CN105488776B - Super-resolution image reconstruction method and device - Google Patents
Super-resolution image reconstruction method and device Download PDFInfo
- Publication number
- CN105488776B CN105488776B CN201410532272.6A CN201410532272A CN105488776B CN 105488776 B CN105488776 B CN 105488776B CN 201410532272 A CN201410532272 A CN 201410532272A CN 105488776 B CN105488776 B CN 105488776B
- Authority
- CN
- China
- Prior art keywords
- resolution
- resolution image
- low
- represent
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 64
- 238000005070 sampling Methods 0.000 claims abstract description 52
- 230000009466 transformation Effects 0.000 claims description 35
- 230000001131 transforming effect Effects 0.000 claims description 35
- 239000011159 matrix material Substances 0.000 claims description 18
- 238000000513 principal component analysis Methods 0.000 claims description 15
- 238000013178 mathematical model Methods 0.000 claims description 14
- 238000001914 filtration Methods 0.000 claims description 12
- 230000014509 gene expression Effects 0.000 claims description 10
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 9
- 230000003321 amplification Effects 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 6
- 238000012549 training Methods 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 5
- 238000000354 decomposition reaction Methods 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims 1
- 230000006870 function Effects 0.000 description 59
- 238000010586 diagram Methods 0.000 description 5
- 238000005457 optimization Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Landscapes
- Image Processing (AREA)
Abstract
The present invention, which provides a kind of super-resolution image reconstruction method and device, this method, to be included：N times are carried out according to the first high-definition picture and obscure down-sampling, generate high resolution image data storehouse, fuzzy down-sampling is carried out using bicubic interpolation algorithm to the first high-definition picture, fuzzy up-sampling obtains the first low-resolution image, carry out n times and obscure down-sampling, generation low resolution image data storehouse；The second low-resolution image is obtained after first high-definition picture is amplified S times, and is divided at least one low-resolution image piece；And the corresponding most like image sheet group of low resolution is obtained in low resolution image data storehouse using approximate KNN searching algorithm；Weight coefficient is calculated using Ridge Regression Method to the most like image sheet group of low resolution and determines high resolution graphics photo.The second high-definition picture is determined using weighted average to high resolution graphics photo, so as to effectively reduce the error of rebuilding super resolution image, improves the quality for rebuilding image in different resolution.
Description
Technical field
The present invention relates to image, field of video processing, more particularly to a kind of super-resolution image reconstruction method and device.
Background technology
With developing rapidly for multimedia technology, visual vivid effect and abundant picture of the people for image and video
The requirement of detailed information is higher and higher, this needs high-resolution image and video, and in actual image processing and analysis
In system, high-resolution image and video are generally also required for.Set however, the resolution ratio of image is generally limited by Image Acquisition
The restriction conditions such as standby, optics, image taking speed and hardware store, what is captured in many imaging applications is all the image of low resolution
And video, for example, digital camera, medical image system and video monitoring system capture usually be all low resolution image and
Video.So in order to obtain high-resolution image and video, it is necessary to which super-resolution technique goes to utilize the low resolution figure obtained
Picture and video reconstruct high-resolution image and video.
Existing super-resolution image reconstruction method has：Method based on interpolation, the method based on reconstruction and Case-based Reasoning
The method of study, wherein the method based on reconstruction needs extra database, for the low-resolution image of acquisition, utilizes image
The priori of a variety of images in database is modeled current low-resolution image, then, by low point after modeling
Resolution image reconstruction is high-resolution image.
But conventionally, as what is utilized when being modeled to super-resolution image is one huge representative
Low resolution and high resolution graphics photo it is as much as possible including all picture structures to data storehouse, the database so that weigh
Build that super-resolution image process is excessively complicated, and usually there will be the feelings that can not restore high frequency detail based on the image data base
Condition.
The content of the invention
The present invention provides a kind of super-resolution image reconstruction method and device, so as to effectively reduce rebuilding super resolution
The error of image, and then improve the quality of reconstruction image.
In a first aspect, the present invention provides a kind of super-resolution image reconstruction method, including：Step S1, by low point of input
Resolution image y is as the first high-definition pictureAccording to first high-definition pictureCarry out n times and obscure down-sampling,
Generate high resolution image data storehouse Dx；Step S2, to first high-definition pictureUsing bicubic interpolation algorithm into
Row, which obscures down-sampling and then carries out fuzzy up-sampling, obtains the first low-resolution imageFor first low resolution
ImageCarry out n times and obscure down-sampling, generation low resolution image data storehouse Dz；Step S3, is calculated using the bicubic interpolation
Method is by first high-definition pictureThe second low-resolution image is obtained after S times of amplificationBy second low resolution
ImageIt is divided at least one low-resolution image piece；Step S4, to each low-resolution image piece using approximate nearest
Adjacent searching algorithm is in the low resolution image data storehouse DzObtain the corresponding most like image sheet group of low resolution；Step S5,
Weight coefficient is calculated using Ridge Regression Method to the most like image sheet group of the low resolution；Step S6, according to the weight coefficient
High resolution graphics photo is determined with the most like image sheet group of high-resolution；Step S7, uses the high resolution graphics photo and adds
Weight average determines the second high-definition pictureAnd by second low-resolution imageIt is added to the database DzIn,
By second high-definition pictureIt is added to the database DxIn.
With reference to first aspect, in the first possible embodiment of first aspect, further include：Step S8, by described
Two high-definition picturesAs the first new high-definition pictureAfter repeating step S3-S7 by M times, obtain
3rd high-definition pictureUsing the bicubic interpolation algorithm to the 3rd high-definition pictureCarry out under obscuring
Sampling, generates initial target high-definition picture
The first possible embodiment with reference to first aspect, in second of possible embodiment of first aspect, also
Including：Step S9, to the initial target high-definition pictureAlternating minimization algorithm is used with the low-resolution image y
Calculate mathematical model：
s.t.hf=hr,
Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition pictureDegenerate matrix H represents mould
Paste down-sampling operator, λ1And λ2Represent regularization parameter, αiIt is the high resolution graphics photo x of target high-resolution image xiIt is dilute
Sparse coefficient, α αiSet, βiIt is αiNon-local mean in sparse coding domain, ο represent the sparse coefficient matrix α of all splicings
With the product of super complete dictionary φ,Represent gradient operation symbol, f is transforming function transformation function, hrBe target high-resolution image reference it is straight
Fang Tu, hfIt is the histogram of transforming function transformation function f.
Second of possible embodiment with reference to first aspect, in the third possible embodiment of first aspect, institute
State to the initial target high-definition pictureMathematical modulo is calculated using alternating minimization algorithm with the low-resolution image y
Type specifically includes：Step S10, super complete dictionary is obtained using K mean cluster algorithm and Principal Component Analysis Algorithm；Step S11, root
Algorithm, which is remained unchanged, according to histogram of gradients updates the transforming function transformation function f；Step S12, according to the super complete dictionary and the change
Exchange the letters number f updates the initial target high-definition picture.
The third possible embodiment with reference to first aspect, in the 4th kind of possible embodiment of first aspect, is adopted
Super complete dictionary is obtained with K mean cluster algorithm and Principal Component Analysis Algorithm, is specifically included：K is randomly selected according to size
The high resolution graphics photo of a initial target high-definition picture is as K initial clustering；According to the high resolution graphics
The distance between photo and the K initial cluster center, the initial target high-definition picture are divided into all
The high resolution graphics photo be divided into corresponding initial clustering；Each initial clustering is used based on singular value point
The corresponding sub- dictionary of Principal Component Analysis training of solution；The K corresponding K sub- dictionary composition of the initial clustering super complete words
Allusion quotation.
