CN105488776A - Super-resolution image reconstruction method and apparatus - Google Patents

Abstract

The invention provides a super-resolution image reconstruction method and apparatus. The method comprises: performing N-time fuzzy down-sampling according to a first high-resolution image to generate a high-resolution image database, performing fuzzy down-sampling and fuzzy up-sampling on the first high-resolution image by adopting a bicubic interpolation algorithm to obtain a first low-resolution image, and performing N-time fuzzy down-sampling to generate a low-resolution image database; amplifying the first high-resolution image by S times to obtain a second low-resolution image, and dividing the second low-resolution image into at least one low-resolution image tile; obtaining a corresponding low-resolution most similar image tile set in the low-resolution image database by adopting an approximate nearest neighbor search algorithm; and calculating a weight coefficient of the low-resolution most similar image tile set with a ridge regression method to determine high-resolution image tiles. A second high-resolution image is determined by adopting weighted average for the high-resolution image tiles, so that the error of super-resolution image reconstruction can be effectively reduced and the quality of super-resolution image reconstruction can be improved.

Description

Super-resolution image reconstruction method and device

Technical field

The present invention relates to image, field of video processing, particularly relate to a kind of super-resolution image reconstruction method and device.

Background technology

Along with developing rapidly of multimedia technology, people are more and more higher with the requirement of abundant picture detail information for the visual vivid effect of image and video, this needs high-resolution image and video, and in the image processing and analysis system of reality, usually also all need high-resolution image and video.But, the resolution of image is limited to the restriction conditions such as image capture device, optics, image taking speed and hardware store usually, what catch in many imaging applications is all image and the video of low resolution, such as, image and the video of low resolution that what digital camera, medical image system and video monitoring system were caught is usually all.So in order to obtain high-resolution image and video, the low-resolution image needing super-resolution technique to go utilization to obtain and video are to reconstruct high-resolution image and video.

Existing super-resolution image reconstruction method has: based on the method for interpolation, the method based on the method for rebuilding and instance-based learning, wherein need extra database based on the method for rebuilding, for the low-resolution image obtained, the priori of the multiple image in image data base is utilized to carry out modeling to current low-resolution image, then, the low-resolution image after modeling is redeveloped into high-resolution image.

But; in the prior art; owing to being that a huge representational low resolution and high resolution graphics photo are to data storehouse to what utilize when super-resolution image modeling; this database is as much as possible comprises all picture structures; make rebuilding super resolution image process too complicated, and usually can there is the situation cannot restoring high frequency detail based on this image data base.

Summary of the invention

The invention provides a kind of super-resolution image reconstruction method and device, thus effectively can reduce the error of rebuilding super resolution image, and then improve the quality of rebuilding image.

First aspect, the invention provides a kind of super-resolution image reconstruction method, comprising: step S1, using the low-resolution image y of input as the first high-definition picture according to described first high-definition picture carry out N fuzzy down-sampling, generate high resolution image data storehouse D _x; Step S2, to described first high-definition picture adopt after bicubic interpolation algorithm carries out fuzzy down-sampling, then carry out fuzzy up-sampling and obtain the first low-resolution image for described first low-resolution image carry out N fuzzy down-sampling, generate low resolution image data storehouse D _z; Step S3, adopts described bicubic interpolation algorithm by described first high-definition picture the second low-resolution image is obtained after amplifying S times by described second low-resolution image be divided at least one low-resolution image sheet; Step S4, adopts approximate KNN searching algorithm at described low resolution image data storehouse D to each low-resolution image sheet _zobtain the most similar diagram photo group of corresponding low resolution; Step S5, adopts Ridge Regression Method to calculate weight coefficient to the most similar diagram photo group of described low resolution; Step S6, according to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power; Step S7, adopts weighted mean to determine the second high-definition picture to described high resolution graphics photo and by described second low-resolution image join described database D _zin, by described second high-definition picture join described database D _xin.

In conjunction with first aspect, in the first possibility embodiment of first aspect, also comprise: step S8, by described second high-definition picture as the first new high-definition picture after M repeated execution of steps S3-S7, obtain third high image in different resolution adopt described bicubic interpolation algorithm to described third high image in different resolution carry out fuzzy down-sampling, generate initial target high-definition picture

In conjunction with the first possibility embodiment of first aspect, in the second possibility embodiment of first aspect, also comprise: step S9, to described initial target high-definition picture alternating minimization algorithm computational mathematics model is adopted with described low-resolution image y:

${\mathrm{\α}}_y=\arg \underset{\mathrm{\φ},\mathrm{\α},f}{\min }\{{\left|\right|y-\mathrm{H\φo\α}\left|\right|}_2^2+{\mathrm{\λ}}_1{\mathrm{\Σ}}_i{\left|\right|{\mathrm{\α}}_i-{\mathrm{\β}}_i\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|f(\▿x|\▿y)-\▿x\left|\right|}^2\},$

s.t.h _f＝h _r,

Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition picture singular matrix H represents fuzzy down-sampling operator, λ ₁and λ ₂represent regularization parameter, α _ithe high resolution graphics photo x of target high-resolution image x _isparse coefficient, α is α _iset, β _iα _iat the non-local mean in sparse coding territory, o represents the sparse coefficient matrix α of all splicings and the product of super complete dictionary φ, represent gradient operation symbol, f is transforming function transformation function, h _rthe reference histograms of target high-resolution image, h _fit is the histogram of transforming function transformation function f.

In conjunction with the second possibility embodiment of first aspect, in the third possibility embodiment of first aspect, described to described initial target high-definition picture adopt alternating minimization algorithm computational mathematics model specifically to comprise with described low-resolution image y: step S10, adopt K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary; Step S11, upgrades described transforming function transformation function f according to the histogram of gradients algorithm that remains unchanged; Step S12, upgrades described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f.

In conjunction with the third possibility embodiment of first aspect, in the 4th kind of possibility embodiment of first aspect, adopt K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary, specifically comprise: according to the high resolution graphics photo of size random selecting K described initial target high-definition picture as K initial clustering; According to the distance between described high resolution graphics photo and a described K initial cluster center, all described high resolution graphics photos be divided into by described initial target high-definition picture are divided in corresponding initial clustering; The corresponding sub-dictionary of principal component analysis (PCA) training based on svd is adopted to each described initial clustering; K the described super complete dictionary of sub-dictionary composition that K described initial clustering is corresponding.

In conjunction with the third possibility embodiment of first aspect, in the 5th kind of possibility embodiment of first aspect, upgrade described transforming function transformation function f according to the histogram of gradients algorithm that remains unchanged, specifically comprise: fuzzy up-sampling is carried out to described low-resolution image y, obtain low-resolution image z; Described target high-resolution image x and described low-resolution image z meets following condition: z=B*x, and wherein B is fuzzy core; Gradient is asked to z=B*x both sides, then wherein b ₀and b _irepresent center coefficient and its surrounding neighbour coefficient of fuzzy core B respectively, represent the gradient image of target high-resolution image x, represent the gradient image of described low-resolution image z, represent the gradient image of described target high-resolution image sheet; Order if $x2=\underset{i}{\mathrm{\Σ}}{b}_i{\▿x}_i$ Be similar to normal distribution, then $h_r=\arg \underset{h_{x1}}{\min }\left\{{\left|\right|h_z-h_{x1}\⊗h_{x2}\left|\right|}^2\right\}.$

By solving-optimizing problem s.t.h _f=h _rupgrade transforming function transformation function f.

Wherein h _rrepresent the terraced histogram of reference of target high-resolution image x, h _zrepresent the histogram of gradients of described low-resolution image z, h _x1the discrete form of the probability density function of stochastic variable x1, h _x2the discrete form of the probability density function of I.i.d. random variables x2, it is convolution operation symbol.

