CN106204451A - The Image Super-resolution Reconstruction method embedded based on the fixing neighborhood of constraint - Google Patents
The Image Super-resolution Reconstruction method embedded based on the fixing neighborhood of constraint Download PDFInfo
- Publication number
- CN106204451A CN106204451A CN201610537501.2A CN201610537501A CN106204451A CN 106204451 A CN106204451 A CN 106204451A CN 201610537501 A CN201610537501 A CN 201610537501A CN 106204451 A CN106204451 A CN 106204451A
- Authority
- CN
- China
- Prior art keywords
- resolution
- image
- block
- represent
- nearest neighbor
- 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
- 238000000034 method Methods 0.000 title claims abstract description 61
- 238000012549 training Methods 0.000 claims abstract description 50
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 claims abstract description 8
- 239000000284 extract Substances 0.000 claims abstract description 6
- 238000012360 testing method Methods 0.000 claims abstract description 5
- 239000000203 mixture Substances 0.000 claims description 14
- 238000005070 sampling Methods 0.000 claims description 7
- 238000004422 calculation algorithm Methods 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 2
- 230000017105 transposition Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 8
- 238000012545 processing Methods 0.000 description 6
- 238000002474 experimental method Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 4
- 238000005457 optimization Methods 0.000 description 4
- 230000003321 amplification Effects 0.000 description 3
- 238000003199 nucleic acid amplification method Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000009414 blockwork Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000007850 degeneration Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
- G06T3/4076—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution using the original low-resolution images to iteratively correct the high-resolution images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Editing Of Facsimile Originals (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of Image Super-resolution Reconstruction method embedded based on the fixing neighborhood of constraint, mainly solve existing method and ignore high-resolution block effect and cause and rebuild image blurring problem.Implementation step is: 1. build high-resolution, low resolution training image set；2. from training image set, build low resolution, high-resolution associating k nearest neighbor group；3. extract luminance picture component and the colourity picture content of test image；4. pair luminance picture component carries out image block, and computing block average, generates the high resoluting characteristic block of initial estimation；5. search the arest neighbors associating k nearest neighbor group of all initial estimation high resoluting characteristic blocks, rebuild high-resolution and estimate image block；6. combine all high-resolution and estimate image block, obtain high-resolution luminance picture；7. merge high-resolution luminance picture and colourity picture content, obtain the full resolution pricture rebuild.The present invention reduces the generation of pseudomorphism, improve the definition of reconstruct image, can be used for high sharpness video and show.
Description
Technical field
The present invention is belonging to technical field of image processing, further relates to a kind of Image Super-resolution Reconstruction method, available
In public safety field, remotely sensed image field, medical imaging field, digital TV in high resolution field.
Background technology
Image is in acquisition process, often by factors such as imaging device, shooting environmental, scene motion changes
Impact, causes the image of actual acquisitions to there is noise, the degradation phenomena such as fuzzy, and image is not enough to the detailed information that offer is enriched,
Many difficulties is brought for image procossing, analysis etc..Super-resolution Reconstruction technology is the method by computer software, from single width or
Several observable low-resolution images reconstruct high-resolution picture rich in detail, are the effective ways improving degraded image resolution
One of, receive significant attention in image processing field.
Existing Image Super-resolution Reconstruction is broadly divided into following three class methods:
One, method based on interpolation.This kind of method generally utilizes spatial neighborhood and the natural image priori of object pixel
Predict object pixel, thus realize the amplification of image, such as bilinear interpolation, bi-cubic interpolation etc..This kind of method speed is fast, but
The image generated is the most smooth, and high frequency detail obscures.
Its two, based on reconstruct method.This kind of method utilizes multiframe to input the complementary information between low resolution image, by image
During priori joins Image Super-resolution Reconstruction as constraint, and then recover the information lost in imaging process.This
Class method relies on the registration accuracy between multiple image and the structure of bound term, rebuild when amplification is bigger effect rapidly under
Fall, is not suitable for the Super-resolution Reconstruction that amplification is bigger.
Its three, the method for instance-based learning.This kind of method is carried out based on image block mostly, utilizes corresponding high score
Distinguish that image block and low resolution image block work in coordination with co-occurrence priori, learn the expression coefficient of the low resolution image block of input, or learn
Practise out low resolution image block and the mapping relations of full resolution pricture block in training set, and then predict loss in low resolution image
Detail of the high frequency, it is achieved the reconstruction of full resolution pricture.Method effect based on study is preferable, is also to study more side at present
Method, be largely divided into method based on probability graph model, method based on manifold learning, method based on rarefaction representation and based on
The method returned.
A) method based on probability graph model, be the earliest by Freeman et al. document " Freeman, William T.,
E.C.Pasztor,and O.T.Carmichael."Learning Low-Level Vision."International
Proposing in, the method uses markov random file Journal of Computer Vision40.1 (2010): 25-47. "
Set up the relational model between low resolution image and high-resolution scene, realized the maximum of full resolution pricture by belief propagation algorithm
Change Posterior estimator.The method depends on the study of great amount of samples, and computation complexity is higher.
B) method based on manifold learning, is the thought of manifold learning to be incorporated in super-resolution, it is assumed that low resolution block stream
Shape and high-resolution block manifold are Local Isometrics." Chang H, Yeung D Y, Xiong Y.Super-such as Chang et al.
Resolution through Neighbor Embedding[J].IEEE Conference on Computer Vision&
Patter Recognition,2010,1:275-282.”.Utilize the collaborative appearance between corresponding low resolution block and high-resolution block
Priori, by finding K neighbour of the low resolution image block of input in training set, tries to achieve the reconstruct minimized under reconstructed error
Weights, and weights are applied in the linear combination of high-resolution block, thus realize between low resolution image and full resolution pricture
Relationship map.Degeneration and the complexity of natural image structure due to low resolution image so that low resolution image block is to high-resolution
Image block presents the corresponding relation of one-to-many, and above-mentioned hypothesis is the most always set up.During this neighbor searching or right-value optimization
The method ignoring full resolution pricture block effect, is easily caused in High resolution reconstruction image the fuzzy and appearance of noise.
C) method based on rarefaction representation, is to be applied in Image Super-resolution Reconstruction, compressive sensing theory such as Yang etc.
" Yang J, Wright J, Huang T S, the et al.Image super-resolution via sparse that people proposes
representation.[J].IEEE Transactions on Image Processing A Publication of the
IEEE Signal Processing Society,2010,19(11):2861-2873.”.It is raw that such method first passes through training
Become one by low resolution, full resolution pricture block to learning the dictionary pair that, then estimate low resolution image block and the training of input
Concentrate the Relationship of Coefficients between the atom in low resolution dictionary, and utilize the Relationship of Coefficients linear combination correspondence high-resolution word of acquisition
Atom in allusion quotation, thus Reconstructing High block.When training dataset selects improper, this kind of method is rebuild in image easy
There is pseudomorphism.
D) based on the method returned, first in the low resolution of training set learning to high-resolution regression function, then utilize
The regression function that study is arrived, is mapped as full resolution pricture by the low resolution image of input." Timofte such as Timofte et al.
R,De V,Van Gool L.Anchored Neighborhood Regression for Fast Example-Based
Super-Resolution[C]//Computer Vision(ICCV),2013IEEE International Conference
on.IEEE,2013:1920-1927.”.The pseudomorphism of reconstruction can be reduced based on the method returned, but due to existing regression function
Determination need to estimate too much parameter, thus cause the generalization ability of algorithm poor, and simple regression function be difficult to right
Full resolution pricture is modeled with the complex mapping relation of low resolution image.
Summary of the invention
It is an object of the invention to overcome above-mentioned existing methodical deficiency, propose a kind of based on constraint fixing neighborhood embedding
Image Super-resolution Reconstruction method, to improve the quality of Reconstructing High.