The third possible embodiment with reference to first aspect, in the 5th kind of possible embodiment of first aspect, according to
Histogram of gradients remains unchanged algorithm and updates the transforming function transformation function f, specifically includes：The low-resolution image y is obscured
Sampling, obtains low-resolution image z；The target high-resolution image x and the low-resolution image z meet following condition：Z=
B*x, wherein B are fuzzy core；Gradient is asked to z=B*x both sides, thenWherein b0And biPoint
Not Biao Shi fuzzy core B center coefficient and its surrounding's neighbour's coefficient,Represent the gradient image of target high-resolution image x,Represent the gradient image of the low-resolution image z,Represent the gradient image of the target high-resolution image sheet；OrderIfIt is similar to normal distribution, then
Pass through solving-optimizing problems.t.hf=hrTo update transforming function transformation function f.
Wherein hrRepresent the terraced histogram of reference of target high-resolution image x, hzRepresent the ladder of the low-resolution image z
Spend histogram, hx1It is the discrete form of the probability density function of stochastic variable x1, hx2It is the general of I.i.d. random variables x2
The discrete form of rate density function,It is convolution operation symbol.
The third possible embodiment with reference to first aspect, in the 6th kind of possible embodiment of first aspect, institute
State and the initial target high-definition picture is updated according to the super complete dictionary and the transforming function transformation function f, specifically include：According to
FormulaDetermine x(t+1/2), wherein x(t)Represent the t times iteration
Target high-resolution Image estimation value, x(t+1/2)Represent the target high-resolution Image estimation value of the t+1/2 times iteration, δ is
Constant,Represent x(t)Gradient map；According toDetermineWhereinRepresent t+1/
The high resolution graphics photo x of 2 iterationiSparse coefficient,Represent the high resolution graphics photo x of the t times iterationiPlace is initial
Cluster corresponding sub- dictionary, RiRepresent from initial target high-definition pictureHigh resolution graphics photo x is obtained at the i of positioni's
Matrix.
According to sparse coefficientWithDetermine the high resolution graphics photo x of the t+1 times iterationiSparse coefficientωiJ-th of component
ωijMeet：Wherein h represents to be used for the control parameter for adjusting the rate of decay, W expression normalizings
Change the factor,Represent high resolution graphics photo xiJ-th of most like image sheet in corresponding most like image sheet group, Sλ/cTable
Show soft-threshold function, c represents regularization parameter；According to formula
Determine x(t+1), wherein x(t+1)Represent the target high-resolution Image estimation value of the t+1 times iteration, φ(t+1)Represent x(t+1)It is corresponding
Super complete dictionary.
The third possible embodiment or the 4th kind of possible embodiment with reference to first aspect or the 5th kind of possible implementation
Mode or the 6th kind of possible embodiment, in the 7th kind of possible embodiment of first aspect, further include：Repeat step
S10-S12, acquisition convergence solution are final goal high-definition picture x '；Using plural impact filtering formula to the final goal
High-definition picture is restored, wherein the plural number impact filtering formula is：
Wherein, x ' expressions final goal high-definition picture, x " represent the final goal high-definition picture after restoring,Represent the gradient map of x', η and ξ represent the gradient direction of image, and Im () represents extraction imaginary part, and a represents to be used to control image
The adjusting parameter of acutance, τ=| τ | exp (i θ) is complex scalar coefficient,It is real number scalar factor.
Second aspect, the present invention provide a kind of super-resolution image reconstruction device, including：Generation module, for that will input
Low-resolution image y as the first high-definition pictureAccording to first high-definition pictureN times are carried out to obscure
Down-sampling, generation high resolution image data storehouse Dx；The generation module, is additionally operable to first high-definition pictureAdopt
Fuzzy down-sampling, which is carried out, with bicubic interpolation algorithm and then carries out fuzzy up-sampling obtaining the first low-resolution imageIt is right
In first low-resolution imageCarry out n times and obscure down-sampling, generation low resolution image data storehouse Dz；Division module,
For using the bicubic interpolation algorithm by first high-definition pictureThe second low resolution figure is obtained after S times of amplification
PictureBy second low-resolution imageIt is divided at least one low-resolution image piece；Acquisition module, for each
A low-resolution image piece is using approximate KNN searching algorithm in the low resolution image data storehouse DzObtain corresponding low
The most like image sheet group of resolution ratio；Computing module, based on using Ridge Regression Method to the most like image sheet group of the low resolution
Calculate weight coefficient；Determining module, for determining high-resolution according to the weight coefficient and the most like image sheet group of high-resolution
Image sheet；The determining module, is additionally operable to determine the second high resolution graphics using weighted average to the high resolution graphics photo
PictureAnd by second low-resolution imageIt is added to the database DzIn, by second high-definition picture
It is added to the database DxIn.
With reference to second aspect, in the first possible embodiment of second aspect, further include：The acquisition module, also
For obtaining the 3rd high-definition pictureThe generation module is also high to the described 3rd using the bicubic interpolation algorithm
Image in different resolutionFuzzy down-sampling is carried out, generates initial target high-definition picture
With reference to the first possible embodiment of second aspect, in second of possible embodiment of second aspect, go back
Including：The computing module, is additionally operable to the initial target high-definition pictureWith the low-resolution image y using friendship
Mathematical model is calculated for algorithm is minimized：
s.t. hf=hr,
Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition pictureDegenerate matrix H represents mould
Paste down-sampling operator, λ1And λ2Represent regularization parameter, αiIt is the high resolution graphics photo x of target high-resolution image xiIt is dilute
Sparse coefficient, α αiSet, βiIt is αiNon-local mean in sparse coding domain, ο represent the sparse coefficient matrix α of all splicings
With the product of super complete dictionary φ,Represent gradient operation symbol, f is transforming function transformation function, hrBe target high-resolution image reference it is straight
Fang Tu, hfIt is the histogram of transforming function transformation function f.
With reference to second of possible embodiment of second aspect, in the third possible embodiment of second aspect, go back
Including：Update module；The acquisition module, is additionally operable to obtain using K mean cluster algorithm and Principal Component Analysis Algorithm super complete
Dictionary；The update module, updates the transforming function transformation function f for remaining unchanged algorithm according to histogram of gradients；The renewal mould
Block, updates the initial target high-definition picture always according to the super complete dictionary and the transforming function transformation function f.
With reference to the third possible embodiment of second aspect, in the 4th kind of possible embodiment of second aspect, institute
Acquisition module is stated to be specifically used for：The high-resolution of the K target high-resolution images is randomly selected according to size
Image sheet is as K initial clustering；According to the distance between the high resolution graphics photo and the K initial cluster center,
All high resolution graphics photos that the target high-resolution image is divided into are divided into corresponding initial clustering
In；To each initial clustering using the corresponding sub- dictionary of Principal Component Analysis training based on singular value decomposition；K described first
Begin to cluster the corresponding K sub- dictionary composition super complete dictionary.
With reference to the third possible embodiment of second aspect, in the 5th kind of possible embodiment of second aspect, institute
Update module is stated to be specifically used for：Fuzzy up-sampling is carried out to the low-resolution image y, obtains low-resolution image z；The mesh
Absolute altitude image in different resolution x and the low-resolution image z meet following condition：Z=B*x, wherein B are fuzzy core；To z=B*x
Gradient is sought on both sides, thenWherein b0And biThe center coefficient of fuzzy core B is represented respectively
And its surrounding's neighbour's coefficient,Represent the gradient image of target high-resolution image x,Represent the low-resolution image
The gradient image of z,Represent the gradient image of the high resolution graphics photo；OrderIfIt is similar to normal distribution, then
Pass through solving-optimizing problems.t. hf=hrTo update transforming function transformation function f.
Wherein hrRepresent the terraced histogram of reference of target high-resolution image x, hzRepresent the ladder of the low-resolution image z
Spend histogram, hx1It is the discrete form of the probability density function of stochastic variable x1, hx2It is the general of I.i.d. random variables x2
The discrete form of rate density function,It is convolution operation symbol.
