CN105809185A - High-spectrum image nonlinear demixing method based on neural network and differential search - Google Patents

Abstract

The invention belongs to the technical filed of image processing, and provides a novel high-spectrum image nonlinear demixing method, for realizing the purpose of completing corresponding demixing work of a high-spectrum image with remarkable high-order nonlinear mapping and multilayer scattering between end members. To this end, the technical scheme adopted by the invention brings forward a high-spectrum image nonlinear demixing method based on a neural network and differential search. The method comprises the following steps: 1, inputting high-spectrum image data; 2, according to a p-order polynomial model, randomly generating a training sample for training a multilayer perceptron neural network (MLP); 3, by use of the trained MLP, extracting a nonlinear order of a single pixel point of an image; 4, determining dimensions and position coding of a search individual through a high-spectrum image end member number R; 5, calculating a fitness value of each individual; 6, obtaining a global optimal position of a whole organism through search; 7, completing demixing of the individual pixel point, and otherwise, continuously performing optimization; and 8, stopping calculation after the demixing, and otherwise, continuously demixing a next pixel. The method provided by the invention is mainly applied to image processing.

Description

Based on the hyperspectral image nonlinear solution mixing method that neutral net and difference are searched for

Technical field

The invention belongs to technical field of image processing, relate to based on high spectrum image solution mixing method, specifically a kind of high spectrum image solution mixing method based on neutral net and difference searching algorithm (DSA).

Background technology

Along with the development of science and technology, Global observation by remote sensing is increasingly mature, has been increasingly becoming one of important means of acquisition spatial geographic information.What high-spectrum remote-sensing device was different from tradition multispectral sensor it is critical only that narrow-band imaging, reaches nanoscale in visible ray near infrared region spectral resolution, can obtain object of study in detail accurate spectral information, and therefore range of application is extremely wide.

Complicated variety due to the spatial resolution limit of high light spectrum image-forming spectrogrph and nature atural object, what make the spectrum reflection obtained at single pixel point place is not necessarily the characteristic of a kind of material, and it being probably the mixing of several different material spectrum, such pixel is referred to as mixed pixel.In order to improve the precision of remote sensing application, it is necessary to solve the resolution problem of mixed pixel, make remote sensing application be reached sub-pixed mapping level by pixel level.

The view data that high-spectrum remote-sensing obtains has tens even up to a hundred wave bands, and this is obtain more, finer Endmember extraction to provide possibility, also makes high-spectrum remote-sensing have bigger advantage in spectrum solution is mixed.Meanwhile, spectrum solution mixes up a kind of preprocessing means for high spectrum image, is not only the important prerequisite realizing atural object precise classification and identification, and is the essential condition that deeply develops to quantification of remote sensing technology.Therefore, how effectively interpretation mixed pixel is the key of remote sensing application.Its solution has very real meaning for aspects such as camouflage identification, sub-pixel target detection, the identification of rock ore deposit, precision agriculture, snow lid charting, urban impervious surface estimation and Biomass estimations.

First spectrum solution is mixed needs to set up the mixed model of spectrum.Mixing according to material and the space scale size of physical distribution, spectral mixing can be divided into linear hybrid and non-linear mixing both of which.Linear hybrid assumes that the photon arriving remote sensor is had an effect with unique atural object (i.e. a spectrum end-member composition).Otherwise, when atural object distribution yardstick is less, photon will be had an effect with many kinds of substance, cause non-linear mixing.Linear spectral mixture model has the advantage that modeling is simple, physical meaning is clear and definite, is currently used Pixel Unmixing Models the most widely.But it is applicable to the atural object substantially belonging to or substantially belonging to close, for the detection of the detailed spectral analysis of atural object on some micro-scales or some small probability targets, it is necessary to nonlinear mixed model is explained.

Bilinearity mixed model is widely studied as the one of nonlinear mixed model, it utilizes the Hadamard product of characteristic spectrum to describe in visual field the scattering between photon, but two-wire mixed model only accounts for the scattering between paired end member, and the high-order nonlinear between end member maps and multilamellar scattering is not quantized progressive die type.