The technical scheme realizing the object of the invention includes the following:
(1) select the clear natural image of N panel height to be downloaded from network, build high-resolution training image set ThWith low point
Distinguish training image set Tl；
(2) from low resolution training image set TlWith high-resolution training image set ThIn, extract M at random identical greatly
Little low resolution, full resolution pricture block pair, composition training image blocks is to setWherein Pi lRepresent the i-th extracted
Low resolution image block, Pi hRepresent and Pi lCorresponding i-th full resolution pricture block, 1≤i≤M；
(3) training image blocks is extracted respectively to setIn low resolution characteristic blockAnd high score
Distinguish characteristic blockObtain the low resolution of correspondence, high resoluting characteristic block pairAgain with all low points obtained
Distinguish, high resoluting characteristic block pairComposition training characteristics block is to setWherein,Represent low resolution image block
Pi lAverage,Represent the i-th low resolution characteristic block extracted,Represent withCorresponding i-th high resoluting characteristic block；
(4) training characteristics block is searched to set according to Euclidean distanceIn each low resolution, high resoluting characteristic
Block pairAssociating k nearest neighbor pair, with these neighbours to composition associating k nearest neighbor group set
WhereinRepresent i-th associating k nearest neighbor group,Represent that i-th group low resolution k nearest neighbor is special
Levy,Represent i-th associating k nearest neighbor group CiIn u low resolution feature；Represent withCorresponding i-th group
High-resolution k nearest neighbor feature,Represent i-th associating k nearest neighbor group CiIn the u high resoluting characteristic,Represent low point of i-th
Distinguish, high resoluting characteristic block pairThe u neighbour coupleAt training characteristics block to setIn
Index, and K ＞ 0,1≤u≤K,
(5) each associating k nearest neighbor group C is calculatediCenterAnd according to Euclidean distance, all in utilizing class
The low resolution center of associating k nearest neighbor group is that each class creates a new K dimension-tree；WhereinRepresent i-th associating k nearest neighbor group
Low resolution center,Represent withThe high-resolution center of corresponding i-th associating k nearest neighbor group；
(6) the low-resolution image Y of given any one test, utilizes the bi-cubic interpolation method low resolution to input
Image Y carries out up-sampling s times, obtains low resolution interpolation image Yb, by interpolation image YbIt is transformed into YCbCr empty from rgb space
Between, obtain its luminance component image YlWith chroma blue component image YCb, red chrominance component image YCr；
(7) by luminance component image YlIt is divided into the luminance component image block that size is identical and overlappedAnd calculate
Each image blockAverageWherein V represents the image block number extracted, 1≤r≤V；
(8) image block obtained according to step (7)With image block averageObtain the feature of high-resolution initial estimation
Block:
(9) to high-resolution initial estimation characteristic blockBeing optimized renewal, the high-resolution after being updated estimates characteristic block
(10) after utilizing renewal, high-resolution estimates characteristic blockWith image block averageObtain the high-resolution brightness rebuild
Component image block:
(11) the high-resolution luminance component image block to all reconstructionsMerge, generate a width and insert corresponding to low resolution
Value luminance component image YlClear full resolution pricture luminance component
(12) by full resolution pricture luminance componentWith chroma blue component YCbWith red chrominance component YCrIt is combined,
Obtain the full resolution pricture under YCbCr spaceAnd by full resolution prictureIt is transformed into rgb space by YCbCr space, obtains
The full resolution pricture X reconstructed under rgb space.
Present invention have the advantage that
First, due to the fact that the associating k nearest neighbor group set first training high-resolution, low resolution characteristic block, then press
According to high-resolution its arest neighbors of block search associating k nearest neighbor group that Euclidean distance is each initial estimation so that in neighbour's selection course
Middle considering low resolution block and the effect of high-resolution block, the neighbour of lookup is closer to practical situation simultaneously；
Second, due to the fact that during right-value optimization, first construct the high-resolution block of initial estimation to its high-resolution K
The distance restraint matrix of neighbour's block, then distance restraint matrix is joined in right-value optimization as bound term so that after renewal
Estimate that high-resolution block details is apparent.
Accompanying drawing explanation
Fig. 1 is the flowchart of the present invention；
Fig. 2 is Experimental comparison's figure of the full resolution pricture rebuild by the present invention and existing two kinds of methods.
Detailed description of the invention
The core concept of the present invention is: the thought embedded by the fixing neighborhood of constraint proposes a kind of Image Super-resolution Reconstruction side
Method, by low resolution block and high-resolution block being optimized neighbour's selection and weights, improves the picture quality of reconstructed results.
With reference to Fig. 1, the enforcement step of the present invention is as follows:
Step 1, builds training image collection.
