}}+C_3}$

Wherein Shown in being defined as follows

${\mathrm{\δ}}_{{x_i}^{\′}{y}^{\′}}=\frac{1}{M-1}\underset{j=1}{\overset{M}{\Σ}}({x_{i_j}}^{\′}-{\mathrm{\μ}}_{{x_i}^{\′}})({y_j}^{\′}-{\mathrm{\μ}}_{y^{\′}})$

(5f) value of SSIM is calculated

SSIM(x_i', y ')=[I (x_i′,y′)]^η[C(x_i′,y′)]^β[S(x_i′,y′)]^γ

η, β, γ are the constant more than zero, control I (x_i′,y′)、C(x_i′,y′)、S(x_i', y ') weight shared when seeking SSIM value The property wanted, SSIM (x_i', y ') i.e. step 5 is designated as Q (x_i', y ') image quality evaluation values.

Hyperspectral image band selection method based on image quality evaluation the most according to claim 1, it is characterised in that Waveband selection described in step (7), its wave band number scale selected is P, realizes as follows:

(7a) Band Set that the N number of wave band in high spectrum image forms is designated as X, null set S=[] is set；

(7b) according to described SSIM value, select a wave band from Band Set X, be added into null set S, selected wave band Set S₁, original wave band collection X removes first wave band selected, obtains Band Set X-S to be selected₁；

The SSIM value of maximum is selected from N number of SSIM value, using the band image of its correspondence as first wave band elected, and This wave band is designated as p₁, this wave band is added null set S, has been selected Band Set S₁=[p₁], in original wave band collection X Remove wave band p₁, obtaining Band Set to be selected is X-S₁, its wave band number is N-1；

(7c) according to the formula improved from described Band Set X-S to be selected₁Second wave band of middle selection, is added into S₁Obtain Select Band Set S₂, at set X-S to be selected₁Second wave band that middle removal is selected, obtains new Band Set X-S to be selected₂；

MRMR band selection method, the mutual information between formula medium wave band used and sample label replaces with quality evaluation value SSIM, the formula of improvement is as follows:

$\underset{{x_k}^{\′}\∈X-S_1}{max}\[SSIM({x_k}^{\′},y^{\′})-\mathrm{\α}\underset{{x_q}^{\′}\∈S_1}{\mathrm{\Σ}}I({x_k}^{\′},{x_q}^{\′})\]$

$I({x_k}^{\′},{x_q}^{\′})=\∫\∫p({x_k}^{\′},{x_q}^{\′})log\frac{p({x_k}^{\′},{x_q}^{\′})}{p\left({x_k}^{\′}\right)p\left({x_q}^{\′}\right)}{{\mathrm{dx}}_k}^{\′}{{\mathrm{dx}}_q}^{\′}$

Second wave band, wherein x is selected according to above formula_k' it is Band Set X-S to be selected₁In a wave band, k=1,2...N-1, x_q' it is to have selected Band Set S₁In a wave band, selecting in only one of which wave band in Band Set, i.e. (7b) now is selected The wave band p come₁, the purpose of this formula be select the big wave band of SSIM value this wave band simultaneously again with select Band Set medium wave band Redundancy is minimum.Second wave band elected is designated as p₂, this wave band is added Band Set S₁, selected Band Set S₂=[p₁p₂], S₂Represent to have selected in Band Set have two wave bands, at X-S₁Set is removed second wave band selected, To new Band Set X-S to be selected₂, its wave band number is N-2, and parameter alpha is used for adjusting image quality evaluation values SSIM and wave band Between importance shared by redundancy；

(7d) according to the formula improved from described Band Set X-S to be selected₂The 3rd wave band of middle selection, is added into S₂Obtain Select Band Set S₃, at Band Set X-S to be selected₂The 3rd wave band that middle removal is selected, obtains new Band Set X-to be selected S₃；

$\underset{{x_k}^{\′}\∈X-S_2}{max}\[SSIM({x_k}^{\′},y^{\′})-\mathrm{\α}\frac{1}{2}\underset{{x_q}^{\′}\∈S_2}{\Σ}I({x_k}^{\′},{x_q}^{\′})\]$

The 3rd wave band, wherein band image x is selected according to above formula_k' for Band Set X-S to be selected₂In a wave band, k=1, 2...N-2, band image x_q' it is to have selected Band Set S₂In a wave band, selecting now has two ripples in Band Set Section, the wave band p i.e. elected in (7b)₁(7c) the wave band p elected in₂, the purpose of this formula is to select the ripple that SSIM value is big Section this wave band simultaneously is minimum with the redundancy selecting Band Set medium wave band again.The 3rd wave band elected is designated as p₃, will This wave band adds S₂, selected Band Set S₃=[p₁p₂p₃], S₃Represent to have selected in Band Set have three wave bands, at X-S₂ Set is removed the 3rd wave band selected, obtains new Band Set X-S to be selected₃, its wave band number is N-3；

(7e) according to the formula improved from described Band Set X-S to be selected_m-1Middle selection m-th wave band, is added into S_m-1Obtain Select Band Set S_m, at Band Set X-S to be selected_m-1The m-th wave band that middle removal is selected, obtains new wave band collection to be selected Close X-S_m；

$\underset{{x_k}^{\′}\∈X-S_{m-1}}{max}\[SSIM({x_k}^{\′},y^{\′})-\mathrm{\α}\frac{1}{m-1}\underset{{x_q}^{\′}\∈S_{m-1}}{\Σ}I({x_k}^{\′},{x_q}^{\′})\]$

M-th wave band, wherein x is selected according to above formula_k' for Band Set X-S to be selected_m-1In a wave band, k=1,2...N- (m-1), band image x_q' it is to have selected Band Set S_m-1In a wave band, selecting now has m-1 ripple in Band Set Section, the wave band p i.e. elected by above-mentioned steps₁p₂...p_m-1, the purpose of this formula be select the big wave band of SSIM value simultaneously this Wave band is minimum with the redundancy of the wave band selecting Band Set again.The m-th wave band elected is designated as p_m, this wave band is added Select Band Set S_m-1In selected Band Set S_m=[p₁p₂p₃...p_m-1p_m],S_mRepresent to have selected in Band Set have m Wave band, at X-S_m-1Set is removed the m-th wave band selected, obtains new Band Set X-S to be selected_m；

(7f): repeated execution of steps (7e), wave band is progressively selected, until reaching required wave band number p, the most selected works to close S_p In be required wave band.