When several end member of different nature tight clusters in one piece of region, multilamellar scattering the nonlinear interaction caused is can not be uncared-for, for instance: three end member scenes, urban architecture scene and multilamellar scene etc..These three scene is frequently encountered by hyperspectral image nonlinear solution is mixed, especially the model of multilamellar scene not only can describe in macro-scale (such as hills or mountain area), it is possible to describes the interaction between the end member occurred in micro-scale (as concentrated containing several mineralogical compositions).The present invention uses nonlinear polynomial model to replace bilinear model, is mapped by high-order nonlinear and multilamellar scattering is taken into account, more conform to interact really between photon.When priori disappearances such as true terrestrial object informations, use and have the neutral net of supervision can estimate the non-linear exponent number in this model.

Some scholars propose the gradient algorithm based on different bilinear models, but it has and is subject to initial value impact and is absorbed in local convergence and the shortcoming being only applicable to single model.The present invention uses the difference searching algorithm having better ability of searching optimum to replace gradient algorithm, and it is as novel swarm intelligence algorithm, the shortcoming that can effectively evade gradient method.

Summary of the invention

For overcoming the deficiencies in the prior art, new hyperspectral image nonlinear solution mixing method is provided, the high-order nonlinear between end member can be effectively realized map and the comparatively significant high spectrum image of multilamellar scattering completes the mixed work of corresponding solution, solve mixed effect to be further enhanced, there is theoretical preferably and use value.For this, the present invention adopts the technical scheme that, based on the hyperspectral image nonlinear solution mixing method that neutral net and difference are searched for, comprises the steps:

Step (1): input hyperspectral image data, VCA (VertexComponentAnalysis) vertex component analysis algorithm extracts the end member of true high spectrum image；

Step (2): randomly generate training sample training multilayer perceptron MLP (Multi-layerPerceptronNeuralNetworks) neutral net according to p rank multinomial model, and test its performance；

Step (3): utilize the MLP trained to extract the non-linear exponent number of the single pixel of image；

Step (4): determined the dimension of Search of Individual and position encoded by high optical spectrum image end member number R, initialization principle according to difference searching algorithm DSA, location parameter abundance a and nonlinear factor β ' being set to organism individual initializes, and produces the Search of Individual of respective numbers in search volume；

Step (5): calculate the position resting ground made new advances according to DSA rule, to position individual after migrating, calculate the fitness value of each individuality, every iteration generation, individual body position after updating need to be carried out constraint consistency, it is ensured that meet and retrain；

Step (6): the Greedy principle according to DSA, search obtains the global optimum position of whole organism；

Step (7): when reaching the maximum iteration time arranged, then the position that in output current search population, optimum search is individual, complete to solve mixed single pixel；Otherwise continue to optimize；

Step (8): repeat to carry out pixels all in image stopping calculating after solution is mixed；Otherwise, return step (4), continue to solve to next pixel and mix.

VCA algorithm in described step (1) changes employing Orthogonal subspace projection (OSP), Pure pixel index (PPI), the replacement of monomorphous growth algorithm.

P rank multinomial model in described step (2) is:

${\stackrel{\‾}{y}}_l=\underset{r=1}{\overset{R}{\Σ}}{a}_{rl}{\stackrel{\‾}{m}}_r+\underset{k=2}{\overset{p}{\Σ}}\underset{r=1}{\overset{R}{\Σ}}{\mathrm{\β}}_{rkl}^{\′}{\stackrel{\‾}{m}}_r^k---\left(1\right)$

In formula, by the r end member at all nonlinear interactions in k rank of l pixel by nonlinear factor β '_rklControlling, for meeting high spectrum image inherence end member mixed nature, model must is fulfilled for following non-negative and the constraint of full additivity:

$\begin{array}{c}{a}_{rl}\≥0,{\mathrm{\β}}_{rkl}^{\′}\≥0,\\ \∀r\∈\{1,...,R\},k\∈\{2,...,p\}.\end{array}\underset{r}{\mathrm{\Σ}}{a}_{rl}+\underset{rk}{\mathrm{\Σ}}{\mathrm{\β}}_{rkl}^{\′}=1---\left(2\right)$

DSA in described step (4) is initialized as: for the mixed problem of high spectrum image solution of R kind end member composition of the present invention, amount abundance a to be asked and nonlinear factor β ' is set to the position of Search of Individual, namely

(a₁,a₂,…,a_R,β′_1,2,β′_1,3,…,β′_1,R,β′_2,3,β′_2,4..., β '_2,R,…,β′_R-1,R)。

In described step (5) according to DSA rule calculate rest ground position S be:

S=X_i+γ·(donor-X_i)(3)

Donor is the individuality randomly selected in organism be find properly rest ground migration target.

