CN105957013B - A kind of single image super-resolution reconstruction method - Google Patents
A kind of single image super-resolution reconstruction method Download PDFInfo
- Publication number
- CN105957013B CN105957013B CN201610394659.9A CN201610394659A CN105957013B CN 105957013 B CN105957013 B CN 105957013B CN 201610394659 A CN201610394659 A CN 201610394659A CN 105957013 B CN105957013 B CN 105957013B
- Authority
- CN
- China
- Prior art keywords
- resolution
- image
- low
- block
- super
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
Abstract
The invention discloses a kind of single image super-resolution reconstruction methods, this method is using mutual corresponding high-resolution and low-resolution image block as training dataset, high-resolution dictionary and low-resolution dictionary are trained from tranining database using alternating direction method, then by the low-resolution image of input, division obtains the set of the low resolution image block of original image;The low resolution image block of original image is then converted into corresponding high-definition picture block using high-resolution dictionary and low-resolution dictionary, high-definition picture block is combined into high-definition picture, high resolution image reconstruction is finally obtained into super-resolution image.This method converts 3 subproblems with analytic solutions for high-resolution and low-resolution dictionary training pattern and iteratively solves, significantly improve computational efficiency under the premise of guaranteeing constringent under alternating direction method theoretical frame.It is rebuild using the rapid super-resolution that single image may be implemented in this method, is of great significance to image procossing and many application fields of display.
Description
Technical field
The present invention relates to field of image processings, more particularly to a kind of single image super-resolution reconstruction method.
Background technique
Visual information is that the mankind obtain the main source of information and the critical function in the human cognitive world from objective world
Means, it accounts for about the 80% of the informational capacity that the mankind are obtained by face by the external world.Therefore, for a long time image procossing all by
To the extensive concern of researcher.With the rapid development of information technology, being led in remotely sensed image, medical image, video monitoring etc.
Domain, Image Super-resolution technology occupy increasingly consequence.In the research of Image Super-resolution technology, single image oversubscription
Distinguish that method for reconstructing has embodied distinctive advantage in practical applications.
The ultra-resolution method of single image is broadly divided into interpolation method, the method for reconstructing based on image prior and based on study
Method.Method based on interpolation, such as closest interpolation method and cube interpolation method carry out will cause image when super-resolution rebuilding
Surface is relatively fuzzyyer, and high-frequency information can be lost seriously, and especially edge sawtooth phenomenon is obvious, seriously affect picture quality.Base
Full resolution pricture edge is caused to have serrating phenomenon to occur due to the image prior artificially forced in the method for reconstruction, and
The quality of reconstruction image is seriously degenerated under the conditions of high amplification factor.Although these image super-resolution methods are in the presence of excessively smooth, edge
There is the defects of serrating phenomenon, but technically achieve certain breakthrough, tended to be mature and is regarded in electronic image, internet
Frequently, the multiple fields such as DTV are widely applied.Image super-resolution method based on study is in recent years by Freeman
Et al. a kind of image resolution method for proposing first, content be to concentrate to obtain priori knowledge as super from a large amount of training sample
The foundation of resolution ratio.Training sample is all to include the image with category information with input picture, using input picture as foundation, with study
The knowledge obtained in the process supplements the information in input picture, and the method based on study takes full advantage of image itself
Priori knowledge is remained to generate new high frequency detail, be obtained than based on reconstruction in the case where not increasing input picture sample size
The better restoration result of algorithm, and can preferably apply to the recovery of the images such as face and text.Therefore based on study
Super-resolution reconstruction method becomes the hot spot of research.
In order to overcome the shortcomings of that conventional learning algorithms computing cost is excessive, sparse learning method becomes mainstream in recent years.
Sparse learning method makes image have rarefaction representation on the dictionary by one excessively complete dictionary of study, thus significantly
Reduce operation scale.
The existing super-resolution reconstruction method based on rarefaction representation, during using training dataset study dictionary,
Model solution mainly uses gradient descent method, and solution efficiency is not high, and computation complexity is larger, affects entire single width
The practical application of Image Super-resolution Reconstruction method.
