CN106408550A - Improved self-adaptive multi-dictionary learning image super-resolution reconstruction method - Google Patents
Improved self-adaptive multi-dictionary learning image super-resolution reconstruction method Download PDFInfo
- Publication number
- CN106408550A CN106408550A CN201610847106.4A CN201610847106A CN106408550A CN 106408550 A CN106408550 A CN 106408550A CN 201610847106 A CN201610847106 A CN 201610847106A CN 106408550 A CN106408550 A CN 106408550A
- Authority
- CN
- China
- Prior art keywords
- image
- dictionary
- matrix
- self
- similarity
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 51
- 230000013016 learning Effects 0.000 title claims abstract description 21
- 239000011159 matrix material Substances 0.000 claims abstract description 23
- 230000006978 adaptation Effects 0.000 claims description 13
- 230000003044 adaptive effect Effects 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 5
- 238000000513 principal component analysis Methods 0.000 claims description 4
- 238000006731 degradation reaction Methods 0.000 abstract 1
- 238000004422 calculation algorithm Methods 0.000 description 56
- 230000000694 effects Effects 0.000 description 17
- 238000012549 training Methods 0.000 description 11
- 230000000007 visual effect Effects 0.000 description 6
- 238000012545 processing Methods 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 3
- 238000000354 decomposition reaction Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 230000006866 deterioration Effects 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000013341 scale-up Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20004—Adaptive image processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20024—Filtering details
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an improved self-adaptive multi-dictionary learning image super-resolution reconstruction method, which comprises the steps of: (1) determining a downsampling matrix D and a fuzzy matrix B according to a quality degradation process of an image; (2) establishing a pyramid by utilizing the self-similarity of the image, regarding an upper-layer image and a natural image of the pyramid as samples of dictionary learning, constructing various types of dictionaries Phi k<i> by adopting a PCA method, and regarding a top-layer image of the pyramid as an initial reconstructed image X<^>; (3) calculating a weight matrix A of nonlocal structural self-similarity of sparse coding; (4) setting an iteration termination error e, a maximum number iteration times Max_Iter, a constant eta controlling nonlocal regularization term contribution amount and a condition P for updating parameters; (5) updating current estimation of the image; (6) updating a sparse representation coefficient; (7) updating current estimation of the image; (8) updating a self-adaptive sparse domain of X if mod(k, P)=0, and using X<^><k+1> for updating the matrix A; (9) and repeating the steps from (5) to (8), and terminating iteration until the iteration meets a condition shown in the description or k>=Max_Iter.
Description
Technical field
The invention belongs to technical field of image processing, it is related to a kind of Image Super-resolution of the many dictionary learnings of improved self adaptation
Rate method for reconstructing, can be used for diagnosis, analysis and the further process of image.
Background technology
High-definition picture comprises a lot of detailed information, and these information are led in medical image, video monitoring, remotely sensed image etc.
Domain is most important, and it is conducive to diagnosis, analysis and the further process of image, so obtaining high-resolution have important meaning
Justice.To obtain high-definition picture, there are two kinds of approach：One is to set about transforming imaging system from hardware aspect, and two is from software
Aspect improves the resolution ratio of image by algorithm.But, due to existing chip and the aspect such as sensor manufacturing process, system cost
Restriction, from hardware aspect improve image resolution ratio highly difficult, so setting about becoming raising image resolution ratio from software aspects
Important means.Image super-resolution (Super-resolution, SR) method be exactly using image processing techniques from a width or
A panel height resolution ratio (High- is recovered in low resolution (Low-resolution, the LR) image of several Same Scene
Resolution, HR) image technology, it by low resolution imaging device obtain image process regard as by HR image deterioration
Obtain, the key of this problem solving be how to make up degrade during lose high-frequency information[1].
Classical SR method for reconstructing[2]It is to be carried using the Displacement existing between the multiple image under Same Scene
For complementary information to rebuild HR image.It is good that this method rebuilds effect, but it by registration accuracy affected larger, operate as
This height, and some fields are difficult to obtain Same Scene, several LR images under close phase.So the SR based on single image
Method for reconstructing gradually grows up, and it mainly uses priori or sample learning to obtain high-frequency information.Among these
It is simply based on interpolation[3]Method, such as bilinear interpolation, bi-cubic interpolation etc., this method calculates simple, and operand is low,
But after rebuilding, image border is fuzzyyer, and there is chessboard and ringing effect.Second is the method based on rebuilding, and it is mainly
Obtain the high-frequency information of loss by priori, such as projections onto convex sets[4], iterative backprojection method[5], regularization method[6]
Deng, this method alleviates the shortcoming of interpolation method to a certain extent, but when extraction yield is larger, rebuild effect can drastically under
Fall.The third is based on the method that the SR method for reconstructing of single image is based on study[7,8], by learning high-resolution and low-resolution image
Between statistical relationship, and this relation is applied to realize in process of reconstruction image SR rebuild, compared to first two rebuild
Method, is obtained in that more detailed information based on the method for study, therefore rebuilds effect more preferable.