The amplitude γ of different Search of Individual change in location is

γ=1/GamRnd (a, b) (4)

It is to be randomly generated by the function GamRnd obeying gamma distribution, and a, b are randoms number.The mode of generation amplitude γ necessarily makes superior biological body relatively radical on change investigation；

And it is necessary when individuality crosses the border individual body position carries out non-negative and full additivity constraint consistency, it answers the multiformity of As soon as possible Promising Policy body position, and guarantee that individuality will not lose directivity in transition process, carry out such as the mapping of formula (5) (6):

x′_ij=abs (x_ij)(j≤D)(5)

$x_{ij}^{\′\′}=\frac{x_{ij}^{\′}}{\underset{j=1}{\overset{D-1}{\mathrm{\Σ}}}{x}_{ij}^{\′}}---\left(6\right).$

The Greedy principle of the DSA in described step (6) is:

$S^{*}=\left\{\begin{array}{cc}S,& ify\left(S\right)<y\left(S^{*}\right);\\ S^{*},& otherwise.\end{array}\right.---\left(7\right)$

In formula, y (S) and y (S^*) respectively rest ground fitness value and currently most value.If organism transform rests position, the superior biological body comprising this organism will continue to migrate towards global optimum place.

The evaluation criterion of the global optimum position in described step (7) is:

If abundance a and nonlinear parameter β ' estimates correct, then a certain pixel y in high spectrum image_lWith reconstructed image vegetarian refreshmentsWill closely, the object function that therefore construction solution is mixed

$J(a,{\mathrm{\&beta;}}^{\&prime;})=\min \left|\right|y_l-{\stackrel{\&OverBar;}{y}}_l|{|}^2---\left(8\right)$

Mixed for hyperspectral image nonlinear solution problem is attributed to the optimization problem for object function, and utilizes DS algorithm that object function is optimized to solve, thus obtaining the abundance vector a and nonlinear factor β ' of single pixel；By all pixels of image are repeated the Optimization Solution process based on DS algorithm, finally realize nonlinear solution and mix；Object function can be converted into optimization problem according to p rank multinomial model:

$\min J(a,{\mathrm{\&beta;}}^{\&prime;})=\left|\right|y_l-\underset{r=1}{\overset{R}{\&Sigma;}}{a}_{rl}{\stackrel{\&OverBar;}{m}}_r-\underset{k=2}{\overset{p}{\&Sigma;}}\underset{r=1}{\overset{R}{\&Sigma;}}{\mathrm{\&beta;}}_{rkl}^{\&prime;}{\stackrel{\&OverBar;}{m}}_r^k|{|}^2---\left(9\right)$

In formula,It is the r end member.

The feature of the present invention and providing the benefit that:

Spectrum solution mixes up a kind of preprocessing means for high spectrum image, is not only the important prerequisite realizing atural object precise classification and identification, and is the essential condition that deeply develops to quantification of remote sensing technology.Therefore, how effectively interpretation mixed pixel is the key of remote sensing application.In order to improve the mixed precision of solution, the present invention is on the basis of p rank multinomial model, it is proposed to a kind of mixed algorithm of solution combined based on MLP estimation model nonlinear exponent number and DSA.Neutral net is utilized can effectively to estimate nonlinear model exponent number, and and DSA combine remote sensing images are carried out nonlinear solution mix, the mixed precision of mixed pixel solution can be improved, especially for nonlinear mapping and the more complicated region of multilamellar scattering, performance has bigger lifting.This is significant for processing actual complex high spectrum image, and provides reference for the further optimized development of high spectrum image analytical technology, effectively supports EO-1 hyperion mixed pixel nonlinear solution and is mixed in the application in civilian and military field.

Accompanying drawing illustrates:

Fig. 1 is the flow chart of hyperspectral image nonlinear solution mixing method of the present invention；

Fig. 2 is that choose from EO-1 hyperion original image one piece of region in truthful data of the present invention experiment is as experimental subject；

Fig. 3 is the curve of spectrum of end member in this region；

Fig. 4 is that MLP uses the mixed performance parameter of solution that different number training sample obtains to compare；

Fig. 5 (a) to Fig. 5 (c) is the abundance estimation figure using the nonlinear solution based on neutral net and difference searching algorithm to mix acquisition.(a) water, (b) soil, (c) vegetation.

Detailed description of the invention

It is an object of the invention to as solving the problems of the prior art, for domestic and international present Research at present, the mixed framework of a kind of hyperspectral image nonlinear solution is proposed, neutral net (utilizing MLP neutral net in the present invention) and difference searching algorithm can be combined, mixed for solution problem is converted into the optimization problem of object function, the parameter to be asked solution sneaked out in journey is mapped as the location parameter in difference search procedure, introduces mapping mechanism and meet the abundance non-negative solving mixed requirement and the constraint of full additivity in search procedure.

The present invention estimates model nonlinear exponent number and difference searching algorithm optimization in conjunction with neutral net, under the mixed effect assessment standard reconstructed error of given traditional high spectrum image solution, the program can be effectively realized the high-order nonlinear between end member and maps and complete the mixed work of corresponding solution in the comparatively significant high spectrum image of multilamellar scattering, solve mixed effect to be further enhanced, there is theoretical preferably and use value.

To achieve these goals, the present invention adopts the following technical scheme that

A kind of hyperspectral image nonlinear solution mixing method based on neutral net and difference searching algorithm, comprises the steps:

Step (1): input hyperspectral image data, VCA (VertexComponentAnalysis) algorithm extracts the end member of true high spectrum image.

Step (2): randomly generate training sample training multilayer perceptron MLP (Multi-layerPerceptronNeuralNetworks) neutral net according to p rank multinomial model, and test its performance.

Step (3): utilize the MLP trained to extract the non-linear exponent number of the single pixel of image.

Step (4): determined the dimension of Search of Individual and position encoded by high optical spectrum image end member number R, initialization principle according to difference searching algorithm DSA, location parameter abundance a and nonlinear factor β ' being set to organism individual initializes, and produces the Search of Individual of respective numbers in search volume.

Step (5): calculate the position resting ground made new advances according to DSA rule, to position individual after migrating, calculate the fitness value of each individuality.Every iteration generation, need to carry out constraint consistency to the individual body position after updating, it is ensured that meet constraint.

Step (6): the Greedy principle according to DSA, search obtains the global optimum position of whole organism.

Step (7): when reaching the maximum iteration time arranged, then the position that in output current search population, optimum search is individual, complete to solve mixed single pixel；Otherwise continue to optimize.

Step (8): repeat to carry out pixels all in image stopping calculating after solution is mixed；Otherwise, return step (4), continue to solve to next pixel and mix.

VCA algorithm in described step (1) is the algorithm that main flow extracts true high-spectral data end member, it is also replaced based on the Endmember extraction algorithm of geometric theory by other, replaces as changed employing Orthogonal subspace projection (OSP), Pure pixel index (PPI), monomorphous growth algorithm etc..

P rank multinomial model in described step (2) is:

${\stackrel{\&OverBar;}{y}}_l=\underset{r=1}{\overset{R}{\&Sigma;}}{a}_{rl}{\stackrel{\&OverBar;}{m}}_r+\underset{k=2}{\overset{p}{\&Sigma;}}\underset{r=1}{\overset{R}{\&Sigma;}}{\mathrm{\&beta;}}_{rkl}^{\&prime;}{\stackrel{\&OverBar;}{m}}_r^k---\left(1\right)$

In formula, by the r end member at all nonlinear interactions in k rank of l pixel by nonlinear factor β '_rklControl.For meeting high spectrum image inherence end member mixed nature, model must is fulfilled for following non-negative and the constraint of full additivity:

$\begin{array}{c}{a}_{rl}\&GreaterEqual;0,{\mathrm{\&beta;}}_{rkl}^{\&prime;}\&GreaterEqual;0,\\ \&ForAll;r\&Element;\{1,...,R\},k\&Element;\{2,...,p\}.\end{array}\underset{r}{\mathrm{\&Sigma;}}{a}_{rl}+\underset{rk}{\mathrm{\&Sigma;}}{\mathrm{\&beta;}}_{rkl}^{\&prime;}=1---\left(2\right)$

DSA in described step (4) is initialized as: for the mixed problem of high spectrum image solution of R kind end member composition of the present invention, amount abundance a to be asked and nonlinear factor β ' is set to the position of Search of Individual, namely

(a₁,a₂,…,a_R,β′_1,2,β′_1,3,…,β′_1,R,β′_2,3,β′_2,4..., β '_2,R,…,β′_R-1,R)。

In described step (5) according to DSA rule calculate make new advances rest ground position S be:

S=X_i+γ·(donor-X_i)(3)

Donor is the individuality randomly selected in organism is find properly to rest the migration target on ground, and the amplitude γ of different Search of Individual change in location is

γ=1/GamRnd (a, b) (4)

It is to be randomly generated by the function obeying gamma distribution, and a, b are randoms number.The mode of generation amplitude γ necessarily makes superior biological body relatively radical on change investigation.

And it is necessary when individuality crosses the border individual body position carries out non-negative and full additivity constraint consistency, it answers the multiformity of As soon as possible Promising Policy body position, and guarantee that individuality will not lose directivity in transition process, can carry out such as the mapping of formula (5) (6):

x′_ij=abs (x_ij)(j≤D)(5)

$x_{ij}^{\&prime;\&prime;}=\frac{x_{ij}^{\&prime;}}{\underset{j=1}{\overset{D-1}{\mathrm{\&Sigma;}}}{x}_{ij}^{\&prime;}}---\left(6\right)$

The Greedy principle of the DSA in described step (6) is:

$S^{*}=\left\{\begin{array}{cc}S,& ify\left(S\right)<y\left(S^{*}\right);\\ S^{*},& otherwise.\end{array}\right.---\left(7\right)$

In formula, y (S) and y (S^*) respectively rest ground fitness value and currently most value.S* meaning is the new position asked for.If the fitness value y (S of current location^*) can not be better than resting the fitness value y (S) on ground, then still select originally to rest the fitness value on ground, otherwise choose the fitness y (S choosing out in current location^*) value is the new ground fitness value that rests, this is the Greedy principle of DSA algorithm.If organism transform rests position, the superior biological body comprising this organism will continue to migrate towards global optimum place.

The evaluation criterion of the global optimum position in described step (7) is:

If abundance a and nonlinear parameter β ' estimates correct, then a certain pixel y in high spectrum image_lWith reconstructed image vegetarian refreshmentsWill closely, therefore can the mixed object function of construction solution

$J(a,{\mathrm{\&beta;}}^{\&prime;})=\min \left|\right|y_l-{\stackrel{\&OverBar;}{y}}_l|{|}^2---\left(8\right)$

Mixed for hyperspectral image nonlinear solution problem is attributed to the optimization problem for object function by the present invention, and utilizes DS algorithm that object function is optimized to solve, thus obtaining the abundance vector a and nonlinear factor β ' of single pixel.By all pixels of image are repeated the Optimization Solution process based on DS algorithm, finally realize nonlinear solution and mix.Object function can be converted into optimization problem according to p rank multinomial model:

$\min J(a,{\mathrm{\&beta;}}^{\&prime;})=\left|\right|y_l-\underset{r=1}{\overset{R}{\&Sigma;}}{a}_{rl}{\stackrel{\&OverBar;}{m}}_r-\underset{k=2}{\overset{p}{\&Sigma;}}\underset{r=1}{\overset{R}{\&Sigma;}}{\mathrm{\&beta;}}_{rkl}^{\&prime;}{\stackrel{\&OverBar;}{m}}_r^k|{|}^2---\left(9\right)$

It is the r end member.

Below in conjunction with accompanying drawing, the invention will be further described with embodiment.

Random synthesis sample training MLP, the MLP trained estimate the non-linear exponent number in each pixel of truthful data.The random synthesis training sample when atural object real information lacks, and the end member of sample is to be obtained by VCA.VCA also can replace with other geometry Endmember extraction algorithms.MLP estimates the maximum non-linear exponent number of each pixel in high spectrum image, and the parameter to be asked solution sneaked out in journey is set to the location parameter in difference search procedure, each pixel in truthful data is run and solves mixed algorithm.

Optimum Theory realizes hyperspectral image nonlinear solution and mixes and need to choose suitable object function, and utilizes optimized algorithm that object function is optimized to solve, obtain abundance a and the value of nonlinear factor β '.For the mixed problem of high spectrum image solution of R kind end member composition of the present invention, amount abundance a to be asked and nonlinear factor β ' is set to the position of Search of Individual, i.e. (a₁,a₂,…,a_R,β′_1,2,β′_1,3,…,β′_1,R,β′_2,3,β′_2,4..., β '_2,R,…,β′_R-1,R)。

Specific embodiment:

The present invention chooses Samson area high spectrum image as test object, and it is sized to 952 × 952, and containing 156 wave bands, wavelength band isSpectral resolution is 3.13nm.Owing to original image is too big, computational complexity is higher, and the present invention starts to intercept from (252,332) pixel of original image, chooses the subimage block of 95 × 95 and is used for testing, as shown in Figure 2.

From high-spectral data used, the curve of spectrum of the water (Water), soil (Soil) and vegetation (Tree) the three kinds of main increased surface coverings that obtain, as shown in Figure 3.VCA endmember spectra curve in advance is basic consistent with real end unit curve, can use as actual value.

The present invention utilizes truthful data experimental verification performance, experiment to adopt reconstructed error (ReconstructionError, RE) to evaluate the reconstruction property of invention.

$RE=\sqrt{\frac{1}{MN}\underset{l=1}{\overset{M}{\&Sigma;}}\left|\right|y_l-{\stackrel{\&OverBar;}{y}}_l|{|}^2}---\left(10\right)$

In formula, M is sum of all pixels, and N is wave band number, y_lWithRespectively real spectrum data and reconstruct data.

For proving that MLP estimates the effectiveness of non-linear exponent number, if the maximum non-linear exponent number of pixel is 5, use the MLP of 500 to 5000 alternate 500 sample trainings to estimate the non-linear exponent number of truthful data respectively, solve mixed data and obtain RE.Experiment repeats to take optimal result 10 times, and result is as shown in Figure 4.

As shown in Figure 4: along with the increase of training sample, MLP can more accurately estimate non-linear exponent number, and the RE that inventive algorithm solution mixes acquisition is less, to the lifting of performance little but when training sample continues to increase.Training sample quantity is the important factor in order affecting data reconstruction performance, but number of samples is too much also easy to cause over-fitting, thus causing hydraulic performance decline.Table 1 compares the performance parameter of inventive algorithm and other main flow algorithms.When sample is 2000, inventive algorithm performance is better than other gradient algorithms.

Fig. 5 (a), Fig. 5 (b), Fig. 5 (c) respectively use inventive algorithm solution to mix the relative abundance figure of the water (Water) of acquisition, soil (Soil), vegetation (Tree)., wherein, 0 is black, and 1 is white, and color is more white, illustrates that in this block region, the relative amount of this kind of material is more many.Such as, Fig. 5 (a) is water (Water) relative abundance figure in the images, and the more white region of color illustrates that the relative amount of moisture is more many.

1 four kinds of average RE of algorithm of table compare

For prove MLP estimate non-linear exponent number p effectiveness, table 4 give algorithm omit MLP estimate non-linear exponent number step, and p is fixed to 2 to 5 acquisition RE compared with primal algorithm after result.As shown in Table 2, MLP can estimate the non-linear exponent number of high spectrum image effectively, it is thus achieved that the RE algorithm than fixing p good.

