CN105069825B - Image super-resolution rebuilding method based on depth confidence network - Google Patents
Image super-resolution rebuilding method based on depth confidence network Download PDFInfo
- Publication number
- CN105069825B CN105069825B CN201510501171.7A CN201510501171A CN105069825B CN 105069825 B CN105069825 B CN 105069825B CN 201510501171 A CN201510501171 A CN 201510501171A CN 105069825 B CN105069825 B CN 105069825B
- Authority
- CN
- China
- Prior art keywords
- image
- resolution
- network
- super
- depth confidence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Abstract
Based on the image super-resolution rebuilding method of depth confidence network, it is related to image procossing.The image of low resolution is obtained, the advanced row interpolation of the image of low resolution is amplified, is amplified to required size；Low resolution luminance graph image block is obtained with the method for repeating piecemeal sampling；Low-resolution image block is inputted, with the advance trained high-resolution image block of depth confidence neural network forecast；Obtained fitting result is subjected to neighborhood regularization Optimization Solution；All high-resolution luminance picture blocks are combined to obtain high-resolution luminance image；The value of other two channels obtained before most heel combines, and the image that reconvert is represented into colored RGB is to get to the high-definition picture predicted.The Super-resolution Reconstruction of single-frame images can be achieved, improve the Y-PSNR of image, obtain reconstruction image sharp edge and abundant texture, available for Video security monitoring, medical digital image, space flight detection etc..
Description
Technical field
The present invention relates to image procossing, more particularly, to monitored available for Video security, medical digital image, space flight detect
Deng the image super-resolution rebuilding method based on depth confidence network.
Background technology
Image super-resolution rebuilding is widely paid close attention in recent years because of its application prospect and actual application value,
And a large amount of outstanding algorithm is emerged.These algorithms can be substantially divided into three categories：Image super-resolution weight based on interpolation
Build algorithm, the Super-Resolution of Images Based based on reconstruct and the image super-resolution rebuilding algorithm based on study.Compared to other two
Class algorithm, the attention rate that the image super-resolution rebuilding algorithm based on study is subject to are higher.It by a large amount of learning training, then
In addition appropriate is prior-constrained, make it that there is more prominent performance compared to traditional interpolation and the method for reconstruct.
In the super-resolution reconstruction method based on study, the Super-resolution Reconstruction based on rarefaction representation is a kind of important oversubscription
Method for reconstructing is widely paid close attention to, and is proved to be a kind of oversubscription method for reconstructing got a good chance of.In addition one kind is paid close attention to
Method be the method based on recurrence, oversubscription problem is converted into regression problem by such method, designs regression optimization model.Such
Method is also proved to be a kind of method that effective oversubscription is rebuild.However, this two major class method is all by top-down
Mode carry out modelling, that is, need to provide model of fit.Currently to the bottom-up less discussion of solution mode.Depth
Habit is a kind of natural bottom-up network architecture, and in recent years, deep learning is in speech recognition, text identification and image classification
In all achieve breakthrough good effect.
Chinese patent CN104778659A discloses a kind of single-frame image super-resolution reconstruction method based on deep learning, packet
Include following steps：1st, it is used to obtain low resolution by two autocoders of training first and corresponds to high-definition picture block
Feature；2nd, based on the feature for having obtained high-resolution and low-resolution image block, one monolayer neuronal network study two of retraining
The Nonlinear Mapping relationship of a feature；3rd, based on two autocoders and monolayer neural networks, three layers of depth network is built,
Using low-resolution image block as input, high-definition picture block finely tunes the parameter of three layer depth networks as output；Step 4,
Single-frame images super-resolution rebuilding is done according to three obtained layer depth networks, using the gray value of low-resolution image block as input,
Obtain the gray value that output is corresponding high-definition picture block.
Invention content
The purpose of the present invention is to provide the Super-resolution Reconstructions of achievable single-frame images, improve the Y-PSNR of image,
Obtain the image super-resolution rebuilding method based on depth confidence network of reconstruction image sharp edge and abundant texture.