Summary of the invention
The purpose of the present invention is in view of the above shortcomings of the prior art, and provides and a kind of change under the premise of guaranteeing constringent
The single image super-resolution reconstruction method of kind computational efficiency.
In order to solve the above technical problems, the technical solution adopted by the present invention is that:
A kind of single image super-resolution reconstruction method, includes the following steps:
S1:High-definition image in collection network data constructs tranining database, and forms high-definition picture block YhWith low point
Resolution image block Yl;
S2:Using alternating direction method, the training high-resolution dictionary D from the tranining databasehWith low-resolution dictionary Dl;
S3:By the low-resolution image of input, division obtains the low resolution image block X of original imagelSet;
S4:Utilize high-resolution dictionary DhWith low-resolution dictionary DlBy the low resolution image block X of the original imagelTurn
Change corresponding high-definition picture block X intoh;
S5:By high-definition picture block XhIt is combined into high-definition picture;
S6:High resolution image reconstruction is obtained into super-resolution image.
In a kind of another embodiment of single image super-resolution reconstruction method of the present invention, the step S2, including such as
Lower step:
S21:Establish training pattern:
Wherein, YhIt is high-definition picture block, YlIt is low-resolution image block, N is that the high-definition picture block is expressed as
The length of one-dimensional vector, M are the length that the low-resolution image block is expressed as one-dimensional vector, DhFor high-resolution word to be asked
Allusion quotation, DlFor low-resolution dictionary to be asked, α is the coefficient of rarefaction representation;
S22:Using alternated process, which is split into 3 subproblem α-subproblem solution, DhSubproblem solution, Dl-
Subproblem solves, and solves until convergence.
It is described to utilize alternated process to the instruction in a kind of another embodiment of single image super-resolution reconstruction method
White silk model splits into 3 sub- problem solving methods and includes:
S22a:α-subproblem solves:
Fixed DhAnd Dl, α-subproblem is solved to:
Wherein, symbol []+The Moore-Penrose generalized inverse of representing matrix, I indicate unit matrix;
S22b:DhSubproblem solves:
Fixed α and Dl, DhSubproblem is solved to:
S22c:DlSubproblem solves:
Fixed α and Dh, DlSubproblem is solved to:
In a kind of another embodiment of single image super-resolution reconstruction method, in step s3, by the low resolution
Rate image extracts the image block of p × p size from left to right, from top to bottom with step-length s, obtains the original of the low-resolution image
The image block X of image low resolutionlSet.
In a kind of another embodiment of single image super-resolution reconstruction method, in step s 4, by the original image
Low-resolution image block XlIt is converted into corresponding high-definition picture block XhMethod it is as follows:
Xh=Dh(Dl TXl)
Wherein,
Xl--- the image block of original low-resolution;
Xh--- high-resolution image block;
Dl--- low-resolution dictionary;
Dh--- high-resolution dictionary.
In a kind of another embodiment of single image super-resolution reconstruction method, in step s 5, rebuilds and obtain oversubscription
Resolution image X*Using following methods model:
Wherein, H is fuzzy operator, and S is next sample operator adjacent in iterative calculation, the empirical parameter that c is positive, X
For super-resolution image block.
The beneficial effects of the invention are as follows:A kind of single image super-resolution alternating direction reconstruction side is provided through the invention
Method, for this method under alternating direction method theoretical frame, converting 3 for high-resolution and low-resolution dictionary training pattern has analytic solutions
Subproblem iterative solution, guarantee it is constringent under the premise of significantly improve computational efficiency.List may be implemented using this method
The rapid super-resolution of width image is rebuild, and is of great significance to image procossing and many application fields of display.
Detailed description of the invention
Fig. 1 is the flow chart of an embodiment of single image super-resolution reconstruction method according to the present invention;
Fig. 2 is the low-resolution image embodiment being originally inputted;
Fig. 3 is that low-resolution image embodiment shown in Fig. 2 is implemented by single image super-resolution reconstruction method one of the present invention
The super-resolution image that example obtains after rebuilding.
Specific embodiment
To facilitate the understanding of the present invention, in the following with reference to the drawings and specific embodiments, the present invention will be described in more detail.
A better embodiment of the invention is given in the attached drawing.But the invention can be realized in many different forms, and unlimited
In this specification described embodiment.On the contrary, purpose of providing these embodiments is makes to the disclosure
Understand more thorough and comprehensive.
It should be noted that unless otherwise defined, all technical and scientific terms used in this specification with belong to
The normally understood meaning of those skilled in the art of the invention is identical.Used term in the description of the invention
It is the purpose in order to describe specific embodiment, is not intended to the limitation present invention.
Fig. 1 is the flow chart for carrying out single image super-resolution reconstruction method according to an embodiment of the present invention.It can be with from Fig. 1
Find out, which starts from beginning, it then proceedes to execute following steps,
Step S1:High-definition image in collection network data constructs tranining database, and forms high-definition picture block YhWith
Low-resolution image block Yl。
Step S2:Using alternating direction method, the training high-resolution dictionary D from the tranining databasehWith low resolution word
Allusion quotation Dl。
Step S3:By the low-resolution image of input, division obtains the low resolution image block X of original imagelSet.
Step S4:Utilize high-resolution dictionary DhWith low-resolution dictionary DlBy the low resolution image block of the original image
XlIt is converted into corresponding high-definition picture block Xh。
Step S5:By high-definition picture block XhIt is combined into high-definition picture.
Step S6:High resolution image reconstruction is obtained into super-resolution image.
Preferably, for step S2, training high-resolution dictionary DhWith low-resolution dictionary Dl, training mould is first established first
Type:
Wherein, YhIt is high-definition picture block, YlIt is low-resolution image block, N is that the high-definition picture block is expressed as
The length of one-dimensional vector, M are the length that the low-resolution image block is expressed as one-dimensional vector, DhFor high-resolution word to be asked
Allusion quotation, DlFor low-resolution dictionary to be asked, α is the coefficient of rarefaction representation.
Then alternated process is utilized, which is split into 3 subproblem α-subproblem solution, DhSubproblem solution, Dl-
Subproblem solves, and solves until convergence, preferred solution formula is as follows:
α-subproblem solves:
Fixed DhAnd Dl, α-subproblem is solved to:
Wherein, symbol [...]+The Moore-Penrose generalized inverse of representing matrix, I indicate unit matrix.
DhSubproblem solves:
Fixed α and Dl, DhSubproblem is solved to:
DlSubproblem solves:
Fixed α and Dh, DlSubproblem is solved to:
High-resolution dictionary D is acquired by above-mentioned formula iteration convergencehWith low-resolution dictionary DlAnd rarefaction representation
Factor alpha.
Preferably, in step s3, input 128 Χ 128 low-resolution image, by the low-resolution image from a left side to
The right side is extracted the image block of 7 Χ, 7 size of one-dimensional vector length=7 with step-length s=1 from top to bottom, obtains the low resolution of original image
The image block X of ratelSet.
Preferably, in step s 4, by original image low-resolution image block XlIt is converted into corresponding high-definition picture
Block XhFormula it is as follows:
Xh=Dh(Dl TXl)
Wherein,
Xl--- the image block of original low-resolution;
Xh--- high-resolution image block;
Dl--- low-resolution dictionary;
Dh--- high-resolution dictionary.
Preferably, in step s 5, it rebuilds and obtains super-resolution image X*It uses with drag:
Wherein H uses size as 5 Χ 5, and the fuzzy operator for the Gaussian matrix that variance is 1, S is adjacent in iterative calculation
Next sample operator, the empirical parameter that c is positive, X be super-resolution image block.
The image that the super-resolution that resolution ratio is 252 Χ 252 is finally obtained by above method is as shown in Figure 3.
The beneficial effects of the invention are as follows:A kind of single image super-resolution alternating direction reconstruction side is provided through the invention
Method, for this method under alternating direction method theoretical frame, converting 3 for high-resolution and low-resolution dictionary training pattern has analytic solutions
Subproblem iterative solution, guarantee it is constringent under the premise of significantly improve computational efficiency.List may be implemented using this method
The rapid super-resolution of width image is rebuild, and is of great significance to image procossing and many application fields of display.
The above description is only an embodiment of the present invention, is not intended to limit the scope of the invention, all to utilize this hair
Equivalent structure transformation made by bright specification and accompanying drawing content is applied directly or indirectly in other relevant technical fields,
It is included within the scope of the present invention.
Claims (4)
1. a kind of single image super-resolution reconstruction method, which is characterized in that include the following steps:
S1:High-definition image in collection network data constructs tranining database, and forms high-definition picture block YhAnd low resolution
Image block Yl;
S2:Using alternating direction method, the training high-resolution dictionary D from the tranining databasehWith low-resolution dictionary Dl;
S3:By the low-resolution image of input, division obtains the low resolution image block X of original imagelSet;
S4:Utilize the high-resolution dictionary DhWith the low-resolution dictionary DlBy the low resolution image block X of the original imagel
It is converted into corresponding high-definition picture block Xh;
S5:By the high-definition picture block XhIt is combined into high-definition picture;
S6:The high resolution image reconstruction is obtained into super-resolution image;
The step S2, includes the following steps:
S21:Establish training pattern:
Wherein, YhIt is high-definition picture block, YlIt is low-resolution image block, it is one-dimensional that N is that the high-definition picture block is expressed as
The length of vector, M are the length that the low-resolution image block is expressed as one-dimensional vector, DhFor high-resolution dictionary to be asked, Dl
For low-resolution dictionary to be asked, α is the coefficient of rarefaction representation;
S22:Using alternated process, the training pattern is split into 3 sub- problem solvings:α-subproblem solution, DhSubproblem
It solves, DlSubproblem solves, and solves until convergence;
3 sub- problem solving methods include:
S22a:α-subproblem solves:
Fixed DhAnd Dl, α-subproblem is solved to:
Wherein, symbol []+The Moore-Penrose generalized inverse of representing matrix, I indicate unit matrix;
S22b:DhSubproblem solves:
Fixed α and Dl, DhSubproblem is solved to:
S22c:DlSubproblem solves:
Fixed α and Dh, DlSubproblem is solved to:
2. single image super-resolution reconstruction method according to claim 1, which is characterized in that in step s3, will be described
Low-resolution image extracts the image block of p × p size from left to right, from top to bottom with step-length s, obtains the low-resolution image
Original image low resolution image block XlSet.
3. single image super-resolution reconstruction method according to claim 2, which is characterized in that in step s 4, will be described
Original image low-resolution image block XlIt is converted into corresponding high-definition picture block XhMethod it is as follows:
Xh=Dh(Dl TXl)
Wherein,
Xl--- the image block of original low-resolution;
Xh--- high-resolution image block;
Dl--- low-resolution dictionary;
Dh--- high-resolution dictionary.
