CN108492252A - Face image super-resolution reconstruction method based on secondary reconstruction - Google Patents
Face image super-resolution reconstruction method based on secondary reconstruction Download PDFInfo
- Publication number
- CN108492252A CN108492252A CN201810229088.2A CN201810229088A CN108492252A CN 108492252 A CN108492252 A CN 108492252A CN 201810229088 A CN201810229088 A CN 201810229088A CN 108492252 A CN108492252 A CN 108492252A
- Authority
- CN
- China
- Prior art keywords
- resolution
- image
- low
- reconstruction
- block
- 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.)
- Withdrawn
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 face image super-resolution reconstruction methods based on secondary reconstruction.The method of the present invention is on the basis of the image super-resolution reconstructing method based on study, secondary reconstruction is carried out to its reconstructed results, based on principle as Local Phase, the formed matrix of similar image block carries out secondary reconstruct by way of carrying out SVD decomposition in neighborhood, the blocking artifact problem of super-resolution reconstruction facial image can be effectively eliminated, face image super-resolution rebuilding effect is improved.The present invention further improves the method for first time super-resolution image reconstruction, high-low resolution transition matrix is obtained during dictionary learning, and the rarefaction representation coefficient of high-definition picture is optimized using high-low resolution transition matrix, to effectively solve the problems, such as that high-low resolution image manifold is inconsistent, further improves the practicability of method and rebuild effect.Compared with prior art, there is the present invention better image to rebuild effect.
Description
Technical field
The present invention relates to image super-resolution rebuilding technical field more particularly to a kind of facial images based on secondary reconstruction
Super resolution ratio reconstruction method.
Background technology
With popularizing for video monitoring system, facial image takes in internet finance, safety precaution, criminal investigation and court
The fields such as card have played increasingly important role.In practical applications it is generally desirable that obtaining high-resolution (High
Resolution, HR) facial image, because the resolution ratio of facial image is higher, the image detail that can be provided is more, to reality
The value of application is higher.But in the video monitoring system of practical application, especially shooting distance personage farther out when, often
What is obtained is low resolution (Low Resolution, LR) facial image.Therefore, how low-resolution image to be surpassed by image
It seems a valuable research contents that resolution ratio (Super Resolution, SR) reconfiguration technique, which constructs high resolution graphics,.
Super-Resolution Image Restoration is broadly divided into 3 types, i.e.,：Based on interpolation, based on reconstruction and based on study
Reconstructing method.Main thought based on interpolation method is：First the relative movement information between each frame image is estimated, to obtain
Pixel value of the HR images on non-homogeneous spacing sampled point is obtained, the pixel on HR grids is obtained by non-homogeneous interpolation in next step
Value restores removal finally by image and obscures and reduce noise.Main thought based on method for reconstructing is：Assuming that super-resolution
Image is under deformation appropriate, translation and sub-sampling and noise jamming, using multiframe low-resolution image as data consistency
Constraint, and the priori of image is combined to be solved.Method based on study is calculated using given training image collection
Then neighborhood relationships between test sample patch and training atlas patch construct best initial weights constraint and know to obtain priori
Know, the high score rate image of final approach test sample.
Method based on study does not need too many Image Priori Knowledge, thus has obtained further development.Document (IEEE
Computer Graphics and Application,2002,22(2):56-65) the Example-based methods proposed,.
Document (Proceeding of IEEE Conf CVPR.Washington, DC:IEEE Press,2004:275-282) propose
Field embedding grammar.Document (Proc.of International Conference on Computer Vision and
Pattern Recognition,2008,2:729-73) the image super-resolution based on overcomplete sparse representation theory proposed
Algorithm for reconstructing.But these super-resolution rebuilding technologies are all based on manifold congruity theory, it is assumed that high-low resolution image it is dilute
It dredges and indicates that coefficient has space geometry consistency.But in practical applications, lead to height point due to being influenced by various interference
Resolution image sparse indicates that the manifold between coefficient is not consistent, this certainly will generate negative effect to the effect of reconstruction.In addition,
Due to being reconstructed using fritter, so can also produce blocking artifact when reconstruct.
