CN102682441A - Hyperspectral image super-resolution reconstruction method based on subpixel mapping - Google Patents

Abstract

The invention discloses a hyperspectral image super to resolution reconstruction method based on the subpixel mapping. The method comprises the following steps of: carrying out the subpixel mapping to acquire proportional values of various ground objects in each pixel element in the original hyperspectral remote sensing image; dividing each pixel element, namely a rough pixel element, in the original hyperspectral remote sensing image into nPoundn subpixels, wherein n is a predefined resolution amplifying factor; assigning each subpixel element to some ground object at random according to the abundance value of each known ground object in each pixel element, so as to obtain the initialized ground object distribution image; then obtaining a ground object distribution image of which resolution is n times of that of the original image by using the simulated annealing algorithm according to the iterative optimization rule; and finally, performing the super to resolution reconstruction based on the subpixel mapping. By considering the partial and overall distribution properties of the ground object at the same time, the hyperspectral image super-resolution reconstruction method based on the subpixel mapping avoids the inauthenticity caused by solving the partial property for the ground object distribution, and thus the hyperspectral remote sensing image can be obtained from the ground object distribution image.

Description

High spectrum image super-resolution method for reconstructing based on inferior pixel mapping

Technical field

The invention belongs to technical field of image processing, be applicable to the high-spectrum remote sensing reconstruction, be specifically related to a kind of high spectrum image super-resolution method for reconstructing based on inferior pixel mapping.

Background technology

In recent years; Satellite remote sensing technology has obtained development at full speed; The new high-spectrum remote-sensing imaging technique that rises can be a lot, obtain face of land observation data on the very narrow and quasi-continuous spectral band; For people provide abundant more atural object observation information, strengthened directly identification and the ability of analyzing the atural object situation from remote sensing images greatly.But high light spectrum image-forming technology spatial resolution when obtaining very high spectral resolution is restricted, and the spatial resolution that therefore how to promote high-spectrum remote sensing becomes problem demanding prompt solution.

For with treat the imaging space plane at a distance of the space remote sensing platform of R, ignoring under the condition that optical system nonlinear distortion and noise influence, its spatial resolution Δ L can be expressed as

$\mathrm{\ΔL}=\frac{\mathrm{WR}}{f}$

Wherein, W is a CCD device array element width, and f is the focal length of optical system.So the spatial resolution that wants to improve remote sensing images can be passed through three kinds of modes, a kind of is to reduce remote-sensing flatform flight track height, but can make the lost of life of this platform like this; Keeping under the constant situation of R, other two kinds of methods that improve resolution are to increase focal length and dwindle array element width, yet; Increasing focal length can make the difficulty of processing of optical element increase; Expense increases, and causes the big Heavy Weight of volume of remote sensor, for the application of reality brings difficulty; The size of CCD array element receives the restriction of technology, because external blockade on new techniques, it is enough little that the CCD aperture that we can access can not be done.To sum up, how can be in distance R, making the remote sensing images spatial resolution get a promotion under the big or small constant prerequisite of focal distance f and CCD aperture W is problem demanding prompt solution.The super-resolution image reconstruction technology is the effective way that addresses this problem, and these techniques make use one or more from different perspectives, the low-resolution image that obtains such as diverse location, different sensors reconstructs the one or more high-definition picture.

The high light spectrum image-forming mode receive the sensor spatial resolution restriction and atural object complex distribution property influence and make that the spatial resolution of gained image is not high; This just causes single pixel institute numerical values recorded information to be mostly the amalgam that multiple type of ground objects spectral signature is constituted, and is called as mixed pixel.The existence of mixed pixel phenomenon has brought great difficulty to the quantitative decipher of image, also brings difficulty for the spatial resolution that improves high spectrum image.Mixed pixel decomposition method is to solve one of this difficult effective way; Its basic meaning is for to extract typical case " pure " object spectrum that is called as end member through mixed pixel spectrum; Estimate each end member shared ratio in mixed pixel simultaneously, this ratio is called the abundance value.For every type of atural object, its abundance value in each pixel forms a width of cloth abundance figure.But this type mixed pixel decomposition method only provides the ratio of every kind of atural object in each pixel, does not provide them and in this pixel, how to distribute.The abundance figure that inferior pixel mapping method utilizes mixed pixel decomposition method to obtain is estimating the distribution of every kind of atural object in each pixel, thereby can access the terrain classification image higher than original image resolution under some restrictive condition.Existing inferior pixel mapping method mainly is divided into two types; Based on study and based on space continuity; The former collects some known high resolving power atural object distributed images as training set; Through being reduced resolution, they obtain abundance figure; Adopt the means training of machine learning to obtain high resolving power atural object distributed image and the relation between the abundance figure accordingly then, and with this relational application on pending high spectrum image abundance figure, thereby obtain needed high resolving power atural object distributed image; Mostly the latter is that based on the algorithm of space continuity distribution has this characteristics proposition different model of local continuity to atural object, distributes through optimizing the optimum atural object that is met this model then.