Rarefaction representation is to represent image with atom as few as possible in given super complete dictionary, and image is through sparse
More succinct representation can be obtained, thus being easier to obtain the information being contained in image, and image is got over after expression
The sparse precision of images finally reconstructing is higher.In recent years, sparse representation model was in the super-resolution rebuilding algorithm based on study
In achieve preferable effect.Yang etc.[9]Propose a kind of super-resolution algorithms based on rarefaction representation, this algorithm is from high and low
Image block is chosen respectively as dictionary training sample, using FSS (Feature-Sign Search) algorithm instruction in image in different resolution
Practise a pair of high-resolution and low-resolution dictionary, then obtain the sparse coefficient of pending LR image using low-resolution dictionary, then profit
Reconstruct HR image with this sparse coefficient and high-resolution dictionary.Zeyde etc.[10]Propose a kind of new on the basis of Yang
The super-resolution algorithms based on rarefaction representation, he proposes to be replaced with K-SVD algorithm the FSS algorithm of Yang in dictionary training,
Both accelerate dictionary training speed and also improve reconstruction effect.Both algorithms be all had under dictionary using image openness
This property, this openness regularization term of regarding as to be constrained uniqueness of solution, thus the SR realizing image rebuilds.Additionally,
The substantial amounts of repeat patterns of generally existing and structure in natural image, can also be carried using this local and non-local self-similarity
High reconstruction effect.Such as Glasner etc.[11]Propose to produce pyramid using image self-similarity, then in pyramidal difference
Layer search similar image block rebuilds high-definition picture.Dong etc.[12]First the image block in image library is clustered, then
Multiple dictionaries are trained to every class sample PCA, process of reconstruction introduces autoregression model and non local structure
Self-similarity, as regularization term, achieves preferable reconstruction effect.In above method, the method for Yang and Zeyde is only all
To train dictionary merely with image library, this Global Dictionary to be referred to as by the dictionary that image library is trained, because image library contains
Abundant high-frequency information, so Global Dictionary can obtain the additional information of abundance, but cannot ensure the standard of these additional informations
Really property and reliability, therefore often have preferable visual effect, but there is larger mean square error.Moreover it is many due to image
, actually there is not the Global Dictionary being capable of all image blocks of rarefaction representation in sample.The method that Glasner proposes is using input
LR image itself self-similarity reconstruction image, this method takes full advantage of the priori of images themselves, ensure that
Obtain accuracy and the reliability of additional information, but the high-frequency information that images themselves are contained is limited, so rebuilding effect
There is also certain limitation.Although the method for Dong make use of the self-similarity information of images themselves, increased acquisition information
Accuracy and reliability, but its remain using image library train Global Dictionary, can not all of image block of rarefaction representation.
Hereafter in order to improve reconstruction effect further, Pan Zongxu etc.[13]Images themselves are combined with image library and proposes one kind and be based on
Self adaptation many dictionaries learning method SR algorithm for reconstructing, LR image is carried out pyramid decomposition by him, and the image block in pyramid is entered
Image block in image library is classified under the guiding of cluster result by row cluster, multiple for the sample structure in all kinds of
Dictionary, so each image block can be suitable for the dictionary of oneself according to the adaptive selection of own characteristic and realize SR reconstruction.
, when training dictionary, the pyramid upper layer images being generated with the self-similarity of pending image are as itself for the present invention
Sample, obtains the image information under more different scales with this；Additionally, when rebuilding using pyramidal top layer as first starting weight
Build image, the non local structure self-similarity of sparse coding to be realized the Super-resolution reconstruction of image as regularization constraint item
Build, so also can be using the analog information under image same scale come reconstruction image.Test result indicate that, rebuild in the present invention
HR image either has preferable effect in subjective assessment or objective evaluation.
The present invention can more fully utilize the analog information of image same scale and different scale, makes the image effect of reconstruction
More preferably, more accurate.
Bibliography：
[1] Pan Zongxu, Yu Jing, Hu Shaoxing, Sun Weidong. the single image super-resolution based on Multi-scale model self-similarity
Algorithm [J]. automation journal, 2014,40 (4)：594-603.Kim S P, Bose N K, and Valenzuela H
M.Recursive reconstruction of high resolution image from noisy undersampled
Multiframes [J] .IEEE Transactions on Acoustics, Speech and Signal
Processing.1990,38 (6)：1013-1027.
[2]Rajan D and Chaudhuri S.Generalized interpolation and its
Application in super-resolution imaging [J] .Image and Vision Computing, 2001,19
(13)：957-969.
[3]Stark H and Oskoui P.High-resolution image recovery from image-
Plane arrays, using convex projections [J] .Journal of the Optical Society of
America A(Optics and Image Science).
[4]Irani M and Peleg S.Motion analysis for image enhancement：
Resolution, occlusion, and transparency [J] .Journal of Visual Communication and
Image Representation, 1993,4 (4)：324-335.