Select the clear natural image of N panel height to be downloaded from network, build high-resolution training image set ThWith low resolution
Training image set Tl, step is as follows:
(1a) from network, the clear natural image of N panel height is downloaded as high-definition picture, and by this N panel height image in different resolution
It is transformed into brightness, chroma blue, red color YCbCr space from red, green, blue rgb space, extracts its luminance component composition high score
Distinguish training image setWherein Te hRepresent that e panel height differentiates training image, 1≤e≤N；
(1b) utilize interpolation method to high-resolution training image set ThIn each image first carry out the down-sampling of s times and obtain
After low resolution image, then the up-sampling carrying out s times obtains low resolution interpolation image；
(1c) low resolution training image set is formed with N width low resolution interpolation imageWherein Te lRepresent e
Width low resolution training image.
Interpolation method in described (1b), including bi-cubic interpolation method and bilinear interpolation method, this example uses
It it is bi-cubic interpolation method.
Step 2, builds training image set of blocks.
From low resolution training image set TlWith high-resolution training image set ThIn, extract M formed objects at random
Low resolution, full resolution pricture block pair, composition training image blocks is to setWherein Pi lRepresent low point of the i-th extracted
Distinguish image block, Pi hRepresent and Pi lCorresponding i-th full resolution pricture block, 1≤i≤M.
Step 3, builds training characteristics set of blocks.
(3a) training image blocks is extracted respectively to setIn low resolution characteristic blockAnd high score
Distinguish characteristic blockObtain the low resolution of correspondence, high resoluting characteristic block pair
(3b) the low resolution that obtains with all, high resoluting characteristic block pairComposition training characteristics block is to setWherein,Represent low resolution image block Pi lAverage,Represent the i-th low resolution characteristic block extracted,Table
Show withCorresponding i-th high resoluting characteristic block.
Step 4, builds associating k nearest neighbor group set.
(4a) with each low resolution, high resoluting characteristic block pairForm a new union feature block Zi；
(4b) by all new union feature block ZiThe union feature set of blocks of compositionUtilize according to Euclidean distance
K-means algorithm clusters；
(4c) setting sample number threshold value A=600 in minimum class, in iterative search class, sample number is less than the class of threshold value A, and deletes
Remove, according still further to Euclidean distance, sample in such is repartitioned in its arest neighbors class, until in the class of all classes sample number is all
No less than threshold value A, iteration stopping；
(4d) utilize sample data in class, be that every class creates a K dimension-tree according to Euclidean distance；
(4e) for each union feature block Zi, according to Euclidean distance, utilize in its generic sample data and should
K dimension-the tree of class, to associating characteristic block ZiCarry out k nearest neighbor search, obtain union feature block ZiK neighbour；
(4f) with the union feature block Z obtainediK neighbour, form associating k nearest neighbor group Ci, then by all associatings obtained
K nearest neighbor group CiComposition associating k nearest neighbor group set
WhereinRepresent i-th associating k nearest neighbor group,Represent that i-th group low resolution k nearest neighbor is special
Levy,Represent i-th associating k nearest neighbor group CiIn u low resolution feature；Represent withCorresponding i-th
Group high-resolution k nearest neighbor feature,Represent i-th associating k nearest neighbor group CiIn the u high resoluting characteristic,Represent that i-th is low
Resolution, high resoluting characteristic block pairThe u neighbour coupleAt training characteristics block to setIn
Index, and K ＞ 0,1≤u≤K,
Step 5, creates new K dimension-tree.
Calculate each associating k nearest neighbor group CiCenterAnd according to Euclidean distance, all associatings in utilizing class
The low resolution center of k nearest neighbor group is that each class creates a new K dimension-tree, whereinRepresent the low of i-th associating k nearest neighbor group
Resolution center,Represent withThe high-resolution center of corresponding i-th associating k nearest neighbor group.
Step 6, carries out pretreatment to the low resolution image of test.
The low-resolution image Y of given any one test, utilizes the bi-cubic interpolation method low resolution figure to input
As Y carries out up-sampling s times, obtain low resolution interpolation image Yb；
By interpolation image YbIt is transformed into YCbCr space from rgb space, obtains its luminance component image YlDivide with chroma blue
Spirogram is as YCb, red chrominance component image YCr。
Step 7, to luminance component image YlDo piecemeal to process.
By luminance component image YlIt is divided into the luminance component image block that size is identical and overlappedAnd calculate every
Individual image block Pr tAverageWherein V represents the image block number extracted, 1≤r≤V.