With reference to the third possible embodiment of second aspect, in the 6th kind of possible embodiment of second aspect, institute
Update module is stated to be specifically used for：According to the iterative solution formula of super-resolution optimization problem
Determine x(t+1/2), wherein x(t)Represent the t times iteration
Target high-resolution Image estimation value, x(t+1/2)Represent the target high-resolution Image estimation value of the t+1/2 times iteration, δ is normal
Number,Represent x(t)Gradient map；According toDetermineWhereinRepresent the t+1/2 times
The high resolution graphics photo x of iterationiSparse coefficient,Represent the high resolution graphics photo x of the t times iterationiPlace initial clustering
Corresponding sub- dictionary, RiRepresent from initial target high-definition pictureHigh resolution graphics photo x is obtained at the i of positioniMatrix；Root
According to sparse coefficientWith
Determine the high resolution graphics photo x of the t+1 times iterationiSparse coefficientωiJ-th of component ωijMeet：Wherein h represents the control parameter for adjusting the rate of decay, and W represents normalization factor,
Represent high resolution graphics photo xiJ-th of most like image sheet in corresponding most like image sheet group, Sλ/cRepresent soft-threshold letter
Number, c represent regularization parameter；According to formulaDetermine x(t+1), its
Middle x(t+1)Represent the target high-resolution Image estimation value of the t+1 times iteration, φ(t+1)Represent x(t+1)Corresponding super complete dictionary.
It may implement with reference to the third possible embodiment of second aspect or the 4th kind of possible embodiment or the 5th kind
Mode or the 6th kind of possible embodiment, in the 7th kind of possible embodiment of second aspect, further include：Filter module；Institute
It is final goal high-definition picture x ' to state acquisition module and obtain convergence solution；The filter module, for being filtered using plural number impact
Ripple formula restores the final goal high-definition picture, wherein the plural number impact filtering formula is：
Wherein, x ' expressions final goal high-definition picture, x " represent the final goal high-definition picture after restoring,Represent the gradient map of x', η and ξ represent the gradient direction of image, and Im () represents extraction imaginary part, and a represents to be used for control figure
As the adjusting parameter of acutance, τ=| τ | exp (i θ) is complex scalar coefficient,It is real number scalar factor.
The present invention provides a kind of super-resolution image reconstruction method, including：Using the low-resolution image of input as first
High-definition picture, carries out n times according to first high-definition picture and obscures down-sampling, generate high resolution image data
Storehouse, carries out fuzzy down-sampling and then obscures to adopt to first high-definition picture using bicubic interpolation algorithm
Sample obtains the first low-resolution image, and carrying out n times for first low-resolution image obscures down-sampling, generates low resolution
Image data base；Second low point is obtained after first high-definition picture is amplified S times using the bicubic interpolation algorithm
Resolution image, at least one low-resolution image piece is divided into by second low-resolution image；To each low resolution
Image sheet is most like in the corresponding low resolution of low resolution image data storehouse acquisition using approximate KNN searching algorithm
Image sheet group；Weight coefficient is calculated using Ridge Regression Method to the most like image sheet group of the low resolution；According to the weight system
Number and the most like image sheet group of high-resolution determine high resolution graphics photo.Weighted average is used to the high resolution graphics photo
Determine the second high-definition picture, initial target high resolution graphics is estimated by the multistage amplifying technique based on Ridge Regression Method
Picture, and comprehensively utilize data fidelity, sparse non local regularization and histogram of gradients regularization priori and carry out mathematical modeling,
Final target high-resolution image is obtained by duty Optimization, edge is further sharpened finally by impact filtering
With recovery picture structure treatment of details so as to effectively reduce the error of rebuilding super resolution image, and then improve reconstruction image
Quality.
Brief description of the drawings
Fig. 1 is a kind of flow chart for super-resolution image reconstruction method that one embodiment of the invention provides；
Fig. 2 is the flow chart for the super-resolution image reconstruction method that another embodiment of the present invention provides；
Fig. 3 is a kind of super-resolution image reconstruction apparatus structure schematic diagram that one embodiment of the invention provides；
Fig. 4 is a kind of super-resolution image reconstruction apparatus structure schematic diagram that another embodiment of the present invention provides.
Embodiment
To make the purpose, technical scheme and advantage of the embodiment of the present invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, the technical solution in the embodiment of the present invention is clearly and completely described, it is clear that described embodiment is
Part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art
All other embodiments obtained without making creative work, belong to the scope of protection of the invention.
Fig. 1 is a kind of flow chart for super-resolution image reconstruction method that one embodiment of the invention provides, wherein this method
Scene suitable for obtaining high-definition picture or video, the executive agent of wherein this method are：Super-resolution image reconstruction
Device, the device can be that intelligent terminal, the super-resolution image reconstruction methods such as computer specifically include following flow：
Step S1, using the low-resolution image of input as the first high-definition picture, according to the first high-definition picture
Carry out n times and obscure down-sampling, generation high resolution image data storehouse.
Wherein, low-resolution image is represented with y, and the first high-definition picture is usedRepresent, high resolution image data storehouse
Use DxRepresent.
Step S2, fuzzy down-sampling and then progress are carried out to the first high-definition picture using bicubic interpolation algorithm
Fuzzy up-sampling obtains the first low-resolution image, and carrying out n times for the first low-resolution image obscures down-sampling, generates low point
Resolution image data base.
Wherein, the first low-resolution image is usedRepresent, low resolution image data storehouse DzRepresent, in same scale level
The first low-resolution image on notWith the first high-definition pictureBetween correspondence establish it is as follows：Wherein, * is convolution operation symbol, and ↑ s is that the up-sampling that scale factor is s operates
Symbol, ↓ s are the down-sampling operators that scale factor is s, and B is fuzzy core, EsIt is the scale factor that first fuzzy down-sampling up-samples again
For the operator that meets of s, then the first low-resolution image is obtainedFor the first low-resolution imageN times are carried out to obscure down
Sampling, generation low resolution image data storehouse Dz。
Step S3, the second low resolution is obtained after the first high-definition picture is amplified S times using bicubic interpolation algorithm
Image, at least one low-resolution image piece is divided into by the second low-resolution image.
Wherein, the second low-resolution image isSpecifically,And by the second low-resolution imageDraw
It is divided at least one low-resolution image piece.
Step S4, to each low-resolution image piece using approximate KNN searching algorithm in low resolution image data
Storehouse obtains the corresponding most like image sheet group of low resolution.
Step S5, weight coefficient is calculated to the most like image sheet group of the low resolution using Ridge Regression Method.
Specifically, approximate KNN searching algorithm is may be by from low resolution for any one low-resolution image piece
Image data base DzThe corresponding most like image sheet group of low resolution of middle search, such as：Assuming that low-resolution image piece ziIt is corresponding
The most like image sheet group of low resolution is Lpi, then using Ridge Regression Modeling Method to ziAnd LpiBetween linear approximate relationship built
The constrained optimization problem that mould is formed is as follows：Wherein, γ is weight coefficient, and τ is to be used to alleviate
Singularity problem and the regularization parameter for stablizing solution.By solving this regular least square regression problem, the present invention gives
Go out the closed solutions of the problem：Wherein, I is unit matrix.
Step S6, high resolution graphics photo is determined according to the weight coefficient and the most like image sheet group of high-resolution.
By z in step S5iWith LpiMapping relations from low resolution image data storehouse DzIt is applied to high-definition picture number
According to storehouse Dx, then high resolution graphics photo piWith the most like image sheet group Hp of high-resolutioniRelational expression be pi=Hpiγ。
Step S7, the second high-definition picture is determined to high resolution graphics photo using weighted mean method, and by described in
Second low-resolution image is added in the database, and second high-definition picture is added in the database.