[5] Han Yubing, Wu Lenan, Zhang Dongqing. the super-resolution rebuilding [J] based on Regularization. electronics and informatics
Report, 2007,29 (7)：1713-1716.doi：10.3724/SP.J.1146.2005.01631.
[6] Chang H, Yeung D, and Xiong Y.Super-resolution through neighbor
embedding[C].Proceedings of the 2004 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, 2004,1：1063-6919.
[7] Freeman W T, Jones T R, and Pasztor E C.Example-based super-
Resolution [J] .IEEE Computer Graphics and Applications, 2002,22 (2)：56-65.
[8] Yang J, Wright J, Huang T S, and Ma Y.Image super-resolution via
Sparse representation [J] .IEEE Transaction on Image Processing, 2010,19 (11)：
2861-2873.
[9] Zeyde R, Elad M, and Protter M.On single image scale-up using
sparse-representation[C].In Proc.of Internation Conferenceon Curves and
Surfaces, 2010,6920：711-730.
[10] Glasner D, Bagon S, and Irani M.Super-resolution from a single
image[C].Proceedings of IEEE Conference on Computer Vision and Pattern
Recognition, 2009,30 (2)：349-356.
[11] Dong W S, Zhang L, Shi G M, and Wu X L.Image deblu rring and super-
resolution by adaptive sparse domain selection and adaptive regularization
[J] .IEEE Transactions on Image Processing, 2011,20 (7)：1838-1857.
[12] Pan Zongxu, Yu Jing, Xiao Chuanbai, Sun Weidong. the single image super-resolution based on the many dictionary learnings of self adaptation
Algorithm [J]. electronic letters, vol, 2015,43 (2)：209-216.doi：10.3969/j.issn.0372-2112.2015.02.001.
[13] Candes E, Romberg J, and Tao T.Robust uncertainty principles：Exact
signal reconstruction from highly incomplete frequency information[J].IEEE
Transactions on Information Theory, 2006,52 (2)：489-509.
[14] Dong W S, Zhang L, Lukac R, and Shi G M.Sparse representation based
image interpolation with nonlocal autoregressive modeling[J].IEEE
Transactions on Image Processing, 2013,22 (4)：1382-1394.
[15] Daubechies I, Defrise M, and De Mol C.An iterative thresholding
algorithm for linear inverse problems with a sparsity constraint[J]
.Communications on Pure and Applied Mathematics, 2004,57 (11)：1413-1457.
[16]Freeman G and Fattal G.Image and Video Upscaling from Local Self-
Examples [J] .ACM Transactions on Graphics, 2011,30 (2)：1-11.
[17] Wang Z, Bovik A C, and Sheikh H R, et al.Image quality assessment：
from error visibility to structural similarity[J].IEEE Transactions on Image
Processing, 2004,13 (4)：600-612.
[18]http：//www.ifp.illinois.edu/～jyang29/
[19]http：//www.cs.technion.ac.il/～elad/software/.
[20]https：//eng.ucmerced.edu/people/cyang35
[21]http：//see.xidian.edu.cn/faculty/wsdong/wsdong_Publication.htm
Content of the invention
In order to improve image super-resolution rebuilding effect, the comprehensive utilization pending image information of itself and natural image storehouse
The high-frequency information providing, proposes a kind of image super-resolution rebuilding method of the many dictionary learnings of improved self adaptation.In training word
During allusion quotation, replaced existing as itself sample with the pyramid upper layer images that the self-similarity of pending low-resolution image generates
The sample being obtained using pyramid decomposition in the many dictionary learnings of self adaptation, and during rebuilding, pyramidal top layer is made
For initial reconstructed image, the non local structure self-similarity of sparse coding to be realized the oversubscription of image as regularization constraint item
Resolution is rebuild.The present invention can more fully utilize the analog information of image same scale and different scale, so that the image of reconstruction is imitated
Fruit is more preferable, more accurate.Test result indicate that, with bicubic interpolation method, Yang algorithm, Zeyde algorithm, Glasner algorithm and
Dong algorithm is compared, the present invention either in visual effect or in Y-PSNR and structure self-similarity quantitative target all
It is significantly increased.Realize the object of the invention technical scheme, comprise the following steps：
Step 1：Down-sampled matrix D and fuzzy matrix B are determined according to the process that degrades of image；
Step 2：Set up pyramid using image self-similarity, using pyramidal upper layer images and natural image as dictionary
The sample of study, builds all kinds of dictionaries with PCA methodAnd using pyramidal top layer images as initial reconstructed image
Step 3：Calculate the weight matrix A of sparse coding non local structure self similarity；
Step 4：Setting iteration ends error e, maximum iteration time Max_Iter, the non local regularization term contribution amount of control
Constant η and undated parameter condition P；
Step 5：The current estimation of more new images：
Step 6：Update rarefaction representation coefficient：
Wherein, soft (, τI, j) it is threshold tauI, jSoft-threshold function；
Soft (x, τ)=sgn (x) max (| x |-τ, 0)
Wherein, sgn (x) is sign function；
Step 7：The current estimation of new images
Step 8：If mod (k, P)=0, then update the adaptive sparse domain of X, useUpdate matrix A；
Step 9：Repeat (5)～(8), until iteration meetsOr k >=Max_Iter iteration ends.