The present invention includes the following steps：
(1) the image Y of low resolution is inputted；
(2) the image Y of low resolution is first subjected to bi-cubic interpolation amplification, is amplified to required size；
(3) image is transformed into brightness from the RGB color space represented with three Color Channels of red, green, blue and blue is red
Colour saturation offset (YCbCr) color space, and subsequent place only is carried out to the brightness value of image i.e. brightness (Y) channel
Reason；
(4) it is slided on the image with sliding window, step-length takes image block for 1, obtains the image block of low resolution
(patch)Yi, the serial number of i expression image blocks；
(5) with the training image blocks of low resolution and corresponding high-definition picture block, limited by training 3
Boltzmann machine (Restricted Boltzmamn Machines, RBM) obtains network parameter { Wi},{ci, wherein, { WiTable
Show the weight vector of the side connection in network, { ciRepresent to be obtained with training next time to the biasing of last layer in network
{Wi},{ciInitialization depth confidence network (Deep Belief Network, DBN) parameter；
(6) training DBN networks (Trained DBN), 3 RBM are gathered into folds, obtain DBN, utilize training obtained above
Image block, with the back propagation of stochastic gradient descent method and error fine tuning parameter { Wi},{ci, until e<Tol or iteration time
Number t>T, e represent error, and tol and T are the constant thresholds given in algorithm；
(7) by the image block of the low resolution of image to be tested (patch blocks) YiAs input, step (6) training is utilized
Good DBN predicts to obtain corresponding high-resolution patch blocks Xi；
(8) the high-resolution patch blocks X that will be obtainedi, the pixel among image block is taken to be assigned a value of full resolution pricture pair
The pixel value of position coordinates is answered, to the patch blocks Y of all low resolutioniIt is handled, obtains required high-resolution figure
Image brightness figure XDBN_SR；
(9) non local similarity constraint, local similarity constraint, the fitting constraint of depth confidence network are added, foundation is based on
Neighborhood relationships regularization and the super-resolution Optimized model of fit correlation regularization；
(10) with the graceful iterative algorithm of Donald Bragg of separation come the Optimized model in solution procedure (9)；
(11) obtained luminance channel (Y) Super-resolution Reconstruction image is combined with the value (CbCr) of other two Color Channels
Get up, reconvert into RGB image to get the high-definition picture X predicted.
In step (5), the method for 3 limited Boltzmann machines of training can be：
(5.1) energy function of RBM known to is E (υ, h；θ)=- hTWυ-bTh-cTυ, θ={ W, b, c }, joint probability is fixed
Justice is,Wherein Z=ΣvΣhexp(-E(υ,h；θ)), v is RBM inputs, and h is
RBM hidden layers export, and W represents the weight vector of the side connection of input layer in a network, and c represents that input layer is to last layer in network
Biasing, b represent biasing of the last layer to next layer in network；Set iteration total degree T, random initializtion parameter θ to be trained0
=(W0,b0,c0) and it is assigned to θt, t expression current iteration numbers；
(5.2) feature vector/image block to be entered is assigned to v0, and utilization P (v | h), P (h | v) iterate n times
Obtain h0, vnAnd hn, v0Represent each component of the 0th iteration input vector, h0Represent each point of the 0th iteration hidden layer
Amount, conditional probability calculate as follows：
P(hi=1 | v, θ)=σ (ci+wiV),
P (v | h)=ΠjP(vi| h),
P(vj=1 | h, θ)=σ (bj+w'jH),
Wherein σ () is activation primitive；
(5.3) it is obtained using stochastic gradient descent method and error back propagation method general by the joint of n times gibbs sampler
Rate distribution gradient value, and update θtThree parameter W of the insidet,bt,ct, obtain θt+1；
(5.4) if iterations t=T or difference value reach certain small degree, EP (end of program)；Otherwise by θt+1It assigns
It is worth to θt, and return to (5.2) step.
In step (6), the method for the trained DBN networks can be：
(6.1) 3 RBM are stacked, i.e. using low-resolution image block as input, operation obtains defeated first RBM
Go out the input obtained as second RBM, the input exported as third RBM of second RBM, the output finally obtained is
XDBN_SR；
(6.2) using stochastic gradient descent method and error back propagation method training network, θ is finely tunedtThree parameters of the inside
Wt,bt,ct, obtain θt+1；
(6.3) if iterations t=T or difference value reach certain small degree, EP (end of program)；Otherwise by θt+1It assigns
It is worth to θt, and return to (6.1).
In step (9), the non local similarity constraint is：In the bigger neighborhood of some pixel, utilize
Similar pixel to constrain current point, that is, by similitude weighted average represents current point, and calculation formula is：
RNL=| | (I-WNL)X||1
Wherein X is the vector form of input picture；I is unit matrix；WNLFor weight matrix, if representing, block i and block j gets over phase
Seemingly, then weighted value is bigger, and the value of each element is provided by following formula：
Wherein,XiRepresent the image block centered on ith pixel, GαTable
Show the Gaussian function that standard deviation is α, its effect is that larger weights, deep position are distributed closer to the position at center
Smaller weights are distributed, a ° expression step-by-step is multiplied, and h is attenuation parameter；
The local similarity is constrained to：Intuitively, each pixel is with surrounding pixel in an image
It is extremely similar, it is possible to make full use of the phenomenon, constrain current point with neighbouring pixel, can make image more in this way
Discontinuous, rough phenomenon caused by non local similarity constraint can smoothly be overcome, the two is known as complementary constraint, uses
The method of controllable kernel regression calculates this constraint, which comes from optimization problem：
Xl=[xi,x2,…,xP]T
K=diag [KH(l1-l),KH(l2-l),…,KH(lP-l)]
Obtained regularization constraint calculation formula is：
RL=| | (I-WL)X||1
Wherein I is unit vector, and WLHave：
XiRepresent the image block centered on i pixels, e1It is 1 there was only first element, other elements are all 0 row
Vector, Ψ are the distance matrix of pixel.
The depth confidence network fitting is constrained to：
The super-resolution Optimized model is expressed as：
Wherein, D represents the matrix of down-sampling, and H represents fuzzy matrix, λNL,λLAnd λDBN_SRRespectively three canonical above
The weighted value of item, for adjusting their proportions.
The method of deep learning is introduced into image super-resolution rebuilding by the present invention, gives full play to deep learning protrusion
Learning ability, to restore the high-frequency information of image, in conjunction with appropriate regularization constraint, it is proposed that super-resolution rebuilding algorithm.
The present invention has the advantages that following prominent：
1. traditional depth confidence network model for classification problem design is transformed by the present invention first, its is defeated
Enter and make real value data into from binaryzation data with output, be finally transformed into the model of suitable super-resolution rebuilding.It is put using depth
The stratification of communication network, the structure of nonlinearity play its processing capacity powerful to data, it is largely high to recover image
Frequency information.
2. designing regularization term, image super-resolution rebuilding model is established.Due to doing super-resolution rebuilding with DBN, do not have
There is the information in view of neighborhood so that the super-resolution rebuilding algorithm based on DBN has certain limitation.It is lacked for this
Point, the present invention add in the regularization constraint that can make full use of neighborhood information, construct the oversubscription based on regularization constraint and DBN
Resolution reconstruction model.
3. design is based on the separation graceful algorithm of Donald Bragg come the numerical algorithm of rapid solving.
Description of the drawings
Fig. 1 is the image super-resolution rebuilding algorithm frame based on depth confidence network of the present invention；
Fig. 2 is using reconstructed results of the image of twice of present invention amplification after depth confidence network processes and its
The comparison of his 3 kinds of methods；
Fig. 3 is that the reconstructed results of twice of the image magnification using the present invention and existing 5 kinds of methods compare.
Specific embodiment
1. with reference to Fig. 1, frame of the invention is:
Step 1, low resolution interpolation image Y is obtained.
It downloads a panel height at random from internet and differentiates luminance picture X, and utilize the imresize letters in Matlab softwares
Number amplifies the low bi-cubic interpolation for differentiating 2 times of luminance picture progress, obtains low-resolution image Y.
Step 2, network output figure X is obtainedDBN_SR。
Image is transformed into YCbCr color spaces, and only to the brightness value of image i.e. Y by (2a) from RGB color space
Channel carries out subsequent processing；
(2b) is slided on the image with sliding window, obtains the patch blocks Y of low resolutioni；
(2c) utilizes advance trained DBN (Trained using the patch blocks of obtained low resolution as input
DBN) predict to obtain high-resolution patch blocks Xi；
The high-resolution patch blocks X that (2d) will be obtainedi, recombinated according to the sequence of script, obtain wanted height
The image brightness picture X of resolution ratioDBN_SR；
Step 3, it rebuilds and obtains high-definition picture.
Y channels and the value of other two channels are combined, for reconvert into the image of RGB, this just obtains what is predicted
High-definition picture.
2. experimental result and interpretation of result：
Experiment one, super-resolution image reconstruction result is carried out with the present invention by depth confidence network.
For the validity of verification algorithm, on test library set9, it is compared with other three kinds of classic algorithms.Fig. 2's
Five width images are respectively：Bilinear is bilinear interpolation, and SCSR represents the super-resolution rebuilding algorithm based on rarefaction representation,
Bicubic represents segmentation cubic interpolation method.For convenience, the algorithm for remembering the present invention is DBN_SR, and Original is original image.
Table 1 is Fig. 2 reconstructed results and structural similarity (the Structural Similarity Index of other algorithms
Measurement, SSIM) and Y-PSNR (Peak Signal to Noise Ratio, PSNR) compare.
The experimental result explanation of Fig. 2：From intuitive visual, algorithm of the invention is handled in detail must be well very much,
It is clear, more careful than other three kinds of algorithms on the texture of parrot eyes.Although bilinear interpolation speed is fast, sawtooth, mould
Paste phenomenon it is serious, although bicubic interpolation than bilinear interpolation at edge more smoothly, still have serious blooming,
SCSR can restore many details, but still have ringing.Generally speaking, algorithm proposed by the present invention, not only in vision
In effect, and the significant effect of than other three kinds typical algorithms is all achieved in objective evaluation standard, present it
Outstanding super-resolution rebuilding performance.
Experiment two, with super-resolution image reconstruction result of the present invention after depth confidence network adds canonical item constraint.
For the validity of verification algorithm, on test library set9, it is compared with other three kinds of classic algorithms.Fig. 3's
Seven width images are respectively that Original is original image, and Bicubic is segmentation cubic interpolation method, and SCSR is based on rarefaction representation
Super-resolution rebuilding algorithm, ANR are returned for anchor point neighbour, and SRCNN is depth convolutional neural networks, and LLE is to be locally linear embedding into,
Ours is the present invention.Table 2 is reconstructed results compared with the SSIM and PSNR of other algorithms.
The experimental result explanation of Fig. 3 and table 2：After three regular terms are continuously added, the either structure in test set
Similitude (SSIM) and Y-PSNR (PSNR) all increase, and after all regular terms are added, have obtained highest
Effect.The experiment shows three kinds of regular terms proposed by the present invention, all increases to the effect of high-definition picture,
It is all useful to illustrate each single item, and the combination of three can effectively improve the quality of final high-definition picture.
Table 1
Table 2
Input for artwork by bicubic down-sampling in the case of tested, enable algorithm with other algorithms same
In one starting point.Table 2 is as can be seen that all higher than other several classic algorithms on Y-PSNR is similar with structure, wherein being based on
The oversubscription result SRCNN that depth convolutional neural networks obtain is that super-resolution rebuilding effect is best in the paper delivered in 2014
Method, but the present invention, after three regularization constraints have been added, image reconstruction effect is more than SRCNN's.Illustrate the present invention's
Algorithm has very strong reconstruction ability.
Claims (7)
1. the image super-resolution rebuilding method based on depth confidence network, it is characterised in that include the following steps：
(1) the image Y of low resolution is inputted；
(2) the image Y of low resolution is first subjected to bi-cubic interpolation amplification, is amplified to required size；
(3) image is transformed into brightness from the RGB color space represented with three Color Channels of red, green, blue and blue is red dense
Offset (YCbCr) color space is spent, and subsequent processing only is carried out to the brightness value of image i.e. brightness (Y) channel；
(4) it is slided on the image with sliding window, step-length takes image block for 1, obtains image block (patch) Y of low resolutioni,
I represents the sequence number of image block；
(5) with the training image blocks of low resolution and corresponding high-definition picture block, pass through 3 limited bohrs of training
Hereby graceful machine obtains network parameter { Wi},{ci, wherein, { WiRepresent the weight vector that the side in network connects, { ciRepresent network
In biasing of next layer to last layer, with the obtained { W of trainingi},{ciInitialization depth confidence network (Deep Belief
Network, DBN) parameter；
(6) training DBN networks (Trained DBN), 3 RBM are gathered into folds, obtain DBN, utilize training image obtained above
Block, with the back propagation of stochastic gradient descent method and error fine tuning parameter { Wi},{ci, until e<Tol or iterations t>
T, e represent error, and tol and T are the constant thresholds given in algorithm；
(7) by the image block patch blocks Y of the low resolution of image to be testediAs input, step (6) trained DBN is utilized
To predict to obtain corresponding high-resolution image block Xi；
(8) the high-resolution image block X that will be obtainedi, the pixel among image block is taken to be assigned a value of full resolution pricture corresponding position
The pixel value of coordinate, to the image block Y of all low resolutioniIt is handled, it is bright to obtain wanted high-resolution image
Degree figure XDBN_SR；
(9) non local similarity constraint, local similarity constraint are added, the fitting constraint of depth confidence network is established based on neighborhood
Relationship regularization and the super-resolution Optimized model of fit correlation regularization；
(10) with the graceful iterative algorithm of Donald Bragg of separation come the Optimized model in solution procedure (9)；
(11) obtained luminance channel (Y) Super-resolution Reconstruction image is combined with the value (CbCr) of other two Color Channels
Come, reconvert into RGB image to get the high-definition picture X predicted.
2. the image super-resolution rebuilding method as described in claim 1 based on depth confidence network, it is characterised in that in step
Suddenly the method for 3 limited Boltzmann machines of training described in (5) is：
(5.1) energy function of RBM known to is E (ν, h；θ)=- hTWν-bTh-cTν, θ=(W, b, c), joint probability is defined asWherein Z=∑sν∑hexp(-E(ν,h；θ)), ν is the input of RBM, and h is RBM hidden
Layer output；W represents the weight vector of the side connection of input layer in a network, and c represents biasing of the input layer to last layer in network,
B represents biasing of the last layer to next layer in network；Set iteration total degree T, random initializtion parameter θ to be trained0=(W0,
b0,c0) and it is assigned to θt, t expression current iteration numbers；
(5.2) feature vector/image block to be entered is assigned to ν0, and using P (ν | h), P (h | the ν) n times that iterate obtain
h0, νnAnd hn, ν0Represent each component of the 0th iteration input vector, h0Represent each component of the 0th iteration hidden layer, item
Part probability calculation is as follows：
P(hi=1 | ν, θ)=σ (ci+wiν),
P (ν | h)=ΠjP(νj| h),
P(νj=1 | h, θ)=σ (bj+w'jH),
Wherein σ () is activation primitive；
(5.3) joint probability point by n times gibbs sampler is obtained using stochastic gradient descent method and error back propagation method
Cloth Grad, and update θtThree parameter W of the insidet、bt、ct, obtain θt+1；
(5.4) if iterations t=T or difference value reach certain small degree, EP (end of program)；Otherwise by θt+1It is assigned to
θt, and return to (5.2) step.
3. the image super-resolution rebuilding method as claimed in claim 2 based on depth confidence network, it is characterised in that in step
(6) in, the method for the trained DBN networks is：
(6.1) 3 RBM are stacked, i.e. first RBM obtains low-resolution image block to be defeated using claim 1 step (4)
Enter, the input that the output that operation obtains is obtained as second RBM, the input exported as third RBM of second RBM,
The output finally obtained is XDBN_SR；
(6.2) using stochastic gradient descent method and error back propagation method training network, θ is finely tunedtThree parameter W of the insidet、bt、
ct, obtain θt+1, wherein parameter θt、Wt、bt、ct、θt+1Claim 2 step (5.1) and (5.3) are shown in definition；
(6.3) if iterations t=T or difference value reach certain small degree, EP (end of program)；Otherwise by θt+1Assignment
To θt, and return to (6.1).
4. the image super-resolution rebuilding method as claimed in claim 3 based on depth confidence network, it is characterised in that in step
(9) in, the non local similarity constraint is：In the bigger neighborhood of some pixel, using similar pixel, come
Current point is constrained, that is, current point is represented by similitude weighted average, calculation formula is：
RNL=| | (I-WNL)X||1
Wherein X is the vector form of input picture, and I is unit matrix；WNLFor weight matrix, represent to get over phase such as fruit block i and block j
Seemingly, then weighted value is bigger, and the value of each element is provided by following formula：
Wherein,XiRepresent the image block centered on ith pixel, GαRepresent mark
The Gaussian function that quasi- difference is α, its effect are that larger weights are distributed closer to the position at center, deep position distribution
Smaller weights,.Represent that step-by-step is multiplied, h is attenuation parameter.
5. the image super-resolution rebuilding method as claimed in claim 4 based on depth confidence network, it is characterised in that in step
(9) in, the local similarity is constrained to：Intuitively, each pixel and its surrounding pixel point are extremely in an image
Similar, it is possible to the phenomenon is made full use of, current point is constrained with neighbouring pixel, can make image more smooth in this way
Discontinuous, rough phenomenon caused by non local similarity constraint can be overcome, the two is known as complementary constraint, with controllable
Kernel regression method calculates this constraint, which comes from optimization problem：
WhereinKH(li- l) it is weight core,It is the near of position l
Neighbour, nearer with l, weight is bigger, CiFor gradient covariance matrix, hkSmoothing parameter for controllable core；
The problem is converted into matrix form：
Wherein,
Xl=[x1,x2,…,xP]T,
K=diag { KH(l1-l),KH(l2-l),…,KH(lP- l) } it is diagonal matrix；
Obtained regularization constraint calculation formula is：
RL=| | { (I-WL)X}||1
Wherein I is unit matrix, and WLHave：
XiRepresent the image block centered on i, e1It is 1 there was only first element, other elements are all 0 column vector, and ψ is picture
The distance matrix of element.
6. the image super-resolution rebuilding method as claimed in claim 5 based on depth confidence network, it is characterised in that in step
(9) it in, defines the fitting of depth confidence network and is constrained to：
7. the image super-resolution rebuilding method as claimed in claim 6 based on depth confidence network, it is characterised in that in step
(9) in, the super-resolution Optimized model is expressed as：
Wherein, D represents the matrix of down-sampling, and H represents fuzzy matrix λNL, λLAnd λDBN_SRThe respectively power of three regular terms above
Weight values, for adjusting the ratio shared by them.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510501171.7A CN105069825B (en) | 2015-08-14 | 2015-08-14 | Image super-resolution rebuilding method based on depth confidence network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510501171.7A CN105069825B (en) | 2015-08-14 | 2015-08-14 | Image super-resolution rebuilding method based on depth confidence network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105069825A CN105069825A (en) | 2015-11-18 |
| CN105069825B true CN105069825B (en) | 2018-06-12 |
Family
ID=54499181
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510501171.7A Expired - Fee Related CN105069825B (en) | 2015-08-14 | 2015-08-14 | Image super-resolution rebuilding method based on depth confidence network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105069825B (en) |
Families Citing this family (35)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105657263B (en) * | 2015-12-31 | 2018-11-02 | 杭州卓腾信息技术有限公司 | A kind of super-resolution digital slices scan method based on area array cameras |
| CN105976318A (en) * | 2016-04-28 | 2016-09-28 | 北京工业大学 | Image super-resolution reconstruction method |
| CN106067161A (en) * | 2016-05-24 | 2016-11-02 | 深圳市未来媒体技术研究院 | A kind of method that image is carried out super-resolution |
| CN107818545B (en) * | 2016-09-12 | 2020-07-14 | 北京大学 | Image processing method and device |
| CN106558021B (en) * | 2016-11-21 | 2020-03-31 | 重庆大学 | Video enhancement method based on super-resolution technology |
| CN106600538A (en) * | 2016-12-15 | 2017-04-26 | 武汉工程大学 | Human face super-resolution algorithm based on regional depth convolution neural network |
| CN106897545B (en) * | 2017-01-05 | 2019-04-30 | 浙江大学 | A kind of tumor prognosis forecasting system based on depth confidence network |
| CN106846286B (en) * | 2017-01-05 | 2019-10-29 | 四川大学 | Video super-resolution algorithm for reconstructing based on a variety of complementary priori |
| CN106886978B (en) * | 2017-02-16 | 2020-01-03 | 清华大学深圳研究生院 | Super-resolution reconstruction method of image |
| CN107018422B (en) * | 2017-04-27 | 2019-11-05 | 四川大学 | Still image compression method based on depth convolutional neural networks |
| CN107302657B (en) * | 2017-05-16 | 2019-07-23 | 苏州大学 | Image capturing system suitable for Internet of Things |
| CN107563965A (en) * | 2017-09-04 | 2018-01-09 | 四川大学 | Jpeg compressed image super resolution ratio reconstruction method based on convolutional neural networks |
| CN109905717A (en) * | 2017-12-11 | 2019-06-18 | 四川大学 | A kind of H.264/AVC Encoding Optimization based on Space-time domain down-sampling and reconstruction |
| CN109961396B (en) * | 2017-12-25 | 2020-12-22 | 中国科学院沈阳自动化研究所 | Image super-resolution reconstruction method based on convolutional neural network |
| CN108416734A (en) * | 2018-02-08 | 2018-08-17 | 西北大学 | Text image super resolution ratio reconstruction method and device based on edge driving |
| US10489943B2 (en) | 2018-02-28 | 2019-11-26 | General Electric Company | System and method for sparse image reconstruction |
| CN108647775B (en) * | 2018-04-25 | 2022-03-29 | 陕西师范大学 | Super-resolution image reconstruction method based on full convolution neural network single image |
| CN108805814B (en) * | 2018-06-07 | 2020-05-19 | 西安电子科技大学 | Image super-resolution reconstruction method based on multi-band deep convolutional neural network |
| CN109580629A (en) * | 2018-08-24 | 2019-04-05 | 绍兴文理学院 | Crankshaft thrust collar intelligent detecting method and system |
| CN109285223B (en) * | 2018-09-07 | 2019-05-31 | 中山大学 | Image reconstruction quadrilateral mesh method and system based on voxel model |
| CN109523470B (en) * | 2018-11-21 | 2023-07-25 | 四川长虹电器股份有限公司 | Depth image super-resolution reconstruction method and system |
| CN109584164B (en) * | 2018-12-18 | 2023-05-26 | 华中科技大学 | Medical image super-resolution three-dimensional reconstruction method based on two-dimensional image transfer learning |
| CN109934773B (en) * | 2019-03-13 | 2023-08-25 | 北京旷视科技有限公司 | Image processing method, device, electronic equipment and computer readable medium |
| CN110223230A (en) * | 2019-05-30 | 2019-09-10 | 华南理工大学 | A kind of more front end depth image super-resolution systems and its data processing method |
| CN112116526A (en) * | 2019-06-19 | 2020-12-22 | 中国石油化工股份有限公司 | Torch smoke image super-resolution method based on deep convolutional neural network |
| CN110490806B (en) * | 2019-08-19 | 2021-07-27 | 福州大学 | Method and system for reconstructing super-resolution image of atomic force microscope |
| US11010871B1 (en) * | 2019-11-05 | 2021-05-18 | Moxa Inc. | Device and method of handling image super-resolution |
| CN110930333A (en) * | 2019-11-22 | 2020-03-27 | 北京金山云网络技术有限公司 | Image restoration method and device, electronic equipment and computer-readable storage medium |
| CN111652354B (en) * | 2020-05-29 | 2023-10-24 | 北京百度网讯科技有限公司 | Method, apparatus, device and storage medium for training super network |
| CN111861881B (en) * | 2020-06-09 | 2022-05-20 | 复旦大学 | Image super-resolution reconstruction method based on CNN interpolation |
| CN112288737A (en) * | 2020-11-18 | 2021-01-29 | 中国人民解放军国防科技大学 | Super-resolution image-based porthole detection method |
| CN113191955A (en) * | 2021-06-17 | 2021-07-30 | 江苏奥易克斯汽车电子科技股份有限公司 | Method and device for reconstructing image super-resolution |
| CN113570701B (en) * | 2021-07-13 | 2023-10-24 | 聚好看科技股份有限公司 | Hair reconstruction method and device |
| CN114820326A (en) * | 2022-05-25 | 2022-07-29 | 厦门大学 | Efficient single-frame image super-resolution method based on adjustable kernel sparsification |
| CN116128717B (en) * | 2023-04-17 | 2023-06-23 | 四川观想科技股份有限公司 | Image style migration method based on neural network |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102629374A (en) * | 2012-02-29 | 2012-08-08 | 西南交通大学 | Image super resolution (SR) reconstruction method based on subspace projection and neighborhood embedding |
| CN103167218A (en) * | 2011-12-14 | 2013-06-19 | 北京大学 | Super-resolution reconstruction method and equipment based on non-locality |
| US8811774B1 (en) * | 2012-01-19 | 2014-08-19 | Pixelworks, Inc. | Super resolution using an interpretive scaler |
| CN104008538A (en) * | 2014-05-12 | 2014-08-27 | 清华大学 | Super-resolution method based on single image |
| CN104408697A (en) * | 2014-10-23 | 2015-03-11 | 西安电子科技大学 | Image super-resolution reconstruction method based on genetic algorithm and regular prior model |
| CN104778659A (en) * | 2015-04-15 | 2015-07-15 | 杭州电子科技大学 | Single-frame image super-resolution reconstruction method on basis of deep learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8594464B2 (en) * | 2011-05-26 | 2013-11-26 | Microsoft Corporation | Adaptive super resolution for video enhancement |
-
2015
- 2015-08-14 CN CN201510501171.7A patent/CN105069825B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103167218A (en) * | 2011-12-14 | 2013-06-19 | 北京大学 | Super-resolution reconstruction method and equipment based on non-locality |
| US8811774B1 (en) * | 2012-01-19 | 2014-08-19 | Pixelworks, Inc. | Super resolution using an interpretive scaler |
| CN102629374A (en) * | 2012-02-29 | 2012-08-08 | 西南交通大学 | Image super resolution (SR) reconstruction method based on subspace projection and neighborhood embedding |
| CN104008538A (en) * | 2014-05-12 | 2014-08-27 | 清华大学 | Super-resolution method based on single image |
| CN104408697A (en) * | 2014-10-23 | 2015-03-11 | 西安电子科技大学 | Image super-resolution reconstruction method based on genetic algorithm and regular prior model |
| CN104778659A (en) * | 2015-04-15 | 2015-07-15 | 杭州电子科技大学 | Single-frame image super-resolution reconstruction method on basis of deep learning |
Non-Patent Citations (6)
| Title |
|---|
| Direct Energy Minimization for Super-Resolution on Nonlinear Manifolds;Tien-Lung Chang 等;《European Conference on Computer Vision》;20060513;第1-14页 * |
| Image Super-Resolution based on Multikernel Regression;Ying Gu 等;《Pattern Recognition》;20121115;第2071-2074页 * |
| Image Super-Resolution Using Deep Belief Networks;Yanwen Zhou 等;《Internet Multimedia Computing and Service》;20140712;第1-4页 * |
| Kernel Regression for Image Processing and Reconstruction;Hiroyuki Takeda 等;《IMAGE PROCESSING》;20070228;第16卷(第2期);第349-366页 * |
| 一种基于正则化技术的超分辨影像重建方法;沈焕锋 等;《中国图象图形学报》;20050430;第10卷(第4期);第436-440页 * |
| 引入格式塔理论的超分辨率图像重建技术;李翠华 等;《厦门大学学报(自然科学版)》;20110331;第50卷(第2期);第261-270页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105069825A (en) | 2015-11-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105069825B (en) | Image super-resolution rebuilding method based on depth confidence network | |
| CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
| CN106683067B (en) | Deep learning super-resolution reconstruction method based on residual sub-images | |
| Wang et al. | Deep networks for image super-resolution with sparse prior | |
| Liu et al. | A spectral grouping and attention-driven residual dense network for hyperspectral image super-resolution | |
| CN103871041B (en) | The image super-resolution reconstructing method built based on cognitive regularization parameter | |
| Dong et al. | Laplacian pyramid dense network for hyperspectral pansharpening | |
| Gao et al. | Enhanced autoencoders with attention-embedded degradation learning for unsupervised hyperspectral image super-resolution | |
| CN109064396A (en) | A kind of single image super resolution ratio reconstruction method based on depth ingredient learning network | |
| CN108734661B (en) | High-resolution image prediction method for constructing loss function based on image texture information | |
| CN107492070A (en) | A kind of single image super-resolution computational methods of binary channels convolutional neural networks | |
| CN108764128A (en) | A kind of video actions recognition methods based on sparse time slice network | |
| CN104574336B (en) | Super-resolution image reconstruction system based on adaptive sub- mould dictionary selection | |
| CN103279933B (en) | A kind of single image super resolution ratio reconstruction method based on bilayer model | |
| CN105550989B (en) | The image super-resolution method returned based on non local Gaussian process | |
| CN110136060A (en) | The image super-resolution rebuilding method of network is intensively connected based on shallow-layer | |
| CN111738268A (en) | Semantic segmentation method and system for high-resolution remote sensing image based on random block | |
| CN109087375A (en) | Image cavity fill method based on deep learning | |
| CN111340696B (en) | Convolutional neural network image super-resolution reconstruction method fused with bionic visual mechanism | |
| Li et al. | Deep hybrid 2-D–3-D CNN based on dual second-order attention with camera spectral sensitivity prior for spectral super-resolution | |
| CN112837224A (en) | Super-resolution image reconstruction method based on convolutional neural network | |
| CN110148088A (en) | Image processing method, image rain removing method, device, terminal and medium | |
| Beaulieu et al. | Deep image-to-image transfer applied to resolution enhancement of sentinel-2 images | |
| Pan et al. | FDPPGAN: remote sensing image fusion based on deep perceptual patchGAN | |
| Dong et al. | Fusion of hyperspectral and panchromatic images using generative adversarial network and image segmentation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180612 Termination date: 20210814 |