May embodiment in conjunction with the third of first aspect, may in embodiment at the 6th kind of first aspect, describedly upgrade described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f, specifically comprise: according to formula $x^{(t+1/2)}=x^{\left(t\right)}+\mathrm{\δ}((H^{T}y-H^T{\mathrm{hx}}^{\left(t\right)})+{\mathrm{\λ}}_2{\▿}^T(f-{\▿x}^{\left(t\right)}))$ Determine x ^(t+1/2), wherein x ^(t)represent the target high-resolution Image estimation value of the t time iteration, x ^(t+1/2)represent the target high-resolution Image estimation value of the t+1/2 time iteration, δ is constant, represent x ^(t)gradient map; According to determine wherein represent the high resolution graphics photo x of the t+1/2 time iteration _isparse coefficient, represent the high resolution graphics photo x of the t time iteration _ithe sub-dictionary that place initial clustering is corresponding, R _irepresent from initial target high-definition picture high resolution graphics photo x is obtained at i place, position _imatrix.

According to sparse coefficient ${\mathrm{\α}}_i^{(t+1)}=S_{\mathrm{\λ}/c}({\mathrm{\φ}}^T{\mathrm{oH}}^T(y-\mathrm{H\φo}{\mathrm{\α}}_i^{(t+1/2)})/c+{\mathrm{\α}}_i^{(t+1/2)}-{\mathrm{\β}}_i^{(t+1/2)})+{\mathrm{\β}}_i^{(t+1/2)}$ With determine the high resolution graphics photo x of the t+1 time iteration _isparse coefficient ω _ia jth component ω _ijmeet: wherein h represents the controling parameters for regulating the rate of decay, and W represents normalized factor, represent high resolution graphics photo x _ithe most similar diagram photo of jth in corresponding most similar diagram photo group, S _λ/crepresent soft-threshold function, c represents regularization parameter; According to formula $x^{(t+1)}={\mathrm{\φ}}^{(t+1)}{\mathrm{o\α}}^{(t+1)}={\left(\underset{i}{\mathrm{\Σ}}{R}_i^T{R}_i\right)}^{-1}\underset{i}{\mathrm{\Σ}}{R}_i^T{\mathrm{\φ}}_i^{(t+1)}{\mathrm{\α}}_i^{(t+1)}$ Determine x ^(t+1), wherein x ^(t+1)represent the target high-resolution Image estimation value of the t+1 time iteration, φ ^(t+1)represent x ^(t+1)corresponding super complete dictionary.

In conjunction with first aspect the third may embodiment or the 4th kind may embodiment or the 5th kind may embodiment or the 6th kind may embodiment, in the 7th kind of possibility embodiment of first aspect, also comprise: repeated execution of steps S10-S12, obtaining convergence solution is final goal high-definition picture x '; Adopt plural impact filtering formula to restore described final goal high-definition picture, wherein said plural impact filtering formula is:

$x^{\′\′}=-\frac{2}{\mathrm{\π}}\arctan \left(\mathrm{a\; Im}\left(\frac{x^{\′}}{\mathrm{\θ}}\right)\right)|\▿x^{\′}|+{\mathrm{\τ}}_{\mathrm{\η\η}}^{\′}+\stackrel{\‾}{\mathrm{\τ}}{x}_{\mathrm{\ξ\ξ}}^{\′}$

Wherein, x ' expression final goal high-definition picture, x " represent the final goal high-definition picture after restoring, represent the gradient map of x', η and ξ represents the gradient direction of image, and Im () represents extraction imaginary part, and a represents the adjustment parameter for controlling image sharpness, τ=| τ | exp (i θ) is complex scalar coefficient, it is real number scalar factor.

Second aspect, the invention provides a kind of super-resolution image reconstruction device, comprising: generation module, for will input low-resolution image y as the first high-definition picture according to described first high-definition picture carry out N fuzzy down-sampling, generate high resolution image data storehouse D _x; Described generation module, also for described first high-definition picture adopt after bicubic interpolation algorithm carries out fuzzy down-sampling, then carry out fuzzy up-sampling and obtain the first low-resolution image for described first low-resolution image carry out N fuzzy down-sampling, generate low resolution image data storehouse D _z; Divide module, for adopting described bicubic interpolation algorithm by described first high-definition picture the second low-resolution image is obtained after amplifying S times by described second low-resolution image be divided at least one low-resolution image sheet; Acquisition module, for adopting approximate KNN searching algorithm at described low resolution image data storehouse D to each low-resolution image sheet _zobtain the most similar diagram photo group of corresponding low resolution; Computing module, for adopting Ridge Regression Method to calculate weight coefficient to the most similar diagram photo group of described low resolution; Determination module, for according to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power; Described determination module, also for adopting weighted mean to determine the second high-definition picture to described high resolution graphics photo and by described second low-resolution image join described database D _zin, by described second high-definition picture join described database D _xin.

In conjunction with second aspect, in the first possibility embodiment of second aspect, also comprise: described acquisition module, also for obtaining third high image in different resolution described generation module, also adopts described bicubic interpolation algorithm to described third high image in different resolution carry out fuzzy down-sampling, generate initial target high-definition picture

May embodiment in conjunction with the first of second aspect, in embodiment, also may comprise: described computation model at the second of second aspect, also for described initial target high-definition picture alternating minimization algorithm computational mathematics model is adopted with described low-resolution image y:

${\mathrm{\α}}_y=\arg \underset{\mathrm{\φ},\mathrm{\α},f}{\min }\{{\left|\right|y-\mathrm{H\φo\α}\left|\right|}_2^2+{\mathrm{\λ}}_1{\mathrm{\Σ}}_i{\left|\right|{\mathrm{\α}}_i-{\mathrm{\β}}_i\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|f(\▿x|\▿y)-\▿x\left|\right|}^2\},$

s.t.h _f＝h _r,

Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition picture singular matrix H represents fuzzy down-sampling operator, λ ₁and λ ₂represent regularization parameter, α _ithe high resolution graphics photo x of target high-resolution image x _isparse coefficient, α is α _iset, β _iα _iat the non-local mean in sparse coding territory, o represents the sparse coefficient matrix α of all splicings and the product of super complete dictionary φ, represent gradient operation symbol, f is transforming function transformation function, h _rthe reference histograms of target high-resolution image, h _fit is the histogram of transforming function transformation function f.

In conjunction with the second possibility embodiment of second aspect, in the third possibility embodiment of second aspect, also comprise: update module; Described acquisition module, also for adopting K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary; Described update module, upgrades described transforming function transformation function f for the algorithm that remains unchanged according to histogram of gradients; Described update module, also upgrades described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f.

In conjunction with the third possibility embodiment of second aspect, may in embodiment at the 4th kind of second aspect, described acquisition module specifically for: according to the described high resolution graphics photo of size random selecting K described target high-resolution image as K initial clustering; According to the distance between described high resolution graphics photo and a described K initial cluster center, all described high resolution graphics photos be divided into by described target high-resolution image are divided in corresponding initial clustering; The corresponding sub-dictionary of principal component analysis (PCA) training based on svd is adopted to each described initial clustering; K the described super complete dictionary of sub-dictionary composition that K described initial clustering is corresponding.