Summary of the invention

In order to overcome the deficiency of above-mentioned prior art, the object of the present invention is to provide a kind of high spectrum image super-resolution method for reconstructing based on inferior pixel mapping, can access the high resolving power high-spectrum remote sensing higher than original image resolution.

To achieve these goals, the technical scheme of the present invention's employing is:

High spectrum image super-resolution method for reconstructing based on inferior pixel mapping may further comprise the steps:

Step 1, inferior pixel mapping;

Step 1.1 adopts mixed pixel decomposition method to obtain in the original high-spectrum remote sensing ratio value of various atural objects in each pixel, is called the abundance value;

Step 1.2 is called thick pixel with each pixel in the original high-spectrum remote sensing, and each thick pixel is divided into n ￡ n inferior pixel, and wherein n is predefined resolution enlargement factor;

Step 1.3 according to the abundance value of every kind of known atural object in each pixel, is a certain atural object with each inferior pixel Random assignment, promptly obtains initialized atural object distributed image;

Step 1.4 according to the iteration optimization criterion, utilizes simulated annealing that initialized atural object distributed image is carried out iteration optimization; Until satisfying stopping criterion for iteration; Described iteration optimization criterion is based on the local continuity of atural object distribution and the overall similarity of atural object distribution, wherein

The local continuity that atural object distributes, the atural object that is meant same classification in subrange than other atural object of variety classes distribute near;

The overall similarity that atural object distributes is meant that the atural object distribution in the thick pixel that in global scope, has similar proportion atural object composition has similarity, and arbitrary thus pixel can be by representing with the similar pixel of its atural object distribution;

Following objective function is as described iteration optimization criterion:

$\min \underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{t}{\mathrm{\Σ}}{\left|\right|x_t^i-{\mathrm{\μ}}_i\left|\right|}_2^2+\mathrm{\η}{\left|\right|p-{\mathrm{\Σ}}_j{\mathrm{\α}}_j{q}_j\left|\right|}_2^2$

Wherein, t is used for belonging in pending pixel of index and the neighborhood thereof the inferior pixel of atural object i, and c refers to atural object classification number, $x_t^i=\left[\begin{array}{c}{x}_t^i\\ y_t^i\end{array}\right],$ ${\mathrm{\μ}}_i=\left[\begin{array}{c}{\mathrm{\μ}}_{\mathrm{Ix}}\\ {\mathrm{\μ}}_{\mathrm{Iy}}\end{array}\right],$

Finger belongs to the horizontal ordinate and the ordinate of the inferior pixel of atural object classification i, μ _Ix, μ _IyRefer to respectively

Mean value, η is the regularization factor, q _jBe the pixel that has similarity with pixel p, it is selected according to being

The proper vector that f (p) is made up of the abundance value of pixel p and neighborhood thereof, th is a preset threshold value, if a certain pixel q _jProper vector belong to preceding J (j≤J) individual with the similar vector of f (p), this pixel q so _jThe similar pixel that just is considered to pixel p, therefore the inferior pixel distribution in pixel p can be by pixel q _jIn the linear combination that distributes of inferior pixel represent, promptly α _jBe er _jNormalization reciprocal.The local continuity that on behalf of atural object, this objective function first

distribute, second

represents the overall similarity of atural object distribution;

Obtaining a width of cloth resolution thus is original image n atural object distributed image doubly;

Step 2 is based on the super-resolution method for reconstructing of inferior pixel mapping;

It is m ￡ m that the atural object distributed image that step 1 is obtained is divided into a series of sizes; The subimage block of 1m＜n; Recomputate the ratio that wherein comprises various atural objects for each subimage block; Multiply by the pure end member object spectrum that obtains in the mixed pixel decomposition method with this ratio value then; And linear, additive obtains the curve of spectrum of this subimage block, thereby obtains the n=m high-spectrum remote sensing doubly that resolution is original image.

Mixed pixel decomposition method in the said step 1.1 is meant the ratio of the every kind of typical feature that obtains forming mixed pixel, and wherein mixed pixel is meant the pixel that includes dissimilar atural objects in the high spectrum image of acquisition.

The iteration optimization process is the distribution that changes atural object through continuous iteration in the said step 1.4; Wherein each iterative process all can make objective function

value reduce, and said end condition is that target function value no longer descends or reached the iterations of being scheduled to.

Because the present invention will improve the resolution of high spectrum image, so can claim that this image of waiting to improve resolution is a low-resolution image.

Compared with prior art, advantage of the present invention is:

1) the present invention sets up in the local continuity that atural object is distributed and has adopted the minimum criterion of dispersion in the class that makes between identical atural object on the model;

2) the present invention has considered the overall similarity that atural object distributes when considering the local continuity that atural object distributes, and has avoided the untrue property of bringing when only finding the solution to the local characteristics of atural object distribution;

3) the present invention proposes the super resolution ratio reconstruction method that shines upon based on inferior pixel, can obtain high-spectrum remote sensing from the atural object distributed image.

Description of drawings

Fig. 1 is the dividing mode synoptic diagram of thick pixel among the present invention.

Fig. 2 is the synoptic diagram as a result that obtains of inferior pixel mapping method.

Embodiment

Below in conjunction with embodiment the present invention is explained further details.

High spectrum image super-resolution method for reconstructing based on inferior pixel mapping comprises the steps:

Step 1, inferior pixel mapping;

Step 1.1 adopts mixed pixel decomposition method to obtain in the original high-spectrum remote sensing ratio value of various atural objects in each pixel, is called the abundance value;

Mixed pixel decomposition method is a kind of treatment technology commonly used in this area, provides one here than mixed pixel decomposition method commonly used.This method is at first chosen the pure atural object of suitable composition high spectrum image, is referred to as end member, secondly extracts the spectrum of end member from image or in the wave spectrum storehouse, adopts linear mixed model to be used as the spectrum mixture model then, i.e. the r=∑ _ia _is _i+ w 0≤a _i≤1, ∑ _ia _i=1, wherein, r is the spectrum of each pixel in the high spectrum image, s _iBe the spectrum of each end member, a _iBe the ratio value of each end member in this pixel, i.e. abundance value, 0≤a _i≤1 is meant that the ratio value of every type of atural object should be between 0 and 1, ∑ _ia _i=1 is meant that the ratio value sum of all atural objects in this pixel should be 1, and w is an error.Thus, according to making the minimum principle of error w, the abundance value of each atural object in each pixel can obtain through finding the solution aforesaid equation.Present embodiment is that example specifies embodiment of the present invention with a simple analog image, chooses two types of atural objects in this example, i.e. target " 0 " and background " 1 ".Pending image size is 9 pixels * 9 pixels, is not that real high-spectrum remote sensing and mixed pixel decomposition method do not belong to content of the present invention owing to what adopt, so set the abundance value of two types of atural objects in each pixel here shown in table 1 and table 2.

Table 1, the abundance value of atural object " 0 " in each pixel

?	1	2	3	4	5	6	7	8	9
										1	0	0	0	0	0	0	0	0	0
2	0	0	0	0	0	0	0	0	0
										3	0	0	0	0.0625	0.2500	0.0625	0	0	0

[0039]?

4	0	0	0.1250	0.9375	1	0.9375	0.1250	0	0
										5	0	0	0.5000	1	1	1	0.5000	0	0
6	0	0	0.3750	1	1	1	0.3750	0	0
										7	0	0	0	0.5000	0.7500	0.5000	0	0	0
8	0	0	0	0	0	0	0	0	0
										9	0	0	0	0	0	0	0	0	0

Table 2, the abundance value of atural object " 1 " in each pixel

?	1	2	3	4	5	6	7	8	9
										1	1	1	1	1	1	1	1	1	1
2	1	1	1	1	1	1	1	1	1
										3	1	1	1	0.9375	0.7500	0.9375	1	1	1
4	1	1	0.8750	0.0625	0	0.0625	0.8750	1	1
										5	1	1	0.5000	0	0	0	0.5000	1	1
6	1	1	0.6250	0	0	0	0.6250	1	1
										7	1	1	1	0.5000	0.2500	0.5000	1	1	1
8	1	1	1	1	1	1	1	1	1
										9	1	1	1	1	1	1	1	1	1

Step 1.2 is called thick pixel with each pixel in the original high-spectrum remote sensing, and each thick pixel is divided into n ￡ n inferior pixel, and wherein n is predefined resolution enlargement factor;

Among the present invention each pixel in the pending high spectrum image is called thick pixel, promptly 9 * 9 pixels all are called thick pixel in the table 1.Setting the resolution enlargement factor in the present embodiment is 4, and promptly n=4 is equally divided into 4 * 4 inferior pixels with each thick pixel then, and is as shown in Figure 1.

Step 1.3 according to the abundance value of every kind of known atural object in each pixel, is a certain atural object with each inferior pixel Random assignment, promptly obtains initialized atural object distributed image;

With the abundance value of the 5th row target " 0 " of the third line in the table 1 is that 0.25 pixel is an example.After thick pixel is divided into 4 * 4 inferior pixels with each, should there be 4 * 4 * 0.25=4 inferior pixel to belong to target " 0 ", in like manner, there are 12 inferior pixels to belong to background " 1 ".Then in these 4 * 4 inferior pixels at random 4 of pickings be assigned as target " 0 ", remaining 12 inferior pixel is assigned as background " 1 ".

Step 1.4 according to the iteration optimization criterion, utilizes simulated annealing that initialized atural object distributed image is carried out iteration optimization; Until satisfying stopping criterion for iteration; Described iteration optimization criterion is based on the local continuity of atural object distribution and the overall similarity of atural object distribution, wherein

The local continuity that atural object distributes, the atural object that is meant same classification in subrange than other atural object of variety classes distribute near;

The overall similarity that atural object distributes is meant that the atural object distribution in the thick pixel that in global scope, has similar proportion atural object composition has similarity, and arbitrary thus pixel can be by representing with the similar pixel of its atural object distribution;

Following objective function is as described iteration optimization criterion:

$\min \underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{t}{\mathrm{\Σ}}{\left|\right|x_t^i-{\mathrm{\μ}}_i\left|\right|}_2^2+\mathrm{\η}{\left|\right|p-{\mathrm{\Σ}}_j{\mathrm{\α}}_j{q}_j\left|\right|}_2^2$

Wherein, t is used for belonging in pending pixel of index and the neighborhood thereof the inferior pixel of atural object i, and c refers to atural object classification number, $x_t^i=\left[\begin{array}{c}{x}_t^i\\ y_t^i\end{array}\right],$ ${\mathrm{\μ}}_i=\left[\begin{array}{c}{\mathrm{\μ}}_{\mathrm{Ix}}\\ {\mathrm{\μ}}_{\mathrm{Iy}}\end{array}\right],$

Finger belongs to the horizontal ordinate and the ordinate of the inferior pixel of atural object classification i, μ _Ix, μ _IyRefer to respectively

Mean value, η is the regularization factor, q _jBe the pixel that has similarity with pixel p, it is selected according to being

The proper vector that f (p) is made up of the abundance value of pixel p and neighborhood thereof, th is a preset threshold value, if a certain pixel q _jProper vector belong to preceding J (j≤J) individual with the similar vector of f (p), this pixel q so _jThe similar pixel that just is considered to pixel p, therefore the inferior pixel distribution in pixel p can be by pixel q _jIn the linear combination that distributes of inferior pixel represent, promptly α _jBe er _jNormalization reciprocal.The local continuity that on behalf of atural object, this objective function first

distribute, second

represents the overall similarity of atural object distribution;

Simulated annealing is general optimized Algorithm, is used to seek the optimum solution of objective function.In the present embodiment; One by one thick pixel is optimized calculating; Promptly at first calculate the value of objective function under the initial situation; Change the atural object classification ownership of this thick pixel Nei Geya pixel then according to simulated annealing, so constantly repeat this iterative step, no longer descend or reached predetermined iterations up to target function value.

Obtain the atural object distributed image that a width of cloth resolution is 4 times of original images thus, as shown in Figure 2, wherein target " 0 " is represented with black, and background 1 usefulness white is represented.

Step 2,

In the present embodiment; Per 2 * 2 pixels in the high-resolution atural object distributed image that step 1 is obtained are looked as a whole, i.e. m=2, thus formed a series of subimage block; The ratio of two types of atural objects that can wherein be comprised again through simple calculating each subimage block; To multiply by the pure atural object that obtains in the mixed pixel decomposition method be the spectrum of end member with this ratio value then, and linear, additive obtains the curve of spectrum of this subimage block, promptly

B wherein _iBe ratio value, s _iBe the end member spectrum that obtains in the mixed pixel decomposition method.Because present embodiment is example with the analog image, so do not provide the high-resolution remote sensing images that obtain at last.

Claims (3)

1. based on the high spectrum image super-resolution method for reconstructing of inferior pixel mapping, it is characterized in that, may further comprise the steps:

Step 1, inferior pixel mapping;

Step 1.1 adopts mixed pixel decomposition method to obtain in the original high-spectrum remote sensing ratio value of various atural objects in each pixel, is called the abundance value;

Step 1.2 is called thick pixel with each pixel in the original high-spectrum remote sensing, and each thick pixel is divided into n ￡ n inferior pixel, and wherein n is predefined resolution enlargement factor;

Step 1.3 according to the abundance value of every kind of known atural object in each pixel, is a certain atural object with each inferior pixel Random assignment, promptly obtains initialized atural object distributed image;

Step 1.4 according to the iteration optimization criterion, utilizes simulated annealing that initialized atural object distributed image is carried out iteration optimization; Until satisfying stopping criterion for iteration; Described iteration optimization criterion is based on the local continuity of atural object distribution and the overall similarity of atural object distribution, wherein

The local continuity that atural object distributes, the atural object that is meant same classification in subrange than other atural object of variety classes distribute near;

The overall similarity that atural object distributes is meant that the atural object distribution in the thick pixel that in global scope, has similar proportion atural object composition has similarity, and arbitrary thus pixel can be by representing with the similar pixel of its atural object distribution;

Following objective function is as described iteration optimization criterion:

Wherein, t is used for belonging in pending pixel of index and the neighborhood thereof the inferior pixel of atural object i, and c refers to atural object classification number,

Finger belongs to the horizontal ordinate and the ordinate of the inferior pixel of atural object classification i, μ _Ix, μ _IyRefer to respectively

Mean value, η is the regularization factor, q _jBe the pixel that has similarity with pixel p, it is selected according to being

The proper vector that f (p) is made up of the abundance value of pixel p and neighborhood thereof, th is a preset threshold value, if a certain pixel q _jProper vector belong to preceding J (j≤J) individual with the similar vector of f (p), this pixel q so _jThe similar pixel that just is considered to pixel p, therefore the inferior pixel distribution in pixel p can be by pixel q _jIn the linear combination that distributes of inferior pixel represent, promptly

α _jBe er _jNormalization reciprocal.The local continuity that on behalf of atural object, this objective function first distribute, second

represents the overall similarity of atural object distribution;

Obtaining a width of cloth resolution thus is original image n atural object distributed image doubly;

Step 2 is based on the super-resolution method for reconstructing of inferior pixel mapping;

It is m ￡ m that the atural object distributed image that step 1 is obtained is divided into a series of sizes; The subimage block of 1m＜n; Recomputate the ratio that wherein comprises various atural objects for each subimage block; Multiply by the pure end member object spectrum that obtains in the mixed pixel decomposition method with this ratio value then; And linear, additive obtains the curve of spectrum of this subimage block, thereby obtains the n=m high-spectrum remote sensing doubly that resolution is original image.

2. according to the said high spectrum image super-resolution method for reconstructing of claim 1 based on inferior pixel mapping; It is characterized in that; Mixed pixel decomposition method in the said step 1.1 is meant the ratio of the every kind of typical feature that obtains forming mixed pixel, and wherein mixed pixel is meant the pixel that includes dissimilar atural objects in the high spectrum image of acquisition.

3. according to the said high spectrum image super-resolution method for reconstructing of claim 1 based on inferior pixel mapping; It is characterized in that; The iteration optimization process is the distribution that changes atural object through continuous iteration in the said step 1.4; Wherein each iterative process all can make objective function

value reduce, and said end condition is that target function value no longer descends or reached the iterations of being scheduled to.

Images