CN104732491A - Edge priority guide single-frame remote sensing image super-resolution processing method - Google Patents

Abstract

The invention provides an edge priority guide single-frame remote sensing image super-resolution processing method. Edge details of an image to be processed serve as priorities high-frequency information restoration is conducted on an intermediate image generated through interpolation, and the details are prevented from being lost or becoming fuzzy. The method is based on the regular optimization theory, a single-frame image super-resolution processing mechanism is provided, the image detail description capacity can be effectively improved, the quality of the image can be improved, the parallel capacity of the remote sensing image which is large in data size is good, and execution efficiency is high.

Description

The single frames Remote Sensing Image Super Resolution disposal route that a kind of edge prior guides

Technical field

The invention belongs to image processing field, relate to a kind of super-resolution processing method of single frames remote sensing images.

Background technology

Resolution weighs an important evaluation index of optical satellite image image quality, the detailed information comprised in remote sensing images and resolution closely related, it directly affects analysis to target, identification and station-keeping ability.Due to the restriction of pixel dimension, add high resolution camera bulky, involve great expense, often can only get low-resolution image.Therefore under existing imaging circumstances (image system hardware condition cannot be changed) and data source condition, how according to information theory, adopting digital image processing techniques to improve the resolution of remote sensing images, is one of important step of satellite ground analysis and treament.

Conventional super-resolution method is summed up and can be divided into two classes: a class is multiframe super-resolution method, namely obtains the known image of relative motion relation, utilizes the sample information of redundancy to build the high resolving power details of overlapping region; Another kind of is single frames super-resolution method, namely only has a width image of target area.In most cases, satellite imaging equipment does not support jog mode, cannot obtain multiple image, or the relative motion relation between multiframe is difficult to solve, and therefore the breakthrough of single frames super-resolution technique is more urgent.

Single frames super-resolution method main at present comprises: the method for interpolation, reconstruct and statistical learning.

Typical interpolation method comprises: arest neighbors interpolation, linear interpolation, bicubic interpolation (Keys, Hou) etc.These class methods are usually fuzzy high-frequency information, especially edge details, and serious blocky effect can be produced, be difficult to stand in remote sensing application.But its computing velocity is fast, real-time good, generally as the basis of complicated oversubscription method.In order to overcome the deficiency of classic method, Li, Allebach etc. propose the edge interpolation algorithm based on least square method, make moderate progress to the edge of interpolation, but cannot overcome the fringe effects of point interblock in the application of significantly remote sensing images.

Method based on reconstruct needs operator to provide certain priori usually, result and the original image of super-resolution is carried out to priori is consistent to project, and practicality is higher.This kind of algorithm mainly contains iterative backprojection method, projections onto convex sets.Iterative backprojection method is proposed by people such as Irani and is incorporated in super-resolution problem, and back projection's error is added to by the method equably rebuilds on image, causes rebuilding image at marginal existence sawtooth effect and marginal information is fuzzy.The people such as Stark propose projections onto convex sets based on sets theory, attempt solution to project on constraint set.The people such as Morse utilize level set to achieve reconstruct.These algorithm mathematicses describe directly perceived, but constraint set is often difficult to definition, and for the process of view picture satellite image, computing velocity needs to improve.

Method based on study needs structure low resolution and high-definition picture Sample Storehouse, obtains both inherent contacts, complete reconstruction by learning sample storehouse.In early days, the people such as Chang propose neighborhood embedded mobile GIS, make the algorithm based on study obtain very large concern.The people such as Yang utilize the thought crossing complete dictionary in rarefaction representation, achieve good effect.But it should be noted that any one learning method all can the sample of heavy dependence study, and calculated amount is huge.

Summary of the invention

The technical matters that the present invention solves is: overcome that complexity when existing single frames super-resolution method is applied to remote sensing images is high, the deficiency such as blocky effect and atural object edge fog; for significantly remote sensing images; provide a kind of single frames super-resolution method for parallel processing based on edge prior constraint; merge gradient fields to shift and projection process method after iteration; the final image edge details rebuild is protected, and the quality index such as signal to noise ratio (S/N ratio) does not decline to a great extent and processing speed is fast simultaneously.

Technical solution of the present invention is: the single frames Remote Sensing Image Super Resolution disposal route that a kind of edge prior guides, and comprises the steps:

(1) original image I is obtained _l, and to I _lcarry out piecemeal, to I _lblock image carry out index and be designated as SubI _l;

(2) according to super-resolution multiple m, to each SubI _lcarry out interpolation, obtain block image SubI _htmp, SubI _htmpset form procedural image I _htemp;

(3) for any SubI _land the SubI after corresponding interpolation _htmp, all perform following operation:

(3.1) by super-resolution multiple m, form fuzzy filter h, utilize h to SubI _land SubI _htmpcarry out convolution respectively, and calculate SubI respectively on the basis of convolution _lline direction outline map SubL _ex, column direction outline map SubL _ey, and SubI _htmpline direction outline map SubHtmp _ex, column direction outline map SubHtmp _ey;

(3.2) to SubL _exaccording to pixels ask absolute value, and carry out linear stretch by maximin and obtain linear edge figure | SubL _ex| _lin,

${|\mathrm{SubL}}_{\mathrm{ex}}{|}_{\mathrm{lin}}=\frac{\left|{\mathrm{SubL}}_{\mathrm{ex}}\right|-{\mathrm{Min}}_{\mathrm{ex}}}{{\mathrm{Max}}_{\mathrm{ex}}-{\mathrm{Min}}_{\mathrm{ex}}}$

Wherein, pixel maximum absolute value value Max _ex=Max (| SubL _ex|), pixel absolute value minimum M in _ex=Min (| SubL _ex|); Then to SubL _ey, SubHtmp _ex, SubHtmp _eyperform identical operation, obtain corresponding linear edge figure | SubL _ey| _lin, | SubHtmp _ex| _linwith | SubHtmp _ey| _lin;

(3.3) linear edge figure is calculated respectively | SubL _ex| _lin, | SubL _ey| _lin, | SubHtmp _ex| _linwith, | SubHtmp _ey| _linvariance L _vx, L _vy, Htmp _vx, Htmp _vy, and ask for gradient fields transfer parameters L _varwith H _var, wherein:

L _vx＝std(|SubL _ex| _lin)

L _vy＝std(|SubL _ey| _lin)

Htmp _vx＝std(|SubHtmp _ex| _lin)

Htmp _vy＝std(|SubHtmp _ey| _lin)

L _var＝0.5(L _vx+L _vy)

H _var＝0.5(Htmp _vx+Htmp _vy)

Wherein std represents that variance calculates;

(3.4) first differential is adopted to calculate SubI _htmpline direction gradient fields Htmp _gxwith column direction gradient fields Htmp _gy, complete gradient fields transfer, wherein:

${\mathrm{Htmp}}_{\mathrm{Gx}}={\mathrm{SubI}}_{H{\mathrm{tmp}}_x}-{\mathrm{SubI}}_{H{\mathrm{tmp}}_{x-1}}$

${\mathrm{Htmp}}_{\mathrm{Gy}}={\mathrm{SubI}}_{H{\mathrm{tmp}}_y}-{\mathrm{SubI}}_{H{\mathrm{tmp}}_{y-1}}$

After gradient fields transfer

${\mathrm{Htmp}}_{\mathrm{Gx}}={\mathrm{Htmp}}_{\mathrm{Gx}}*\mathrm{varf}*e^{\frac{-{\mathrm{dist}}_x^2}{L_{\mathrm{var}}}}/e^{\frac{-{\mathrm{dist}}_x^2}{H_{\mathrm{var}}}}$

$\mathrm{varf}=\frac{H_{\mathrm{var}}}{L_{\mathrm{var}}},{\mathrm{dist}}_x=0.5\left(\right|{\mathrm{SubHtmp}}_{{\mathrm{ex}}_x}{|}_{\mathrm{lin}}+\left|{\mathrm{SubHtmp}}_{{\mathrm{ex}}_{x-1}}{|}_{\mathrm{lin}}\right)$