Wherein, the second high-definition picture is
The present invention provides a kind of super-resolution image reconstruction method, including：Using the low-resolution image of input as first
High-definition picture, carries out n times according to first high-definition picture and obscures down-sampling, generate high resolution image data
Storehouse, carries out fuzzy down-sampling and then obscures to adopt to first high-definition picture using bicubic interpolation algorithm
Sample obtains the first low-resolution image, and carrying out n times for first low-resolution image obscures down-sampling, generates low resolution
Image data base；Second low point is obtained after first high-definition picture is amplified S times using the bicubic interpolation algorithm
Resolution image, at least one low-resolution image piece is divided into by second low-resolution image；To each low resolution
Image sheet is most like in the corresponding low resolution of low resolution image data storehouse acquisition using approximate KNN searching algorithm
Image sheet group；Weight coefficient is calculated using Ridge Regression Method to the most like image sheet group of the low resolution；According to the weight system
Number and the most like image sheet group of high-resolution determine high resolution graphics photo.Weighted average is used to the high resolution graphics photo
Method determines the second high-definition picture, estimates initial target high resolution graphics by the multistage amplifying technique based on Ridge Regression Method
Picture, so as to effectively reduce the error of rebuilding super resolution image, and then improves the quality of reconstruction image.
Fig. 2 is the flow chart for the super-resolution image reconstruction method that another embodiment of the present invention provides, which is to build
Stand on upper embodiment basis, performed after upper embodiment step S7, wherein specifically including：
Step S8, using the second high-definition picture as the first new high-definition picture, step is repeated by M times
After S3-S7, the 3rd high-definition picture is obtained, the 3rd high-definition picture is obscured down using bicubic interpolation algorithm
Sampling, generates initial target high-definition picture.
Wherein, the 3rd high-definition picture isInitial target high-definition pictureIt is it is noted that of the invention
Employing and repeating in step S3-S7 M times is based on when scale factor is smaller, low-resolution image and high-definition picture
What this more like factor considered.Therefore, in view of natural image in scale domain and spatial domain there are a large amount of self-similarity redundancies,
Different from traditional bicubic interpolation method, method of the present invention is using the multistage amplifying technique based on Ridge Regression Method come just
Beginningization target high-resolution image.It is to use successive ignition magnification scheme, when single amplification method is taken by multistage amplifying technique
Dai Hou, the scale factor in every grade of amplification procedure of the present invention is just very small, so as to be searched from low-resolution image storehouse
Rope is used for high resolution image reconstruction to more similar diagram photos.Low-resolution image y for width input and given
General size factor d, the cascade in multistage amplification procedure of the present invention are confirmed as
Wherein s is the scale factor in every grade of amplification.
Step S9, mathematics is calculated to initial target high-definition picture and low-resolution image using alternating minimization algorithm
Model：
s.t. hf=hr,
Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition pictureDegenerate matrix H represents mould
Paste down-sampling operator, λ1And λ2Represent regularization parameter, αiIt is the high resolution graphics photo x of target high-resolution image xiIt is dilute
Sparse coefficient, α αiSet, βiIt is αiNon-local mean in sparse coding domain, ο represent the sparse coefficient matrix α of all splicings
With the product of super complete dictionary φ,Represent gradient operation symbol, f is transforming function transformation function, hrBe target high-resolution image reference it is straight
Fang Tu, hfIt is the histogram of transforming function transformation function f.
Specifically, the initial target high-definition picture determined in step S1-S8May not be most for the mathematical model
Excellent solution, it is therefore desirable to continuous iteration, untill obtaining convergence solution.Wherein, it is described to the target high-resolution image x and institute
Low-resolution image y is stated to specifically include using alternating minimization algorithm calculating mathematical model：
Step S10, super complete dictionary is obtained using K mean cluster algorithm and Principal Component Analysis Algorithm.
Specifically, super complete dictionary is obtained using K mean cluster algorithm and Principal Component Analysis Algorithm, specifically included：According to
Size randomly selects the high resolution graphics photo of the K target high-resolution images as K initial clustering；Root
According to the distance between the high resolution graphics photo and the K initial cluster center, by the target high-resolution image institute
All high resolution graphics photos being divided into are divided into corresponding initial clustering；Each initial clustering is used
The corresponding sub- dictionary of Principal Component Analysis training based on singular value decomposition；The corresponding K sub- dictionary composition of the described initial clusterings of K
The super complete dictionary.
Step S11, algorithm renewal transforming function transformation function is remained unchanged according to histogram of gradients.
Wherein, transforming function transformation function f, alternatively, remains unchanged algorithm according to histogram of gradients and updates the transforming function transformation function f,
Specifically include：Fuzzy up-sampling is carried out to the low-resolution image y, obtains low-resolution image z；The target high-resolution
Image x and the low-resolution image z meet following condition：Z=B*x, wherein B are fuzzy core；Gradient is asked to z=B*x both sides,
ThenWherein b0And biRepresent respectively fuzzy core B center coefficient and it around it is near
Adjacent coefficient,Represent the gradient image of target high-resolution image x,Represent the gradient image of the low-resolution image z,Represent the gradient image of the high resolution graphics photo；OrderIfClosely
Normal distribution is similar to, then
Pass through solving-optimizing problems.t.hf=hrTo update transforming function transformation function f；Wherein hrTable
Show the terraced histogram of reference of target high-resolution image x, hzRepresent the histogram of gradients of the low-resolution image z, hx1Be with
The discrete form of the probability density function of machine variable x1, hx2It is the discrete of the probability density function of I.i.d. random variables x2
Form,It is convolution operation symbol.
Step S12, updates target high-resolution image according to super complete dictionary and transforming function transformation function.
It is specifically, described that the initial target high-definition picture is updated according to the super complete dictionary and the transforming function transformation function f,
Specifically include：According to the iterative solution formula of super-resolution optimization problem
Determine x(t+1/2), wherein x(t)Represent the target high-resolution Image estimation value of the t times iteration, x(t+1/2)Represent the t+1/2 times repeatedly
The target high-resolution Image estimation value in generation, δ is constant,Represent x(t)Gradient map；According to
DetermineWhereinRepresent the high resolution graphics photo x of the t+1/2 times iterationiSparse coefficient,Represent the t times
The high resolution graphics photo x of iterationiThe corresponding sub- dictionary of place initial clustering, RiRepresent from initial target high-definition picture
High resolution graphics photo x is obtained at the i of positioniMatrix.
According to sparse coefficientWithDetermine the high resolution graphics photo x of the t+1 times iterationiSparse coefficientωiJ-th of component
ωijMeet：Wherein h represents to be used for the control parameter for adjusting the rate of decay, W expression normalizings
Change the factor,Represent high resolution graphics photo xiJ-th of most like image in corresponding most like image sheet group, Sλ/cRepresent
Soft-threshold function, c represent regularization parameter；According to formulaReally
Determine x(t+1), wherein x(t+1)Represent the target high-resolution Image estimation value of the t+1 times iteration, φ(t+1)Represent x(t+1)It is corresponding super
Complete dictionary.Circulation performs S10-S12, untill the target high-resolution image of acquisition is convergence solution.
Further, step S10-S12 is repeated, acquisition convergence solution is final goal high-definition picture x；Using multiple
Number impact filtering formula restores the final goal high-definition picture, wherein the plural number impact filtering formula is：
Wherein, x ' expressions final goal high-definition picture, x " represent the final goal high-definition picture after restoring,Represent the gradient map of x', η and ξ represent the gradient direction of image, and Im () represents extraction imaginary part, and a represents to be used to control image
The adjusting parameter of acutance, τ=| τ | exp (i θ) is complex scalar coefficient,It is real number scalar factor.
Super-resolution image reconstruction method provided in this embodiment comprehensively utilizes data fidelity, sparse non local regularization
Mathematical modeling is carried out with histogram of gradients regularization priori, final target high-resolution is obtained by duty Optimization
Image, edge is further sharpened finally by impact filtering and recovers picture structure treatment of details.
Fig. 3 is a kind of super-resolution image reconstruction apparatus structure schematic diagram that one embodiment of the invention provides, the wherein dress
Put including：Generation module 301, for using the low-resolution image y of input as the first high-definition pictureAccording to described
One high-definition pictureCarry out n times and obscure down-sampling, generation high resolution image data storehouse Dx；The generation module 301, also
For to first high-definition pictureFuzzy down-sampling is carried out using bicubic interpolation algorithm and then is carried out on fuzzy
Sampling obtains the first low-resolution imageFor first low-resolution imageCarry out n times and obscure down-sampling, generate
Low resolution image data storehouse Dz；Division module 302, for using the bicubic interpolation algorithm by first high-resolution
ImageThe second low-resolution image is obtained after S times of amplificationBy second low-resolution imageIt is divided at least one
Low-resolution image piece；Acquisition module 303, for being existed to each low-resolution image piece using approximate KNN searching algorithm
The low resolution image data storehouse DzObtain the corresponding most like image sheet group of low resolution；Computing module 304, for institute
State low resolution most like image sheet group and weight coefficient is calculated using Ridge Regression Method；Determining module 305, for according to the weight
Coefficient and the most like image sheet group of high-resolution determine high resolution graphics photo；The determining module 305, is additionally operable to the height
Resolution chart photo determines the second high-definition picture using weighted mean methodAnd by second low-resolution image
It is added to the database DzIn, by second high-definition pictureIt is added to the database DxIn.
Super-resolution image reconstruction device provided in this embodiment, can be used for performing the super-resolution figure corresponding to Fig. 1
As the implementation technical solution of method for reconstructing, its implementing principle and technical effect is similar, and details are not described herein again.
Fig. 4 is a kind of super-resolution image reconstruction apparatus structure schematic diagram that another embodiment of the present invention provides, at Fig. 3 pairs
Answer on embodiment basis, the acquisition module 303, be additionally operable to obtain the 3rd high-definition pictureThe generation module
301, also using the bicubic interpolation algorithm to the 3rd high-definition pictureFuzzy down-sampling is carried out, generates initial mesh
Absolute altitude image in different resolutionThe computing module 304, is additionally operable to the initial target high-definition pictureWith described low point
Resolution image y calculates mathematical model using alternating minimization algorithm：
s.t. hf=hr,
Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition pictureDegenerate matrix H is represented
Fuzzy down-sampling operator, λ1And λ2Represent regularization parameter, αiIt is the high resolution graphics photo x of target high-resolution image xi's
Sparse coefficient, α αiSet, βiIt is αiNon-local mean in sparse coding domain, ο represent the sparse coefficient square of all splicings
The product of battle array α and super complete dictionary φ,Represent gradient operation symbol, f is transforming function transformation function, hrIt is the ginseng of target high-resolution image
Examine histogram, hfIt is the histogram of transforming function transformation function f.
Further, which further includes：Update module 306；The acquisition module 303, is additionally operable to use K mean cluster
Algorithm and Principal Component Analysis Algorithm obtain super complete dictionary；The update module 306, for being remained unchanged according to histogram of gradients
Algorithm updates the transforming function transformation function f；The update module 306, updates always according to the super complete dictionary and the transforming function transformation function f
The initial target high-definition picture.
Alternatively, the acquisition module 303 is specifically used for：The K target high-resolution are randomly selected according to size
The high resolution graphics photo of rate image is as K initial clustering；According to described to the high resolution graphics photo and the K
The distance between a initial cluster center, all high resolution graphics that the target high-resolution image is divided into
Photo is divided into corresponding initial clustering；Principal Component Analysis based on singular value decomposition is used to each initial clustering
The corresponding sub- dictionary of training；The K corresponding K sub- dictionary composition of the initial clustering super complete dictionaries.
Alternatively, the update module 306 is specifically used for：Fuzzy up-sampling is carried out to the low-resolution image y, is obtained
Low-resolution image z；The target high-resolution image x and the low-resolution image z meet following condition：Z=B*x, wherein B
For fuzzy core；Gradient is asked to z=B*x both sides, thenWherein b0And biRepresent fuzzy core B's respectively
Center coefficient and its surrounding's neighbour's coefficient,Represent the gradient image of target high-resolution image x,Represent described low point
The gradient image of resolution image z,Represent the gradient image of the high resolution graphics photo；Order
IfIt is similar to normal distribution, then
Pass through solving-optimizing problems.t. hf=hrTo update transforming function transformation function f；
Wherein hrRepresent the terraced histogram of reference of target high-resolution image x, hzRepresent the ladder of the low-resolution image z
Spend histogram, hx1It is the discrete form of the probability density function of stochastic variable x1, hx2It is the general of I.i.d. random variables x2
The discrete form of rate density function,It is convolution operation symbol.
Further, the update module 306 is specifically used for：According to iterative solution formula
Determine x(t+1/2), wherein x(t)Represent the t times iteration
Target high-resolution Image estimation value, x(t+1/2)Represent the target high-resolution Image estimation value of the t+1/2 times iteration, δ is
Constant,Represent x(t)Gradient map；According toDetermineWhereinRepresent t+1/2
The high resolution graphics photo x of secondary iterationiSparse coefficient,Represent the high resolution graphics photo x of the t times iterationiPlace is initial
Cluster corresponding sub- dictionary, RiRepresent from initial target high-definition pictureHigh resolution graphics photo x is obtained at the i of positioni's
Matrix；According to sparse coefficient correction formula
WithDetermine the high resolution graphics photo x of the t+1 times iterationiSparse coefficientωiJ-th of component
ωijMeet：Wherein h represents to be used for the control parameter for adjusting the rate of decay, W expression normalizings
Change the factor,Represent high resolution graphics photo xiJ-th of most like image sheet in corresponding most like image sheet group, Sλ/cTable
Show soft-threshold function, c represents regularization parameter；According to formula
Determine x(t+1), wherein x(t+1)Represent the target high-resolution Image estimation value of the t+1 times iteration, φ(t+1)Represent x(t+1)It is corresponding
Super complete dictionary.
Further, further include：Filter module 307；It is final goal high score that the acquisition module 303, which obtains convergence solution,
Resolution image x '；The filter module 307, for using plural impact filtering formula to the final goal high-definition picture
Restored, wherein the plural number impact filtering formula is：
Wherein, x ' expressions final goal high-definition picture, x " represent the final goal high-definition picture after restoring,Represent the gradient map of x', η and ξ represent the gradient direction of image, and Im () represents extraction imaginary part, and a represents to be used to control image
The adjusting parameter of acutance, τ=| τ | exp (i θ) is complex scalar coefficient,It is real number scalar factor.
Super-resolution image reconstruction device provided in this embodiment, can be used for performing the super-resolution figure corresponding to Fig. 2
As the implementation technical solution of method for reconstructing, its implementing principle and technical effect is similar, and details are not described herein again.
One of ordinary skill in the art will appreciate that：Realizing all or part of step of above-mentioned each method embodiment can lead to
The relevant hardware of programmed instruction is crossed to complete.Foregoing program can be stored in a computer read/write memory medium.The journey
Sequence upon execution, execution the step of including above-mentioned each method embodiment；And foregoing storage medium includes：ROM, RAM, magnetic disc or
Person's CD etc. is various can be with the medium of store program codes.
Finally it should be noted that：The above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
The present invention is described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that：It still may be used
To modify to the technical solution described in foregoing embodiments, or equivalent substitution is carried out to which part technical characteristic；
And these modification or replace, do not make appropriate technical solution essence depart from various embodiments of the present invention technical solution spirit and
Scope.
Claims (16)
- A kind of 1. super-resolution image reconstruction method, it is characterised in that including：Step S1, using the low-resolution image y of input as the first high-definition pictureAccording to first high resolution graphics PictureCarry out n times and obscure down-sampling, generation high resolution image data storehouse Dx；Step S2, to first high-definition pictureUsing bicubic interpolation algorithm carry out fuzzy down-sampling and then into The fuzzy up-sampling of row obtains the first low-resolution imageFor first low-resolution imageProgress n times, which obscure down, adopts Sample, generation low resolution image data storehouse Dz；Step S3, using the bicubic interpolation algorithm by first high-definition pictureIt is low that second is obtained after S times of amplification Image in different resolutionBy second low-resolution imageIt is divided at least one low-resolution image piece；Step S4, to each low-resolution image piece using approximate KNN searching algorithm in the low resolution image data Storehouse DzObtain the corresponding most like image sheet group of low resolution；Step S5, weight coefficient is calculated to the most like image sheet group of the low resolution using Ridge Regression Method；Step S6, high resolution graphics photo is determined according to the weight coefficient and the most like image sheet group of high-resolution；Step S7, the second high-definition picture is determined to the high resolution graphics photo using weighted mean methodAnd by institute State the second low-resolution imageIt is added to the database DzIn, by second high-definition pictureIt is added to the number According to storehouse DxIn.
- 2. according to the method described in claim 1, it is characterized in that, further include：Step S8, by second high-definition pictureAs the first new high-definition pictureAfter repeating step S3-S7 by M times, the 3rd high-definition picture is obtainedUsing the bicubic interpolation algorithm to the 3rd high-definition pictureFuzzy down-sampling is carried out, generates initial target High-definition picture
- 3. according to the method described in claim 2, it is characterized in that, further include：Step S9, to the initial target high-definition pictureAlternating minimization algorithm meter is used with the low-resolution image y Calculate mathematical model：s.t.hf=hr,Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition pictureUnder degenerate matrix H represents fuzzy Sample operator, λ1And λ2Represent regularization parameter, αiIt is the high resolution graphics photo x of target high-resolution image xiSparse system Number, α αiSet, βiIt is αiNon-local mean in sparse coding domain,Represent the sparse coefficient matrix α of all splicings and surpass The product of complete dictionary φ,Represent gradient operation symbol, f is transforming function transformation function, hrIt is the reference Nogata of target high-resolution image Figure, hfIt is the histogram of transforming function transformation function f.
- It is 4. according to the method described in claim 3, it is characterized in that, described to the initial target high-definition pictureAnd institute Low-resolution image y is stated to specifically include using alternating minimization algorithm calculating mathematical model：Step S10, super complete dictionary is obtained using K mean cluster algorithm and Principal Component Analysis Algorithm；Step S11, remains unchanged algorithm according to histogram of gradients and updates the transforming function transformation function f；Step S12, updates the initial target high-definition picture according to the super complete dictionary and the transforming function transformation function f.
- 5. according to the method described in claim 4, it is characterized in that, obtained using K mean cluster algorithm and Principal Component Analysis Algorithm Super complete dictionary is taken, is specifically included：According to size randomly select the high resolution graphics photos of K initial target high-definition pictures as K it is a at the beginning of Begin cluster；According to the distance between the high resolution graphics photo and the K initial cluster center, by the initial target high score All high resolution graphics photos that resolution image is divided into are divided into corresponding initial clustering；To each initial clustering using the corresponding sub- dictionary of Principal Component Analysis training based on singular value decomposition；The K corresponding K sub- dictionary composition of the initial clustering super complete dictionaries.
- 6. according to the method described in claim 4, it is characterized in that, remaining unchanged algorithm according to histogram of gradients updates the change Exchange the letters number f, specifically includes：Fuzzy up-sampling is carried out to the low-resolution image y, obtains low-resolution image z；The target high-resolution image x and the low-resolution image z meet following condition：Z=B*x, wherein B are fuzzy Core；Gradient is asked to z=B*x both sides, thenWherein b0And biFuzzy core B is represented respectively Center coefficient and its surrounding's neighbour's coefficient,Represent the gradient image of target high-resolution image x,Represent described low The gradient image of image in different resolution z,Represent the gradient image of the high resolution graphics photo；OrderIfIt is similar to normal distribution, thenPass through solving-optimizing problems.t.hf=hrTo update transforming function transformation function f；Wherein hrRepresent the terraced histogram of reference of target high-resolution image x, hzRepresent that the gradient of the low-resolution image z is straight Fang Tu, hx1It is the discrete form of the probability density function of stochastic variable x1, hx2Be I.i.d. random variables x2 probability it is close The discrete form of function is spent,It is convolution operation symbol.
- It is 7. according to the method described in claim 4, it is characterized in that, described according to the super complete dictionary and the transforming function transformation function F updates the initial target high-definition picture, specifically includes：According to formulaDetermine x(t+1/2), wherein x(t)Represent the The target high-resolution Image estimation value of t iteration, x(t+1/2)Represent the target high-resolution Image estimation of the t+1/2 times iteration Value, δ is constant,Represent x(t)Gradient map；According toDetermineWhereinRepresent the high-definition picture of the t+1/2 times iteration Piece xiSparse coefficient,Represent the high resolution graphics photo x of the t times iterationiThe corresponding sub- dictionary of place initial clustering, RiTable Show from initial target high-definition pictureHigh resolution graphics photo x is obtained at the i of positioniMatrix；According to sparse coefficientWith Determine the high resolution graphics photo x of the t+1 times iterationiSparse coefficientωiJ-th of component ωijMeet：Wherein h represents the control parameter for adjusting the rate of decay, and W represents normalization factor, Represent high resolution graphics photo xiJ-th of most like image sheet in corresponding most like image sheet group, Sλ/cRepresent soft-threshold letter Number, c represent regularization parameter；According to formulaDetermine x(t+1), wherein x(t+1)Represent t+1 The target high-resolution Image estimation value of secondary iteration, φ(t+1)Represent x(t+1)Corresponding super complete dictionary.
- 8. according to claim 4-7 any one of them methods, it is characterised in that further include：Step S10-S12 is repeated, acquisition convergence solution is final goal high-definition picture x'；The final goal high-definition picture is restored using plural impact filtering formula, wherein the plural number impact filter Ripple formula is：<mrow> <msup> <mi>x</mi> <mrow> <mo>&prime;</mo> <mo>&prime;</mo> </mrow> </msup> <mo>=</mo> <mo>-</mo> <mfrac> <mn>2</mn> <mi>&pi;</mi> </mfrac> <mi>a</mi> <mi>r</mi> <mi>c</mi> <mi>t</mi> <mi>a</mi> <mi>n</mi> <mrow> <mo>(</mo> <mi>a</mi> <mi> </mi> <mi>Im</mi> <mo>(</mo> <mfrac> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>&theta;</mi> </mfrac> <mo>)</mo> <mo>)</mo> </mrow> <mo>|</mo> <mrow> <mo>&dtri;</mo> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> </mrow> <mo>|</mo> <mo>+</mo> <msubsup> <mi>&tau;x</mi> <mrow> <mi>&eta;</mi> <mi>&eta;</mi> </mrow> <mo>&prime;</mo> </msubsup> <mo>+</mo> <mover> <mi>&tau;</mi> <mo>&OverBar;</mo> </mover> <msubsup> <mi>x</mi> <mrow> <mi>&xi;</mi> <mi>&xi;</mi> </mrow> <mo>&prime;</mo> </msubsup> </mrow>Wherein, x ' expressions final goal high-definition picture, x " represent the final goal high-definition picture after restoring,Table Show the gradient map of x', η and ξ represent the gradient direction of image, and Im () represents extraction imaginary part, and a represents to be used to control image sharpness Adjusting parameter, τ=| τ | exp (i θ) is complex scalar coefficient,It is real number scalar factor.
- A kind of 9. super-resolution image reconstruction device, it is characterised in that including：Generation module, for using the low-resolution image y of input as the first high-definition pictureAccording to first high score Resolution imageCarry out n times and obscure down-sampling, generation high resolution image data storehouse Dx；The generation module, is additionally operable to first high-definition pictureUsing bicubic interpolation algorithm adopt under obscuring Sample and then carry out fuzzy up-sampling and obtain the first low-resolution imageFor first low-resolution imageCarry out N times obscure down-sampling, generation low resolution image data storehouse Dz；Division module, for using the bicubic interpolation algorithm by first high-definition pictureObtained after S times of amplification Second low-resolution imageBy second low-resolution imageIt is divided at least one low-resolution image piece；Acquisition module, for using approximate KNN searching algorithm in the low resolution figure to each low-resolution image piece As database DzObtain the corresponding most like image sheet group of low resolution；Computing module, for calculating weight coefficient using Ridge Regression Method to the most like image sheet group of the low resolution；Determining module, for determining high resolution graphics photo according to the weight coefficient and the most like image sheet group of high-resolution；The determining module, is additionally operable to determine the second high resolution graphics using weighted mean method to the high resolution graphics photo PictureAnd by second low-resolution imageIt is added to the database DzIn, by second high-definition picture It is added to the database DxIn.
- 10. device according to claim 9, it is characterised in that further include：The acquisition module, is additionally operable to obtain the 3rd high-definition pictureThe generation module, is also used using the bicubic interpolation algorithm to the 3rd high-definition pictureCarry out under obscuring Sampling, generates initial target high-definition picture
- 11. device according to claim 10, it is characterised in that further include：The computing module, is additionally operable to the initial target high-definition pictureWith the low-resolution image y using alternating Minimize algorithm and calculate mathematical model：s.t.hf=hr,Wherein, the initial value of the x in above-mentioned mathematical model is initial target high-definition pictureUnder degenerate matrix H represents fuzzy Sample operator, λ1And λ2Represent regularization parameter, αiIt is the high resolution graphics photo x of target high-resolution image xiSparse system Number, α αiSet, βiIt is αiNon-local mean in sparse coding domain,Represent the sparse coefficient matrix α of all splicings and surpass The product of complete dictionary φ,Represent gradient operation symbol, f is transforming function transformation function, hrIt is the reference Nogata of target high-resolution image Figure, hfIt is the histogram of transforming function transformation function f.
- 12. according to the devices described in claim 11, it is characterised in that further include：Update module；The acquisition module, is additionally operable to obtain super complete dictionary using K mean cluster algorithm and Principal Component Analysis Algorithm；The update module, updates the transforming function transformation function f for remaining unchanged algorithm according to histogram of gradients；The update module, updates the initial target high-resolution always according to the super complete dictionary and the transforming function transformation function f Image.
- 13. according to the devices described in claim 11, it is characterised in that the acquisition module is specifically used for：According to size randomly select the high resolution graphics photos of K target high-resolution images as K it is a at the beginning of Begin cluster；According to the distance between the high resolution graphics photo and the K initial cluster center, by the target high-resolution All high resolution graphics photos that image is divided into are divided into corresponding initial clustering；To each initial clustering using the corresponding sub- dictionary of Principal Component Analysis training based on singular value decomposition；The K corresponding K sub- dictionary composition of the initial clustering super complete dictionaries.
- 14. device according to claim 12, it is characterised in that the update module is specifically used for：Fuzzy up-sampling is carried out to the low-resolution image y, obtains low-resolution image z；The target high-resolution image x and the low-resolution image z meet following condition：Z=B*x, wherein B are fuzzy Core；Gradient is asked to z=B*x both sides, thenWherein b0And biFuzzy core B is represented respectively Center coefficient and its surrounding's neighbour's coefficient,Represent the gradient image of target high-resolution image x,Represent described low The gradient image of image in different resolution z,Represent the gradient image of the high resolution graphics photo；OrderIfIt is similar to normal distribution, thenPass through solving-optimizing problems.t.hf=hrTo update transforming function transformation function f；Wherein hrRepresent the terraced histogram of reference of target high-resolution image x, hzRepresent that the gradient of the low-resolution image z is straight Fang Tu, hx1It is the discrete form of the probability density function of stochastic variable x1, hx2Be I.i.d. random variables x2 probability it is close The discrete form of function is spent,It is convolution operation symbol.
- 15. device according to claim 12, it is characterised in that the update module is specifically used for：According to formulaDetermine x(t+1/2), wherein x(t)Represent The target high-resolution Image estimation value of the t times iteration, x(t+1/2)Represent that the target high-resolution image of the t+1/2 times iteration is estimated Evaluation, δ are constant,Represent x(t)Gradient map；According toDetermineWhereinRepresent the high-definition picture of the t+1/2 times iteration Piece xiSparse coefficient,Represent the high resolution graphics photo x of the t times iterationiThe corresponding sub- dictionary of place initial clustering, RiTable Show from initial target high-definition pictureHigh resolution graphics photo x is obtained at the i of positioniMatrix；According to sparse coefficientWith Determine the high resolution graphics photo x of the t+1 times iterationiSparse coefficientωiJ-th of component ωijMeet：Wherein h represents the control parameter for adjusting the rate of decay, and W represents normalization factor, Represent high resolution graphics photo xiJ-th of most like image sheet in corresponding most like image sheet group, Sλ/cRepresent soft-threshold letter Number, c represent regularization parameter；According to formulaDetermine x(t+1), wherein x(t+1)Represent t+1 The target high-resolution Image estimation value of secondary iteration, φ(t+1)Represent x(t+1)Corresponding super complete dictionary.
- 16. according to claim 9-15 any one of them devices, it is characterised in that further include：Filter module；It is final goal high-definition picture x' that the acquisition module, which obtains convergence solution,；The filter module, for being restored using plural impact filtering formula to the final goal high-definition picture, Wherein described plural impact filtering formula is：<mrow> <msup> <mi>x</mi> <mrow> <mo>&prime;</mo> <mo>&prime;</mo> </mrow> </msup> <mo>=</mo> <mo>-</mo> <mfrac> <mn>2</mn> <mi>&pi;</mi> </mfrac> <mi>a</mi> <mi>r</mi> <mi>c</mi> <mi>t</mi> <mi>a</mi> <mi>n</mi> <mrow> <mo>(</mo> <mi>a</mi> <mi> </mi> <mi>Im</mi> <mo>(</mo> <mfrac> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>&theta;</mi> </mfrac> <mo>)</mo> <mo>)</mo> </mrow> <mo>|</mo> <mrow> <mo>&dtri;</mo> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> </mrow> <mo>|</mo> <mo>+</mo> <msubsup> <mi>&tau;x</mi> <mrow> <mi>&eta;</mi> <mi>&eta;</mi> </mrow> <mo>&prime;</mo> </msubsup> <mo>+</mo> <mover> <mi>&tau;</mi> <mo>&OverBar;</mo> </mover> <msubsup> <mi>x</mi> <mrow> <mi>&xi;</mi> <mi>&xi;</mi> </mrow> <mo>&prime;</mo> </msubsup> </mrow>Wherein, x ' expressions final goal high-definition picture, x " represent the final goal high-definition picture after restoring,Table Show the gradient map of x', η and ξ represent the gradient direction of image, and Im () represents extraction imaginary part, and a represents to be used to control image sharpness Adjusting parameter, τ=| τ | exp (i θ) is complex scalar coefficient,It is real number scalar factor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410532272.6A CN105488776B (en) | 2014-10-10 | 2014-10-10 | Super-resolution image reconstruction method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410532272.6A CN105488776B (en) | 2014-10-10 | 2014-10-10 | Super-resolution image reconstruction method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105488776A CN105488776A (en) | 2016-04-13 |
| CN105488776B true CN105488776B (en) | 2018-05-08 |
Family
ID=55675742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410532272.6A Expired - Fee Related CN105488776B (en) | 2014-10-10 | 2014-10-10 | Super-resolution image reconstruction method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105488776B (en) |
Families Citing this family (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106204445A (en) * | 2016-06-30 | 2016-12-07 | 北京大学 | Image/video super-resolution method based on structure tensor total variation |
| CN106339984B (en) * | 2016-08-27 | 2019-09-13 | 中国石油大学(华东) | Distributed image ultra-resolution method based on K mean value driving convolutional neural networks |
| CN106851321A (en) * | 2017-01-15 | 2017-06-13 | 四川精目科技有限公司 | A kind of least square regression high speed camera compresses image rebuilding method |
| CN107239781B (en) * | 2017-05-03 | 2020-07-28 | 北京理工大学 | Hyperspectral reflectivity reconstruction method based on RGB image |
| CN107680037B (en) * | 2017-09-12 | 2020-09-29 | 河南大学 | Improved face super-resolution reconstruction method based on nearest characteristic line manifold learning |
| CN108416734A (en) * | 2018-02-08 | 2018-08-17 | 西北大学 | Text image super resolution ratio reconstruction method and device based on edge driving |
| CN109978785B (en) * | 2019-03-22 | 2020-11-13 | 中南民族大学 | Image super-resolution reconstruction system and method based on multi-level recursive feature fusion |
| CN110148091A (en) * | 2019-04-10 | 2019-08-20 | 深圳市未来媒体技术研究院 | Neural network model and image super-resolution method based on non local attention mechanism |
| CN111986078B (en) * | 2019-05-21 | 2023-02-10 | 四川大学 | Multi-scale core CT image fusion reconstruction method based on guide data |
| CN110363235B (en) * | 2019-06-29 | 2021-08-06 | 苏州浪潮智能科技有限公司 | High-resolution image matching method and system |
| WO2022036556A1 (en) * | 2020-08-18 | 2022-02-24 | 香港中文大学（深圳） | Image processing method and apparatus, computer device, and storage medium |
| CN112767427A (en) * | 2021-01-19 | 2021-05-07 | 西安邮电大学 | Low-resolution image recognition algorithm for compensating edge information |
| CN112950476A (en) * | 2021-03-12 | 2021-06-11 | 广州冠图视觉科技有限公司 | Method for improving resolution and definition of picture |
| CN115829842B (en) * | 2023-01-05 | 2023-04-25 | 武汉图科智能科技有限公司 | Device for realizing super-resolution reconstruction of picture based on FPGA |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102136144A (en) * | 2011-04-11 | 2011-07-27 | 北京大学 | Image registration reliability model and reconstruction method of super-resolution image |
| CN102354394A (en) * | 2011-09-22 | 2012-02-15 | 中国科学院深圳先进技术研究院 | Image super-resolution method and system |
| CN102968766A (en) * | 2012-11-23 | 2013-03-13 | 上海交通大学 | Dictionary database-based adaptive image super-resolution reconstruction method |
| CN103871041A (en) * | 2014-03-21 | 2014-06-18 | 上海交通大学 | Image super-resolution reconstruction method based on cognitive regularization parameters |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2009087641A2 (en) * | 2008-01-10 | 2009-07-16 | Ramot At Tel-Aviv University Ltd. | System and method for real-time super-resolution |
-
2014
- 2014-10-10 CN CN201410532272.6A patent/CN105488776B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102136144A (en) * | 2011-04-11 | 2011-07-27 | 北京大学 | Image registration reliability model and reconstruction method of super-resolution image |
| CN102354394A (en) * | 2011-09-22 | 2012-02-15 | 中国科学院深圳先进技术研究院 | Image super-resolution method and system |
| CN102968766A (en) * | 2012-11-23 | 2013-03-13 | 上海交通大学 | Dictionary database-based adaptive image super-resolution reconstruction method |
| CN103871041A (en) * | 2014-03-21 | 2014-06-18 | 上海交通大学 | Image super-resolution reconstruction method based on cognitive regularization parameters |
Non-Patent Citations (3)
| Title |
|---|
| Anchored Neighborhood Regression for Fast Example-Based Super-Resolution;Radu Timofte 等;《2013 IEEE International Conference on Computer Vision(ICCV)》;20140303;1920-1927 * |
| Image Super-Resolution Via Sparse Representation;Jianchao Yang 等;《IEEE TRANSACTIONS ON IMAGE PROCESSING》;20101130;第19卷(第11期);2861-2873 * |
| 基于显著性稀疏表示的图像超分辨率算法;白蔚 等;《中国科技论文》;20140131;第9卷(第1期);103-107 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105488776A (en) | 2016-04-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105488776B (en) | Super-resolution image reconstruction method and device | |
| Xiao et al. | Satellite video super-resolution via multiscale deformable convolution alignment and temporal grouping projection | |
| Yang et al. | DRFN: Deep recurrent fusion network for single-image super-resolution with large factors | |
| CN101950365B (en) | Multi-task super-resolution image reconstruction method based on KSVD dictionary learning | |
| Sun et al. | Lightweight image super-resolution via weighted multi-scale residual network | |
| CN109146788A (en) | Super-resolution image reconstruction method and device based on deep learning | |
| CN105590304B (en) | Super-resolution image reconstruction method and device | |
| CN111369487A (en) | Hyperspectral and multispectral image fusion method, system and medium | |
| CN104599242A (en) | Multi-scale non-local regularization blurring kernel estimation method | |
| CN106408550A (en) | Improved self-adaptive multi-dictionary learning image super-resolution reconstruction method | |
| Li et al. | A lightweight multi-scale channel attention network for image super-resolution | |
| CN111861886B (en) | Image super-resolution reconstruction method based on multi-scale feedback network | |
| CN114418853B (en) | Image super-resolution optimization method, medium and equipment based on similar image retrieval | |
| CN111951167A (en) | Super-resolution image reconstruction method, super-resolution image reconstruction device, computer equipment and storage medium | |
| CN113421187B (en) | Super-resolution reconstruction method, system, storage medium and equipment | |
| Xu et al. | AutoSegNet: An automated neural network for image segmentation | |
| CN113469884A (en) | Video super-resolution method, system, equipment and storage medium based on data simulation | |
| CN113888491A (en) | Multilevel hyperspectral image progressive and hyper-resolution method and system based on non-local features | |
| Xia et al. | Meta-learning-based degradation representation for blind super-resolution | |
| Shen et al. | RSHAN: Image super-resolution network based on residual separation hybrid attention module | |
| CN113723472A (en) | Image classification method based on dynamic filtering equal-variation convolution network model | |
| CN116797456A (en) | Image super-resolution reconstruction method, system, device and storage medium | |
| CN116188272B (en) | Two-stage depth network image super-resolution reconstruction method suitable for multiple fuzzy cores | |
| CN111311732A (en) | 3D human body grid obtaining method and device | |
| CN114155560B (en) | Light weight method of high-resolution human body posture estimation model based on space dimension reduction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20220621 Address after: 100871 No. 5, the Summer Palace Road, Beijing, Haidian District Patentee after: Peking University Patentee after: New founder holdings development Co.,Ltd. Patentee after: BEIJING FOUNDER ELECTRONICS Co.,Ltd. Address before: 100871 No. 5, the Summer Palace Road, Beijing, Haidian District Patentee before: Peking University Patentee before: PEKING UNIVERSITY FOUNDER GROUP Co.,Ltd. Patentee before: BEIJING FOUNDER ELECTRONICS Co.,Ltd. |
|
| TR01 | Transfer of patent right | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180508 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |