CN107610048A - A kind of image super-resolution method returned based on projection dictionary learning and neighbour - Google Patents
A kind of image super-resolution method returned based on projection dictionary learning and neighbour Download PDFInfo
- Publication number
- CN107610048A CN107610048A CN201710679613.6A CN201710679613A CN107610048A CN 107610048 A CN107610048 A CN 107610048A CN 201710679613 A CN201710679613 A CN 201710679613A CN 107610048 A CN107610048 A CN 107610048A
- Authority
- CN
- China
- Prior art keywords
- msubsup
- mrow
- msub
- resolution
- neighbour
- 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
Landscapes
- Transforming Electric Information Into Light Information (AREA)
Abstract
The invention discloses a kind of image super-resolution method returned based on projection dictionary learning and neighbour.The present invention obtains synthesis and analytic type dictionary using projection dictionary learning method, and synthesis type dictionary base tissue neighbour is gathered, realizes the division finer to feature space.The optimal model based on 2 norms is solved, accelerates speed.When neighbour is organized, the local geometric characteristic of data is considered, more accurately recover the high-frequency information of high-definition picture, obtain higher-quality high-definition picture.
Description
Technical field
It is more particularly to a kind of based on projection dictionary learning the invention belongs to computer vision and technical field of image processing
The image super-resolution method returned with neighbour.
Background technology
Image Super-resolution belongs to computer vision and image processing field, is a classical image processing problem, has
Important science and industrial research value.The target of Image Super-resolution is exactly, by given low-resolution image, to reconstruct its phase
The high-definition picture answered so that in the case where reconstructed error is as small as possible, visual effect is good as far as possible.Main flow at present
Image super-resolution method can be divided into three major types:Method based on interpolation;Method based on reconstruct;Method based on study.
Method based on interpolation, is a kind of basic ultra-resolution method, and its processing procedure would generally use local covariance
Coefficient, fixing function core or adaptive structure core, are widely used because of the characteristics of its is simple and quick.But many situations
Under, result caused by this kind of method can produce visual artifact with the increase of multiplication factor, such as:Sawtooth effect and fuzzy effect
Should.Method based on reconstruct, it is assumed that low resolution image is obtained by high-definition picture by several degeneration factors, such as:
Down-sampled and blurring.This kind of method emphasizes the importance of Reconstruction Constraints during super-resolution, thus, its obtained high-resolution
Rate image often has excessively smooth and unnatural edge and ringing effect is produced near image border.Based on study
Method, because using machine learning techniques from training focusing study to a large amount of prioris, thus obtain more preferable result.
But this kind of method usually requires to solve and is based on l0Norm or l1The optimization problem of norm, its processing speed are very slow.
The content of the invention
In order to solve the technical problem that above-mentioned background technology proposes, the present invention is intended to provide a kind of based on projection dictionary learning
The image super-resolution method returned with neighbour, the defects of overcoming prior art to exist, obtain higher-quality high-definition picture.
In order to realize above-mentioned technical purpose, the technical scheme is that:
A kind of image super-resolution method returned based on projection dictionary learning and neighbour, is comprised the following steps:
(1) concentrate to obtain full resolution pricture set from natural image common dataIts down-sampling is obtained corresponding
Low-resolution image setWherein, subscript i represents the i-th width image in set;
(2) respectively fromWithHigh-resolution features set corresponding to middle extractionWith low resolution characteristic setClustered in low resolution feature space, closed for every a pair of cluster setsObtained using projection study dictionary
To corresponding synthesis type dictionary DcWith analytic type dictionary Ωc, wherein, c=1,2,3 ..., C, C are the number of cluster centre;
(3) basisAnd Dc, calculate high-resolution neighbour setLow resolution neighbour gathersAnd the mapping matrix set of high-low resolution feature
(4) for each low resolution featureUsing DcAnd ΩcIt is rightClassified;
(5) basisClassification results, from corresponding DcIt is middle search withImmediate dictionary base, recycling are reflected accordingly
Penetrate matrix FcReconstruct high-resolution features
(6) according to the high-resolution features set reconstructedRecover corresponding high-definition picture
Further, in step (2), by low-resolution imageUsing interpolation algorithm up-sample to corresponding high score
Distinguish the equal sized of image;Then fromMiddle extraction high-definition picture set of blocksWith high-resolution features setAgain fromIn relevant position extract low-resolution image set of blocksWith low resolution characteristic set
In above formula, f1And f2For gradient high-pass filter horizontally and vertically, f3And f4For horizontal direction and hang down
Nogata to Laplacian high-pass filter, " * " represent convolution algorithm.
Further, in step (6), by the high-resolution features set of reconstructPlus low-resolution image block collection
CloseHigh-definition picture set of blocks corresponding to obtainingWillIn all fragments compositings into a high score rate
Image
Further, in step (2), the optimization problem of projection dictionary learning is established:
In above formula, | | | |FExpression takes norm,λ is constant, dc,kFor
Synthesis type dictionary DcK-th of dictionary base;
By the above-mentioned optimization problem of solution by iterative method, synthesis type dictionary D is obtainedcWith analytic type dictionary Ωc。
Further, in step (3), for synthesis type dictionary DcK-th of dictionary base dc,k,In look for max near
Neighbour, compositionAgain fromIn look for max neighbours, formFor low resolution neighbourKth column vector,
For high-resolution neighbourKth column vector;Then the mapping matrix F of high-low resolution feature is calculated according to following formulacKth
Column vector
In above formula,ForTransposition, λ1For constant, I is unit matrix.
Further, in step (4), using following formula to a low resolution featureClassified:
In above formula, | | | |2Expression takes 2 norms, and my the optimal solution c obtained isAffiliated classification number.
Further, in step (5), in low resolution featureSynthesis type dictionary D corresponding to affiliated classification ccIn
Search withImmediate dictionary base, according to mapping matrix F corresponding to the dictionary basecColumn vectorReconstruct corresponding high score
Resolution feature
The beneficial effect brought using above-mentioned technical proposal:
The present invention obtains synthesis and analytic type dictionary using projection dictionary learning method, to synthesis type dictionary base tissue neighbour
Set, realizes the division finer to feature space.The optimal model based on 2 norms is solved, accelerates speed.In tissue neighbour
When, consider the local geometric characteristic of data, more accurately recover the high-frequency information of high-definition picture, obtain higher quality
High-definition picture.
Brief description of the drawings
Fig. 1 is the overall flow figure of the present invention.
Embodiment
Below with reference to accompanying drawing, technical scheme is described in detail.
A kind of image super-resolution method returned based on projection dictionary learning and neighbour, as shown in figure 1, specific steps are such as
Under.
Step 1:Full resolution pricture set is obtained from natural image common data sets (such as Image Net data sets)Its down-sampling is obtained into corresponding low-resolution image setWherein, subscript i represents the i-th width image in set.
Step 2:Respectively fromWithHigh-resolution features set corresponding to middle extractionWith low resolution feature set
CloseClustered in low resolution feature space, closed for every a pair of cluster setsDictionary is learnt using projection
Synthesis type dictionary D corresponding to obtainingcWith analytic type dictionary Ωc, wherein, c=1,2,3 ..., C, C are the number of cluster centre.
By low-resolution imageUp-sampled using interpolation algorithm to equal sized with corresponding full resolution pricture;Then
FromMiddle extraction high-definition picture set of blocksWith high-resolution features setAgain fromIn relevant position
Extract low-resolution image set of blocksWith low resolution characteristic set
In above formula, f1And f2For gradient high-pass filter horizontally and vertically, f3And f4For horizontal direction and hang down
Nogata to Laplacian high-pass filter, " * " represent convolution algorithm.
Establish the optimization problem of projection dictionary learning:
In above formula, | | | |FExpression takes norm,λ is constant, dc,kFor
Synthesis type dictionary DcK-th of dictionary base.
By the above-mentioned optimization problem of solution by iterative method, synthesis type dictionary D is obtainedcWith analytic type dictionary Ωc.Introduced for this
Intermediate variable Ac, by above-mentioned optimization problem, relaxation is:
In above formula, τ is constant.It is divided into two parts during iterative.
1st, fixed Dc、Ωc, update Ac:
2nd, fixed Ac, update Dc、Ωc:
Iteration 1,2, solves Dc、Ωc。
Step 3:According toAnd Dc, calculate high-resolution neighbour setLow resolution neighbour gathersAnd the mapping matrix set of high-low resolution feature
For synthesis type dictionary DcK-th of dictionary base dc,k,In look for max neighbours, formAgain fromIn look for
Max neighbours, compositionFor low resolution neighbourKth column vector,For high-resolution neighbourKth
Column vector.
In order to obtain the coefficient x required for reconstruct high-resolution features, following optimization objective function is first solved:
There are analytic solutions in the problem, concrete form is:
So corresponding high-resolution features can be tried to achieve by following formula:
The major part of above-mentioned formula is and inputIt is unrelated, thus can be come out with off-line calculation, as mapping matrix:
Wherein,ForTransposition, λ1For constant, I is unit matrix.
Step 4:For each low resolution featureUsing DcAnd ΩcIt is rightClassified.
Using following formula to a low resolution featureClassified:
In above formula, | | | |2Expression takes 2 norms, and my the optimal solution c obtained isAffiliated classification number.
Step 5:According toClassification results, from corresponding DcIt is middle search withImmediate dictionary base, recycle corresponding
Mapping matrix FcReconstruct high-resolution features
In low resolution featureSynthesis type dictionary D corresponding to affiliated classification ccIt is middle search withImmediate dictionary
Base, according to mapping matrix corresponding to the dictionary baseReconstruct corresponding high-resolution features
Step 6:According to the high-resolution features set reconstructedRecover corresponding high-definition picture
By the high-resolution features set of reconstructPlus low-resolution image set of blocksObtain corresponding high score
Resolution image block setWillIn all fragments compositings into a high score rate image
The technological thought of embodiment only to illustrate the invention, it is impossible to protection scope of the present invention is limited with this, it is every according to
Technological thought proposed by the present invention, any change done on the basis of technical scheme, each falls within the scope of the present invention.
Claims (7)
- A kind of 1. image super-resolution method returned based on projection dictionary learning and neighbour, it is characterised in that:Comprise the following steps:(1) concentrate to obtain full resolution pricture set from natural image common dataIts down-sampling is obtained into corresponding low resolution Rate image collectionWherein, subscript i represents the i-th width image in set;(2) respectively fromWithHigh-resolution features set corresponding to middle extractionWith low resolution characteristic set Clustered in low resolution feature space, closed for every a pair of cluster setsObtained correspondingly using projection study dictionary Synthesis type dictionary DcWith analytic type dictionary Ωc, wherein, c=1,2,3 ..., C, C are the number of cluster centre;(3) basisAnd Dc, calculate high-resolution neighbour setLow resolution neighbour gathersAnd Mapping matrix set { the F of high-low resolution featurec};(4) for each low resolution featureUsing DcAnd ΩcIt is rightClassified;(5) basisClassification results, from corresponding DcIt is middle search withImmediate dictionary base, recycle corresponding mapping square Battle array FcReconstruct high-resolution features(6) according to the high-resolution features set reconstructedRecover corresponding high-definition picture
- 2. the image super-resolution method returned according to claim 1 based on projection dictionary learning and neighbour, it is characterised in that: In step (2), by low-resolution imageUp-sampled using interpolation algorithm to equal sized with corresponding full resolution pricture; Then fromMiddle extraction high-definition picture set of blocksWith high-resolution features setAgain fromIn it is corresponding Position extracts low-resolution image set of blocksWith low resolution characteristic set<mrow> <msubsup> <mi>p</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>=</mo> <msubsup> <mi>p</mi> <mi>H</mi> <mi>i</mi> </msubsup> <mo>-</mo> <msubsup> <mi>y</mi> <mi>H</mi> <mi>i</mi> </msubsup> </mrow><mrow> <msubsup> <mi>y</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>=</mo> <mo>&lsqb;</mo> <msub> <mi>f</mi> <mn>1</mn> </msub> <mo>*</mo> <msubsup> <mi>p</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>;</mo> <msub> <mi>f</mi> <mn>2</mn> </msub> <mo>*</mo> <msubsup> <mi>p</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>;</mo> <msub> <mi>f</mi> <mn>3</mn> </msub> <mo>*</mo> <msubsup> <mi>p</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>;</mo> <msub> <mi>f</mi> <mn>4</mn> </msub> <mo>*</mo> <msubsup> <mi>p</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>&rsqb;</mo> </mrow>In above formula, f1And f2For gradient high-pass filter horizontally and vertically, f3And f4For horizontal direction and Vertical Square To Laplacian high-pass filter, " * " represent convolution algorithm.
- 3. the image super-resolution method returned according to claim 2 based on projection dictionary learning and neighbour, it is characterised in that: In step (6), by the high-resolution features set of reconstructPlus low-resolution image set of blocksCorresponding to obtaining High-definition picture set of blocksWillIn all fragments compositings into a high score rate image
- 4. the image super-resolution method returned according to claim 1 based on projection dictionary learning and neighbour, it is characterised in that: In step (2), the optimization problem of projection dictionary learning is established:<mrow> <munder> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> <mrow> <msub> <mi>D</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>&Omega;</mi> <mi>i</mi> </msub> </mrow> </munder> <munderover> <mo>&Sigma;</mo> <mi>c</mi> <mi>C</mi> </munderover> <mo>|</mo> <mo>|</mo> <msubsup> <mi>Y</mi> <mi>L</mi> <mi>c</mi> </msubsup> <mo>-</mo> <msub> <mi>D</mi> <mi>c</mi> </msub> <msub> <mi>&Omega;</mi> <mi>c</mi> </msub> <msubsup> <mi>Y</mi> <mi>L</mi> <mi>c</mi> </msubsup> <mo>|</mo> <msubsup> <mo>|</mo> <mi>F</mi> <mn>2</mn> </msubsup> <mo>+</mo> <mi>&lambda;</mi> <mo>|</mo> <mo>|</mo> <msub> <mi>&Omega;</mi> <mi>c</mi> </msub> <msubsup> <mover> <mi>Y</mi> <mo>&OverBar;</mo> </mover> <mi>L</mi> <mi>c</mi> </msubsup> <mo>|</mo> <msubsup> <mo>|</mo> <mi>F</mi> <mn>2</mn> </msubsup> </mrow><mfenced open = "" close = ""> <mtable> <mtr> <mtd> <mrow> <mi>s</mi> <mo>.</mo> <mi>t</mi> <mo>.</mo> </mrow> </mtd> <mtd> <mrow> <mo>|</mo> <mo>|</mo> <msub> <mi>d</mi> <mrow> <mi>c</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>|</mo> <msubsup> <mo>|</mo> <mi>F</mi> <mn>2</mn> </msubsup> <mo>&le;</mo> <mn>1</mn> </mrow> </mtd> </mtr> </mtable> </mfenced>In above formula, | | | |FExpression takes norm,λ is constant, dc,kFor synthesis type Dictionary DcK-th of dictionary base;By the above-mentioned optimization problem of solution by iterative method, synthesis type dictionary D is obtainedcWith analytic type dictionary Ωc。
- 5. the image super-resolution method returned according to claim 4 based on projection dictionary learning and neighbour, it is characterised in that: In step (3), for synthesis type dictionary DcK-th of dictionary base dc,k,In look for max neighbours, formAgain from In look for max neighbours, form For low resolution neighbourKth column vector,For high-resolution neighbour's Kth column vector;Then the mapping matrix F of high-low resolution feature is calculated according to following formulacKth column vector<mrow> <msubsup> <mi>F</mi> <mi>k</mi> <mi>c</mi> </msubsup> <mo>=</mo> <msubsup> <mi>N</mi> <mrow> <mi>H</mi> <mo>,</mo> <mi>k</mi> </mrow> <mi>c</mi> </msubsup> <msup> <mrow> <mo>(</mo> <msup> <msubsup> <mi>N</mi> <mrow> <mi>L</mi> <mo>,</mo> <mi>k</mi> </mrow> <mi>c</mi> </msubsup> <mi>T</mi> </msup> <msubsup> <mi>N</mi> <mrow> <mi>L</mi> <mo>,</mo> <mi>k</mi> </mrow> <mi>c</mi> </msubsup> <mo>+</mo> <msub> <mi>&lambda;</mi> <mn>1</mn> </msub> <mi>I</mi> <mo>)</mo> </mrow> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msup> <msubsup> <mi>N</mi> <mrow> <mi>L</mi> <mo>,</mo> <mi>k</mi> </mrow> <mi>c</mi> </msubsup> </mrow>In above formula,ForTransposition, λ1For constant, I is unit matrix.
- 6. the image super-resolution method returned according to claim 1 based on projection dictionary learning and neighbour, it is characterised in that: In step (4), using following formula to a low resolution featureClassified:<mrow> <munder> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> <mi>c</mi> </munder> <mo>|</mo> <mo>|</mo> <msubsup> <mi>y</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>-</mo> <msub> <mi>D</mi> <mi>c</mi> </msub> <msub> <mi>&Omega;</mi> <mi>c</mi> </msub> <msubsup> <mi>y</mi> <mi>L</mi> <mi>i</mi> </msubsup> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> </mrow>In above formula, | | | |2Expression takes 2 norms, and my the optimal solution c obtained isAffiliated classification number.
- 7. the image super-resolution method returned according to claim 1 based on projection dictionary learning and neighbour, it is characterised in that: In step (5), in low resolution featureSynthesis type dictionary D corresponding to affiliated classification ccIt is middle search withImmediate word Allusion quotation base, according to mapping matrix F corresponding to the dictionary basecColumn vectorReconstruct corresponding high-resolution features
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710679613.6A CN107610048A (en) | 2017-08-10 | 2017-08-10 | A kind of image super-resolution method returned based on projection dictionary learning and neighbour |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710679613.6A CN107610048A (en) | 2017-08-10 | 2017-08-10 | A kind of image super-resolution method returned based on projection dictionary learning and neighbour |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107610048A true CN107610048A (en) | 2018-01-19 |
Family
ID=61065176
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710679613.6A Pending CN107610048A (en) | 2017-08-10 | 2017-08-10 | A kind of image super-resolution method returned based on projection dictionary learning and neighbour |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107610048A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113628108A (en) * | 2021-07-05 | 2021-11-09 | 上海交通大学 | Image super-resolution method and system based on discrete representation learning and terminal |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104899830A (en) * | 2015-05-29 | 2015-09-09 | 清华大学深圳研究生院 | Image super-resolution method |
CN105787899A (en) * | 2016-03-03 | 2016-07-20 | 河海大学 | Rapid image super-resolution method based on self-adaptive regression |
-
2017
- 2017-08-10 CN CN201710679613.6A patent/CN107610048A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104899830A (en) * | 2015-05-29 | 2015-09-09 | 清华大学深圳研究生院 | Image super-resolution method |
CN105787899A (en) * | 2016-03-03 | 2016-07-20 | 河海大学 | Rapid image super-resolution method based on self-adaptive regression |
Non-Patent Citations (1)
Title |
---|
YIHUI FENG 等: "Single Image Super-Resolution via Projective Dictionary Learning with Anchored Neighborhood Regression", 《VCIP》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113628108A (en) * | 2021-07-05 | 2021-11-09 | 上海交通大学 | Image super-resolution method and system based on discrete representation learning and terminal |
CN113628108B (en) * | 2021-07-05 | 2023-10-27 | 上海交通大学 | Image super-resolution method and system based on discrete representation learning and terminal |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Fang et al. | Soft-edge assisted network for single image super-resolution | |
CN104899830B (en) | A kind of image super-resolution method | |
CN106204447A (en) | The super resolution ratio reconstruction method with convolutional neural networks is divided based on total variance | |
CN102156875B (en) | Image super-resolution reconstruction method based on multitask KSVD (K singular value decomposition) dictionary learning | |
CN101950365B (en) | Multi-task super-resolution image reconstruction method based on KSVD dictionary learning | |
CN103136727B (en) | Based on the super resolution image reconstruction method of gradient consistency and anisotropy regularization | |
CN110136062B (en) | Super-resolution reconstruction method combining semantic segmentation | |
CN105488776B (en) | Super-resolution image reconstruction method and device | |
CN103366347B (en) | Image super-resolution rebuilding method based on rarefaction representation | |
CN105023240A (en) | Dictionary-type image super-resolution system and method based on iteration projection reconstruction | |
CN111626927B (en) | Binocular image super-resolution method, system and device adopting parallax constraint | |
CN104992407B (en) | A kind of image super-resolution method | |
CN107784628A (en) | A kind of super-resolution implementation method based on reconstruction optimization and deep neural network | |
Zhang et al. | Efficient sparse representation based image super resolution via dual dictionary learning | |
CN111861886B (en) | Image super-resolution reconstruction method based on multi-scale feedback network | |
Chen et al. | Single image super resolution using local smoothness and nonlocal self-similarity priors | |
Gendy et al. | Lightweight image super-resolution based on deep learning: State-of-the-art and future directions | |
CN109191377A (en) | A kind of image magnification method based on interpolation | |
Li et al. | Multiple degradation and reconstruction network for single image denoising via knowledge distillation | |
Yang et al. | An image super-resolution network based on multi-scale convolution fusion | |
CN108765287B (en) | Image super-resolution method based on non-local mean value | |
JP2023532755A (en) | Computer-implemented method, computer program product, and system for processing images | |
CN107610048A (en) | A kind of image super-resolution method returned based on projection dictionary learning and neighbour | |
CN102842123A (en) | Sparse-region residual error compensating and revising method for improving marginal definition during image sampling | |
CN104123707B (en) | Local rank priori based single-image super-resolution reconstruction method |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180119 |