Facial image is the image of a kind of specific area.In terms of face image super-resolution rebuilding, document
(Proceedings of the 4th International Conference on Automatic Face and
Gesture Recognition,Washington D C,2000:" illusory face " algorithm 83-89) is proposed for the first time, introduces image
Gradient prior information, i.e., the laplacian pyramid of image, the single order of gaussian pyramid and second order gradient are as feature space
It is trained, document (Proceedings of IEEE Int Conf on Computer Vision and Pattern
Recognition.Hawaii,USA,2001:192-198) propose the algorithm that global information and local message are combined.Document
(Beijing Institute of Technology's journal, 2012,32 (4):The face image super-resolution reconstruct based on study 386-389) is proposed to calculate
Method.But main problem is existing for these face image super-resolution rebuilding algorithms：It can only be to the front Jing Guo image registration
Facial image carries out super-resolution rebuilding.
In conclusion existing face image super-resolution technology is primarily present following problem：
(1) face image super-resolution rebuilding algorithm is for the front face image Jing Guo image registration, i.e. people mostly
Face is basic position, and carries out position correction according to eyes coordinate.This kind of inconvenient face of the processing with decoration of algorithm
Image, such as：People often wear glasses, with the just not so good processing of these algorithms of the facial image of glasses, so these are calculated
The practicability of method is not satisfactory.It is actually rare that the valuable achievement of super-resolution reconstruction is directly carried out to facial image, is needed more
More further investigations.
(2) the super-resolution rebuilding technology based on study is typically based on manifold congruity theory, it is assumed that high-low resolution figure
The rarefaction representation coefficient of picture has space geometry consistency.But in practical applications, cause due to being influenced by various interference
High-low resolution image sparse indicates that the manifold between coefficient is not consistent, this certainly will generate the effect of reconstruction passive shadow
It rings.In addition, due to being reconstructed using block form, so also will produce blocking artifact when reconstruct.
Invention content
The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and to provide a kind of people based on secondary reconstruction
Face image super resolution ratio reconstruction method can effectively eliminate the blocking artifact problem of super-resolution reconstruction facial image, improvement method
Practicability.
The present invention specifically uses following technical scheme to solve above-mentioned technical problem：
Based on the face image super-resolution reconstruction method of secondary reconstruction, first with the image super-resolution based on study
Reconstructing method carries out first time super-resolution image reconstruction to low-resolution face image to be reconstructed, obtains initial high resolution
Human face rebuilding image；Then secondary reconstruction is carried out to initial high resolution human face rebuilding image using following methods, obtained final
Reconstructing high-resolution human face image：
Step 1, by initial high resolution human face rebuilding image segmentation be the identical image block of one group of size；
Step 2, for each image block, find out most like with it c image block in its neighborhood, while record is each
The position of image block in the picture, and this c most like image blocks are reconstructed in accordance with the following methods：By this c most phases
As image block change into one-dimensional vector respectively, then c one-dimensional vector is combined into a matrix；Then to the matrix that is formed into
Row singular value decomposition, and the singular value to absolute value less than threshold value T is zeroed out processing；It is restored into row matrix, and by reduction
Matrix becomes c image block again, and c image block is put back into image home position；C is preset positive integer；
Step 3, by initial high resolution human face rebuilding image it is all participated in reconstruct pixel, with its reconstruct after pixel
The mean value of value is its progress again assignment to get to final reconstructing high-resolution human face image.
Further, in order to solve the problems, such as that high-low resolution image manifold is inconsistent, image reconstruction effect is further increased
Fruit, the present invention improve the method for first time super-resolution image reconstruction, specifically：
The image super-resolution reconstructing method based on study includes study stage and image reconstruction stage；
The study stage includes the following steps：
Step A, training sample set is built in accordance with the following methods：
Step A1, it is low-resolution image sample by high-definition picture sample dimensionality reduction；
Step A2, feature is carried out respectively to the low-resolution image sample using a different set of feature extraction parameter to carry
It takes, obtains one group of characteristic image for corresponding respectively to different characteristic extracting parameter；
Step A3, the high-definition picture sample and each characteristic image are carried out respectively according to identical partitioned mode
Each characteristic image block of piecemeal, each high-definition picture block and corresponding position constitutes a joint training sample；
Step B, the training sample set is trained, obtains high-resolution dictionary and low-resolution dictionary；
Step C, high-low resolution transition matrix W is obtained in accordance with the following methods：
Step C1, it is low-resolution face image sample by high-resolution human face image pattern dimensionality reduction；
Step C2, using a different set of feature extraction parameter described in step A2 to the low-resolution face image sample
Feature extraction is carried out respectively, obtains the lineup's face characteristic image for corresponding respectively to different characteristic extracting parameter；
Step C3, according to partitioned mode described in step A3 to high-resolution human face image and each face characteristic image respectively into
Row piecemeal, all high-resolution human face image blocks constitute high resolution graphics image set, and all face characteristic image blocks constitute low resolution
Rate image set；
Step C4, rarefaction representation coefficient collection A of the high resolution graphics image set in the high-resolution dictionary is obtainedh, and
Rarefaction representation coefficient collection A of the low-resolution image collection in the low-resolution dictionaryl；
Step C5, by being solved to following formula, high-low resolution transition matrix W is obtained：
Wherein, λ is regularization coefficient；
Described image phase of regeneration includes the following steps：
Step D, using a different set of feature extraction parameter described in step A2 to low-resolution face image to be reconstructed
Feature extraction is carried out respectively, obtains one group of low resolution face characteristic image for corresponding respectively to different characteristic extracting parameter；So
Piecemeal is carried out respectively to each low resolution face characteristic image according to partitioned mode described in step A3 afterwards, obtains several low resolution
Face characteristic image block；
Step E, rarefaction representation system of each low resolution face characteristic image block in the low-resolution dictionary is obtained
Number, and it is multiplied by high-low resolution transition matrix W respectively, LS-SVM sparseness then is carried out to obtained each product, i.e.,
Obtain sparse table of the full resolution pricture block in the high-resolution dictionary corresponding to each low resolution face characteristic image block
Show coefficient；The LS-SVM sparseness refers to that the coefficient by each absolute value in obtained product less than predetermined threshold value is zeroed
Processing；
Step F, the full resolution pricture block corresponding to all low resolution face characteristic image blocks is reconstructed, and is combined
At initial high resolution human face rebuilding image.
Preferably, in stepb, the training sample set is trained using alternating iteration method, obtains high-resolution
Rate dictionary and low-resolution dictionary.
Preferably, the specific method is as follows for the dimensionality reduction：High-definition picture is first subjected to down-sampling, then uses bilinearity
Interpolation algorithm amplifies.
Preferably, the method for the feature extraction is specific as follows：Believed using First-order Gradient traffic filter and second order gradient
Number filter carries out image lateral filtering, longitudinal filtering respectively.
Preferably, coefficient of utilization coupling estimation model method solves high-low resolution transition matrix W.
Preferably, the value range of predetermined threshold value described in LS-SVM sparseness is 0.01-0.02.
Preferably, the value range of threshold value T is 38-42.
Preferably, the value range of c is 56-64.
Preferably, the tile size in secondary reconstruction process is 64 pixels.
Compared with prior art, it the present invention and its is further improved scheme and has the advantages that：
The present invention carries out its reconstructed results secondary heavy on the basis of the image super-resolution reconstructing method based on study
It builds, is based on principle as Local Phase, by carrying out SVD (Singular Value to the formed matrix of similar image block in neighborhood
Decomposition, singular value) decompose mode be reconstructed, can effectively eliminate super-resolution reconstruction facial image block effect
Problem is answered, face image super-resolution rebuilding effect is improved.
The present invention further improves the method for first time super-resolution image reconstruction, during dictionary learning
High-low resolution transition matrix is obtained, and using high-low resolution transition matrix come the rarefaction representation coefficient to high-definition picture
Optimize, to effectively solving the problems, such as that high-low resolution image manifold is inconsistent, further increase image reconstruction effect with
The practicability of method.
Description of the drawings
Fig. 1 is the basic procedure schematic diagram of super-resolution face image method of the present invention；
Fig. 2 is the generating principle schematic diagram of training sample in specific implementation mode；
Fig. 3 is the principle schematic that similar image block matrix is built in specific implementation mode；
Fig. 4 is the reconstruction Contrast on effect of the method for the present invention and the prior art.
Specific implementation mode
The facial image reconstruction stage of the prior art is to carry out piecemeal reconstruct to image, and this method will produce blocking artifact.
Solving the problems, such as this usual practice at present is：When image block, allows between adjacent image fritter and overlap；However this
The effect of kind method is not satisfactory.For the reconstruction image blocking artifact problem present in the prior art, thinking of the invention be
On the basis of image super-resolution reconstructing method based on study, secondary reconstruction is carried out to its reconstructed results, based on as Local Phase
Principle, the formed matrix of similar image block is reconstructed by way of carrying out SVD decomposition in neighborhood.
Technical scheme of the present invention is specific as follows：
Low-resolution face image to be reconstructed is carried out first with the image super-resolution reconstructing method based on study
First time super-resolution image reconstruction obtains initial high resolution human face rebuilding image；Then utilize following methods to initial high
Resolution ratio human face rebuilding image carries out secondary reconstruction, obtains final reconstructing high-resolution human face image：
Step 1, by initial high resolution human face rebuilding image segmentation be the identical image block of one group of size；
Step 2, for each image block, find out most like with it c image block in its neighborhood, while recording and often scheming
As the position of block in the picture, and this c most like image blocks are reconstructed in accordance with the following methods：It is most like by this c
Image block change into one-dimensional vector respectively, then c one-dimensional vector is combined into a matrix；Then the matrix formed is carried out
Singular value decomposition, and the singular value to absolute value less than threshold value T is zeroed out processing；It is restored into row matrix, and by the square of reduction
Battle array becomes c image block again, and c image block is put back into image home position；C is preset positive integer；
Step 3, by initial high resolution human face rebuilding image it is all participated in reconstruct pixel, with its reconstruct after pixel
The mean value of value is its progress again assignment to get to final reconstructing high-resolution human face image.
For the ease of public understanding, come below with a specific embodiment and in conjunction with attached drawing to technical scheme of the present invention into
Row is further described.
The basic procedure that human face super-resolution in the present embodiment is rebuild is as shown in Figure 1, specific as follows：
Low-resolution face image to be reconstructed is carried out first with the image super-resolution reconstructing method based on study
First time super-resolution image reconstruction obtains initial high resolution human face rebuilding image.
1, the preparation stage
Input：Sample set
Output：Dh、Dl
(1) sample collection
Training sample set Y={ y1, y2 ..., yN }, N indicate the number of sample.Each sample yi is by high resolution graphics
As block and the joint generation of low-resolution image characteristic block, as shown in Figure 2.
It is low-resolution face image sample by high-resolution human face image pattern dimensionality reduction, then utilizes a different set of spy
Sign extracting parameter carries out feature extraction respectively to the low-resolution image sample, obtains corresponding respectively to different characteristic extraction ginseng
One group of several characteristic images.Then piecemeal, each high-resolution are carried out to high-definition picture sample and characteristic image sample
Rate image block and the characteristic image block of corresponding position constitute a joint training sample.
High-definition picture block in the present embodiment is to progressively scan to obtain according to stationary window.Low-resolution image feature
The acquisition methods of block are：First with following methods to high-definition picture sample dimensionality reduction：High-definition picture is subjected to down-sampling,
Then amplified with bilinear interpolation algorithm；The single order and second order gradient signal filter for using formula (1) again extract characteristic image, so
Stationary window progressive scan is pressed afterwards and obtains four characteristic image blocks, and is combined into a vector.
(2) training dictionary
Shown in dictionary training pattern such as formula (2)
Wherein：D is high-low resolution joint dictionary, DhIt is high-resolution dictionary, DlIt is low-resolution dictionary, Y is height point
Resolution joint training sample set, YhIt is high-resolution training sample set, YlIt is low resolution training sample set, F is filter group, A
For rarefaction representation coefficient collection, t is degree of rarefication (what is selected in the present embodiment is 10).
Formula (2) is solved using the method for alternating iteration, it is specific as follows.Dictionary D fixed first, is asked using orthogonal matching pursuit
Solve rarefaction representation coefficient A；Then rarefaction representation coefficient A is fixed, using SVD isolations update dictionary D.Alternately and repeatedly iteration,
Until reaching end condition, then terminates iterative operation, obtain dictionary D, then dictionary D is split as DhAnd Dl。
2, high-low resolution transition matrix learns the stage
Input：Face sample set, Dh、Dl
Output：High-low resolution transition matrix W
The D that preparation stage obtainshAnd DlDictionary is for natural image, if for the super of this kind of specific image of face
Resolution reconstruction can have that high-low resolution image manifold is inconsistent, and in order to solve this problem, the present invention devises
One high-low resolution transition matrix.Specific modeling is as follows：
(1) collect high-resolution facial image database, then press " sample collection " method of preparation stage to face database into
Row sample collection obtains high-resolution human face image set FhWith low resolution eigenface image set Fl, then utilize formula (3)
(4) A is acquiredhAnd Al。
Finally utilize AhAnd AlW is solved, method is as follows：
The solution of the methods of gradient descent method, Newton method, extremum method can be used in formula (5).Formula (5) is that typical double optimization is asked
Topic, there are analytic solutions, solves it using extremum method in the present embodiment, obtains following analytic solutions：
3, the first reconstruction stage of facial image
Input：Low-resolution face image Yl, Dh、Dl、W
Output：Initial high-resolution human face image
Step 1, by low-resolution image YlFeature extraction is carried out, several characteristic image block y are obtainedl；
Step 2 obtains y using formula (4)lIn dictionary DlOn rarefaction representation coefficient αl；
Step 3 obtains the rarefaction representation coefficient α of high-definition picture block using formula (7)h。
αh=W αl (7)
This step obtains αhNot sparse, because will appear many very small coefficients in coefficient, these coefficients can be with
The interference being considered in calculating process, therefore can be removed.Specific method is：1 suitable threshold value e is set, then will
Coefficient of the absolute value less than e is reset.The preferred value ranges of e are 0.01-0.02.
Step 4 reconstructs initial high-resolution image block using formula (8)
Step 5 finds out all high-definition picture blocksAfterwards, then it is reassembled into initial high-resolution image
4, it is based on secondary reconstruction stage as Local Phase
Input：Initial high-resolution human face image
Output：Secondary reconstruct high-resolution human face image
Facial image can centainly find the similar fritter of c structure in subrange, these blocks are formed a square
Battle array, as shown in Figure 3.After doing SVD decomposition to this matrix, it can be reconstructed with a few larger singular value.According to
This characteristic carries out secondary reconstruction to initial high resolution human face rebuilding image using following methods, obtains final high score
Resolution human face rebuilding image：
Step 1, by initial high resolution human face rebuilding image segmentation be the identical image block of one group of size；
Step 2, for each image block, find out in its neighborhood (assuming that Size of Neighborhood be L × L) c most like with it
A image block, while the often position as block in the picture is recorded, and image block most like to this c in accordance with the following methods carries out
Reconstruct：This c most like image blocks are changed into one-dimensional vector respectively, then c one-dimensional vector is combined into a matrix；Then
Singular value decomposition is carried out to the matrix formed, and (T is an empirical value, preferred value model less than threshold value T to absolute value
Enclose for 38-42) singular value be zeroed out processing (its essence is the singular value less than T regard as reconstruction stage blocking artifact produce
Raw noise is handled)；It is restored into row matrix, and becomes the matrix of reduction again c image block, c image block is put back to
To image home position；C is preset positive integer, and preferred value range is 56-64；
Step 3, by initial high resolution human face rebuilding image it is all participated in reconstruct pixel, with its reconstruct after pixel
The mean value of value is its progress again assignment to get to final reconstructing high-resolution human face image.
, can be with parallel processing to the reconstruct of each image block neighborhood similar image block in above-mentioned secondary reconstruction process, it can also
It sequentially handles, specifically can flexibly be chosen according to hardware environment.A kind of specific implementation algorithm of above-mentioned secondary method for reconstructing is as follows：
1st step：Initializing variable：Image block label i, window size L, tile size s, similar number of blocks c, threshold value T.
2nd step：By imageIt is divided into N number of size to be according to the sequence of progressive scanImage fritter, image
There is the overlapping of 1 pixel between fritter.
3rd step, to i-th of image block yi∈RsIt is handled, in image block yiL × L neighborhoods in, it is similar to search out c
Image block, be then combined into a matrix, be expressed as
4th step is rightDo SVD decomposition
Singular value by absolute value less than threshold value T is reset
Restructuring matrix
5th step, willMatrix becomes again as image block, then it is put back into the position in original image.
6th step：Compare the value of i.If the value of i is greater than or equal to N, the 7th step is turned to；Otherwise, the value of i is added 1, then returned
Return to the 3rd step.
7th step：The reconstruct number of obtained image divided by each pixel final image has just been obtained into
In order to verify the effect of the method for the present invention, following compliance test result experiment has been carried out：
Test image selects ORL standard pictures library, and the size of low-resolution image block is 3 × 3 pixels, amplification factor 3.
When secondary reconstruct, the size of image block is 64, and window size 20, threshold value T is 40, and the quantity c of similar image block is 60.
Experimental result and bilinear interpolation algorithm and document (Proc.of International Conference on
Computer Vision and Pattern Recognition,2008,2:729-73) the SRSR algorithms proposed are compared.
The objective evaluation index of picture quality uses 2 indexs of Y-PSNR (PSNR) and structural similarity (SSIM).Comparison result
See Tables 1 and 2.The partial enlargement image of 3 kinds of algorithm reconstruction images is given in Fig. 4, is respectively from left to right low resolution in Fig. 4
Rate image, interpolation reconstruction image, SRSR algorithms reconstructed image, the method for the present invention reconstruction image and original high-resolution image.
1 PSNR values of table compare
Table 2 SSIM comparisons
The super-resolution facial image proposed by the invention based on secondary reconstruction is can be seen that according to 1,2 and Fig. 4 of table
The image reconstruction result of method for reconstructing is substantially better than existing bilinear interpolation algorithm and SRSR algorithms.
Claims (10)
1. the face image super-resolution reconstruction method based on secondary reconstruction, which is characterized in that first with the figure based on study
As ultra-resolution ratio reconstructing method carries out first time super-resolution rebuilding to low-resolution face image to be reconstructed, initial height is obtained
Resolution ratio human face rebuilding image；Then secondary reconstruction is carried out to initial high resolution human face rebuilding image using following methods, obtained
To final reconstructing high-resolution human face image：
Step 1, by initial high resolution human face rebuilding image segmentation be the identical image block of one group of size；
Step 2, for each image block, find out most like with it c image block in its neighborhood, while recording each image
The position of block in the picture, and this c most like image blocks are reconstructed in accordance with the following methods：It is most like by this c
Image block changes into one-dimensional vector respectively, then c one-dimensional vector is combined into a matrix；Then the matrix formed is carried out strange
Different value is decomposed, and the singular value to absolute value less than threshold value T is zeroed out processing；It is restored into row matrix, and by the matrix of reduction
It becomes c image block again, c image block is put back into image home position；C is preset positive integer；
Step 3, by all pixels for participating in reconstruct in initial high resolution human face rebuilding image, with pixel value after its reconstruct
Mean value is its progress again assignment to get to final reconstructing high-resolution human face image.
2. method as described in claim 1, which is characterized in that the image super-resolution reconstructing method based on study includes learning
Habit stage and image reconstruction stage；
The study stage includes the following steps：
Step A, training sample set is built in accordance with the following methods：
Step A1, it is low-resolution image sample by high-definition picture sample dimensionality reduction；
Step A2, feature extraction is carried out respectively to the low-resolution image sample using a different set of feature extraction parameter,
Obtain corresponding respectively to one group of characteristic image of different characteristic extracting parameter；
Step A3, the high-definition picture sample and each characteristic image are divided respectively according to identical partitioned mode
Each characteristic image block of block, each high-definition picture block and corresponding position constitutes a joint training sample；
Step B, the training sample set is trained, obtains high-resolution dictionary and low-resolution dictionary；
Step C, high-low resolution transition matrix W is obtained in accordance with the following methods：
Step C1, it is low-resolution face image sample by high-resolution human face image pattern dimensionality reduction；
Step C2, the low-resolution face image sample is distinguished using a different set of feature extraction parameter described in step A2
Feature extraction is carried out, the lineup's face characteristic image for corresponding respectively to different characteristic extracting parameter is obtained；
Step C3, high-resolution human face image and each face characteristic image are divided respectively according to partitioned mode described in step A3
Block, all high-resolution human face image blocks constitute high resolution graphics image set, and all face characteristic image blocks constitute low resolution figure
Image set；
Step C4, rarefaction representation coefficient collection A of the high resolution graphics image set in the high-resolution dictionary is obtainedhAnd low resolution
Rarefaction representation coefficient collection A of the rate image set in the low-resolution dictionaryl；
Step C5, by being solved to following formula, high-low resolution transition matrix W is obtained：
Wherein, λ is regularization coefficient；
Described image phase of regeneration includes the following steps：
Step D, low-resolution face image to be reconstructed is distinguished using a different set of feature extraction parameter described in step A2
Feature extraction is carried out, one group of low resolution face characteristic image for corresponding respectively to different characteristic extracting parameter is obtained；Then it presses
Piecemeal is carried out respectively to each low resolution face characteristic image according to partitioned mode described in step A3, obtains several low resolution faces
Characteristic image block；
Step E, rarefaction representation coefficient of each low resolution face characteristic image block in the low-resolution dictionary is obtained, and
It is multiplied by high-low resolution transition matrix W respectively, LS-SVM sparseness then is carried out to get to every to obtained each product
Rarefaction representation coefficient of the full resolution pricture block in the high-resolution dictionary corresponding to a low resolution face characteristic image block；
The LS-SVM sparseness refers to the coefficient progress return-to-zero that each absolute value in obtained product is less than to predetermined threshold value；
Step F, the full resolution pricture block corresponding to all low resolution face characteristic image blocks is reconstructed, and is combined into first
Beginning reconstructing high-resolution human face image.
3. method as claimed in claim 2, which is characterized in that in stepb, using alternating iteration method to the training sample
Collection is trained, and obtains high-resolution dictionary and low-resolution dictionary.
4. method as claimed in claim 2, which is characterized in that the specific method is as follows for the dimensionality reduction：First by high-definition picture
Down-sampling is carried out, is then amplified with bilinear interpolation algorithm.
5. method as claimed in claim 2, which is characterized in that the method for the feature extraction is specific as follows：Utilize First-order Gradient
Traffic filter and second order gradient signal filter carry out image lateral filtering, longitudinal filtering respectively.
6. method as claimed in claim 2, which is characterized in that coefficient of utilization coupling estimation model method solves high-low resolution and turns
Change matrix W.
7. method as claimed in claim 2, which is characterized in that the value range of predetermined threshold value described in LS-SVM sparseness is
0.01-0.02。
8. method as claimed in claim 1 or 2, which is characterized in that the value range of threshold value T is 38-42.
9. method as claimed in claim 1 or 2, which is characterized in that the value range of c is 56-64.
10. method as claimed in claim 1 or 2, which is characterized in that the tile size in secondary reconstruction process is 64 pictures
Element.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810229088.2A CN108492252A (en) | 2018-03-20 | 2018-03-20 | Face image super-resolution reconstruction method based on secondary reconstruction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810229088.2A CN108492252A (en) | 2018-03-20 | 2018-03-20 | Face image super-resolution reconstruction method based on secondary reconstruction |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108492252A true CN108492252A (en) | 2018-09-04 |
Family
ID=63318681
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810229088.2A Withdrawn CN108492252A (en) | 2018-03-20 | 2018-03-20 | Face image super-resolution reconstruction method based on secondary reconstruction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108492252A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110287818A (en) * | 2019-06-05 | 2019-09-27 | 广州市森锐科技股份有限公司 | Face feature vector optimization method based on layered vectorization |
| CN110443885A (en) * | 2019-07-18 | 2019-11-12 | 西北工业大学 | Three-dimensional number of people face model reconstruction method based on random facial image |
| CN110738601A (en) * | 2019-10-23 | 2020-01-31 | 智慧视通(杭州)科技发展有限公司 | low-resolution face image super-resolution reconstruction method based on three-dimensional face model |
| CN116008911A (en) * | 2022-12-02 | 2023-04-25 | 南昌工程学院 | Orthogonal matching pursuit sound source identification method based on novel atomic matching criteria |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105844590A (en) * | 2016-03-23 | 2016-08-10 | 武汉理工大学 | Image super-resolution reconstruction method and system based on sparse representation |
| CN107424119A (en) * | 2017-04-12 | 2017-12-01 | 广西大学 | A kind of super-resolution method of single image |
-
2018
- 2018-03-20 CN CN201810229088.2A patent/CN108492252A/en not_active Withdrawn
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105844590A (en) * | 2016-03-23 | 2016-08-10 | 武汉理工大学 | Image super-resolution reconstruction method and system based on sparse representation |
| CN107424119A (en) * | 2017-04-12 | 2017-12-01 | 广西大学 | A kind of super-resolution method of single image |
Non-Patent Citations (3)
| Title |
|---|
| HUASHENG ZHU.ETC: "Manifold inconsistency constrained sparse coding for image super-resolution reconstruction", 《2017 10TH INTERNATIONAL CONGRESS ON IMAGE AND SIGNAL PROCESSING, BIOMEDICAL ENGINEERING AND INFORMATICS》 * |
| HUASHENG ZHU.ETC: "Super-Resolution Reconstruction of Face Image with Improved Sparse Constraint", 《2016 6TH INTERNATIONAL CONFERENCE ON DIGITAL HOME》 * |
| JINSHENG XIAO.ETC: "Detail enhancement of image super-resolution based on detail synthesis", 《SIGNAL PROCESSING: IMAGE COMMUNICATION》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110287818A (en) * | 2019-06-05 | 2019-09-27 | 广州市森锐科技股份有限公司 | Face feature vector optimization method based on layered vectorization |
| CN110287818B (en) * | 2019-06-05 | 2024-01-16 | 广州市森锐科技股份有限公司 | Hierarchical vectorization-based face feature vector optimization method |
| CN110443885A (en) * | 2019-07-18 | 2019-11-12 | 西北工业大学 | Three-dimensional number of people face model reconstruction method based on random facial image |
| CN110443885B (en) * | 2019-07-18 | 2022-05-03 | 西北工业大学 | Three-dimensional human head and face model reconstruction method based on random human face image |
| CN110738601A (en) * | 2019-10-23 | 2020-01-31 | 智慧视通(杭州)科技发展有限公司 | low-resolution face image super-resolution reconstruction method based on three-dimensional face model |
| CN116008911A (en) * | 2022-12-02 | 2023-04-25 | 南昌工程学院 | Orthogonal matching pursuit sound source identification method based on novel atomic matching criteria |
| CN116008911B (en) * | 2022-12-02 | 2023-08-22 | 南昌工程学院 | Orthogonal matching pursuit sound source identification method based on novel atomic matching criteria |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
| Zhang et al. | Image super-resolution via dual-dictionary learning and sparse representation | |
| Gao et al. | Image super-resolution with sparse neighbor embedding | |
| CN106920214B (en) | Super-resolution reconstruction method for space target image | |
| CN108492252A (en) | Face image super-resolution reconstruction method based on secondary reconstruction | |
| CN107240066A (en) | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks | |
| CN103279933A (en) | Method for reconstructing single-image super-resolution based on double-layer model | |
| CN107833182A (en) | The infrared image super resolution ratio reconstruction method of feature based extraction | |
| CN103020940B (en) | Local feature transformation based face super-resolution reconstruction method | |
| CN106709874B (en) | Compressed low-resolution face image restoration method based on face structure correlation | |
| CN104299193A (en) | Image super-resolution reconstruction method based on high-frequency information and medium-frequency information | |
| Wang et al. | Medical image super-resolution analysis with sparse representation | |
| CN106254720B (en) | A kind of video super-resolution method for reconstructing based on joint regularization | |
| Chen et al. | Guided dual networks for single image super-resolution | |
| CN103632357B (en) | A kind of image super-resolution Enhancement Method separated based on illumination | |
| CN106846286B (en) | Video super-resolution algorithm for reconstructing based on a variety of complementary priori | |
| Wang | Single image super-resolution with u-net generative adversarial networks | |
| Miura et al. | Image quality improvement for learning-based super-resolution with PCA | |
| Wang et al. | Research on image super-resolution reconstruction based on transformer | |
| Rohit et al. | Modified dictionary learning method for sparsity based single image super-resolution | |
| Yang et al. | Sparse representation based super resolution using saliency and edge information | |
| CN111292238A (en) | Face image super-resolution reconstruction method based on orthogonal partial least squares | |
| Nie et al. | Image restoration from patch-based compressed sensing measurement | |
| Zhu et al. | Super-resolution image reconstruction via patch haar wavelet feature extraction combined with sparse coding | |
| Seetharamaswamy et al. | Super resolution image reconstruction via dual dictionary learning in sparse environment |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180904 |