Step 8, the image block P obtained according to step 7r tWith image block averageGenerate the feature of high-resolution initial estimation
Block:
Step 9, rebuilds high-resolution and estimates characteristic block
(9a) according to Euclidean distance, for each high-resolution initial estimation characteristic blockFind out arest neighbors class；
(9b) for high-resolution initial estimation characteristic blockAccording to Euclidean distance, in utilizing its generic in low resolution
Calculation evidence and such new K dimension-tree carry out nearest neighbor search, obtain the associating k nearest neighbor group of arest neighbors
WhereinRepresent high-resolution initial estimation characteristic blockLow resolution k nearest neighbor,Represent high score
Distinguish initial estimation characteristic blockU low resolution neighbour；Represent high-resolution initial estimation characteristic block's
High-resolution k nearest neighbor,Represent high-resolution initial estimation characteristic blockThe u high-resolution neighbour, WrRepresent that high-resolution is initially estimated
Meter characteristic blockArest neighbors associating k nearest neighbor groupIn associating k nearest neighbor group setIn index, 1≤Wr≤ M, 1≤u
≤K；
(9c) high-resolution initial estimation characteristic block is calculatedWith its u high-resolution neighbourEuclidean distance:
Wherein | | | |2Represent 2-norm；
(9d) K the Euclidean distance obtained with step (9c) is as the elements in a main diagonal, and remaining element is 0, and forming size is K
The distance restraint matrix Γ of × K；
(9e) according to step (9b) and the result of (9d), rebuild high-resolution and estimate characteristic block
Wherein T represents that transposition, λ represent balance factor.
Step 10, generates high-resolution luminance component image block
The high-resolution rebuild is utilized to estimate characteristic blockWith image block averageGeneration high-resolution luminance component image block:
Step 11, generates high-resolution luminance component image
High-resolution luminance component image block to all generationsMerge, generate a width corresponding to low resolution interpolation graphs
Image brightness component YlClear full resolution pricture luminance component
Step 12, generates the full resolution pricture X rebuild.
(12a) by full resolution pricture luminance componentWith chroma blue component YCbWith red chrominance component YCrIt is combined,
Obtain the full resolution pricture under YCbCr space
(12b) by full resolution prictureIt is transformed into rgb space by YCbCr space, obtains the height reconstructed under rgb space
Resolution image X.
The effect of the present invention can be described further by following emulation experiment.
1. simulated conditions
The present invention is to be that Intel (R) Core i7-4790 3.6GHZ, internal memory 16G, WINDOWS 7 grasp at central processing unit
Make in system, use the MATLAB software of Mathworks company of U.S. exploitation, Andrea Vedaldi and Brian
The emulation that the image procossing storehouse VLFeat that increases income that Mr. Fulkerson creates is carried out.
Control methods used in experiment includes following 2 kinds:
One is method for reconstructing based on Beta combine processes dictionary learning BPJDL, is designated as BPJDL in experiment；List of references
For He L, Qi H, Zaretzki R.Beta Process Joint Dictionary Learning for Coupled
Feature Spaces with Application to Single Image Super-Resolution[C]//IEEE
Conference on Computer Vision and Pattern Recognition.IEEE Computer Society,
2013:345-352；
Two is method for reconstructing based on Statistical Prediction Model SPM, is designated as SPM in experiment；List of references is Peleg T,
Elad M.A Statistical Prediction Model Based on Sparse Representations for
Single Image Super-Resolution[J].IEEE Transactions on Image Processing,2014,
23(6):2569-82.
2. emulation content
Experiment: low resolution image is carried out Super-resolution Reconstruction
By the inventive method and existing BPJDL method, SPM method, the low resolution image randomly choosed is carried out super-resolution
Rebuild, experimental result such as Fig. 2, wherein:
Fig. 2 (a) is original full resolution pricture；
Fig. 2 (b) is the full resolution pricture that BPJDL method is rebuild；
Fig. 2 (c) is the full resolution pricture that SPM method is rebuild；
Fig. 2 (d) is the full resolution pricture that the present invention rebuilds.
It can be seen that fix the thought that neighborhood embeds, by low owing to the present invention by means of constraint from the Comparative result of Fig. 2
The effect differentiating block and high-resolution block considers to select and during right-value optimization neighbour simultaneously so that the image of reconstruction is in suppression
More visible image detail is provided while pseudomorphism, rebuilds effect and be better than other Image Super-resolution Reconstruction method, demonstrate this
The advance of invention.