The RE of the non-linear exponent number that the mixed MLP estimation data of non-linear exponent number solution fixed by table 2 compares

When carrying out truthful data experiment, DS algorithm parameter is: population scale N=30, maximum iteration time G=80, dimension D=6, upper limit u_j=1, lower limit l_j=0.Experiment hardware environment is MATLAB2014a, 2.80GHzInteli5CPU and 4GB internal memory.For quantitative measurement algorithm complex, table 3 gives the present invention and calculates the operation time needed for the MLP different number of sample training of use.The MLP training time is longer, and DS algorithm searching process computational complexity is higher, but the multi-core CPU parallel algorithms provided in MATLAB can be adopted can to overcome this shortcoming.Four core CPU parallel computations can reduce by the time of original 3/4ths.Follow-up will use solves disadvantages mentioned above based on the parallel computation of GPU (GraphicsProcessingUnits).

Table 3 uses different number of training sample to estimate the time used of non-linear exponent number

In order to improve the mixed precision of solution, the present invention is on the basis of p rank multinomial model, it is proposed to a kind of mixed algorithm of solution combined based on MLP estimation model nonlinear exponent number and DS algorithm.Experiments show that, utilize neutral net can effectively estimate nonlinear model exponent number, and and DS algorithm combine and remote sensing images are carried out nonlinear solution mix, it is possible to increase mixed pixel solution mixes precision, especially for the region that nonlinear mapping and multilamellar scattering are more complicated, performance has bigger lifting.This is significant for processing actual complex high spectrum image.Follow-up work of the present invention, by testing the effect of inventive algorithm under different test environments more, optimizes MLP and estimates the step of non-linear exponent number, reduces the quantity of unnecessary training sample and improves arithmetic speed based on GPU parallel computation.

The specific embodiment of the present invention is described in conjunction with accompanying drawing although above-mentioned; but non-limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme, those skilled in the art need not pay various amendments or deformation that creative work can make still within protection scope of the present invention.

Draw attention to:

Summary formula 3 upper row eliminates " newly " word, because S* meaning is the new position asked for.Current location is S.

Claims (5)

1. the hyperspectral image nonlinear solution mixing method searched for based on neutral net and difference, is characterized in that, comprise the steps:

Step (1): input hyperspectral image data, VCA (VertexComponentAnalysis) vertex component analysis algorithm extracts the end member of true high spectrum image；

Step (2): randomly generate training sample training multilayer perceptron MLP (Multi-layerPerceptronNeuralNetworks) neutral net according to p rank multinomial model, and test its performance；

Step (3): utilize the MLP trained to extract the non-linear exponent number of the single pixel of image；

Step (4): determined the dimension of Search of Individual and position encoded by high optical spectrum image end member number R, initialization principle according to difference searching algorithm DSA, location parameter abundance a and nonlinear factor β ' being set to organism individual initializes, and produces the Search of Individual of respective numbers in search volume；

Step (5): calculate the position resting ground made new advances according to DSA rule, to position individual after migrating, calculate the fitness value of each individuality, every iteration generation, individual body position after updating need to be carried out constraint consistency, it is ensured that meet and retrain；

Step (6): the Greedy principle according to DSA, search obtains the global optimum position of whole organism；

Step (7): when reaching the maximum iteration time arranged, then the position that in output current search population, optimum search is individual, complete to solve mixed single pixel；Otherwise continue to optimize；

Step (8): repeat to carry out pixels all in image stopping calculating after solution is mixed；Otherwise, return step (4), continue to solve to next pixel and mix.

2. the hyperspectral image nonlinear solution mixing method searched for based on neutral net and difference as claimed in claim 1, it is characterized in that, the VCA algorithm in described step (1) changes a kind of replacement adopted in Orthogonal subspace projection (OSP), Pure pixel index (PPI), monomorphous growth algorithm.