May embodiment in conjunction with the third of second aspect, may in embodiment at the 5th kind of second aspect, described update module specifically for: fuzzy up-sampling is carried out to described low-resolution image y, obtains low-resolution image z; Described target high-resolution image x and described low-resolution image z meets following condition: z=B*x, and wherein B is fuzzy core; Gradient is asked to z=B*x both sides, then wherein b ₀and b _irepresent center coefficient and its surrounding neighbour coefficient of fuzzy core B respectively, represent the gradient image of target high-resolution image x, represent the gradient image of described low-resolution image z, represent the gradient image of described high resolution graphics photo; Order $x1=b_0\▿x,x2=\underset{i}{\mathrm{\Σ}}{b}_i{\▿x}_i,$ If $x2=\underset{i}{\mathrm{\Σ}}{b}_i{\▿x}_i$ Be similar to normal distribution, then $h_r=\arg \underset{h_{x1}}{\min }\left\{{\left|\right|h_z-h_{x1}\⊗h_{x2}\left|\right|}^2\right\}.$

By solving-optimizing problem s.t.h _f=h _rupgrade transforming function transformation function f.

Wherein h _rrepresent the terraced histogram of reference of target high-resolution image x, h _zrepresent the histogram of gradients of described low-resolution image z, h _x1the discrete form of the probability density function of stochastic variable x1, h _x2the discrete form of the probability density function of I.i.d. random variables x2, it is convolution operation symbol.

May embodiment in conjunction with the third of second aspect, may in embodiment at the 6th kind of second aspect, described update module specifically for: according to the iterative formula of super-resolution optimization problem $x^{(t+1/2)}=x^{\left(t\right)}+\mathrm{\δ}((H^{T}y-H^T{\mathrm{hx}}^{\left(t\right)})+{\mathrm{\λ}}_2{\▿}^T(f-{\▿x}^{\left(t\right)}))$ Determine x ^(t+1/2), wherein x ^(t)represent the target high-resolution Image estimation value of the t time iteration, x ^(t+1/2)represent the target high-resolution Image estimation value of the t+1/2 time iteration, δ is constant, represent x ^(t)gradient map; According to determine wherein represent the high resolution graphics photo x of the t+1/2 time iteration _isparse coefficient, represent the high resolution graphics photo x of the t time iteration _ithe sub-dictionary that place initial clustering is corresponding, R _irepresent from initial target high-definition picture high resolution graphics photo x is obtained at i place, position _imatrix; According to sparse coefficient ${\mathrm{\α}}_i^{(t+1)}=S_{\mathrm{\λ}/c}({\mathrm{\φ}}^T{\mathrm{oH}}^T(y-\mathrm{H\φo}{\mathrm{\α}}_i^{(t+1/2)})/c+{\mathrm{\α}}_i^{(t+1/2)}-{\mathrm{\β}}_i^{(t+1/2)})+{\mathrm{\β}}_i^{(t+1/2)}$ With determine the high resolution graphics photo x of the t+1 time iteration _isparse coefficient ω _ia jth component ω _ijmeet: wherein h represents the controling parameters for regulating the rate of decay, and W represents normalized factor, represent high resolution graphics photo x _ithe most similar diagram photo of jth in corresponding most similar diagram photo group, S _λ/crepresent soft-threshold function, c represents regularization parameter; According to formula $x^{(t+1)}={\mathrm{\φ}}^{(t+1)}{\mathrm{o\α}}^{(t+1)}={\left(\underset{i}{\mathrm{\Σ}}{R}_i^T{R}_i\right)}^{-1}\underset{i}{\mathrm{\Σ}}{R}_i^T{\mathrm{\φ}}_i^{(t+1)}{\mathrm{\α}}_i^{(t+1)}$ Determine x ^(t+1), wherein x ^(t+1)represent the target high-resolution Image estimation value of the t+1 time iteration, φ ^(t+1)represent x ^(t+1)corresponding super complete dictionary.

In conjunction with second aspect the third may embodiment or the 4th kind may embodiment or the 5th kind may embodiment or the 6th kind may embodiment, in embodiment, also may comprise: filtration module at the 7th kind of second aspect; It is final goal high-definition picture x ' that described acquisition module obtains convergence solution; Described filtration module, for adopting plural impact filtering formula to restore described final goal high-definition picture, wherein said plural impact filtering formula is:

$x^{\′\′}=-\frac{2}{\mathrm{\π}}\arctan \left(\mathrm{a\; Im}\left(\frac{x^{\′}}{\mathrm{\θ}}\right)\right)|\▿x^{\′}|+{\mathrm{\τ}}_{\mathrm{\η\η}}^{\′}+\stackrel{\‾}{\mathrm{\τ}}{x}_{\mathrm{\ξ\ξ}}^{\′}$

Wherein, x ' expression final goal high-definition picture, x " represent the final goal high-definition picture after restoring, represent the gradient map of x', η and ξ represents the gradient direction of image, and Im () represents extraction imaginary part, and a represents the adjustment parameter for controlling image sharpness, τ=| τ | exp (i θ) is complex scalar coefficient, it is real number scalar factor.

The invention provides a kind of super-resolution image reconstruction method, comprise: using the low-resolution image of input as the first high-definition picture, N fuzzy down-sampling is carried out according to described first high-definition picture, generate high resolution image data storehouse, after fuzzy down-sampling is carried out to described first high-definition picture employing bicubic interpolation algorithm, carry out fuzzy up-sampling again and obtain the first low-resolution image, N fuzzy down-sampling is carried out for described first low-resolution image, generates low resolution image data storehouse; Obtain the second low-resolution image after adopting described bicubic interpolation algorithm that described first high-definition picture is amplified S times, described second low-resolution image is divided at least one low-resolution image sheet; Approximate KNN searching algorithm is adopted to obtain the most similar diagram photo group of corresponding low resolution in described low resolution image data storehouse to each low-resolution image sheet; Ridge Regression Method is adopted to calculate weight coefficient to the most similar diagram photo group of described low resolution; According to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power.Weighted mean is adopted to determine the second high-definition picture to described high resolution graphics photo, initial target high-definition picture is estimated by the multistage amplifying technique based on Ridge Regression Method, and fully utilize data fidelity, sparse non local regularization and histogram of gradients regularization priori carry out mathematical modeling, final target high-resolution image is obtained by duty Optimization, the error of rebuilding super resolution image effectively can be reduced finally by the further sharpen edges of impact filtering and the process of Recovery image CONSTRUCTED SPECIFICATION, and then improve the quality of rebuilding image.

Accompanying drawing explanation

The process flow diagram of a kind of super-resolution image reconstruction method that Fig. 1 provides for one embodiment of the invention;

The process flow diagram of the super-resolution image reconstruction method that Fig. 2 provides for another embodiment of the present invention;

A kind of super-resolution image reconstruction apparatus structure schematic diagram that Fig. 3 provides for one embodiment of the invention;

A kind of super-resolution image reconstruction apparatus structure schematic diagram that Fig. 4 provides for another embodiment of the present invention.

Embodiment

For making the object of the embodiment of the present invention, technical scheme and advantage clearly, below in conjunction with the accompanying drawing in the embodiment of the present invention, technical scheme in the embodiment of the present invention is clearly and completely described, obviously, described embodiment is the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.

The process flow diagram of a kind of super-resolution image reconstruction method that Fig. 1 provides for one embodiment of the invention, wherein the method is applicable to the scene obtaining high-definition picture or video, wherein the executive agent of the method is: super-resolution image reconstruction device, this device can be the intelligent terminals such as computing machine, and super-resolution image reconstruction method specifically comprises following flow process:

Step S1, using the low-resolution image of input as the first high-definition picture, carries out N fuzzy down-sampling according to the first high-definition picture, generates high resolution image data storehouse.

Wherein, low-resolution image y represents, the first high-definition picture is used represent, high resolution image data storehouse D _xrepresent.

Step S2, after fuzzy down-sampling is carried out to the first high-definition picture employing bicubic interpolation algorithm, carry out fuzzy up-sampling again and obtain the first low-resolution image, N fuzzy down-sampling is carried out for the first low-resolution image, generate low resolution image data storehouse.

Wherein, the first low-resolution image is used represent, low resolution image data storehouse D _zrepresent, the first low-resolution image in same scale rank with the first high-definition picture between corresponding relation set up as follows: wherein, * is convolution operation symbol, the up-sampling operational character of ↑ s to be scale factor be s, and the down-sampling operational character of ↓ s to be scale factor be s, B is fuzzy core, E _sto be the scale factor of first fuzzy down-sampling up-sampling be again s meets operational character, then obtain the first low-resolution image for the first low-resolution image carry out N fuzzy down-sampling, generate low resolution image data storehouse D _z.

Step S3, obtains the second low-resolution image after adopting bicubic interpolation algorithm that the first high-definition picture is amplified S times, the second low-resolution image is divided at least one low-resolution image sheet.

Wherein, the second low-resolution image is particularly, and by the second low-resolution image be divided at least one low-resolution image sheet.

Step S4, adopts approximate KNN searching algorithm to obtain the most similar diagram photo group of corresponding low resolution in low resolution image data storehouse to each low-resolution image sheet.

Step S5, adopts Ridge Regression Method to calculate weight coefficient to the most similar diagram photo group of described low resolution.

Particularly, approximate KNN searching algorithm can be utilized from low resolution image data storehouse D for any one low-resolution image sheet _zthe most similar diagram photo group of low resolution corresponding to middle search, such as: suppose low-resolution image sheet z _ithe most similar diagram photo group of corresponding low resolution is Lp _i, then adopt Ridge Regression Modeling Method to z _iand Lp _ibetween linear approximate relationship to carry out the constrained optimization problem of modeling formation as follows: wherein, γ is weight coefficient, and τ is for alleviating singularity problem and the stable regularization parameter solved.By solving this regular least square regression problem, the present invention provides the closed solutions of this problem: wherein, I is unit matrix.

Step S6, according to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power.

By z in step S5 _iwith Lp _imapping relations from low resolution image data storehouse D _zbe applied to high resolution image data storehouse D _x, then high resolution graphics photo p _isimilar diagram photo group Hp most with high resolving power _irelational expression be p _i=Hp _iγ.

Step S7, adopts method of weighted mean to determine the second high-definition picture to high resolution graphics photo, and joins in described database by described second low-resolution image, joined in described database by described second high-definition picture.

Wherein, the second high-definition picture is

The invention provides a kind of super-resolution image reconstruction method, comprise: using the low-resolution image of input as the first high-definition picture, N fuzzy down-sampling is carried out according to described first high-definition picture, generate high resolution image data storehouse, after fuzzy down-sampling is carried out to described first high-definition picture employing bicubic interpolation algorithm, carry out fuzzy up-sampling again and obtain the first low-resolution image, N fuzzy down-sampling is carried out for described first low-resolution image, generates low resolution image data storehouse; Obtain the second low-resolution image after adopting described bicubic interpolation algorithm that described first high-definition picture is amplified S times, described second low-resolution image is divided at least one low-resolution image sheet; Approximate KNN searching algorithm is adopted to obtain the most similar diagram photo group of corresponding low resolution in described low resolution image data storehouse to each low-resolution image sheet; Ridge Regression Method is adopted to calculate weight coefficient to the most similar diagram photo group of described low resolution; According to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power.Method of weighted mean is adopted to determine the second high-definition picture to described high resolution graphics photo, initial target high-definition picture is estimated by the multistage amplifying technique based on Ridge Regression Method, thus effectively can reduce the error of rebuilding super resolution image, and then improve the quality of rebuilding image.

The process flow diagram of the super-resolution image reconstruction method that Fig. 2 provides for another embodiment of the present invention, this embodiment is based upon on an embodiment basis, performs, wherein specifically comprise after a upper embodiment step S7:

Step S8, using the second high-definition picture as the first new high-definition picture, after M repeated execution of steps S3-S7, obtain third high image in different resolution, adopt bicubic interpolation algorithm to carry out fuzzy down-sampling to third high image in different resolution, generate initial target high-definition picture.

Wherein, third high image in different resolution is initial target high-definition picture it is worth mentioning that, present invention employs in repeated execution of steps S3-S7 is for M time based on when scale factor is less, and low-resolution image this factor more similar to high-definition picture is considered.Therefore, in view of natural image exists a large amount of self-similarity redundancy in scale domain and spatial domain, different from traditional bicubic interpolation method, method of the present invention adopts carrys out initialized target high-definition picture based on the multistage amplifying technique of Ridge Regression Method.Namely be adopt successive ignition to amplify scheme, after single amplification method is replaced by multistage amplifying technique, scale factor in every grade of amplification procedure of the present invention is just very little, thus can search more similar diagram photo for high resolution image reconstruction from low-resolution image storehouse.For low-resolution image y and the given general size factor d of a width input, the cascade in multistage amplification procedure of the present invention is confirmed as wherein s is the scale factor in every grade of amplification.

Step S9, adopts alternating minimization algorithm computational mathematics model to initial target high-definition picture and low-resolution image:

${\mathrm{\α}}_y=\arg \underset{\mathrm{\φ},\mathrm{\α},f}{\min }\{{\left|\right|y-\mathrm{H\φo\α}\left|\right|}_2^2+{\mathrm{\λ}}_1{\mathrm{\Σ}}_i{\left|\right|{\mathrm{\α}}_i-{\mathrm{\β}}_i\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|f(\▿x|\▿y)-\▿x\left|\right|}^2\},$

s.t.h _f＝h _r,

Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition picture singular matrix H represents fuzzy down-sampling operator, λ ₁and λ ₂represent regularization parameter, α _ithe high resolution graphics photo x of target high-resolution image x _isparse coefficient, α is α _iset, β _iα _iat the non-local mean in sparse coding territory, o represents the sparse coefficient matrix α of all splicings and the product of super complete dictionary φ, represent gradient operation symbol, f is transforming function transformation function, h _rthe reference histograms of target high-resolution image, h _fit is the histogram of transforming function transformation function f.

Particularly, the initial target high-definition picture determined in step S1-S8 may not be optimum solution for this mathematical model, therefore need continuous iteration, until obtain convergence solution.Wherein, describedly alternating minimization algorithm computational mathematics model is adopted specifically to comprise to described target high-resolution image x and described low-resolution image y:

Step S10, adopts K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary.

Particularly, adopt K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary, specifically comprise: according to the described high resolution graphics photo of size random selecting K described target high-resolution image as K initial clustering; According to the distance between described high resolution graphics photo and a described K initial cluster center, all described high resolution graphics photos be divided into by described target high-resolution image are divided in corresponding initial clustering; The corresponding sub-dictionary of principal component analysis (PCA) training based on svd is adopted to each described initial clustering; K the described super complete dictionary of sub-dictionary composition that K described initial clustering is corresponding.

Step S11, according to histogram of gradients remain unchanged algorithm upgrade transforming function transformation function.

Wherein, transforming function transformation function is f, alternatively, upgrades described transforming function transformation function f, specifically comprise: carry out fuzzy up-sampling to described low-resolution image y according to the histogram of gradients algorithm that remains unchanged, obtain low-resolution image z; Described target high-resolution image x and described low-resolution image z meets following condition: z=B*x, and wherein B is fuzzy core; Gradient is asked to z=B*x both sides, then wherein b ₀and b _irepresent center coefficient and its surrounding neighbour coefficient of fuzzy core B respectively, represent the gradient image of target high-resolution image x, represent the gradient image of described low-resolution image z, represent the gradient image of described high resolution graphics photo; Order $x1=b_0\▿x,x2=\underset{i}{\mathrm{\Σ}}{b}_i{\▿x}_i,$ If $x2=\underset{i}{\mathrm{\Σ}}{b}_i{\▿x}_i$ Be similar to normal distribution, then $h_r=\arg \underset{h_{x1}}{\min }\left\{{\left|\right|h_z-h_{x1}\⊗h_{x2}\left|\right|}^2\right\}.$

By solving-optimizing problem s.t.h _f=h _rupgrade transforming function transformation function f; Wherein h _rrepresent the terraced histogram of reference of target high-resolution image x, h _zrepresent the histogram of gradients of described low-resolution image z, h _x1the discrete form of the probability density function of stochastic variable x1, h _x2the discrete form of the probability density function of I.i.d. random variables x2, it is convolution operation symbol.

Step S12, upgrades target high-resolution image according to super complete dictionary and transforming function transformation function.

Particularly, describedly upgrade described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f, specifically comprise: according to the iterative formula of super-resolution optimization problem $x^{(t+1/2)}=x^{\left(t\right)}+\mathrm{\δ}((H^{T}y-H^T{\mathrm{hx}}^{\left(t\right)})+{\mathrm{\λ}}_2{\▿}^T(f-{\▿x}^{\left(t\right)}))$ Determine x ^(t+1/2), wherein x ^(t)represent the target high-resolution Image estimation value of the t time iteration, x ^(t+1/2)represent the target high-resolution Image estimation value of the t+1/2 time iteration, δ is constant, represent x ^(t)gradient map; According to determine wherein represent the high resolution graphics photo x of the t+1/2 time iteration _isparse coefficient, represent the high resolution graphics photo x of the t time iteration _ithe sub-dictionary that place initial clustering is corresponding, R _irepresent from initial target high-definition picture high resolution graphics photo x is obtained at i place, position _imatrix.

According to sparse coefficient ${\mathrm{\α}}_i^{(t+1)}=S_{\mathrm{\λ}/c}({\mathrm{\φ}}^T{\mathrm{oH}}^T(y-\mathrm{H\φo}{\mathrm{\α}}_i^{(t+1/2)})/c+{\mathrm{\α}}_i^{(t+1/2)}-{\mathrm{\β}}_i^{(t+1/2)})+{\mathrm{\β}}_i^{(t+1/2)}$ With determine the high resolution graphics photo x of the t+1 time iteration _isparse coefficient ω _ia jth component ω _ijmeet: wherein h represents the controling parameters for regulating the rate of decay, and W represents normalized factor, represent high resolution graphics photo x _ithe most similar image of jth in corresponding most similar diagram photo group, S _λ/crepresent soft-threshold function, c represents regularization parameter; According to formula $x^{(t+1)}={\mathrm{\φ}}^{(t+1)}{\mathrm{o\α}}^{(t+1)}={\left(\underset{i}{\mathrm{\Σ}}{R}_i^T{R}_i\right)}^{-1}\underset{i}{\mathrm{\Σ}}{R}_i^T{\mathrm{\φ}}_i^{(t+1)}{\mathrm{\α}}_i^{(t+1)}$ Determine x ^(t+1), wherein x ^(t+1)represent the target high-resolution Image estimation value of the t+1 time iteration, φ ^(t+1)represent x ^(t+1)corresponding super complete dictionary.Circulation performs S10-S12, until the target high-resolution image obtained is convergence solution.

Further, repeated execution of steps S10-S12, obtaining convergence solution is final goal high-definition picture x '; Adopt plural impact filtering formula to restore described final goal high-definition picture, wherein said plural impact filtering formula is:

$x^{\′\′}=-\frac{2}{\mathrm{\π}}\arctan \left(\mathrm{a\; Im}\left(\frac{x^{\′}}{\mathrm{\θ}}\right)\right)|\▿x^{\′}|+{\mathrm{\τ}}_{\mathrm{\η\η}}^{\′}+\stackrel{\‾}{\mathrm{\τ}}{x}_{\mathrm{\ξ\ξ}}^{\′},$

Wherein, x ' expression final goal high-definition picture, x " represent the final goal high-definition picture after restoring, represent the gradient map of x', η and ξ represents the gradient direction of image, and Im () represents extraction imaginary part, and a represents the adjustment parameter for controlling image sharpness, τ=| τ | exp (i θ) is complex scalar coefficient, it is real number scalar factor.

The super-resolution image reconstruction method comprehensive utilization data fidelity that the present embodiment provides, sparse non local regularization and histogram of gradients regularization priori carry out mathematical modeling, final target high-resolution image is obtained, finally by the further sharpen edges of impact filtering and the process of Recovery image CONSTRUCTED SPECIFICATION by duty Optimization.

A kind of super-resolution image reconstruction apparatus structure schematic diagram that Fig. 3 provides for one embodiment of the invention, wherein this device comprises: generation module 301, for will input low-resolution image y as the first high-definition picture according to described first high-definition picture carry out N fuzzy down-sampling, generate high resolution image data storehouse D _x; Described generation module 301, also for described first high-definition picture adopt after bicubic interpolation algorithm carries out fuzzy down-sampling, then carry out fuzzy up-sampling and obtain the first low-resolution image for described first low-resolution image carry out N fuzzy down-sampling, generate low resolution image data storehouse D _z; Divide module 302, for adopting described bicubic interpolation algorithm by described first high-definition picture the second low-resolution image is obtained after amplifying S times by described second low-resolution image be divided at least one low-resolution image sheet; Acquisition module 303, for adopting approximate KNN searching algorithm at described low resolution image data storehouse D to each low-resolution image sheet _zobtain the most similar diagram photo group of corresponding low resolution; Computing module 304, for adopting Ridge Regression Method to calculate weight coefficient to the most similar diagram photo group of described low resolution; Determination module 305, for according to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power; Described determination module 305, also for adopting method of weighted mean to determine the second high-definition picture to described high resolution graphics photo and by described second low-resolution image join described database D _zin, by described second high-definition picture join described database D _xin.

The super-resolution image reconstruction device that the present embodiment provides, may be used for the enforcement technical scheme of the super-resolution image reconstruction method performed corresponding to Fig. 1, it realizes principle and technique effect is similar, repeats no more herein.

A kind of super-resolution image reconstruction apparatus structure schematic diagram that Fig. 4 provides for another embodiment of the present invention, on the corresponding embodiment basis of Fig. 3, described acquisition module 303, also for obtaining third high image in different resolution described generation module 301, also adopts described bicubic interpolation algorithm to described third high image in different resolution carry out fuzzy down-sampling, generate initial target high-definition picture described computation model 304, also for described initial target high-definition picture alternating minimization algorithm computational mathematics model is adopted with described low-resolution image y:

${\mathrm{\α}}_y=\arg \underset{\mathrm{\φ},\mathrm{\α},f}{\min }\{{\left|\right|y-\mathrm{H\φo\α}\left|\right|}_2^2+{\mathrm{\λ}}_1{\mathrm{\Σ}}_i{\left|\right|{\mathrm{\α}}_i-{\mathrm{\β}}_i\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|f(\▿x|\▿y)-\▿x\left|\right|}^2\},$

s.t.h _f＝h _r,

Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition picture singular matrix H represents fuzzy down-sampling operator, λ ₁and λ ₂represent regularization parameter, α _ithe high resolution graphics photo x of target high-resolution image x _isparse coefficient, α is α _iset, β _iα _iat the non-local mean in sparse coding territory, o represents the sparse coefficient matrix α of all splicings and the product of super complete dictionary φ, represent gradient operation symbol, f is transforming function transformation function, h _rthe reference histograms of target high-resolution image, h _fit is the histogram of transforming function transformation function f.

Further, this device also comprises: update module 306; Described acquisition module 303, also for adopting K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary; Described update module 306, upgrades described transforming function transformation function f for the algorithm that remains unchanged according to histogram of gradients; Described update module 306, also upgrades described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f.

Alternatively, described acquisition module 303 specifically for: according to the described high resolution graphics photo of size random selecting K described target high-resolution image as K initial clustering; According to described to the distance between described high resolution graphics photo and a described K initial cluster center, all described high resolution graphics photos be divided into by described target high-resolution image are divided in corresponding initial clustering; The corresponding sub-dictionary of principal component analysis (PCA) training based on svd is adopted to each described initial clustering; K the described super complete dictionary of sub-dictionary composition that K described initial clustering is corresponding.

Alternatively, described update module 306 specifically for: fuzzy up-sampling is carried out to described low-resolution image y, obtains low-resolution image z; Described target high-resolution image x and described low-resolution image z meets following condition: z=B*x, and wherein B is fuzzy core; Gradient is asked to z=B*x both sides, then wherein b ₀and b _irepresent center coefficient and its surrounding neighbour coefficient of fuzzy core B respectively, represent the gradient image of target high-resolution image x, represent the gradient image of described low-resolution image z, represent the gradient image of described high resolution graphics photo; Order $x1=b_0\▿x,x2=\underset{i}{\mathrm{\Σ}}{b}_i{\▿x}_i,$ If $x2=\underset{i}{\mathrm{\Σ}}{b}_i{\▿x}_i$ Be similar to normal distribution, then $h_r=\arg \underset{h_{x1}}{\min }\left\{{\left|\right|h_z-h_{x1}\⊗h_{x2}\left|\right|}^2\right\};$

By solving-optimizing problem s.t.h _f=h _rupgrade transforming function transformation function f;

Wherein h _rrepresent the terraced histogram of reference of target high-resolution image x, h _zrepresent the histogram of gradients of described low-resolution image z, h _x1the discrete form of the probability density function of stochastic variable x1, h _x2the discrete form of the probability density function of I.i.d. random variables x2, it is convolution operation symbol.

Further, described update module 306 specifically for: according to iterative formula $x^{(t+1/2)}=x^{\left(t\right)}+\mathrm{\δ}((H^{T}y-H^T{\mathrm{hx}}^{\left(t\right)})+{\mathrm{\λ}}_2{\▿}^T(f-{\▿x}^{\left(t\right)}))$ Determine x ^(t+1/2), wherein x ^(t)represent the target high-resolution Image estimation value of the t time iteration, x ^(t+1/2)represent the target high-resolution Image estimation value of the t+1/2 time iteration, δ is constant, represent x ^(t)gradient map; According to determine wherein represent the high resolution graphics photo x of the t+1/2 time iteration _isparse coefficient, represent the high resolution graphics photo x of the t time iteration _ithe sub-dictionary that place initial clustering is corresponding, R _irepresent from initial target high-definition picture high resolution graphics photo x is obtained at i place, position _imatrix; According to sparse coefficient correction formula ${\mathrm{\α}}_i^{(t+1)}=S_{\mathrm{\λ}/c}({\mathrm{\φ}}^T{\mathrm{oH}}^T(y-\mathrm{H\φo}{\mathrm{\α}}_i^{(t+1/2)})/c+{\mathrm{\α}}_i^{(t+1/2)}-{\mathrm{\β}}_i^{(t+1/2)})+{\mathrm{\β}}_i^{(t+1/2)}$ With determine the high resolution graphics photo x of the t+1 time iteration _isparse coefficient ω _ia jth component ω _ijmeet: wherein h represents the controling parameters for regulating the rate of decay, and W represents normalized factor, represent high resolution graphics photo x _ithe most similar diagram photo of jth in corresponding most similar diagram photo group, S _λ/crepresent soft-threshold function, c represents regularization parameter; According to formula $x^{(t+1)}={\mathrm{\φ}}^{(t+1)}{\mathrm{o\α}}^{(t+1)}={\left(\underset{i}{\mathrm{\Σ}}{R}_i^T{R}_i\right)}^{-1}\underset{i}{\mathrm{\Σ}}{R}_i^T{\mathrm{\φ}}_i^{(t+1)}{\mathrm{\α}}_i^{(t+1)}$ Determine x ^(t+1), wherein x ^(t+1)represent the target high-resolution Image estimation value of the t+1 time iteration, φ ^(t+1)represent x ^(t+1)corresponding super complete dictionary.

Further, also comprise: filtration module 307; It is final goal high-definition picture x ' that described acquisition module 303 obtains convergence solution; Described filtration module 307, for adopting plural impact filtering formula to restore described final goal high-definition picture, wherein said plural impact filtering formula is:

$x^{\′\′}=-\frac{2}{\mathrm{\π}}\arctan \left(\mathrm{a\; Im}\left(\frac{x^{\′}}{\mathrm{\θ}}\right)\right)|\▿x^{\′}|+{\mathrm{\τ}}_{\mathrm{\η\η}}^{\′}+\stackrel{\‾}{\mathrm{\τ}}{x}_{\mathrm{\ξ\ξ}}^{\′}$

Wherein, x ' expression final goal high-definition picture, x " represent the final goal high-definition picture after restoring, represent the gradient map of x', η and ξ represents the gradient direction of image, and Im () represents extraction imaginary part, and a represents the adjustment parameter for controlling image sharpness, τ=| τ | exp (i θ) is complex scalar coefficient, it is real number scalar factor.

The super-resolution image reconstruction device that the present embodiment provides, may be used for the enforcement technical scheme of the super-resolution image reconstruction method performed corresponding to Fig. 2, it realizes principle and technique effect is similar, repeats no more herein.

One of ordinary skill in the art will appreciate that: all or part of step realizing above-mentioned each embodiment of the method can have been come by the hardware that programmed instruction is relevant.Aforesaid program can be stored in a computer read/write memory medium.This program, when performing, performs the step comprising above-mentioned each embodiment of the method; And aforesaid storage medium comprises: ROM, RAM, magnetic disc or CD etc. various can be program code stored medium.

Last it is noted that above embodiment is only in order to illustrate technical scheme of the present invention, be not intended to limit; Although with reference to previous embodiment to invention has been detailed description, those of ordinary skill in the art is to be understood that: it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein portion of techniques feature; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the spirit and scope of various embodiments of the present invention technical scheme.

Claims (16)

1. a super-resolution image reconstruction method, is characterized in that, comprising:

Step S1, using the low-resolution image y of input as the first high-definition picture according to described first high-definition picture carry out N fuzzy down-sampling, generate high resolution image data storehouse D _x;

Step S2, to described first high-definition picture adopt after bicubic interpolation algorithm carries out fuzzy down-sampling, then carry out fuzzy up-sampling and obtain the first low-resolution image for described first low-resolution image carry out N fuzzy down-sampling, generate low resolution image data storehouse D _z;

Step S3, adopts described bicubic interpolation algorithm by described first high-definition picture the second low-resolution image is obtained after amplifying S times by described second low-resolution image be divided at least one low-resolution image sheet;

Step S4, adopts approximate KNN searching algorithm at described low resolution image data storehouse D to each low-resolution image sheet _zobtain the most similar diagram photo group of corresponding low resolution;

Step S5, adopts Ridge Regression Method to calculate weight coefficient to the most similar diagram photo group of described low resolution;

Step S6, according to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power;

Step S7, adopts method of weighted mean to determine the second high-definition picture to described high resolution graphics photo and by described second low-resolution image join described database D _zin, by described second high-definition picture join described database D _xin.

2. method according to claim 1, is characterized in that, also comprises:

Step S8, by described second high-definition picture as the first new high-definition picture

After M repeated execution of steps S3-S7, obtain third high image in different resolution

Adopt described bicubic interpolation algorithm to described third high image in different resolution carry out fuzzy down-sampling, generate initial target high-definition picture

3. method according to claim 2, is characterized in that, also comprises:

Step S9, to described initial target high-definition picture alternating minimization algorithm computational mathematics model is adopted with described low-resolution image y:

${\mathrm{\α}}_y=\arg \underset{\mathrm{\φ},\mathrm{\α},f}{\min }\{{\left|\right|y-\mathrm{H\φo\α}\left|\right|}_2^2+{\mathrm{\λ}}_1{\mathrm{\Σ}}_i{\left|\right|{\mathrm{\α}}_i-{\mathrm{\β}}_i\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|f(\▿x|\▿y)-\▿x\left|\right|}^2\},$

s.t.h _f＝h _r,

Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition picture singular matrix H represents fuzzy down-sampling operator, λ ₁and λ ₂represent regularization parameter, α _ithe high resolution graphics photo x of target high-resolution image x _isparse coefficient, α is α _iset, β _iα _iat the non-local mean in sparse coding territory, o represents the sparse coefficient matrix α of all splicings and the product of super complete dictionary φ, represent gradient operation symbol, f is transforming function transformation function, h _rthe reference histograms of target high-resolution image, h _fit is the histogram of transforming function transformation function f.

4. method according to claim 3, is characterized in that, described to described initial target high-definition picture alternating minimization algorithm computational mathematics model is adopted specifically to comprise with described low-resolution image y:

Step S10, adopts K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary;

Step S11, upgrades described transforming function transformation function f according to the histogram of gradients algorithm that remains unchanged;

Step S12, upgrades described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f.

5. method according to claim 4, is characterized in that, adopts K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary, specifically comprises:

According to the high resolution graphics photo of size random selecting K described initial target high-definition picture as K initial clustering;

According to the distance between described high resolution graphics photo and a described K initial cluster center, all described high resolution graphics photos be divided into by described initial target high-definition picture are divided in corresponding initial clustering;

The corresponding sub-dictionary of principal component analysis (PCA) training based on svd is adopted to each described initial clustering;

K the described super complete dictionary of sub-dictionary composition that K described initial clustering is corresponding.

6. method according to claim 4, is characterized in that, upgrades described transforming function transformation function f, specifically comprise according to the histogram of gradients algorithm that remains unchanged:

Fuzzy up-sampling is carried out to described low-resolution image y, obtains low-resolution image z;

Described target high-resolution image x and described low-resolution image z meets following condition: z=B*x, and wherein B is fuzzy core;

Gradient is asked to z=B*x both sides, then wherein b ₀and b _irepresent center coefficient and its surrounding neighbour coefficient of fuzzy core B respectively, represent the gradient image of target high-resolution image x, represent the gradient image of described low-resolution image z, represent the gradient image of described high resolution graphics photo;

Order $x1=b_0\▿x,x2=\underset{i}{\mathrm{\Σ}}{b}_i\▿x_i,$ If $x2=\underset{i}{\mathrm{\Σ}}{b}_i\▿x_i$ Be similar to normal distribution, then $h_r=\arg \underset{h_{x1}}{\min }\left\{{\left|\right|h_z-h_{x1}\⊗h_{x2}\left|\right|}^2\right\};$

By solving-optimizing problem s.t.h _f=h _rupgrade transforming function transformation function f;

Wherein h _rrepresent the terraced histogram of reference of target high-resolution image x, h _zrepresent the histogram of gradients of described low-resolution image z, h _x1the discrete form of the probability density function of stochastic variable x1, h _x2the discrete form of the probability density function of I.i.d. random variables x2, it is convolution operation symbol.

7. method according to claim 4, is characterized in that, describedly upgrades described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f, specifically comprises:

According to formula $x^{(t+1/2)}=x^{\left(t\right)}+\mathrm{\δ}((H^{T}y-H^T{\mathrm{hx}}^{\left(t\right)})+{\mathrm{\λ}}_2{\▿}^T(f-\▿x^{\left(t\right)}))$ Determine x ^(t+1/2), wherein x ^(t)represent the target high-resolution Image estimation value of the t time iteration, x ^(t+1/2)represent the target high-resolution Image estimation value of the t+1/2 time iteration, δ is constant, represent x ^(t)gradient map;

According to determine wherein represent the high resolution graphics photo x of the t+1/2 time iteration _isparse coefficient, represent the high resolution graphics photo x of the t time iteration _ithe sub-dictionary that place initial clustering is corresponding, R _irepresent from initial target high-definition picture high resolution graphics photo x is obtained at i place, position _imatrix;

According to sparse coefficient ${\mathrm{\α}}_i^{(t+1)}=S_{\mathrm{\λ}/c}({\mathrm{\φ}}^T{\mathrm{oH}}^T(y-\mathrm{H\φo}{\mathrm{\α}}_i^{(t+1/2)})/c+{\mathrm{\α}}_i^{(t+1/2)}-{\mathrm{\β}}_i^{(t+1/2)})+{\mathrm{\β}}_i^{(t+1/2)}$ With determine the high resolution graphics photo x of the t+1 time iteration _isparse coefficient ω _ia jth component ω _ijmeet: wherein h represents the controling parameters for regulating the rate of decay, and W represents normalized factor, represent high resolution graphics photo x _ithe most similar diagram photo of jth in corresponding most similar diagram photo group, S _λ/crepresent soft-threshold function, c represents regularization parameter;

According to formula $x^{(t+1)}={\mathrm{\φ}}^{(t+1)}o{\mathrm{\α}}^{(t+1)}={\left(\underset{i}{\mathrm{\Σ}}{R}_i^T{R}_i\right)}^{-1}\underset{i}{\mathrm{\Σ}}{R}_i^T{\mathrm{\φ}}_i^{(t+1)}{\mathrm{\α}}_i^{(t+1)}$ Determine wherein x ^(t+1)represent the target high-resolution Image estimation value of the t+1 time iteration, φ ^(t+1)represent x ^(t+1)corresponding super complete dictionary.

8. the method according to any one of claim 4-7, is characterized in that, also comprises:

Repeated execution of steps S10-S12, obtaining convergence solution is final goal high-definition picture x';

Adopt plural impact filtering formula to restore described final goal high-definition picture, wherein said plural impact filtering formula is:

$x^{\′\′}=-\frac{2}{\mathrm{\π}}\arctan \left(\mathrm{aIm}\left(\frac{x^{\′}}{\mathrm{\θ}}\right)\right)|\▿x^{\′}|+\mathrm{\τ}{x}_{\mathrm{\η\η}}^{\′}+\stackrel{-}{\mathrm{\τ}}{x}_{\mathrm{\ξ\ξ}}^{\′}$

Wherein, x ' expression final goal high-definition picture, x " represent the final goal high-definition picture after restoring, represent the gradient map of x', η and ξ represents the gradient direction of image, and Im () represents extraction imaginary part, and a represents the adjustment parameter for controlling image sharpness, τ=| τ | exp (i θ) is complex scalar coefficient, it is real number scalar factor.

9. a super-resolution image reconstruction device, is characterized in that, comprising:

Generation module, for will input low-resolution image y as the first high-definition picture according to described first high-definition picture carry out N fuzzy down-sampling, generate high resolution image data storehouse D _x;

Described generation module, also for described first high-definition picture adopt after bicubic interpolation algorithm carries out fuzzy down-sampling, then carry out fuzzy up-sampling and obtain the first low-resolution image for described first low-resolution image carry out N fuzzy down-sampling, generate low resolution image data storehouse D _z;

Divide module, for adopting described bicubic interpolation algorithm by described first high-definition picture the second low-resolution image is obtained after amplifying S times by described second low-resolution image be divided at least one low-resolution image sheet;

Acquisition module, for adopting approximate KNN searching algorithm at described low resolution image data storehouse D to each low-resolution image sheet _zobtain the most similar diagram photo group of corresponding low resolution;

Computing module, for adopting Ridge Regression Method to calculate weight coefficient to the most similar diagram photo group of described low resolution;

Determination module, for according to described weight coefficient and the most similar diagram photo group determination high resolution graphics photo of high resolving power;

Described determination module, also for adopting method of weighted mean to determine the second high-definition picture to described high resolution graphics photo and by described second low-resolution image join described database D _zin, by described second high-definition picture join described database D _xin.

10. device according to claim 9, is characterized in that, also comprises:

Described acquisition module, also for obtaining third high image in different resolution

Described generation module, also with adopting described bicubic interpolation algorithm to described third high image in different resolution carry out fuzzy down-sampling, generate initial target high-definition picture

11. devices according to claim 10, is characterized in that, also comprise:

Described computation model, also for described initial target high-definition picture alternating minimization algorithm computational mathematics model is adopted with described low-resolution image y:

${\mathrm{\α}}_y=\arg \underset{\mathrm{\φ},\mathrm{\α},f}{\min }\{{\left|\right|y-\mathrm{H\φo\α}\left|\right|}_2^2+{\mathrm{\λ}}_1{\mathrm{\Σ}}_i{\left|\right|{\mathrm{\α}}_i-{\mathrm{\β}}_i\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|f(\▿x|\▿y)-\▿x\left|\right|}^2\},$

s.t.h _f＝h _r,

Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition picture singular matrix H represents fuzzy down-sampling operator, λ ₁and λ ₂represent regularization parameter, α _ithe high resolution graphics photo x of target high-resolution image x _isparse coefficient, α is α _iset, β _iα _iat the non-local mean in sparse coding territory, o represents the sparse coefficient matrix α of all splicings and the product of super complete dictionary φ, represent gradient operation symbol, f is transforming function transformation function, h _rthe reference histograms of target high-resolution image, h _fit is the histogram of transforming function transformation function f.

12. devices according to claim 11, is characterized in that, also comprise: update module;

Described acquisition module, also for adopting K means clustering algorithm and Principal Component Analysis Algorithm to obtain super complete dictionary;

Described update module, upgrades described transforming function transformation function f for the algorithm that remains unchanged according to histogram of gradients;

Described update module, also upgrades described initial target high-definition picture according to described super complete dictionary and described transforming function transformation function f.

13. devices according to claim 11, is characterized in that, described acquisition module specifically for:

According to the described high resolution graphics photo of size random selecting K described target high-resolution image as K initial clustering;

According to the distance between described high resolution graphics photo and a described K initial cluster center, all described high resolution graphics photos be divided into by described target high-resolution image are divided in corresponding initial clustering;

The corresponding sub-dictionary of principal component analysis (PCA) training based on svd is adopted to each described initial clustering;

K the described super complete dictionary of sub-dictionary composition that K described initial clustering is corresponding.

14. devices according to claim 11, is characterized in that, described update module specifically for:

Fuzzy up-sampling is carried out to described low-resolution image y, obtains low-resolution image z;

Described target high-resolution image x and described low-resolution image z meets following condition: z=B*x, and wherein B is fuzzy core;

Gradient is asked to z=B*x both sides, then wherein b ₀and b _irepresent center coefficient and its surrounding neighbour coefficient of fuzzy core B respectively, represent the gradient image of target high-resolution image x, represent the gradient image of described low-resolution image z, represent the gradient image of described high resolution graphics photo;

Order $x1=b_0\▿x,x2=\underset{i}{\mathrm{\Σ}}{b}_i\▿x_i,$ If $x2=\underset{i}{\mathrm{\Σ}}{b}_i\▿x_i$ Be similar to normal distribution, then $h_r=\arg \underset{h_{x1}}{\min }\left\{{\left|\right|h_z-h_{x1}\⊗h_{x2}\left|\right|}^2\right\};$

By solving-optimizing problem s.t.h _f=h _rupgrade transforming function transformation function f;

Wherein h _rrepresent the terraced histogram of reference of target high-resolution image x, h _zrepresent the histogram of gradients of described low-resolution image z, h _x1the discrete form of the probability density function of stochastic variable x1, h _x2the discrete form of the probability density function of I.i.d. random variables x2, it is convolution operation symbol.

15. devices according to claim 11, is characterized in that, described update module specifically for:

According to formula $x^{(t+1/2)}=x^{\left(t\right)}+\mathrm{\δ}((H^{T}y-H^T{\mathrm{hx}}^{\left(t\right)})+{\mathrm{\λ}}_2{\▿}^T(f-\▿x^{\left(t\right)}))$ Determine x ^(t+1/2), wherein x ^(t)represent the target high-resolution Image estimation value of the t time iteration, x ^(t+1/2)represent the target high-resolution Image estimation value of the t+1/2 time iteration, δ is constant, represent x ^(t)gradient map;

According to determine wherein represent the high resolution graphics photo x of the t+1/2 time iteration _isparse coefficient, represent the high resolution graphics photo x of the t time iteration _ithe sub-dictionary that place initial clustering is corresponding, R _irepresent from initial target high-definition picture high resolution graphics photo x is obtained at i place, position _imatrix;

According to sparse coefficient ${\mathrm{\α}}_i^{(t+1)}=S_{\mathrm{\λ}/c}({\mathrm{\φ}}^T{\mathrm{oH}}^T(y-\mathrm{H\φo}{\mathrm{\α}}_i^{(t+1/2)})/c+{\mathrm{\α}}_i^{(t+1/2)}-{\mathrm{\β}}_i^{(t+1/2)})+{\mathrm{\β}}_i^{(t+1/2)}$ With determine the high resolution graphics photo x of the t+1 time iteration _isparse coefficient ω _ia jth component ω _ijmeet: wherein h represents the controling parameters for regulating the rate of decay, and W represents normalized factor, represent high resolution graphics photo x _ithe most similar diagram photo of jth in corresponding most similar diagram photo group, S _λ/crepresent soft-threshold function, c represents regularization parameter;

According to formula $x^{(t+1)}={\mathrm{\φ}}^{(t+1)}o{\mathrm{\α}}^{(t+1)}={\left(\underset{i}{\mathrm{\Σ}}{R}_i^T{R}_i\right)}^{-1}\underset{i}{\mathrm{\Σ}}{R}_i^T{\mathrm{\φ}}_i^{(t+1)}{\mathrm{\α}}_i^{(t+1)}$ Determine x ^(t+1), wherein x ^(t+1)represent the target high-resolution Image estimation value of the t+1 time iteration, φ ^(t+1)represent x ^(t+1)corresponding super complete dictionary.

16. devices according to any one of claim 9-15, is characterized in that, also comprise: filtration module;

It is final goal high-definition picture x' that described acquisition module obtains convergence solution;

Described filtration module, for adopting plural impact filtering formula to restore described final goal high-definition picture, wherein said plural impact filtering formula is:

$x^{\′\′}=-\frac{2}{\mathrm{\π}}\arctan \left(\mathrm{aIm}\left(\frac{x^{\′}}{\mathrm{\θ}}\right)\right)|\▿x^{\′}|+\mathrm{\τ}{x}_{\mathrm{\η\η}}^{\′}+\stackrel{-}{\mathrm{\τ}}{x}_{\mathrm{\ξ\ξ}}^{\′}$

Wherein, x ' expression final goal high-definition picture, x " represent the final goal high-definition picture after restoring, represent the gradient map of x', η and ξ represents the gradient direction of image, and Im () represents extraction imaginary part, and a represents the adjustment parameter for controlling image sharpness, τ=| τ | exp (i θ) is complex scalar coefficient, it is real number scalar factor.