Claims (4)
1. the Image Super-resolution Reconstruction method embedded based on the fixing neighborhood of constraint, including:
(1) select the clear natural image of N panel height to be downloaded from network, build high-resolution training image set ThInstruction is differentiated with low
Practice image collection Tl；
(2) from low resolution training image set TlWith high-resolution training image set ThIn, extract M formed objects at random
Low resolution, full resolution pricture block pair, composition training image blocks is to setWherein Pi lRepresent low point of the i-th extracted
Distinguish image block, Pi hRepresent and Pi lCorresponding i-th full resolution pricture block, 1≤i≤M；
(3) training image blocks is extracted respectively to setIn low resolution characteristic blockAnd high resoluting characteristic
BlockObtain the low resolution of correspondence, high resoluting characteristic block pairThe low resolution that obtains with all again, high score
Distinguish characteristic block pairComposition training characteristics block is to setWherein,Represent low resolution image block Pi lEqual
Value,Represent the i-th low resolution characteristic block extracted,Represent withCorresponding i-th high resoluting characteristic block；
(4) training characteristics block is searched to set according to Euclidean distanceIn each low resolution, high resoluting characteristic block pairAssociating k nearest neighbor pair, with these neighbours to composition associating k nearest neighbor group set
WhereinRepresent i-th associating k nearest neighbor group,Represent i-th group low resolution k nearest neighbor feature,
Represent i-th associating k nearest neighbor group CiIn u low resolution feature；Represent withI-th group of corresponding high-resolution
K nearest neighbor feature,Represent i-th associating k nearest neighbor group CiIn the u high resoluting characteristic,Represent the low resolution of i-th, high score
Distinguish characteristic block pairThe u neighbour coupleAt training characteristics block to setIn index, and K
＞ 0,1≤u≤K,
(5) each associating k nearest neighbor group C is calculatediCenterAnd according to Euclidean distance, all associating K in utilizing class
The low resolution center of neighbor group is that each class creates a new K dimension-tree；WhereinRepresent low point of i-th associating k nearest neighbor group
Distinguish center,Represent withThe high-resolution center of corresponding i-th associating k nearest neighbor group；
(6) the low-resolution image Y of given any one test, utilizes the bi-cubic interpolation method low-resolution image to input
Y carries out up-sampling s times, obtains low resolution interpolation image Yb, by interpolation image YbIt is transformed into YCbCr space from rgb space,
To its luminance component image YlWith chroma blue component image YCb, red chrominance component image YCr；
(7) by luminance component image YlIt is divided into the luminance component image block that size is identical and overlappedAnd calculate each
Image blockAverageWherein V represents the image block number extracted, 1≤r≤V；
(8) image block obtained according to step (7)With image block averageObtain the characteristic block of high-resolution initial estimation:
(9) to high-resolution initial estimation characteristic blockBeing optimized renewal, the high-resolution after being updated estimates characteristic block
(10) after utilizing renewal, high-resolution estimates characteristic blockWith image block averageObtain the high-resolution luminance component figure rebuild
As block:
(11) the high-resolution luminance component image block to all reconstructionsMerge, generate a width corresponding to low resolution interpolation graphs
Image brightness component YlClear full resolution pricture luminance component
(12) by full resolution pricture luminance componentWith chroma blue component YCbWith red chrominance component YCrIt is combined, obtains
Full resolution pricture under YCbCr spaceAnd by full resolution prictureIt is transformed into rgb space by YCbCr space, obtains at RGB
The full resolution pricture X reconstructed under space.
2. according to the Image Super-resolution Reconstruction method embedded based on the fixing neighborhood of constraint in claim 1, wherein structure in step (1)
Build high-resolution training image set ThWith low resolution training image set Tl, carry out as follows:
(1a) from network download the clear natural image of N panel height as high-definition picture, and by this N panel height image in different resolution from
Red, green, blue rgb space is transformed into brightness, chroma blue, red color YCbCr space, extracts its luminance component composition high-resolution
Training image setWhereinRepresent that e panel height differentiates training image, 1≤e≤N；
(1b) utilize bi-cubic interpolation method to high-resolution training image set ThIn each image first carry out the down-sampling of s times
After obtaining low resolution image, then the up-sampling carrying out s times obtains low resolution interpolation image；
(1c) low resolution training image set is formed with N width low resolution interpolation imageWhereinRepresent low point of e width
Distinguish training image.