3. the hyperspectral image nonlinear solution mixing method searched for based on neutral net and difference as claimed in claim 1, is characterized in that, the p rank multinomial model in described step (2) is:

${\stackrel{\&OverBar;}{y}}_l=\underset{r=1}{\overset{R}{\&Sigma;}}{a}_{rl}{\stackrel{\&OverBar;}{m}}_r+\underset{k=2}{\overset{p}{\&Sigma;}}\underset{r=1}{\overset{R}{\&Sigma;}}{\mathrm{\&beta;}}_{rkl}^{\&prime;}{\stackrel{\&OverBar;}{m}}_r^k---\left(1\right)$

In formula, by the r end member at all nonlinear interactions in k rank of l pixel by nonlinear factor β '_rklControlling, for meeting high spectrum image inherence end member mixed nature, model must is fulfilled for following non-negative and the constraint of full additivity:

a_rl≥0,β′_rkl≥0、

$\underset{r}{\&Sigma;}{a}_{rl}+\underset{rk}{\&Sigma;}{\mathrm{\&beta;}}_{rkl}^{\&prime;}=1---\left(2\right)$

$\&ForAll;r\&Element;\{1,...,R\},k\&Element;\{2,...,p\}.$

DSA in described step (4) is initialized as: for the mixed problem of high spectrum image solution of R kind end member composition of the present invention, amount abundance a to be asked and nonlinear factor β ' is set to the position of Search of Individual, namely

(a₁,a₂,…,a_R,β′_1,2,β′_1,3,…,β′_1,R,β′_2,3,β′_2,4..., β '_2,R,…,β′_R-1,R)。

4. the as claimed in claim 1 hyperspectral image nonlinear solution mixing method searched for based on neutral net and difference, is characterized in that, the position S on ground of resting calculated according to DSA rule in described step (5) is:

S=X_i+γ·(donor-X_i)(3)

Donor is the individuality randomly selected in organism be find properly rest ground migration target；The amplitude γ of different Search of Individual change in location is:

γ=1/GamRnd (a, b) (4)

Being randomly generated by the function GamRnd obeying gamma distribution, a, b are randoms number.The mode of generation amplitude γ necessarily makes superior biological body relatively radical on change investigation；

And it is necessary when individuality crosses the border individual body position carries out non-negative and full additivity constraint consistency, it answers the multiformity of As soon as possible Promising Policy body position, and guarantee that individuality will not lose directivity in transition process, carry out such as the mapping of formula (5) (6):

x′_ij=abs (x_ij)(j≤D)(5)

$x_{ij}^{\&prime;\&prime;}=\frac{x_{ij}^{\&prime;}}{\underset{j=1}{\overset{D-1}{\&Sigma;}}{x}_{ij}^{\&prime;}}---\left(6\right).$

The Greedy principle of the DSA in described step (6) is:

$S^{*}=\left\{\begin{array}{cc}S,& ify\left(S\right)<y\left(S^{*}\right);\\ S^{*},& otherwise.\end{array}\right.---\left(7\right)$

In formula, y (S) and y (S^*) respectively rest ground fitness value and currently most value.If organism transform rests position, the superior biological body comprising this organism will continue to migrate towards global optimum place.

5. the hyperspectral image nonlinear solution mixing method searched for based on neutral net and difference as claimed in claim 1, is characterized in that, the evaluation criterion of the global optimum position in described step (7) is:

If abundance a and nonlinear parameter β ' estimates correct, then a certain pixel y in high spectrum image_lWith reconstructed image vegetarian refreshmentsWill closely, the object function that therefore construction solution is mixed

$J(a,{\mathrm{\&beta;}}^{\&prime;})=min\left|\right|y_l-{\stackrel{\&OverBar;}{y}}_l|{|}^2---\left(8\right)$

Mixed for hyperspectral image nonlinear solution problem is attributed to the optimization problem for object function, and utilizes DS algorithm that object function is optimized to solve, thus obtaining the abundance vector a and nonlinear factor β ' of single pixel；By all pixels of image are repeated the Optimization Solution process based on DS algorithm, finally realize nonlinear solution and mix；Object function can be converted into optimization problem according to p rank multinomial model:

$\min J(a,{\mathrm{\&beta;}}^{\&prime;})=\left|\right|y_l-\underset{r=1}{\overset{R}{\&Sigma;}}{a}_{rl}{\stackrel{\&OverBar;}{m}}_r-\underset{k=2}{\overset{p}{\&Sigma;}}\underset{r=1}{\overset{R}{\&Sigma;}}{\mathrm{\&beta;}}_{rkl}^{\&prime;}{\stackrel{\&OverBar;}{m}}_r^k|{|}^2---\left(9\right)$

In formula,It is the r end member.