4. single image super-resolution reconstruction method according to claim 3, which is characterized in that in step s 5, reconstruction obtains
Obtain super-resolution image X*Using following methods:
Wherein, H is fuzzy operator, and S is next sample operator adjacent in iterative calculation, and the empirical parameter that c is positive, X is super
Image in different resolution block.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610394659.9A CN105957013B (en) | 2016-06-06 | 2016-06-06 | A kind of single image super-resolution reconstruction method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610394659.9A CN105957013B (en) | 2016-06-06 | 2016-06-06 | A kind of single image super-resolution reconstruction method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105957013A CN105957013A (en) | 2016-09-21 |
CN105957013B true CN105957013B (en) | 2018-11-23 |
Family
ID=56907826
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610394659.9A Expired - Fee Related CN105957013B (en) | 2016-06-06 | 2016-06-06 | A kind of single image super-resolution reconstruction method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105957013B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107203968B (en) * | 2017-05-25 | 2021-07-13 | 四川大学 | Single image super-resolution reconstruction method based on improved subspace tracking algorithm |
CN109754447B (en) | 2018-12-28 | 2021-06-22 | 上海联影智能医疗科技有限公司 | Image generation method, device, equipment and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103077511A (en) * | 2013-01-25 | 2013-05-01 | 西安电子科技大学 | Image super-resolution reconstruction method based on dictionary learning and structure similarity |
CN105335929A (en) * | 2015-09-15 | 2016-02-17 | 清华大学深圳研究生院 | Depth map super-resolution method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4600209B2 (en) * | 2005-08-18 | 2010-12-15 | ソニー株式会社 | Image processing method, image processing apparatus, program, and recording medium |
-
2016
- 2016-06-06 CN CN201610394659.9A patent/CN105957013B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103077511A (en) * | 2013-01-25 | 2013-05-01 | 西安电子科技大学 | Image super-resolution reconstruction method based on dictionary learning and structure similarity |
CN105335929A (en) * | 2015-09-15 | 2016-02-17 | 清华大学深圳研究生院 | Depth map super-resolution method |
Non-Patent Citations (1)
Title |
---|
基于改进的粒子滤波非线性约束优化算法;张凌晓 等;《计算机应用研究》;20141130;第31卷(第11期);第3265-3272页 * |
Also Published As
Publication number | Publication date |
---|---|
CN105957013A (en) | 2016-09-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Sun et al. | Deep RGB-D saliency detection with depth-sensitive attention and automatic multi-modal fusion | |
Liu et al. | Sketch2Photo: Synthesizing photo-realistic images from sketches via global contexts | |
CN109389556A (en) | The multiple dimensioned empty convolutional neural networks ultra-resolution ratio reconstructing method of one kind and device | |
WO2018120329A1 (en) | Single-frame super-resolution reconstruction method and device based on sparse domain reconstruction | |
Zhou et al. | Efficient image super-resolution using vast-receptive-field attention | |
CN105631807B (en) | The single-frame image super-resolution reconstruction method chosen based on sparse domain | |
WO2018103538A1 (en) | Deep learning method and device for analysis of high-dimensional medical data | |
CN102142137A (en) | High-resolution dictionary based sparse representation image super-resolution reconstruction method | |
CN101901472B (en) | Method for aligning non-rigid robust batch images based on matrix rank minimization | |
CN103295192B (en) | The image real-time super-resolution method for reconstructing accelerating based on GPU | |
CN105095857B (en) | Human face data Enhancement Method based on key point perturbation technique | |
Chen et al. | Single image super-resolution using deep CNN with dense skip connections and inception-resnet | |
CN105957013B (en) | A kind of single image super-resolution reconstruction method | |
Xue et al. | Automated tongue segmentation in Chinese medicine based on deep learning | |
Yi et al. | Research and improvement of convolutional neural network | |
CN104091364A (en) | Single-image super-resolution reconstruction method | |
CN115587930B (en) | Image color style migration method, device and medium | |
CN105225215A (en) | A kind of single-frame image super-resolution reconstruction method based on group rarefaction representation | |
CN102842123A (en) | Sparse-region residual error compensating and revising method for improving marginal definition during image sampling | |
Cheng et al. | TRL: Transformer based refinement learning for hybrid-supervised semantic segmentation | |
CN115661612A (en) | General climate data downscaling method based on element migration learning | |
Yin et al. | Fractional-order difference curvature-driven fractional anisotropic diffusion equation for image super-resolution | |
Wei et al. | Regression-based line detection network for delineation of largely deformed brain midline | |
Zhang et al. | Deep residual network for image super-resolution reconstruction | |
Qin et al. | Spatial-temporal feature refine network for single image super-resolution |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20181123 Termination date: 20210606 |
|
CF01 | Termination of patent right due to non-payment of annual fee |