Compared with prior art, the invention has the beneficial effects as follows：
1. the present invention is the improvement to the algorithm that Pan's ancestor's sequence proposes, when training dictionary, with the self similarity of pending image
Property the pyramid upper layer images that generate replace Pan's ancestor's sequence to propose as itself sample the many dictionary learnings of self adaptation in using golden word
The sample that tower decomposition obtains, obtains the image information under more different scales with this；Additionally, when rebuilding by pyramidal top
Layer, as initial reconstructed image, the non local structure self-similarity of sparse coding to be realized image as regularization constraint item
Super-resolution rebuilding, so also can be using the analog information under image same scale come reconstruction image.
2. reconstructed results of the present invention are more accurate, and image border is more sharp, and details is more rich, compared with other algorithms, this
Invent the result visual effect reconstructing preferably, recover more details, image becomes apparent from.
Brief description
Fig. 1：The flow chart of the present invention；
Fig. 2：A () Butterfly original image, (b) is Butterfly low-resolution image, (c) bicubic interpolation algorithm
Reconstructed results figure, the reconstructed results figure of (d) Yang algorithm, the reconstructed results figure of (e) Zeyde algorithm, (f) Glasner algorithm
Reconstructed results figure, the reconstructed results figure of (g) Dong algorithm, (h) be the present invention result figure；
Fig. 3：(a) hat original image, (b) hat low-resolution image, the reconstructed results figure of (c) bicubic interpolation algorithm,
The reconstructed results figure of (d) Yang algorithm, the reconstructed results figure of (e) Zeyde algorithm, the reconstructed results figure of (f) Glasner algorithm,
G the reconstructed results figure of () Dong algorithm, (h) is the result figure of the present invention.
Specific embodiment
With reference to specific embodiment, the present invention is described in further detail.
The present invention achieves the image oversubscription based on the many dictionary learnings of self adaptation and structure self-similarity by following steps
Resolution method for reconstructing, comprises the following steps that：
Step 1：Down-sampled matrix D and fuzzy matrix B are determined according to the process that degrades of image；
Step 2：Set up pyramid using image self-similarity, using pyramidal upper layer images and natural image as dictionary
The sample of study, builds all kinds of dictionaries with PCA methodAnd using pyramidal top layer images as initial reconstructed imageFor
Obtain the high-frequency information lost, need to obtain the dictionary comprising high-frequency information.The sample that the present invention is used for dictionary training is to treat
The image itself processing and natural image storehouse, wherein natural image storehouse are mainly chooses the high resolution graphics that grain details are enriched
Picture, utilizes pending image self-similarity to generate image pyramid simultaneously, chooses pyramidal upper layer images as dictionary training
When for providing the training sample of image self information.The principle generating image pyramid using self-similarity is as follows：
Low-resolution image will be inputted as intermediate layer Iin, then it is reduced z times step by step, obtains pyramidal lower floor figure
As Iin-1, Iin-2..., Iin-k(k is the series reducing).For IinIn image block M it is assumed that can be in Iin-1And Iin-2In search
Rope is to the similar image block P of M1And P2, then can determine P1And P2In IinIn respective regions Q1And Q2, then by Q1And Q2As
Image block M is in Iin+1And Iin+2In respective regions D1And D2Estimation, according to this principle block-by-block calculate, thus trying to achieve golden word
The upper layer images I of towerin+1, Iin+2..., Iin+j(j is the series rebuild).If the similar image block of same image block have multiple,
The image block that then can be rebuild by weighted calculation, i.e. Iin+jIn image block DiIt is by Iin+j-1Middle MiSimilar image block P1~
PmCorresponding Q1～QmWeighting obtains, and formula is as follows：
Di=∑ ωmQm
Wherein, weights
ωm=exp (- | | Mi-Pm||2/σ2)
In formula, σ is the weights controlling similarity.By all of DiSuperposition has just obtained pyramidal upper layer images Iin+j.So
And not all of image block can find its similar image block, so the present invention utilizes backprojection algorithm in pyramid[5]
To fill up these regions, to improve the resolution ratio of image.The training principle of the many dictionaries of self adaptation is as follows：
First, pending image is processed using pyramid upper layer images obtained above as dictionary training sample
S is it is contemplated that the vision system of people is only sensitive to high-frequency information, so the present invention carries out high-pass filtering to these images extracts them
High-frequency information, then they are cut intoThe image block of size, in order to ensure to exclude smooth region, only retains edge
Structure, only selects image block standard deviation to be more than the image block of Δ here.Note si, i=1,2 ..., m is the sample image of selection
Block, gi, i=1,2 ..., m is the radio-frequency component after its corresponding high-pass filtering, by K-means algorithm by giIt is polymerized to K class,mkFor the sample number that comprises in every class it is ensured thatCalculate in the class of every class
The heartWith class radius
Secondly, to the image in natural image storehouseEqually carry out high-pass filtering, be then cut intoThe block of size, choosing
Select the block that image block standard deviation is more than Δ, noteFor select sample image block,For its corresponding height
Radio-frequency component after pass filter.CalculateWith cluster all kinds of barycenter μ obtaining abovekDistance, be designated asIfThen willIt is added to Ck, otherwise give up this image block, wherein, δ is used to control the phase of image block and class center
Parameter like degree.After aforesaid operations, sample is expanded, and the sample after every class expands is designated as
qkFor number of samples after expanding.
Then, according to radio-frequency component sample CkFind corresponding image block sample matrixK=1,
2 ..., K, dictionary learning is carried out according to following formula, obtains K sub- dictionary：
Wherein, ΛkFor sparse coefficient matrix, λ is the parameter controlling sparse degree.In order to reduce amount of calculation, the present invention adopts
Carry out dictionary learning with PCA (principal component analysis, PCA).If SkCovariance matrix
For Ωk, to ΩkCarry out principal component analysis, obtain an orthogonal transform matrix PkIt is assumed that PkFor dictionary, then sparse coefficientIt is desirable thatIn order to preferably balance sparse degree and data fidelity item it
Between relation, the present invention only selects PkMiddle r most important characteristic vector forms dictionary Φr=[p1, p2..., pr], thenWhen r reduces,In reconstruction errorWill increase
Plus, | | Λr||1Value will reduce.Therefore, for each class in K class, each find an optimum r value, be designated as r0, use
Carry out Equilibrium fitting errorWith | | Λr||1Between relation, formula is as follows：
Finally try to achieve each category dictionary
Finally, in ΦkThe corresponding self-adapting dictionary of image block to be reconstructed is found in (k=1,2 ..., K).The present invention is by profit
The pyramidal top layer being generated with self-similarity is as initial reconstructed imageWillIt is divided intoImage block to be reconstructed
Its corresponding radio-frequency component isAccording to formulaTry to achieve kiIt is assured that expression image blockDictionary
Step 3：Calculate the weight matrix A of sparse coding non local structure self similarity；
Step 4：Setting iteration ends error e, maximum iteration time Max_Iter, the non local regularization term contribution amount of control
Constant η and undated parameter condition P；
Step 5：The current estimation of more new images：
Step 6：Update rarefaction representation coefficient：
Wherein, soft (, τI, j) it is threshold tauI, jSoft-threshold function；
Soft (x, τ)=sgn (x) max (| x |-τ, 0)
Wherein, sgn (x) is sign function；
Step 7：The current estimation of new images
Step 8：If mod (k, P)=0, then update the adaptive sparse domain of X, useUpdate matrix A；
Step 9：Repeat 5～8, until iteration meetsOr k >=Max_Iter iteration ends.
In conjunction with accompanying drawing, whole process is described in detail：
1., in testing, using original image as high-resolution reference picture, then it is carried out obscuring, down-sampled operation generates
Low-resolution image to be reconstructed, Gaussian Blur core a size of 7 × 7, standard deviation is 1.6, and the down-sampled factor is 3.When rebuilding
Pending low resolution coloured image is converted to YCbCr form, only luminance component is rebuild, other two components by
Bicubic interpolation is tried to achieve.By inventive algorithm reconstructed results respectively with bicubic interpolation method, Yang algorithm, Zeyde algorithm,
The result that Glasner algorithm, Dong algorithm obtain is compared that (program of rear four kinds of algorithms all can be downloaded from its corresponding website
[19-22]).Experiment relative parameters setting is as follows：Sample in natural image storehouse is to randomly select from Yang algorithm Sample Storehouse
20 width images；When setting up pyramid, z=1.25, k=4, j=5；Tile size n=36；Δ=4.4 when choosing sample；
Element number in dictionary is 36；Similar image block number is 12；During reconstruction, maximum iteration time is 999；Iteration ends error
For 2 × 10-6.Fig. 2, Fig. 3 be respectively this two width image of Butterfly and hat various algorithm reconstructed results figures, for convenience than
Relatively, low-resolution image is tuned into onesize with original image, simultaneously in order to more clearly embody comparing result, by two width
The subregion of image is amplified, and Fig. 2 is exaggerated the texture of butterfly's wing, and Fig. 3 is exaggerated the letter on cap.By Fig. 2 (c)
With Fig. 3 (c) as can be seen that the image being reconstructed by bicubic interpolation method is more fuzzy, edge is smoother, and visual effect is
Difference, all can not tell the letter on cap.Fig. 2 (d) and Fig. 3 (d) for Yang algorithm reconstruct as a result, it is possible to find out reconstruction
Result has recovered part details, good compared with the result that bicubic interpolation method is rebuild, but image is still relatively fuzzyyer, equally can not tell
Letter on cap.Fig. 2 (e) and Fig. 3 (e) be Zeyde algorithm reconstructed results it can be seen that earlier above two methods reconstructed results whole
Body is clear, but the texture on butterfly's wing is still relatively fuzzyyer, and the letter on cap has artifact, and Zeyde is described
Algorithm introduces deceptive information.Fig. 2 (f) and Fig. 3 (f) is Glasner algorithm reconstructed results, and overall visual effect is calculated with Zeyde
Method is close, but by local magnification region can be seen that Glasner algorithm reconstructed results more accurate, the letter on cap is not
There is artifact, this is because it only make use of image itself to be rebuild during rebuilding, do not introduce external image.Figure
2 (g) and Fig. 3 (g) be Dong algorithm reconstructed results it can be seen that compared with algorithm above, image border is more sharp, details
More rich, substantially can tell the letter on cap.Fig. 2 (h) and Fig. 3 (h) is inventive algorithm reconstructed results, with other
Algorithm is compared, and the result visual effect that inventive algorithm reconstructs preferably, has recovered more details, image becomes apparent from.
2. parameter measurements analysis
For the advantage of quantam of proof inventive algorithm, have chosen 10 different types of natural images, they come respectively
From in document [12], document [15] and document [17], table 1 lists the Y-PSNR (PSNR) of reconstruction image under various algorithms
With the value of structure self-similarity (SSIM) [18], wherein PSNR value bigger explanation reconstruction image is closer to original image, reconstruction effect
Fruit is better, and SSIM is the index of evaluation criterion high-definition picture and the architectural difference of super-resolution rebuilding image, is worth bigger theory
Bright reconstruction effect is better.
By obtaining with the contrast of bicubic interpolation, Yang algorithm, Zeyde algorithm, Glasner algorithm and Dong algorithm
Go out, with respect to the reconstruction effect of this five kinds of algorithms, the PSNR value of reconstructed results of the present invention respectively average improve 4.12dB,
3.18dB, 1.37dB, 3.84dB and 0.56dB, SSIM value is average respectively to improve 0.1072,0.0681,0.0169,0.0656 and
0.0093.This absolutely proves that the image effect that the present invention reconstructs is more preferable.
Various algorithm reconstruction image PSNR (the dB)/SSIM value of table 1 compares
Claims (4)
1. a kind of image super-resolution rebuilding method of the many dictionary learnings of improved self adaptation, comprises the following steps：
Step 1：Down-sampled matrix D and fuzzy matrix B are determined according to the process that degrades of image；
Step 2：Set up pyramid using image self-similarity, using pyramidal upper layer images and natural image as dictionary learning
Sample, build all kinds of dictionaries with PCA methodAnd using pyramidal top layer images as initial reconstructed imageStep 3：
Calculate the weight matrix A of sparse coding non local structure self similarity；
Step 4：Setting iteration ends error e, maximum iteration time Max_Iter, control non local regularization term contribution amount normal
Number η and condition P of undated parameter；
Step 5：The current estimation of more new images：
Step 6：Update rarefaction representation coefficient：
Wherein, soft (, τI, j) it is threshold tauI, jSoft-threshold function；
Soft (x, τ)=sgn (x) max (| x |-τ, 0)
Wherein, sgn (x) is sign function；
Step 7：The current estimation of more new images
Step 8：If mod (k, P)=0, then update the adaptive sparse domain of X, useUpdate matrix A；
Step 9：Repeat (5)～(8), until iteration meetsOr k >=Max_Iter iteration ends.
2. the image super-resolution rebuilding method of many dictionary learnings of improved self adaptation according to claim 1, its feature exists
In to the image in natural image storehouse in step 2Carry out high-pass filtering, be then cut intoThe block of size, selects image block
Standard deviation is more than the block of Δ, noteFor select sample image block,For its corresponding high-pass filtering
Radio-frequency component afterwards；CalculateObtain with pyramid upper layer images sample clusteringK class barycenter μk's
Distance, is designated asIfThen willIt is added to Ck, otherwise give up this image block, wherein, δ is used to control figure
Parameter as the similarity degree at Kuai Yulei center；After aforesaid operations, sample is expanded, and the sample after every class expands is designated asqkFor number of samples after expanding.
3. according to claim 1 the many dictionary learnings of improved self adaptation image super-resolution rebuilding method it is characterised in that
In step 2, dictionary learning is carried out using PCA (principal component analysis, PCA)；If SkAssociation side
Difference matrix is Ωk, to ΩkCarry out principal component analysis, obtain an orthogonal transform matrix PkIt is assumed that PkFor dictionary, then sparse coefficientSelect PkMiddle r most important characteristic vector forms formulaIn
Dictionary Φr=[p1, p2..., pr], then sparse coefficient matrix
4. the image super-resolution rebuilding method of many dictionary learnings of improved self adaptation according to claim 1, its feature exists
In, in step 2 by the use of self-similarity generate pyramidal top layer as initial reconstructed imageWillIt is divided intoTreat
Reconstruction image blockIts corresponding radio-frequency component isAccording to formulaTry to achieve kiTo determine expression image
BlockDictionary
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610847106.4A CN106408550A (en) | 2016-09-22 | 2016-09-22 | Improved self-adaptive multi-dictionary learning image super-resolution reconstruction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610847106.4A CN106408550A (en) | 2016-09-22 | 2016-09-22 | Improved self-adaptive multi-dictionary learning image super-resolution reconstruction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106408550A true CN106408550A (en) | 2017-02-15 |
Family
ID=57997414
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610847106.4A Pending CN106408550A (en) | 2016-09-22 | 2016-09-22 | Improved self-adaptive multi-dictionary learning image super-resolution reconstruction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106408550A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107146263A (en) * | 2017-04-27 | 2017-09-08 | 浙江大学 | A kind of dynamic PET images method for reconstructing constrained based on tensor dictionary |
| CN107169925A (en) * | 2017-04-21 | 2017-09-15 | 西安电子科技大学 | The method for reconstructing of stepless zooming super-resolution image |
| CN107481189A (en) * | 2017-06-28 | 2017-12-15 | 西安邮电大学 | A kind of super-resolution image reconstruction method of the rarefaction representation based on study |
| CN107948635A (en) * | 2017-11-28 | 2018-04-20 | 厦门大学 | It is a kind of based on degenerate measurement without refer to sonar image quality evaluation method |
| CN108986027A (en) * | 2018-06-26 | 2018-12-11 | 大连大学 | Depth image super-resolution reconstruction method based on improved joint trilateral filter |
| CN109035360A (en) * | 2018-07-31 | 2018-12-18 | 四川大学华西医院 | A kind of compressed sensing based CBCT image rebuilding method |
| CN109035144A (en) * | 2018-07-19 | 2018-12-18 | 广东工业大学 | super-resolution image construction method |
| CN109064406A (en) * | 2018-08-26 | 2018-12-21 | 东南大学 | A kind of rarefaction representation image rebuilding method that regularization parameter is adaptive |
| CN109146785A (en) * | 2018-08-02 | 2019-01-04 | 华侨大学 | A kind of image super-resolution method based on the sparse autocoder of improvement |
| CN109741263A (en) * | 2019-01-11 | 2019-05-10 | 四川大学 | Remote sensed image super-resolution reconstruction algorithm based on adaptive combined constraint |
| CN111210390A (en) * | 2019-10-15 | 2020-05-29 | 杭州电子科技大学 | Motion blur restoration method based on Golay sequence complementary code word set |
| CN114972339A (en) * | 2022-07-27 | 2022-08-30 | 金成技术股份有限公司 | Data enhancement system for bulldozer structural member production abnormity detection |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101398886A (en) * | 2008-03-17 | 2009-04-01 | 杭州大清智能技术开发有限公司 | Rapid three-dimensional face identification method based on bi-eye passiveness stereo vision |
| CN103020909A (en) * | 2012-12-06 | 2013-04-03 | 清华大学 | Single-image super-resolution method based on multi-scale structural self-similarity and compressive sensing |
| CN104156957A (en) * | 2014-08-06 | 2014-11-19 | 昆山天工智能科技有限公司 | Stable and high-efficiency high-resolution stereo matching method |
| WO2015180054A1 (en) * | 2014-05-28 | 2015-12-03 | 北京大学深圳研究生院 | Video coding and decoding methods and apparatuses based on image super-resolution |
-
2016
- 2016-09-22 CN CN201610847106.4A patent/CN106408550A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101398886A (en) * | 2008-03-17 | 2009-04-01 | 杭州大清智能技术开发有限公司 | Rapid three-dimensional face identification method based on bi-eye passiveness stereo vision |
| CN103020909A (en) * | 2012-12-06 | 2013-04-03 | 清华大学 | Single-image super-resolution method based on multi-scale structural self-similarity and compressive sensing |
| WO2015180054A1 (en) * | 2014-05-28 | 2015-12-03 | 北京大学深圳研究生院 | Video coding and decoding methods and apparatuses based on image super-resolution |
| CN104156957A (en) * | 2014-08-06 | 2014-11-19 | 昆山天工智能科技有限公司 | Stable and high-efficiency high-resolution stereo matching method |
Non-Patent Citations (6)
| Title |
|---|
| JIAN LU 等: "Context-aware single image super-resolution using sparse representation and cross-scale similarity", 《SIGNAL PROCESSING: IMAGE COMMUNICATION》 * |
| 李海斌: "基于稀疏表示的图像超分辨率重建研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 潘宗序 等: "基于多尺度结构自相似性的单幅图像超分辨率算法", 《自动化学报》 * |
| 潘宗序 等: "基于多尺度非局部约束的单幅图像超分辨率算法", 《自动化学报》 * |
| 潘宗序 等: "基于自适应多字典学习的单幅图像超分辨率算法", 《电子学报》 * |
| 陈少冲: "一种自适应学习的图像超分辨率重建算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169925A (en) * | 2017-04-21 | 2017-09-15 | 西安电子科技大学 | The method for reconstructing of stepless zooming super-resolution image |
| CN107169925B (en) * | 2017-04-21 | 2019-10-22 | 西安电子科技大学 | The method for reconstructing of stepless zooming super-resolution image |
| CN107146263A (en) * | 2017-04-27 | 2017-09-08 | 浙江大学 | A kind of dynamic PET images method for reconstructing constrained based on tensor dictionary |
| CN107146263B (en) * | 2017-04-27 | 2019-11-01 | 浙江大学 | A kind of dynamic PET images method for reconstructing based on the constraint of tensor dictionary |
| CN107481189A (en) * | 2017-06-28 | 2017-12-15 | 西安邮电大学 | A kind of super-resolution image reconstruction method of the rarefaction representation based on study |
| CN107948635B (en) * | 2017-11-28 | 2019-09-27 | 厦门大学 | It is a kind of based on degenerate measurement without reference sonar image quality evaluation method |
| CN107948635A (en) * | 2017-11-28 | 2018-04-20 | 厦门大学 | It is a kind of based on degenerate measurement without refer to sonar image quality evaluation method |
| CN108986027A (en) * | 2018-06-26 | 2018-12-11 | 大连大学 | Depth image super-resolution reconstruction method based on improved joint trilateral filter |
| CN109035144A (en) * | 2018-07-19 | 2018-12-18 | 广东工业大学 | super-resolution image construction method |
| CN109035144B (en) * | 2018-07-19 | 2023-03-17 | 广东工业大学 | Super-resolution image construction method |
| CN109035360A (en) * | 2018-07-31 | 2018-12-18 | 四川大学华西医院 | A kind of compressed sensing based CBCT image rebuilding method |
| CN109146785A (en) * | 2018-08-02 | 2019-01-04 | 华侨大学 | A kind of image super-resolution method based on the sparse autocoder of improvement |
| CN109064406A (en) * | 2018-08-26 | 2018-12-21 | 东南大学 | A kind of rarefaction representation image rebuilding method that regularization parameter is adaptive |
| CN109741263A (en) * | 2019-01-11 | 2019-05-10 | 四川大学 | Remote sensed image super-resolution reconstruction algorithm based on adaptive combined constraint |
| CN109741263B (en) * | 2019-01-11 | 2019-10-11 | 四川大学 | Remote sensed image super-resolution reconstruction method based on adaptive combined constraint |
| CN111210390A (en) * | 2019-10-15 | 2020-05-29 | 杭州电子科技大学 | Motion blur restoration method based on Golay sequence complementary code word set |
| CN114972339A (en) * | 2022-07-27 | 2022-08-30 | 金成技术股份有限公司 | Data enhancement system for bulldozer structural member production abnormity detection |
| CN114972339B (en) * | 2022-07-27 | 2022-10-21 | 金成技术股份有限公司 | Data enhancement system for bulldozer structural member production abnormity detection |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106408550A (en) | Improved self-adaptive multi-dictionary learning image super-resolution reconstruction method | |
| CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
| Li et al. | Survey of single image super‐resolution reconstruction | |
| Yang et al. | Single-image super-resolution reconstruction via learned geometric dictionaries and clustered sparse coding | |
| Kim et al. | Iterative kernel principal component analysis for image modeling | |
| CN111369440B (en) | Model training and image super-resolution processing method, device, terminal and storage medium | |
| Yu et al. | A unified learning framework for single image super-resolution | |
| Zhang et al. | Image super-resolution based on structure-modulated sparse representation | |
| CN110992270A (en) | Multi-scale residual attention network image super-resolution reconstruction method based on attention | |
| Hu et al. | SERF: A simple, effective, robust, and fast image super-resolver from cascaded linear regression | |
| Li et al. | FilterNet: Adaptive information filtering network for accurate and fast image super-resolution | |
| CN107392865B (en) | Restoration method of face image | |
| Chen et al. | Convolutional neural network based dem super resolution | |
| CN112541864A (en) | Image restoration method based on multi-scale generation type confrontation network model | |
| CN115564649B (en) | Image super-resolution reconstruction method, device and equipment | |
| Zhang et al. | MR image super-resolution reconstruction using sparse representation, nonlocal similarity and sparse derivative prior | |
| CN111861886B (en) | Image super-resolution reconstruction method based on multi-scale feedback network | |
| CN113205449A (en) | Expression migration model training method and device and expression migration method and device | |
| Li et al. | A self-learning image super-resolution method via sparse representation and non-local similarity | |
| Dou et al. | Medical image super-resolution via minimum error regression model selection using random forest | |
| CN116977674A (en) | Image matching method, related device, storage medium and program product | |
| Uddin et al. | A perceptually inspired new blind image denoising method using $ L_ {1} $ and perceptual loss | |
| CN115293966A (en) | Face image reconstruction method and device and storage medium | |
| Gao et al. | Bayesian image super-resolution with deep modeling of image statistics | |
| Shi et al. | Exploiting multi-scale parallel self-attention and local variation via dual-branch transformer-cnn structure for face super-resolution |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170215 |