CN108171659A - A kind of image repair method based on K-SVD dictionaries - Google Patents
A kind of image repair method based on K-SVD dictionaries Download PDFInfo
- Publication number
- CN108171659A CN108171659A CN201711247000.1A CN201711247000A CN108171659A CN 108171659 A CN108171659 A CN 108171659A CN 201711247000 A CN201711247000 A CN 201711247000A CN 108171659 A CN108171659 A CN 108171659A
- Authority
- CN
- China
- Prior art keywords
- image
- svd
- dictionary
- algorithms
- pixel
- 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
- 230000008439 repair process Effects 0.000 title claims abstract description 43
- 238000000034 method Methods 0.000 title claims abstract description 23
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 69
- 239000011159 matrix material Substances 0.000 claims description 7
- 238000009499 grossing Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 12
- 238000012549 training Methods 0.000 abstract description 3
- 230000006872 improvement Effects 0.000 description 8
- 230000006378 damage Effects 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 230000007812 deficiency Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000001914 filtration Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 239000006002 Pepper Substances 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000009415 formwork Methods 0.000 description 1
- 230000001771 impaired effect Effects 0.000 description 1
- 210000003127 knee Anatomy 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000007420 reactivation Effects 0.000 description 1
- 238000011160 research Methods 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
- G06T5/77—Retouching; Inpainting; Scratch removal
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of image repair methods based on K SVD dictionaries, the algorithm is based on sparse representation theory simultaneously using the similitude between image block as theoretical foundation, method using control kernel regression clusters foundation as image block, ready-portioned a few class regions are trained respectively using K SVD dictionaries, dictionary after training with their sparse coefficient is multiplied, repairs the damaged area in image.Step is as follows:The first step:Complex pattern to be repaired is divided into image fritter and clusters all image fritters using the method for control kernel regression weights;Second step:All kinds of images obtained in the first step are trained respectively using K svd algorithms;Third walks:Iteration updates, until the image-region of all categories is fully completed update.The present invention has more apparent repairing effect compared to the TV image repairs algorithm based on Total Variation and the Criminisi image repair algorithms based on texture, has in terms of the reparation details of image compared to traditional K svd algorithms and significantly improves.
Description
Technical field
The invention belongs to image repair field, more particularly to a kind of image repair method based on K-SVD dictionaries.
Background technology
With the continuous development of computer and multimedia technology, the information based on image becomes the mainstream matchmaker of information exchange
Body greatly affected the life style of the people.Important research branch of the image repair as image processing field, defends in medical treatment
The numerous areas such as life, military security, public administration have a wide range of applications.Image repair refers to utilize the presence damaged in image
Some intact information, fill up zone of ignorance or be removed the extra object in image, image approaches after making reparation
Or reach the visual effect of original image, ensure that image information still can completely give expression to the content that it contains.Repair part
Impaired image is to rebuild or restore original image processing as far as possible using the priori of image.Traditional image repair algorithm
Including the reparation algorithm (TV algorithms, BSCB algorithms and CCD algorithms etc.) based on partial differential equation and the reparation based on texture information
Algorithm (Criminisi algorithms).In recent years, rarefaction representation receives more and more favors, makes its fortune in image repair technology
It uses and receives more concerns.It is represented image sparse using the openness of signal, after original image recovered by reconstruct.
Traditional image repair algorithm, which exists, repairs edge blurry, and repairing effect is poor, image detail partial reduction degree is not high
Deng.It is our problems to be solved that these shortcomings how to be overcome, which to carry out algorithm improvement,.
Invention content
Objects of the present invention:Picture signal is a kind of 2D signal containing a large amount of signals and abundant content, is being used
During rarefaction representation carries out image repair, excessively complete dictionary is difficult the feature for representing the various complexity in image, thereby increases and it is possible to
It is not inconsistent with the overall structure of image.The use of traditional K-SVD dictionaries is from global angle mostly, is calculated with rarefaction representation
Method handles entire image.Then rarefaction representation is carried out to all fritters in image by this dictionary.But it if needs at present
The image of rarefaction representation is wanted to contain abundant grain details information, then a variety of image fritters can not be met by being only applicable in single dictionary
Demand.
In view of the importance of dictionary selection, the present invention have carried out algorithm steps on the basis of traditional K-SVD algorithms
It improves, first in view of the similitude between image block, by image clustering, is divided into multiple regions with similar characteristic, later
The excessively complete word that can give full expression to this kind of image-region is obtained using K-SVD image repair algorithms to each image-region again
Allusion quotation.Rarefaction representation coefficient is obtained according to dictionary, updates each image-region.After the completion of iteration, according to the index generated when clustering
Each image block is put back to, is finally completed the reparation of broken image.
In order to solve the above-mentioned technical problem, the technical solution adopted by the present invention is:A kind of image based on K-SVD dictionaries
Restorative procedure carries out in accordance with the following steps:(1) it inputs:Image X to be repaired;(2) image X to be repaired is divided into pixel
The pixel of each image fritter is lined up one by nonoverlapping multiple images fritter according to sequence from left to right from top to bottom
Row, the row sequence that all image fritters are generated form index matrix;(3) all image fritters are used into control core
The method for returning weights is clustered, according to formulaIt is divided
The ready-portioned different subregion X of good K classesk;(4) using K-SVD algorithms:
Wherein D(k)For with the corresponding dictionary of each subclass, xkTo belong to the image fritter of K class image-regions, αkFor
The rarefaction representation of this kind of image fritter.During K-SVD methods are used, first have to assign dictionary D(k)One fixed dictionary,
The initial value of dictionary used complete DCT dictionaries, solved the rarefaction representation under the dictionary using BP greedy algorithms later:
The K class images obtained in above-mentioned steps are trained respectively;(5) iteration updates:Update is per a kind of dictionary D(k)If iterations are less than the number of iterations of setting, then return to step (4) otherwise stops iteration, obtains finally being suitble to K
The dictionary D of class image-regionk;(6) by dictionary DkCorresponding rarefaction representation coefficient is multiplied, and obtains K class images;(7) weight
Multiple step (4) is to step (6), until the image-region of all categories is fully completed update;(8) according to the index matrix of step 2
Image-region is put back in position, the image finally repaired.
The step (3) will control kernel regression function to draw as the foundation of image clustering according to each pixel is corresponding
Weight feature vector is led, clustering is carried out to image, guiding weight is defined as:
Wherein i, j ∈ NiFor the pixel in image, NiFor the local neighborhood centered on pixel i, piAnd pjFor pixel
The position of point i and pixel j in the picture, h be control Gaussian kernel support global smoothing parameter, CjFor to the vertical of pixel j
The symmetric gradient variance matrix that gradient and horizontal gradient are estimated combines rotation operator, extension operator and scale operator, makes
The structure for obtaining image can be consistent with control Gaussian kernel,
U in formulaθjFor twiddle factor, for by Gaussian function and its edge direction θjAlignment, ΛjRepresent extension matrix, γj
Represent scale parameter.
The object function of rarefaction representation image repair algorithm of the step (4) based on K-SVD dictionaries is:
X in formula0For reconstructed image, λ and μlFor weight, l=1,2 ... N.
The beneficial effects of the present invention are:(1) for traditional single dictionary rarefaction representation the shortcomings that, is returned using control core
The mode of function is returned to be split cluster to image, entire image is represented jointly in the form of multiword allusion quotation.
(2) visual communication error peace sliding formwork paste square is susceptible in image block Breakage Processes are repaired compared to TV algorithms
Deficiency, using based on rarefaction representation improve K-SVD image repair algorithms can be good at overcoming this shortcoming.
(3) although can preferably repair breakage for Criminisi algorithms, in the mistake that texture information is extended
Cheng Zhong, picture structure aspect will appear the deficiency of manifest error, can be very using the image repair algorithm for improving K-SVD dictionaries
The destruction to picture structure is reduced in big degree.
(4) the shortcomings that bad to the processing of image detail part compared to traditional K-SVD algorithms, one kind proposed by the invention
Image repair algorithm based on K-SVD dictionaries, is improved on the basis of original algorithm.Similitude between image block is made
For theoretical foundation, the method using control kernel regression clusters foundation as image block, ready-portioned a few class regions is used respectively
K-SVD dictionaries are trained, and the dictionary after training is multiplied with their sparse coefficient, repair the damaged area in image.Energy
Enough reach and preferable effect is repaired to image detail.
Description of the drawings:
Fig. 1 be add noise each algorithm image repair comparison diagram, wherein (a) represent artwork, (b) represent noise pattern,
(c) represent that median filtering algorithm, (d) represent that TV algorithms, (e) represent K-SVD dictionaries, (f) inventive algorithm;
Fig. 2 is blocky damaged each algorithm image repair comparison diagram, wherein (a) represents that artwork, (b) represent blocky damaged figure
Picture, (c) represent that TV algorithms, (d) represent that Criminisi algorithms, (e) represent K-SVD algorithms, (f) inventive algorithm;
Fig. 3 is each algorithm image repair comparison diagram there are cut, wherein (a) original image, (b) represent that there are cuts
Image, (c) represent that TV algorithms, (d) represent that Criminisi algorithms, (e) represent K-SVD algorithms, (f) inventive algorithm.
Specific embodiment
Based on the image repair algorithm of improved K-SVD dictionaries, it will cluster and pass using kernel regression is controlled to carry out image block
The image repair algorithm based on K-SVD dictionaries of system is combined, and can reach preferable image repair effect.
MATLAB R2016b are used below, are emulated under processor i3,3.3GHzCPU environment and with reference to several tools
Body embodiment and attached drawing are described in further detail the present invention.
Embodiment 1:
Fig. 1 carries out emulation experiment using fruit pictures, and former clear picture is added in the salt-pepper noise that noise density is 0.1,
Using median filtering algorithm, TV algorithms, K-SVD dictionaries and these four algorithms pair of improvement K-SVD dictionaries proposed by the invention
Denoising result carries out contrast experiment.
From the point of view of simulation result, these four image repair methods have carried out noise effectively to go to a certain extent
It removes, substantially reduces the original appearance of image, but all not fully reach the effect of reduction artwork completely.It can from Fig. 1 (c)
Although intuitively finding out that median filtering algorithm overall effect is good, speed is repaired, from the part extraction figure in the upper left corner
It can be seen that it has lost a large amount of details during denoising, good denoising effect can not be reached.It can from Fig. 1 (d)
To find out that the image after the reparation of TV algorithms still has apparent noise.It is repaiied as K-SVD algorithms traditional Fig. 1 (e) carry out image
Although the details of meeting lost part, can reach and noise is preferably removed when multiple.It is and proposed by the invention based on improvement K-SVD
The image repair algorithm of dictionary is good to the repairing effect of image, not only ensure that the clarity of image, but also effective eliminate is made an uproar
Sound, image detail partial information also preserve more intact.
Embodiment 2:
TV algorithms are employed for blocky damaged image in Fig. 2, the Criminisi algorithms based on texture information, K-
Svd algorithm and improvement K-SVD algorithms proposed by the invention carry out reparation contrast experiment to blocky corrupted picture.
From the point of view of the simulation result in figure, these four image repair methods all reduce the original of image to a certain extent
Looks, but all it is unable to reach the effect for being reduced into artwork completely.It can be seen that although TV algorithms can be to breakage from Fig. 2 (c)
Reparation substantially can be carried out on position, but since the information of damage location is insufficient, vision company is susceptible in repair process
The error of logical aspect, also will appear smoothly fuzzy square sometimes.From Fig. 2 (d) it can be seen that Criminisi algorithms are to larger
In the reparation of area damaged area, there is more apparent improvement compared to TV algorithms, it is according to damage location week to repair thinking
The texture structure on side is repaired, and when to damaged central repair, the selection of optimized image block can have image repair very big
Influence.Although the repairing effect of the K-SVD algorithms of Fig. 2 (e) still has fuzzy at knee and pigtail, damage location is repaiied
Reactivation is enough met the requirements, and has preferable performance.Image repair method proposed by the invention based on improvement K-SVD dictionaries is to figure
As treatment of details is preferable, reparation result at details and artwork substantially close to.
Embodiment 3:
Fig. 3 is directed to there are the picture of cut using TV algorithms, Criminisi algorithms, traditional K-SVD algorithms and the present invention
The improvement K-SVD algorithms of proposition carry out reparation contrast experiment to image.
From the point of view of simulation result, these four image repair methods all reduce the original appearance of image to a certain extent.From figure
3 (c) it can be seen that when damaged smaller, TV algorithms are removed cut by way of diffusion, but due to information
Deficiency, there is no repair completely, and other parts repairing effect is nor highly desirable for the cut on cap.And from Fig. 3 (d)
It can be seen that Criminisi algorithms are chosen optimal image block and are repaired, but match nothing by being matched to peripheral information
Method ensures the continuity of image, therefore occurs fracture at cap details in Fig. 3 (d).Traditional K-SVD algorithms such as Fig. 3
(e) shown in, still will appear at the details of part fuzzy.And the image proposed by the invention based on improvement K-SVD dictionaries is repaiied
Double calculation method remains raw information in classification with as much as possible in training process, and blurred block is less, compares other algorithm effects
More preferably.
Claims (3)
1. a kind of image repair method based on K-SVD dictionaries, carries out in accordance with the following steps:
(1) it inputs:Image X to be repaired.
(2) image X to be repaired is divided into the nonoverlapping multiple images fritter of pixel, the pixel of each image fritter is pressed
It forms a line according to sequence from left to right from top to bottom, the row sequence that all image fritters are generated forms index square
Battle array;
(3) all image fritters are clustered using the method for control kernel regression weights, according to formula
Obtain the ready-portioned different subregion X of K classesk;
(4) using K-SVD algorithms:
Wherein D(k)For with the corresponding dictionary of each subclass, xkTo belong to the image fritter of K class image-regions, αkIt is this kind of
The rarefaction representation of image fritter.During K-SVD methods are used, first have to assign dictionary D(k)One fixed dictionary, dictionary
Initial value used complete DCT dictionaries, rarefaction representation under the dictionary is solved using BP greedy algorithms later:
The K class images obtained in above-mentioned steps are trained respectively;
(5) iteration updates:Update is per a kind of dictionary D(k)If iterations are less than the number of iterations of setting, then return to step
(4), otherwise stop iteration, obtain the dictionary D of finally suitable K class image-regionsk;
(6) by dictionary DkCorresponding rarefaction representation coefficient is multiplied, and obtains K class images;
(7) step (4) to step (6) is repeated, until the image-region of all categories is fully completed update;
(8) image-region is put back to according to the index matrix position of step 2, the image Y finally repaired.
2. a kind of image repair method based on K-SVD dictionaries according to claim 1, it is characterized in that:The step (3)
Foundation of the kernel regression function as image clustering will be controlled, weight feature vector is guided according to each pixel is corresponding,
Clustering is carried out to image, guiding weight is defined as:
Wherein i, j ∈ NiFor the pixel in image, NiFor the local neighborhood centered on pixel i, piAnd pjFor pixel i and
The positions of pixel j in the picture, h be control Gaussian kernel support global smoothing parameter, CjFor the vertical gradient to pixel j and
The symmetric gradient variance matrix that horizontal gradient is estimated combines rotation operator, extension operator and scale operator so that image
Structure can with control Gaussian kernel it is consistent,
In formulaFor twiddle factor, for by Gaussian function and its edge direction θjAlignment, ΛjRepresent extension matrix, γjIt represents
Scale parameter.
3. a kind of image repair method based on K-SVD dictionaries according to claim 1, it is characterized in that:The step (4)
The object function of rarefaction representation image repair algorithm based on K-SVD dictionaries is:
X in formula0For reconstructed image, λ and μlFor weight, l=1,2 ... N.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711247000.1A CN108171659A (en) | 2017-12-01 | 2017-12-01 | A kind of image repair method based on K-SVD dictionaries |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711247000.1A CN108171659A (en) | 2017-12-01 | 2017-12-01 | A kind of image repair method based on K-SVD dictionaries |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108171659A true CN108171659A (en) | 2018-06-15 |
Family
ID=62524940
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711247000.1A Pending CN108171659A (en) | 2017-12-01 | 2017-12-01 | A kind of image repair method based on K-SVD dictionaries |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108171659A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109615576A (en) * | 2018-06-28 | 2019-04-12 | 西安工程大学 | The single-frame image super-resolution reconstruction method of base study is returned based on cascade |
CN113592728A (en) * | 2021-07-01 | 2021-11-02 | 温州理工学院 | Image restoration method, system, processing terminal and computer medium |
CN113609928A (en) * | 2021-07-19 | 2021-11-05 | 杨薇 | Smart city management system based on cloud computing and image recognition |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103700065A (en) * | 2013-12-03 | 2014-04-02 | 杭州电子科技大学 | Structure sparsity propagation image repairing method adopting characteristic classified learning |
CN103886557A (en) * | 2014-03-28 | 2014-06-25 | 北京工业大学 | Denoising method of depth image |
-
2017
- 2017-12-01 CN CN201711247000.1A patent/CN108171659A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103700065A (en) * | 2013-12-03 | 2014-04-02 | 杭州电子科技大学 | Structure sparsity propagation image repairing method adopting characteristic classified learning |
CN103886557A (en) * | 2014-03-28 | 2014-06-25 | 北京工业大学 | Denoising method of depth image |
Non-Patent Citations (2)
Title |
---|
PRIYAM CHATTERJEE: "Clustering-Based Denoising With Locally Learned Dictionaries", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
黄江林 等: "一种改进的基于K-SVD字典的图像修复算法", 《安徽大学学报(自然科学版)》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109615576A (en) * | 2018-06-28 | 2019-04-12 | 西安工程大学 | The single-frame image super-resolution reconstruction method of base study is returned based on cascade |
CN113592728A (en) * | 2021-07-01 | 2021-11-02 | 温州理工学院 | Image restoration method, system, processing terminal and computer medium |
CN113592728B (en) * | 2021-07-01 | 2024-04-05 | 温州理工学院 | Image restoration method, system, processing terminal and computer medium |
CN113609928A (en) * | 2021-07-19 | 2021-11-05 | 杨薇 | Smart city management system based on cloud computing and image recognition |
CN113609928B (en) * | 2021-07-19 | 2022-12-20 | 广州市雅天网络科技有限公司 | Smart city management system based on cloud computing and image recognition |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
He et al. | Arch++: Animation-ready clothed human reconstruction revisited | |
CN114140353B (en) | Swin-Transformer image denoising method and system based on channel attention | |
CN105957063B (en) | CT image liver segmentation method and system based on multiple dimensioned weighting similarity measure | |
CN108198184B (en) | Method and system for vessel segmentation in contrast images | |
WO2020181900A1 (en) | Image processing method and apparatus, image device and storage medium | |
CN109658344A (en) | Image de-noising method, device, equipment and storage medium based on deep learning | |
CN108932536A (en) | Human face posture method for reconstructing based on deep neural network | |
CN111028343B (en) | Three-dimensional face model generation method, device, equipment and medium | |
Tang et al. | Single image dehazing via lightweight multi-scale networks | |
CN108171659A (en) | A kind of image repair method based on K-SVD dictionaries | |
CN109766866B (en) | Face characteristic point real-time detection method and detection system based on three-dimensional reconstruction | |
CN110197462A (en) | A kind of facial image beautifies in real time and texture synthesis method | |
CN110363719A (en) | A kind of cell layered image processing method and system | |
CN110009722A (en) | Three-dimensional rebuilding method and device | |
CN110298804A (en) | One kind is based on generation confrontation network and the decoded medical image denoising method of 3D residual coding | |
CN108648256A (en) | A kind of gray scale core three-dimensional method for reconstructing based on super dimension | |
CN110335344A (en) | Three-dimensional rebuilding method based on 2D-3D attention mechanism neural network model | |
CN110570394B (en) | Medical image segmentation method, device, equipment and storage medium | |
CN104504672B (en) | Low-rank sparse neighborhood insertion ultra-resolution method based on NormLV features | |
CN107909588A (en) | Partition system under MRI cortex based on three-dimensional full convolutional neural networks | |
CN108399608A (en) | Dimensional images denoising method based on tensor dictionary and full variation | |
CN113052976A (en) | Single-image large-pose three-dimensional color face reconstruction method based on UV position map and CGAN | |
CN110859642A (en) | Method, device, equipment and storage medium for realizing medical image auxiliary diagnosis based on AlexNet network model | |
CN110047054A (en) | A kind of GAN medical image denoising method for extracting feature based on VGG-19 | |
CN114897694A (en) | Image super-resolution reconstruction method based on mixed attention and double-layer supervision |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | 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: 20180615 |