${\mathrm{Htmp}}_{\mathrm{Gy}}={\mathrm{Htmp}}_{\mathrm{Gy}}*\mathrm{varf}*e^{\frac{-{\mathrm{dist}}_y^2}{L_{\mathrm{var}}}}/e^{\frac{-{\mathrm{dist}}_y^2}{H_{\mathrm{var}}}}$

${\mathrm{dist}}_y=0.5\left(\right|{\mathrm{SubHtmp}}_{{\mathrm{ex}}_y}{|}_{\mathrm{lin}}+\left|{\mathrm{SubHtmp}}_{{\mathrm{ey}}_{y-1}}{|}_{\mathrm{lin}}\right)$

In formula, footmark x and x-1 represents the adjacent rows of line direction, and footmark y and y-1 represents adjacent two row of column direction;

(3.5) according to the gradient fields after transfer, utilize sparse matrix solver to resolve standard P oisson image equation, obtain new high-definition picture sub-block SubI _c;

(3.6) adopt iterative backward projection to SubI _cwith SubI _lcarry out iterative projection, obtain oversubscription result images sub-block and be designated as SubI _h;

(4) to whole SubI that step (3) obtains _hcarry out image mosaic, obtain I _lsuper-resolution result images I _h.

In described step (3.1), the calculating of outline map adopts normalization Sobel operator, and the pixel of operator is long and wide all gets [2R (m+1)+1], and R is floor operation.

The method of carrying out iterative projection in described step (3.6) is:

SubI _C(i+1)＝SubI _C(i)*(1+β*RMS)，i＝1，2，…，N _s

Iteration performs N _sstep, wherein $\mathrm{RMS}={\left|\right|{\mathrm{SubI}}_L-{d\left({\mathrm{SubI}}_{C\left(i\right)\⊗}h\right)}_m)\left|\right|}^2,$ table convolution operation, d (I) _mrepresenting doubly carries out down-sampled by image I by m, and β is Lagrangian balance parameters.β＝0.2。

The present invention's advantage is compared with prior art:

(1) the inventive method combines the advantage of interpolation and reconstruct, by introducing edge and gradient fields priori, ensure that the high-frequency information retentivities such as the result images atural object edge that oversubscription discerns are good;

(2) process that projects after have employed iteration of the inventive method, avoids the blocky effect in conventional reconstruction method and learning method;

(3) the inventive method executed in parallel efficiency is high, and computation complexity is little, and simultaneously the quality index such as signal to noise ratio (S/N ratio) does not decline to a great extent, and is specially adapted to single frames significantly remote sensing images.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the inventive method;

Fig. 2 is the present invention's significantly remote sensing images block parallel process schematic diagram.

Embodiment

From remote sensing images single frames super-resolution trend, the reconstructing method of edge and details is kept to have huge advantage in rapidity and validity.

Edge refers to that in image, pixel grey scale has the set of the pixel of Spline smoothing, it is present in object and background, target and target, between region and region, and relevant with the uncontinuity of the first order derivative (gradient fields) of brightness of image or brightness of image, thus show as step edge and line edge.Because the marginal information in remote sensing images often correspond to real atural object, as road, river etc.Therefore the priori value that the present invention is based on the single frames super-resolution processing method of reconstruct will adopt gradient fields and the marginal information of pending image.

It should be noted that at this, space floating data that the image described in the present invention is [0,1].

Note original low-resolution image is I _l, increase resolution multiple is m, and super-resolution result images of the present invention is the solution I of following canonical equation (maximum likelihood problem) _h:

Namely at known low-resolution image I _land the gradient fields of high-definition picture or marginal information ▽ I _hin situation, obtain posterior probability upper maximum high-definition picture result.

Above formula can adopt Lagrange's equation to be expressed as further:

$I_H={\arg \min }_{I_{H*}}{\left|\right|I_L-d{(I_H^{*}\⊗h)}_m\left|\right|}^2+\mathrm{\β}{\left|\right|\▿I_H-\▿I_H^{*}\left|\right|}^2---\left(2\right)$

Wherein, that represent is I _hto I _lrear orientation projection's error, represent convolution operation, h is fuzzy filter, for independent variable, || || ²for standard two norm, represent gradient fields conformity error, d (u) _mrepresenting doubly carries out down-sampled by image u by m, and β is Lagrangian balance parameters, β ∈ (0,1).Namely at known I _lwhen, obtain the gradient fields priori of high-definition picture, and find iteration result by rear orientation projection and gradient fields Transfer Error.

By the super-resolution using method used at present, get I _linterpolation result as I _hinitial solution, in view of the performance of bicubic interpolation is higher, the present invention recommend adopt bicubic interpolation, to reduce the complexity of successive iterations process.Need to find edge and gradient fields information to complete guiding by the description of formula (1) and (2).In conjunction with the multithreading characteristic of current disposal system, original image can be carried out piecemeal process.By this sequence of operations, both complete reappear image edge details and there is not block striped between image block, to make again between the result sub-image that obtains also excessively nature.

Describe technical scheme of the present invention in detail below in conjunction with accompanying drawing, idiographic flow as shown in Figure 1.

1) the fuzzy filter h in formula (1) is relevant to concrete imaging circumstances, usually considers band limit and the smoothness properties of image, generally uses normalization dimensional Gaussian convolution mask.The present invention does not limit choosing h, can be arbitrary normalization two dimension pattern plate, but other two dimension pattern plate also easily causes edge transition sharpening.

The embodiment of the present invention is according to super-resolution multiple m, and adopt the normalization Gaussian template h of [2R (m+1)+1] × [2R (m+1)+1], wherein R is floor operation.

2) according to super-resolution multiple m, the size of super-resolution result images is calculated, to I _lcarry out bicubic interpolation and obtain image high-resolution procedural image I _htemp.

If original image I _lin, the figure image width represented with pixel and height are respectively I _wand I _h, then super-resolution result images I _hthe wide and height of pixel be respectively R (L _w* m) and R (L _h* m), wherein R is floor operation.

Because remote sensing images fabric width is very large, often exceed 10000*10000 pixel, can not disposable integral carry out Memory Allocation, therefore press the method for partition parallel processing shown in accompanying drawing 2.

Get subw=subh=500 (wherein each image block is designated as subw and subh respectively in the length of Width and short transverse) and piecemeal is carried out to image, obtain image I by following multiple threads _lhigh resolving power procedural image I after bicubic interpolation _htemp:

If a) R (L _w* m) can be divided exactly by subw and R (L _h* m) can be divided exactly by subh, then by image integral block, and carry out bicubic interpolation, result directly writes back;

If b) R (L _w* m) can not be divided exactly by subw or R (L _h* m) can not be divided exactly by subh, then press aliquot number by image integral block, and carry out bicubic interpolation.Aliquant width segments extends to subw left, and height component upwards extends to subh process, then carries out bicubic interpolation, and result substep writes back.

Following treatment step for be each image block and possible extension blocks, independently can carry out during process, when final thread merges, be written out to net result file.

3) for original image I _lpiecemeal SubI _lwith procedural image I _htmppiecemeal SubI _htmp, with step 1) in template h convolution, with the impact of noise decrease.Calculate convolution results line direction and column direction outline map respectively.

Note SubI _lline direction outline map be designated as SubL _ex, column direction outline map is designated as SubL _ey.Correspondingly, SubI _htmpbe SubHtmp _ex, SubHtmp _ey.The calculating of outline map adopts normalization Sobel operator, and the pixel of operator is long and wide to be taken as [2R (m+1)+1] accordingly.

4) to SubL _ex, SubL _eyand SubHtmp _ex, SubHtmp _eyaccording to pixels ask absolute value, and carry out linear stretch by maximin and obtain linear edge figure | SubL _ex| _lin, | SubL _ey| _linwith | SubHtmp _ex| _lin, | SubHtmp _ey| _lin.

With | SubL _ex| _linbe calculated as example, pixel maximum absolute value value Max _ex=Max (| SubL _ex|), pixel absolute value minimum M in _ex=Min (| SubL _ex|), then:

${|\mathrm{SubL}}_{\mathrm{ex}}{|}_{\mathrm{lin}}=\frac{{|\mathrm{SubL}}_{\mathrm{ex}}|-{\mathrm{Min}}_{\mathrm{ex}}}{{\mathrm{Max}}_{\mathrm{ex}}-{\mathrm{Min}}_{\mathrm{ex}}}$

In like manner obtain linear edge figure | SubL _ey| _linwith | SubHtmp _ex| _lin, | SubHtmp _ey| _lin.

5) linear edge figure is calculated | SubL _ex| _lin, | SubL _ey| _linwith | SubHtmp _ex| _lin, | SubHtmp _ey| _linvariance L _vx, L _vyand Htmp _vx, Htmp _vy, ask for gradient fields transfer parameters L _varwith H _var.

Particularly:

L _vx＝std(SubL _ex| _lin)

L _vy＝std(|SubL _ey| _lin)

Htmp _vx＝std(|SubHtmp _ex| _lin)

Htmp _vy＝std(|SubHtmp _ey| _lin)

Wherein std represents that variance calculates.Get variance L _vx, L _vyaverage L _var=0.5 (L _vx+ L _vy), Htmp _vx, Htmp _vyaverage H _var=0.5 (Htmp _vx+ Htmp _vy) as step 6) and input.

6) first differential is adopted to calculate SubI _htmpline direction gradient fields Htmp _gxwith column direction gradient fields Htmp _gy, complete gradient fields transfer.

${\mathrm{Htmp}}_{\mathrm{Gx}}={\mathrm{SubI}}_{H{\mathrm{tmp}}_x}-{\mathrm{SubI}}_{{\mathrm{Htmp}}_{x-1}},$ I.e. adjacent column difference;

${\mathrm{Htmp}}_{\mathrm{Gy}}={\mathrm{SubI}}_{H{\mathrm{tmp}}_y}-{\mathrm{SubI}}_{{\mathrm{Htmp}}_{y-1}},$ I.e. adjacent lines difference.

Complete gradient fields transfer by the following method, in perfect (2) subitem:

A) line direction, ${\mathrm{Htmp}}_{\mathrm{Gx}}={\mathrm{Htmp}}_{\mathrm{Gx}}*\mathrm{varf}*e^{\frac{-{\mathrm{dist}}_x^2}{L_{\mathrm{var}}}}/e^{\frac{-{\mathrm{dist}}_x^2}{H_{\mathrm{var}}}},\mathrm{varf}=\frac{H_{\mathrm{var}}}{L_{\mathrm{var}}},$ ${\mathrm{dist}}_x=0.5\left(\right|{\mathrm{SubHtmp}}_{{\mathrm{ex}}_x}{|}_{\mathrm{lin}}+\left|{\mathrm{SubHtmp}}_{{\mathrm{ex}}_{x-1}}{|}_{\mathrm{lin}}\right),$ The i.e. mean value of linear edge figure adjacent column pixel;

B) column direction, ${\mathrm{Htmp}}_{\mathrm{Gy}}={\mathrm{Htmp}}_{\mathrm{Gy}}*\mathrm{varf}*e^{\frac{-{\mathrm{dist}}_y^2}{L_{\mathrm{var}}}}/e^{\frac{-{\mathrm{dist}}_y^2}{H_{\mathrm{var}}}},\mathrm{varf}=\frac{H_{\mathrm{var}}}{L_{\mathrm{var}}},$ ${\mathrm{dist}}_y=0.5\left(\right|{\mathrm{SubHtmp}}_{{\mathrm{ey}}_y}{|}_{\mathrm{lin}}+\left|{\mathrm{SubHtmp}}_{{\mathrm{ey}}_{y-1}}{|}_{\mathrm{lin}}\right),$ The i.e. mean value of linear edge figure adjacent lines pixel.

Above-mentioned steps is exactly the transfer realizing edge prior according to the edge between low resolution image and full resolution pricture and gradient fields difference, thus improves edge fog that interpolation brings and high frequency is lost.

7) according to the gradient fields after transfer, utilize sparse matrix solver to resolve standard P oisson image equation, obtain new high-definition picture sub-block SubI _c.

The documents such as Prez are shown in by the structure of Poisson equation, and sparse matrix solver can adopt realization of increasing income, as Eigen.

This process, by while guarantee original image and target image gradient information, can well be merged background information, avoid occurring block line during piecemeal process.

8) for avoiding the blocky effect after significantly remote sensing images piecemeal process further, adopt iterative backward projection to SubI _cwith SubI _lcarry out iterative projection.

Iteration performs N _sstep,

SubI _C(i+1)＝SubI _C(i)*(1+β*RMS)，i＝1，2，…，N _s

Wherein $\mathrm{RMS}={\left|\right|{\mathrm{SubI}}_L-{d\left({\mathrm{SubI}}_{C\left(i\right)\⊗}h\right)}_m)\left|\right|}^2,$ table convolution operation, d (I) _mrepresenting doubly carries out down-sampled by image I by m.β is formula 2) in balance parameters, choose β=0.2 according to visual effect.

Above-mentioned realization is in formula (2) subitem.

RMS and N _ssize control the result of iterative projection, RMS is less, N _sthe blocky effect of larger then result of calculation is more not obvious, but time complexity rises greatly.Quantize remotely-sensed data to main at present 10, example residual error RMS of the present invention recommends threshold value to be taken as 0.0001, and the remotely-sensed data of other quantization digits can be changed by digit difference.According to results of calculation, iterations N _sbe traditionally arranged to be 100.If namely RMS < 0.0001 or i > 100, then stop iteration, otherwise, continue step 8).

The oversubscription result images sub-block finally obtained is designated as SubI _h.

9) all SubI that each sub-block is corresponding are collected _h, result is written out to file, is low-resolution image I _lsuper-resolution result I _h.

Find the analysis of inventive embodiments, the image level after super-resolution is comparatively clear, and every quality index such as signal to noise ratio (S/N ratio) is improved or does not decline.

The content be not described in detail in instructions of the present invention belongs to the known technology of those skilled in the art.

Claims (4)

1. a single frames Remote Sensing Image Super Resolution disposal route for edge prior guiding, is characterized in that comprising the steps:

(1) original image I is obtained _l, and to I _lcarry out piecemeal, to I _lblock image carry out index and be designated as SubI _l;

(2) according to super-resolution multiple m, to each SubI _lcarry out interpolation, obtain block image SubI _htmp, SubI _htmpset form procedural image I _htemp;

(3) for any SubI _land the SubI after corresponding interpolation _htmp, all perform following operation:

(3.1) by super-resolution multiple m, form fuzzy filter h, utilize h to SubI _land SubI _htmpcarry out convolution respectively, and calculate SubI respectively on the basis of convolution _lline direction outline map SubL _ex, column direction outline map SubL _ey, and SubI _htmpline direction outline map SubHtmp _ex, column direction outline map SubHtmp _ey;

(3.2) to SubL _exaccording to pixels ask absolute value, and carry out linear stretch by maximin and obtain linear edge figure | SubL _ex| _lin,

$|{\mathrm{SubL}}_{\mathrm{ex}}{|}_{\mathrm{lin}}=\frac{\left|{\mathrm{SubL}}_{\mathrm{ex}}\right|-{\mathrm{Min}}_{\mathrm{ex}}}{{\mathrm{Max}}_{\mathrm{ex}}-{\mathrm{Min}}_{\mathrm{ex}}}$

Wherein, pixel maximum absolute value value Max _ex=Max (| SubL _ex|), pixel absolute value minimum M in _ex=Min (| SubL _ex|); Then to SubL _ey, SubHtmp _ex, SubHtmp _eyperform identical operation, obtain corresponding linear edge figure | SubL _ey| _lin, | SubHtmp _ex| _linwith | SubHtmp _ey| _lin;

(3.3) linear edge figure is calculated respectively | SubL _ex| _lin, | SubL _ey| _lin, | SubHtmp _ex| _linwith, | SubHtmp _ey| _linvariance L _vx, L _vy, Htmp _vx, Htmp _vy, and ask for gradient fields transfer parameters L _varwith H _var, wherein:

L _vx＝std(|SubL _ex| _Lin)

L _vy＝std(|SubL _ey| _lin)

Htmp _vx＝std(|SubHtmp _ex| _lin)

Htmp _vy＝std(|SubHtmp _ey| _lin)

L _var＝0.5(L _vx+L _vy)

H _var＝0.5(Htmp _vx+Htmp _vy)

Wherein std represents that variance calculates;

(3.4) first differential is adopted to calculate SubI _htmpline direction gradient fields Htmp _gxwith column direction gradient fields Htmp _gy, complete gradient fields transfer, wherein:

${\mathrm{Htmp}}_{\mathrm{Gx}}={\mathrm{SubI}}_{{\mathrm{Htmp}}_x}-{\mathrm{SubI}}_{{\mathrm{Htmp}}_{x-1}}$

${\mathrm{Htmp}}_{\mathrm{Gy}}={\mathrm{SubI}}_{{\mathrm{Htmp}}_y}-{\mathrm{SubI}}_{{\mathrm{Htmp}}_{y-1}}$

After gradient fields transfer

${\mathrm{Htmp}}_{\mathrm{Gx}}={\mathrm{Htmp}}_{\mathrm{Gx}}*\mathrm{varf}*e^{\frac{-{\mathrm{dist}}_x^2}{L_{\mathrm{var}}}}/e^{\frac{-{\mathrm{dist}}_x^2}{H_{\mathrm{var}}}}$

$\mathrm{varf}=\frac{H_{\mathrm{var}}}{L_{\mathrm{var}}},{\mathrm{dist}}_x=0.5\left(\right|{\mathrm{SubHtmp}}_{{\mathrm{ex}}_x}{|}_{\mathrm{lin}}+\left|{\mathrm{SubHtmp}}_{{\mathrm{ex}}_{x-1}}{|}_{\mathrm{lin}}\right)$

${\mathrm{Htmp}}_{\mathrm{Gy}}={\mathrm{Htmp}}_{\mathrm{Gy}}*\mathrm{varf}*e^{\frac{-{\mathrm{dist}}_y^2}{L_{\mathrm{var}}}}/e^{\frac{-{\mathrm{dist}}_y^2}{H_{\mathrm{var}}}}$

${\mathrm{dist}}_y=0.5\left(\right|{\mathrm{SubHtmp}}_{{\mathrm{ey}}_y}{|}_{\mathrm{lin}}+\left|{\mathrm{SubHtmp}}_{{\mathrm{ey}}_{y-1}}{|}_{\mathrm{lin}}\right)$

In formula, footmark x and x-1 represents the adjacent rows of line direction, and footmark y and y-1 represents adjacent two row of column direction;

(3.5) according to the gradient fields after transfer, utilize sparse matrix solver to resolve standard P oisson image equation, obtain new high-definition picture sub-block SubI _c;

(3.6) adopt iterative backward projection to SubI _cwith SubI _lcarry out iterative projection, obtain oversubscription result images sub-block and be designated as SubI _h;

(4) to whole SubI that step (3) obtains _hcarry out image mosaic, obtain I _lsuper-resolution result images I _h.

2. the single frames Remote Sensing Image Super Resolution disposal route of a kind of edge prior guiding according to claim 1, it is characterized in that: in described step (3.1), the calculating of outline map adopts normalization Sobel operator, the pixel of operator is long and wide all gets [2R (m+1)+1], and R is floor operation.

3. the single frames Remote Sensing Image Super Resolution disposal route of a kind of edge prior guiding according to claim 1 and 2, is characterized in that: the method for carrying out iterative projection in described step (3.6) is:

SubI _C(i+1)＝SubI _C(i)*(1+β*RMS)，i＝1，2，…，N _s

Iteration performs N _sstep, wherein $\mathrm{RMS}=\left|\right|{\mathrm{SubI}}_L-d\left({\left({\mathrm{SubI}}_{C\left(i\right)\&CircleTimes;}h\right)}_m\right){\left|\right|}^2,$ table convolution operation, d (I) _mrepresenting doubly carries out down-sampled by image I by m, and β is Lagrangian balance parameters.

4. the single frames Remote Sensing Image Super Resolution disposal route of a kind of edge prior guiding according to claim 3, is characterized in that: described β=0.2.