3. according to the Image Super-resolution Reconstruction method embedded based on the fixing neighborhood of constraint in claim 1, wherein in step (4),
Training characteristics block is searched to set according to Euclidean distanceIn each low resolution, high resoluting characteristic block pair
Associating k nearest neighbor pair, carry out as follows:
(4a) with each low resolution, high resoluting characteristic block pairForm a new union feature block Zi；
(4b) by all new union feature block ZiComposition union feature set of blocksK-means is utilized according to Euclidean distance
Algorithm is to associating characteristic block setCluster；In iterative search class, sample number is less than sample number threshold value A in minimum class
Class is also deleted；According still further to Euclidean distance, sample in such is repartitioned in its arest neighbors class, until sample in the class of all classes
This number is all no less than threshold value A, iteration stopping；
(4c) utilize sample data in class, be that every class creates a K dimension-tree according to Euclidean distance；
(4d) for each union feature block Zi, according to Euclidean distance, utilize sample data and step (4c) in its generic
Such K dimension-tree created, to associating characteristic block ZiCarry out k nearest neighbor search, K the neighbour obtained, form an associating K near
Adjacent group Ci, then these are combined k nearest neighbor group CiComposition associating k nearest neighbor group set
4. according to the Image Super-resolution Reconstruction method embedded based on the fixing neighborhood of constraint in claim 1, it is characterised in that wherein
To high-resolution initial estimation characteristic block in step (9)Being optimized renewal, the high-resolution after being updated estimates characteristic block
Its step is as follows:
(9a) according to Euclidean distance, for each high-resolution initial estimation characteristic blockFind out arest neighbors class；
(9b) for high-resolution initial estimation characteristic blockAccording to Euclidean distance, calculation in low resolution in utilizing its generic
Carry out nearest neighbor search according to the new K dimension-tree created with step (5), obtain the associating k nearest neighbor group of arest neighbors
WhereinRepresent high-resolution initial estimation characteristic blockLow resolution k nearest neighbor,Represent that high-resolution is initial
Estimate characteristic blockU low resolution neighbour；Represent high-resolution initial estimation characteristic blockHigh-resolution K
Neighbour,Represent high-resolution initial estimation characteristic blockThe u high-resolution neighbour, WrRepresent high-resolution initial estimation feature
BlockArest neighbors associating k nearest neighbor groupIn associating k nearest neighbor group setIn index, 1≤Wr≤ M, 1≤u≤K；
(9c) high-resolution initial estimation characteristic block is calculatedWith its u high-resolution neighbourEuclidean distance:
Wherein | | | |2Represent 2-norm；
(9d) K the Euclidean distance obtained with step (9c) is as the elements in a main diagonal, and remaining element is 0, and forming size is K × K
Distance restraint matrix Γ；
(9e) according to step (9b) and the result of (9d), calculate high-resolution and estimate characteristic block
Wherein T represents that transposition, λ represent balance factor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610537501.2A CN106204451B (en) | 2016-07-08 | 2016-07-08 | Based on the Image Super-resolution Reconstruction method for constraining fixed neighborhood insertion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610537501.2A CN106204451B (en) | 2016-07-08 | 2016-07-08 | Based on the Image Super-resolution Reconstruction method for constraining fixed neighborhood insertion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106204451A true CN106204451A (en) | 2016-12-07 |
| CN106204451B CN106204451B (en) | 2019-04-23 |
Family
ID=57473740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610537501.2A Active CN106204451B (en) | 2016-07-08 | 2016-07-08 | Based on the Image Super-resolution Reconstruction method for constraining fixed neighborhood insertion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106204451B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110782904A (en) * | 2019-11-07 | 2020-02-11 | 四川长虹电器股份有限公司 | User account switching method of intelligent voice equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101216889A (en) * | 2008-01-14 | 2008-07-09 | 浙江大学 | A face image super-resolution method with the amalgamation of global characteristics and local details information |
| CN102800076A (en) * | 2012-07-16 | 2012-11-28 | 西安电子科技大学 | Image super-resolution reconstruction method based on double-dictionary learning |
| CN103824272A (en) * | 2014-03-03 | 2014-05-28 | 武汉大学 | Face super-resolution reconstruction method based on K-neighboring re-recognition |
| CN103854268A (en) * | 2014-03-26 | 2014-06-11 | 西安电子科技大学 | Image super-resolution reconstruction method based on multi-core gaussian process regression |
| CN104881449A (en) * | 2015-05-15 | 2015-09-02 | 西安电子科技大学 | Image retrieval method based on manifold learning data compression hash |
-
2016
- 2016-07-08 CN CN201610537501.2A patent/CN106204451B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101216889A (en) * | 2008-01-14 | 2008-07-09 | 浙江大学 | A face image super-resolution method with the amalgamation of global characteristics and local details information |
| CN102800076A (en) * | 2012-07-16 | 2012-11-28 | 西安电子科技大学 | Image super-resolution reconstruction method based on double-dictionary learning |
| CN103824272A (en) * | 2014-03-03 | 2014-05-28 | 武汉大学 | Face super-resolution reconstruction method based on K-neighboring re-recognition |
| CN103854268A (en) * | 2014-03-26 | 2014-06-11 | 西安电子科技大学 | Image super-resolution reconstruction method based on multi-core gaussian process regression |
| CN104881449A (en) * | 2015-05-15 | 2015-09-02 | 西安电子科技大学 | Image retrieval method based on manifold learning data compression hash |
Non-Patent Citations (3)
| Title |
|---|
| XINBO GAO等: "Image Super-Resolution With Sparse Neighbor Embedding", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
| 李强等: "改进的基于邻域嵌入的图像超分辨率重构", 《激光技术》 * |
| 虞涛: "基于邻域嵌入的图像超分辨率重建研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110782904A (en) * | 2019-11-07 | 2020-02-11 | 四川长虹电器股份有限公司 | User account switching method of intelligent voice equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106204451B (en) | 2019-04-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Cao et al. | Video super-resolution transformer | |
| CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
| Engin et al. | Cycle-dehaze: Enhanced cyclegan for single image dehazing | |
| CN109671023B (en) | Face image super-resolution secondary reconstruction method | |
| CN108122197B (en) | Image super-resolution reconstruction method based on deep learning | |
| CN102902961B (en) | Face super-resolution processing method based on K neighbor sparse coding average value constraint | |
| CN110163801B (en) | Image super-resolution and coloring method, system and electronic equipment | |
| CN111127374B (en) | Pan-sharing method based on multi-scale dense network | |
| CN110136062B (en) | Super-resolution reconstruction method combining semantic segmentation | |
| CN103093444B (en) | Image super-resolution reconstruction method based on self-similarity and structural information constraint | |
| CN105825477B (en) | The Remote sensed image super-resolution reconstruction method merged based on more dictionary learnings with non-local information | |
| Cai et al. | FCSR-GAN: Joint face completion and super-resolution via multi-task learning | |
| CN105631807B (en) | The single-frame image super-resolution reconstruction method chosen based on sparse domain | |
| CN111325165B (en) | Urban remote sensing image scene classification method considering spatial relationship information | |
| CN104657962B (en) | The Image Super-resolution Reconstruction method returned based on cascading linear | |
| Zhang et al. | Semantic-aware dehazing network with adaptive feature fusion | |
| CN102915527A (en) | Face image super-resolution reconstruction method based on morphological component analysis | |
| CN105550989B (en) | The image super-resolution method returned based on non local Gaussian process | |
| CN113139898A (en) | Light field image super-resolution reconstruction method based on frequency domain analysis and deep learning | |
| CN105447840A (en) | Image super-resolution method based on active sampling and Gaussian process regression | |
| CN103049340A (en) | Image super-resolution reconstruction method of visual vocabularies and based on texture context constraint | |
| CN106981046B (en) | Single image super resolution ratio reconstruction method based on multi-gradient constrained regression | |
| CN116934592A (en) | Image stitching method, system, equipment and medium based on deep learning | |
| Liu et al. | Single image super-resolution using a deep encoder–decoder symmetrical network with iterative back projection | |
| CN109615576B (en) | Single-frame image super-resolution reconstruction method based on cascade regression basis learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |