CN109766863A - A kind of face image super-resolution method based on local and sparse non local canonical - Google Patents
A kind of face image super-resolution method based on local and sparse non local canonical Download PDFInfo
- Publication number
- CN109766863A CN109766863A CN201910051451.0A CN201910051451A CN109766863A CN 109766863 A CN109766863 A CN 109766863A CN 201910051451 A CN201910051451 A CN 201910051451A CN 109766863 A CN109766863 A CN 109766863A
- Authority
- CN
- China
- Prior art keywords
- local
- resolution
- block
- dictionary
- definition picture
- 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.)
- Granted
Links
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a kind of face image super-resolution methods based on local and sparse non local canonical, comprising the following steps: step 1: obtaining the image block of test image and each location of pixels of training sample image；Step 2: using part PCA dictionary learning method, uses K mean cluster algorithm by image block partition clustering, one part PCA dictionary of each clustering learning training sample image block；Step 3: to low-quality image block, best expression coefficient vector is solved with based on local restriction and sparse non local double-core norm canonical algorithm；Step 4: it indicates that coefficient vector synthesizes high-definition picture block on high-resolution dictionary using best, updates non local code coefficient, updated coefficient and high-definition picture block are put into step 3 and carry out next iteration；It updates to obtain high-definition picture block by successive ignition；Step 5: output high-definition picture.The present invention has the advantages that improve picture quality.
Description
Technical field
The present invention relates to technical field of image processing, and in particular to a kind of face based on local and sparse non local canonical
Image super-resolution method.
Background technique
With the development of information technology, people increasingly improve the processing requirement of facial image, for example, in intelligent monitoring system
In system, there is the high quality graphic of abundant details extremely important, they can effectively improve system performance, will obtain in monitoring environment
Low-resolution image become high-definition picture also just it is most important.It is this to restore high-definition picture from low-resolution image
Method, be exactly human face super-resolution technology.
Existing face image super-resolution method is broadly divided into two classes, and one kind is the technology based on reconstruction, another kind of to be
Technology based on study.Compared with the technology based on reconstruction, as long as one group of training sample, the technology based on study can also be obtained more
Stablize, better performance.
Super-resolution method based on study is according to low quality and high quality training sample to predicting a high quality
Image.Wherein, the method based on study can be divided into again based on the overall situation and the technology based on part.It may based on global technology
The details that low quality input picture can be lost, such as edge and texture information.And the technology based on part is extracted from general image
Small image block can save more details.
It can be mainly divided into currently based on the face super-resolution method of part: least square error model, sparse coding mould
Type, nuclear norm reconstruction model.Least square error model describes error distribution to all training samples.Initial based on office
Be widely used in the technology in portion, but it is very weak to the robustness of noise, usually can be affected by noise and generate fuzzy.It is sparse
Encoding model only needs part sample to participate in the synthesis of high-definition picture, also there is stronger robustness to noise.Nuclear norm mould
Type is to the summation of the singular value of the error matrix of composograph, and error can be described in more detail below in this method first two that compares
Distribution, and there is stronger robustness.
Certain methods before assume that low-resolution image block and high-definition picture block have identical topology, so will
The code coefficient that low resolution test image block generates is directly used in synthesis high-definition picture block, but is ignored as high-resolution in this way
Rate image block topology, this may be such that coefficient is distorted, influence picture quality.
Summary of the invention
The object of the present invention is to provide a kind of detailed information that loss can be compensated in image synthesizing procedure, improve image matter
The face image super-resolution method based on local restriction and sparse non local double-core norm canonical of amount.
To achieve the above object, present invention employs following technical solutions: described one kind is based on part and sparse non-office
The face image super-resolution method of portion's canonical, comprising the following steps:
Step 1: centered on location of pixels each in image, test image and each pixel of training sample image are obtained
The image block of position；
Step 2: using part PCA dictionary learning method, will using K mean cluster algorithm for training sample image block
Image block partition clustering, one part PCA dictionary of each clustering learning, while calculating the center of each cluster；For every
A image block to be synthesized, with the dictionary encoding with its maximally related cluster；
Step 3: to the low-quality image block of input, with based on local restriction and sparse non local double-core norm canonical
Algorithm solves best expression coefficient vector；
Step 4: it indicates that coefficient vector synthesizes high-definition picture block on high-resolution dictionary using best, keeps low
Quality signal is constant, updates non local code coefficient, by updated coefficient and high-definition picture block be put into step 3 into
Row next iteration；Best expression coefficient vector is updated by successive ignition, obtains the high-definition picture block of the position；
Step 5: each high-definition picture block opsition dependent composograph is averaged lap, obtains final defeated
High-definition picture out.
Further, a kind of face image super-resolution method based on local and sparse non local canonical above-mentioned,
In: in step 3, the best method for solving for indicating coefficient vector s is as follows:
Face image super-resolution algorithm based on local restriction and sparse non local double-core norm canonical, to low resolution
The double-core norm regular function of coupling coding is carried out with high-definition picture block:
Wherein, | | | |*The sum of all singular values of the nuclear norm of representing matrix, i.e. matrix,Product code in indicating, i.e., pair
Answer element multiplication；yLAnd yHRespectively indicate the low-resolution image block and high-definition picture block of input；DL,DHIt is low point respectively
Resolution and high-resolution dictionary, wherein D (s)=s1D1+s2D2+L+sNDNIt is from spaceIt arrivesOne linearly reflect
It penetrates, N is the number of plies of dictionary；λ is the first regularization parameter, and α is the second regularization parameter, and u is non-local code coefficient vector；β
It is third regularization parameter, c is the Euclidean distance vector between the low-resolution image block and each dictionary atom of input；
Model above can further indicate that are as follows:
s.t.EL=yL-DL(s),EH=yH-DH(s)
Its Lagrangian indicates are as follows:
Wherein Z1, Z2It is Lagrange multiplier, μ is the 4th regularization parameter；
Model is solved using alternating direction multipliers method ADMM, detailed process are as follows:
Step (a): fixed EL, EH, update s:
Then:
Introduce following auxiliary variable:
AL=vec (DL), AH=vec (DH)
Then:
It is indicated to simplify, is introduced back into auxiliary variable:
Then:
Wherein, C is the diagonal matrix of a K × K, and Cmm=cm；F=UTU, and U=B-G1T；
Step (b): fixed s, EH, update EL:
It is solved using singular value Thresholding optimal
Wherein,It is E after the K+1 times iterationLValue；
It is rightSingular value decomposition is carried out, is acquired:
Wherein,R is the order of diagonal matrix Σ；
Step (c): fixed s, EL, update EH:
It is solved using singular value Thresholding optimal
Wherein,It is E after the K+1 times iterationHValue；
Step (d): Lagrange multiplier is updated:
Wherein,It is Z after K+1 iteration respectively1,Z2Value.
Further, a kind of face image super-resolution method based on local and sparse non local canonical above-mentioned,
In: in step 4, use the best specific table of high-definition picture block for indicating coefficient vector and synthesizing on high-resolution dictionary
It is shown as:Wherein,Indicate the high-definition picture block updated, AHIndicate the high-resolution dictionary of vectorization,Table
Show the expression coefficient of update.
Further, a kind of face image super-resolution method based on local and sparse non local canonical above-mentioned,
In: in step 4, non local code coefficient is specifically expressed as follows:Wherein,Indicate the non local coding updated
Coefficient, AHIndicate the high-resolution dictionary of vectorization,Indicate the high-definition picture block updated.
Through the implementation of the above technical solution, the invention has the advantages that: high-resolution is added in loss function
Rate image block regularization, so as to compensate the detailed information of loss in image synthesizing procedure, and will the sparse and non-office in part
Self similarity two priori knowledges in portion's permeate a regularization term, reach better binding effect, improve picture quality.
Detailed description of the invention
Fig. 1 is a kind of face image super-resolution method based on local and sparse non local canonical of the present invention
Flow chart.
Specific embodiment
Technical solution of the present invention is described in further detail in the following with reference to the drawings and specific embodiments:
As shown in Figure 1, a kind of face image super-resolution method based on local and sparse non local canonical, packet
Include following steps:
Step 1: centered on location of pixels each in image, test image and each pixel of training sample image are obtained
The image block of position；
Step 2: using part PCA dictionary learning method, will using K mean cluster algorithm for training sample image block
Image block partition clustering, one part PCA dictionary of each clustering learning, while calculating the center of each cluster；For every
A image block to be synthesized, with the dictionary encoding with its maximally related cluster；
Step 3: to the low-quality image block of input, with based on local restriction and sparse non local double-core norm canonical
Algorithm solves best expression coefficient vector；
Wherein, most preferably indicate that the method for solving of coefficient vector s is as follows:
Face image super-resolution algorithm based on local restriction and sparse non local double-core norm canonical, to low resolution
The double-core norm regular function of coupling coding is carried out with high-definition picture block:
Wherein, | | | |*The sum of all singular values of the nuclear norm of representing matrix, i.e. matrix,Product code in indicating, i.e., pair
Answer element multiplication；yLAnd yHRespectively indicate the low-resolution image block and high-definition picture block of input；DL,DHIt is low point respectively
Resolution and high-resolution dictionary, wherein D (s)=s1D1+s2D2+L+sNDNIt is from spaceIt arrivesOne linearly reflect
It penetrates, N is the number of plies of dictionary；λ is the first regularization parameter, and α is the second regularization parameter, and u is non-local code coefficient vector；β
It is third regularization parameter, c is the Euclidean distance vector between the low-resolution image block and each dictionary atom of input；
Model above can further indicate that are as follows:
s.t.EL=yL-DL(s),EH=yH-DH(s)
Its Lagrangian indicates are as follows:
Wherein Z1, Z2It is Lagrange multiplier, μ is the 4th regularization parameter；
Model is solved using alternating direction multipliers method ADMM, detailed process are as follows:
Step (a): fixed EL, EH, update s:
Then:
Introduce following auxiliary variable:
AL=vec (DL), AH=vec (DH)
Then:
It is indicated to simplify, is introduced back into auxiliary variable:
Then:
Wherein, C is the diagonal matrix of a K × K, and Cmm=cm；F=UTU, and U=B-G1T；
Step (b): fixed s, EH, update EL:
It is solved using singular value Thresholding optimal
Wherein,It is E after the K+1 times iterationLValue；
It is rightSingular value decomposition is carried out, is acquired:
Wherein,R is the order of diagonal matrix Σ；
Step (c): fixed s, EL, update EH:
It is solved using singular value Thresholding optimal
Wherein,It is E after the K+1 times iterationHValue；
Step (d): Lagrange multiplier is updated:
Wherein,It is Z after K+1 iteration respectively1,Z2Value.
Step 4: indicate that coefficient vector synthesizes high-definition picture block on high-resolution dictionary using bestIt keeps low-quality signal constant, updates non local code coefficientBy updated coefficientAnd high score
Resolution image blockIt is put into step 3 and carries out next iteration；Best expression coefficient vector is updated by successive ignition?
To the high-definition picture block of the position；
Wherein, it is specifically indicated using the best high-definition picture block for indicating that coefficient vector synthesizes on high-resolution dictionary
Are as follows:Wherein,Indicate the high-definition picture block updated, AHIndicate the high-resolution dictionary of vectorization,It indicates
The expression coefficient of update；
Wherein, non local code coefficient is specifically expressed as follows:Wherein,Indicate the non local coding system updated
Number, AHIndicate the high-resolution dictionary of vectorization,Indicate the high-definition picture block updated；
Step 5: each high-definition picture block opsition dependent composograph is averaged lap, obtains final defeated
High-definition picture out.
Present invention has the advantage that the regularization of high-definition picture block is added, in loss function so as in image
The detailed information of loss is compensated in synthesis process, and sparse and non local two priori knowledges of self similarity in part are permeated
Item regularization term, reaches better binding effect, improves picture quality.
The above, the only specific embodiment in the present invention, but scope of protection of the present invention is not limited thereto, appoints
What is familiar with the people of the technology within the technical scope disclosed by the invention, it will be appreciated that expects transforms or replaces, and should all cover
Within scope of the invention, therefore, the scope of protection of the invention shall be subject to the scope of protection specified in the patent claim.
Claims (4)
1. a kind of face image super-resolution method based on local and sparse non local canonical, it is characterised in that: including following
Step:
Step 1: centered on location of pixels each in image, test image and each location of pixels of training sample image are obtained
Image block；
Step 2: using part PCA dictionary learning method, for training sample image block, using K mean cluster algorithm by image
Block partition clustering, one part PCA dictionary of each clustering learning, while calculating the center of each cluster；For it is each to
Composograph block, with the dictionary encoding with its maximally related cluster；
Step 3: to the low-quality image block of input, with based on local restriction and sparse non local double-core norm canonical algorithm
Solve best expression coefficient vector；
Step 4: it indicates that coefficient vector synthesizes high-definition picture block on high-resolution dictionary using best, keeps low quality
Signal is constant, updates non local code coefficient, and updated coefficient and high-definition picture block are put into step 3 and are carried out down
An iteration；Best expression coefficient vector is updated by successive ignition, obtains the high-definition picture block of the position；
Step 5: each high-definition picture block opsition dependent composograph is averaged lap, obtains final output
High-definition picture.
2. a kind of face image super-resolution method based on local and sparse non local canonical according to claim 1,
It is characterized by: the best method for solving for indicating coefficient vector s is as follows in step 3:
Face image super-resolution algorithm based on local restriction and sparse non local double-core norm canonical, to low resolution and height
Image in different resolution block carries out the double-core norm regular function of coupling coding:
Wherein, | | | |*The sum of all singular values of the nuclear norm of representing matrix, i.e. matrix,Product code in indicating, i.e. corresponding element
Element is multiplied；yLAnd yHRespectively indicate the low-resolution image block and high-definition picture block of input；DL,DHIt is low resolution respectively
With high-resolution dictionary, wherein D (s)=s1D1+s2D2+L+sNDNIt is from spaceIt arrivesA Linear Mapping, N is
The number of plies of dictionary；λ is the first regularization parameter, and α is the second regularization parameter, and u is non-local code coefficient vector；β is third
Regularization parameter, c are the Euclidean distance vectors between the low-resolution image block and each dictionary atom of input；
Model above can further indicate that are as follows:
s.t.EL=yL-DL(s),EH=yH-DH(s)
Its Lagrangian indicates are as follows:
Wherein Z1, Z2It is Lagrange multiplier, μ is the 4th regularization parameter；
Model is solved using alternating direction multipliers method ADMM, detailed process are as follows: step (a): fixed EL, EH, update s:
Then:
Introduce following auxiliary variable:
AL=vec (DL), AH=vec (DH)
Then:
It is indicated to simplify, is introduced back into auxiliary variable:
Then:
Wherein, C is the diagonal matrix of a K × K, and Cmm=cm；F=UTU, and U=B-G1T；
Step (b): fixed s, EH, update EL:
It is solved using singular value Thresholding optimal
Wherein,It is E after the K+1 times iterationLValue；
It is rightSingular value decomposition is carried out, is acquired:
Wherein,R is the order of diagonal matrix Σ；
Step (c): fixed s, EL, update EH:
It is solved using singular value Thresholding optimal
Wherein,It is E after the K+1 times iterationHValue；
Step (d): Lagrange multiplier is updated:
Wherein,It is Z after K+1 iteration respectively1,Z2Value.
3. a kind of face image super-resolution method based on local and sparse non local canonical according to claim 1,
It is characterized by: using the best high-definition picture for indicating coefficient vector and synthesizing on high-resolution dictionary in step 4
Block is embodied as:Wherein,Indicate the high-definition picture block updated, AHIndicate the high-resolution of vectorization
Dictionary,Indicate the expression coefficient updated.
4. a kind of face image super-resolution method based on local and sparse non local canonical according to claim 1,
It is characterized by: non local code coefficient is specifically expressed as follows in step 4:Wherein,It indicates to update non-
Local code coefficient, AHIndicate the high-resolution dictionary of vectorization,Indicate the high-definition picture block updated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910051451.0A CN109766863B (en) | 2019-01-18 | 2019-01-18 | Face image super-resolution method based on local and sparse non-local regularization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910051451.0A CN109766863B (en) | 2019-01-18 | 2019-01-18 | Face image super-resolution method based on local and sparse non-local regularization |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109766863A true CN109766863A (en) | 2019-05-17 |
| CN109766863B CN109766863B (en) | 2022-09-06 |
Family
ID=66454709
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910051451.0A Active CN109766863B (en) | 2019-01-18 | 2019-01-18 | Face image super-resolution method based on local and sparse non-local regularization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109766863B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110288508A (en) * | 2019-06-26 | 2019-09-27 | 复旦大学 | Towards being uniformly distributed superelevation partial image locally sparse accelerated processing method and application |
| CN110443255A (en) * | 2019-06-14 | 2019-11-12 | 深圳大学 | The local retentivity homing method of relaxation for image characteristics extraction |
| CN110619603A (en) * | 2019-08-29 | 2019-12-27 | 浙江师范大学 | Single image super-resolution method for optimizing sparse coefficient |
| CN110648276A (en) * | 2019-09-25 | 2020-01-03 | 重庆大学 | High-dimensional image data dimension reduction method based on manifold mapping and dictionary learning |
| CN111340697A (en) * | 2020-02-16 | 2020-06-26 | 西安工程大学 | Clustering regression-based image super-resolution method |
| CN113033602A (en) * | 2021-02-05 | 2021-06-25 | 西北民族大学 | Image clustering method based on tensor low-rank sparse representation |
| CN113628109A (en) * | 2021-07-16 | 2021-11-09 | 上海交通大学 | Human face five sense organs super-resolution method, system and medium based on learnable dictionary |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103295196A (en) * | 2013-05-21 | 2013-09-11 | 西安电子科技大学 | Super-resolution image reconstruction method based on non-local dictionary learning and biregular terms |
| CN107194873A (en) * | 2017-05-11 | 2017-09-22 | 南京邮电大学 | Low-rank nuclear norm canonical facial image ultra-resolution method based on coupling dictionary learning |
| US20180232857A1 (en) * | 2015-11-04 | 2018-08-16 | Peking University Shenzhen Graduate School | Method and device for super-resolution image reconstruction based on dictionary matching |
-
2019
- 2019-01-18 CN CN201910051451.0A patent/CN109766863B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103295196A (en) * | 2013-05-21 | 2013-09-11 | 西安电子科技大学 | Super-resolution image reconstruction method based on non-local dictionary learning and biregular terms |
| US20180232857A1 (en) * | 2015-11-04 | 2018-08-16 | Peking University Shenzhen Graduate School | Method and device for super-resolution image reconstruction based on dictionary matching |
| CN107194873A (en) * | 2017-05-11 | 2017-09-22 | 南京邮电大学 | Low-rank nuclear norm canonical facial image ultra-resolution method based on coupling dictionary learning |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110443255B (en) * | 2019-06-14 | 2022-04-15 | 深圳大学 | Image recognition method for image feature extraction |
| CN110443255A (en) * | 2019-06-14 | 2019-11-12 | 深圳大学 | The local retentivity homing method of relaxation for image characteristics extraction |
| CN110288508A (en) * | 2019-06-26 | 2019-09-27 | 复旦大学 | Towards being uniformly distributed superelevation partial image locally sparse accelerated processing method and application |
| CN110288508B (en) * | 2019-06-26 | 2023-02-24 | 复旦大学 | Uniform distribution-oriented ultra-high resolution image local sparse acceleration processing method and application |
| CN110619603A (en) * | 2019-08-29 | 2019-12-27 | 浙江师范大学 | Single image super-resolution method for optimizing sparse coefficient |
| CN110619603B (en) * | 2019-08-29 | 2023-11-10 | 浙江师范大学 | Single image super-resolution method for optimizing sparse coefficient |
| CN110648276A (en) * | 2019-09-25 | 2020-01-03 | 重庆大学 | High-dimensional image data dimension reduction method based on manifold mapping and dictionary learning |
| CN111340697B (en) * | 2020-02-16 | 2023-08-25 | 浙江昕微电子科技有限公司 | Image super-resolution method based on clustered regression |
| CN111340697A (en) * | 2020-02-16 | 2020-06-26 | 西安工程大学 | Clustering regression-based image super-resolution method |
| CN113033602B (en) * | 2021-02-05 | 2022-07-05 | 西北民族大学 | Image clustering method based on tensor low-rank sparse representation |
| CN113033602A (en) * | 2021-02-05 | 2021-06-25 | 西北民族大学 | Image clustering method based on tensor low-rank sparse representation |
| CN113628109A (en) * | 2021-07-16 | 2021-11-09 | 上海交通大学 | Human face five sense organs super-resolution method, system and medium based on learnable dictionary |
| CN113628109B (en) * | 2021-07-16 | 2023-11-28 | 上海交通大学 | Face five sense organs super-resolution method, system and medium based on learning dictionary |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109766863B (en) | 2022-09-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109766863A (en) | A kind of face image super-resolution method based on local and sparse non local canonical | |
| CN109035142B (en) | Satellite image super-resolution method combining countermeasure network with aerial image prior | |
| CN113177882B (en) | Single-frame image super-resolution processing method based on diffusion model | |
| CN112541864A (en) | Image restoration method based on multi-scale generation type confrontation network model | |
| CN109598220B (en) | People counting method based on multi-input multi-scale convolution | |
| CN113033570A (en) | Image semantic segmentation method for improving fusion of void volume and multilevel characteristic information | |
| CN110363068B (en) | High-resolution pedestrian image generation method based on multiscale circulation generation type countermeasure network | |
| CN110060286B (en) | Monocular depth estimation method | |
| CN111340697B (en) | Image super-resolution method based on clustered regression | |
| CN111696038A (en) | Image super-resolution method, device, equipment and computer-readable storage medium | |
| CN113628201A (en) | Deep learning-based pathological section analysis method, electronic device and readable storage medium | |
| CN111402138A (en) | Image super-resolution reconstruction method of supervised convolutional neural network based on multi-scale feature extraction fusion | |
| CN114360067A (en) | Dynamic gesture recognition method based on deep learning | |
| CN112614070B (en) | defogNet-based single image defogging method | |
| CN107194873A (en) | Low-rank nuclear norm canonical facial image ultra-resolution method based on coupling dictionary learning | |
| CN110009018A (en) | A kind of image generating method, device and relevant device | |
| CN112950480A (en) | Super-resolution reconstruction method integrating multiple receptive fields and dense residual attention | |
| CN110796022A (en) | Low-resolution face recognition method based on multi-manifold coupling mapping | |
| CN111611962A (en) | Face image super-resolution identification method based on fractional order multi-set partial least square | |
| CN109947960B (en) | Face multi-attribute joint estimation model construction method based on depth convolution | |
| CN110020986B (en) | Single-frame image super-resolution reconstruction method based on Euclidean subspace group double-remapping | |
| CN116597142A (en) | Satellite image semantic segmentation method and system based on full convolution neural network and converter | |
| CN116385281A (en) | Remote sensing image denoising method based on real noise model and generated countermeasure network | |
| CN112164078B (en) | RGB-D multi-scale semantic segmentation method based on encoder-decoder | |
| CN113313127B (en) | Text image recognition method and device, computer equipment and storage medium |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |