CN108734659A - A kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label - Google Patents
A kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label Download PDFInfo
- Publication number
- CN108734659A CN108734659A CN201810473375.8A CN201810473375A CN108734659A CN 108734659 A CN108734659 A CN 108734659A CN 201810473375 A CN201810473375 A CN 201810473375A CN 108734659 A CN108734659 A CN 108734659A
- Authority
- CN
- China
- Prior art keywords
- resolution
- image
- super
- sub
- pix
- 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 60
- 238000005070 sampling Methods 0.000 claims abstract description 21
- 238000000605 extraction Methods 0.000 claims abstract description 19
- 108091006146 Channels Proteins 0.000 claims description 101
- 238000012549 training Methods 0.000 claims description 35
- 230000006870 function Effects 0.000 claims description 20
- 230000004913 activation Effects 0.000 claims description 18
- 230000003321 amplification Effects 0.000 claims description 17
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 17
- 230000008569 process Effects 0.000 claims description 15
- 238000000354 decomposition reaction Methods 0.000 claims description 10
- 230000004927 fusion Effects 0.000 claims description 8
- 230000008707 rearrangement Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 2
- 238000012360 testing method Methods 0.000 abstract description 2
- 238000011156 evaluation Methods 0.000 abstract 1
- 238000013441 quality evaluation Methods 0.000 abstract 1
- 238000012545 processing Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000010606 normalization Methods 0.000 description 6
- 239000000203 mixture Substances 0.000 description 5
- 230000008901 benefit Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 2
- 230000004438 eyesight Effects 0.000 description 2
- 239000004744 fabric Substances 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000001133 acceleration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 239000008199 coating composition Substances 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000004377 improving vision Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000002156 mixing Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000005215 recombination Methods 0.000 description 1
- 230000006798 recombination Effects 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 238000005303 weighing Methods 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
- G06T3/4069—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 by subpixel displacements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label, including：It establishes and trains a sub-pix convolutional network based on multiple dimensioned label being made of feature extraction structure, residual error module, up-sampling structure, feature reconstruction structure and across scale jump connection structure, and complete the super-resolution rebuilding work of image using the network.Input picture is transformed into YCbCr color spaces by RGB color.Wherein, super-resolution rebuilding work is completed in two channels Cb, Cr using the method that bicubic interpolation up-samples.The channels Y are sent in trained network, the super-resolution rebuilding image in the channels output Y.The super-resolution rebuilding image for merging the channel Y, Cb, Cr, obtains final high-definition picture.The present invention can rapidly and accurately obtain super-resolution image, and obtained super-resolution image either in subjectivity evaluation and test or in terms of Objective image quality evaluation, can obtain good effect.
Description
Technical field
The invention belongs to the crossing domains of Digital Image Processing, deep learning and pattern-recognition, more particularly, to one
Sub-pix convolved image super resolution ratio reconstruction method of the kind based on multiple dimensioned label.
Background technology
The resolution ratio for improving image has a very important significance in Digital Image Processing related field.However obtain image
Resolution sizes have close relationship, the wherein parameter and Optical manufacture technology of imaging sensor to determine with the facility for obtaining image
The resolution sizes of image can bring huge economic cost by improving vision facilities level of hardware.Obtain picture quality
It is also influenced by uncontrollable factors such as shooting environmental distances, by improving the level of hardware of shooting and improving the environment of shooting
Mode is limited to solve the problems, such as the ability of image resolution ratio.
Image resolution ratio is improved except through hardware mode, digital picture can also obtained and then by low resolution
Image reconstructs its corresponding high-definition picture using algorithm.Super-resolution technique advantage is to have broken away from image acquisition equipment
Hardware condition limitation, can use software method rebuild high-definition picture.Super-resolution technique is substantially to image
The prediction of high-frequency information has natural advantage in terms of information prediction, obtains due to the particularity of depth convolutional network structure
The high-definition picture arrived has good visual effect.Although traditional depth convolutional network can complete the weight of super-resolution
Work is built, but there are prediction results caused by the supervision of single label not to be inconsistent with true picture information, network structure is excessively complicated,
The problems such as calculating speed is slow.
Invention content
For the disadvantages described above or Improvement requirement of the prior art, the present invention provides a kind of sub- pictures based on multiple dimensioned label
Plain convolved image super resolution ratio reconstruction method, thus the solution prior art is not high to image high-frequency information predictablity rate, network
It is complicated, the slow problem of calculating speed.
The present invention provides a kind of the sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label, feature
It is, includes the following steps：
(1) Y channel images, Cb channel images and the channels Cr figure are obtained after carrying out channel decomposition to input low-resolution image
Picture；
(2) channels Cb are obtained after being rebuild respectively to the Cb channel images and the Cr channel images rebuilds high-resolution
Rate image and the channels Cr rebuild high-definition picture；
(3) using trained, the sub-pix convolutional network realization based on multiple dimensioned label surpasses the Y channel images
Resolution reconstruction；
(4) channels the Y figure after high-definition picture, the channels Cr reconstruction high-definition picture and reconstruction is rebuild to the channels Cb
As carrying out fusion treatment, the sub-pix convolved image super-resolution image based on multiple dimensioned label is obtained.
Further, step (3) is specially：
(3.1) primary features figure is obtained after carrying out feature extraction to Y channel images；
(3.2) characteristics of image is further extracted to primary features figure and obtains advanced features figure；
(3.3) up-sampling treatment is carried out to advanced features figure；
(3.4) the advanced features figure after up-sampling treatment is rebuild, obtains 2 times of super-resolution rebuilding results；
(3.5) advanced residual error feature is obtained after carrying out residual error resume module to the advanced features figure after one times of increase resolution
Figure；
(3.6) it is the primary features that the convolution mask of (2n-1) × (2n-1) obtains step (3.1) to use 256 sizes
Figure carries out convolution operation, obtains the characteristic pattern with 256 channels；Characteristic pattern with 256 channels is subjected to sub-pix volume
Product processing, i.e., carry out feature rearrangement to 256 channels, obtains the characteristic pattern with 64 channels, and this feature figure length and width respectively expand
One times, 2 times of characteristic pattern increase resolution；
(3.7) the advanced residual error for obtaining the characteristic pattern with 64 channels obtained in step (3.6) with step (3.5)
Characteristic pattern is merged, and the characteristic pattern after fusion is carried out up-sampling treatment, realizes the super-resolution to the Y channel images
It rebuilds.Due to having merged the characteristic pattern of different scale during this, therefore it is known as jump connection structure.
Further, primary features figure described in step (3.1) refers to that low-resolution image passes through feature extraction structure
It is obtained after extraction feature to have same size, the characteristic pattern with 64 channels with input picture.
Further, in step (3.1), feature extraction is carried out to Y channel images by the way of convolution, it is specific to wrap
It includes：The convolution that step-length is 1 is carried out on low-resolution image using the convolution mask that 64 sizes are (2n-1) × (2n-1) to grasp
Make, and is reused after being filled with 0 to the edge of input picture after PReLu functions carry out nonlinear activation and obtain primary features
Figure；Wherein, n is convolution kernel size；PReLu functions are：xiIt is inputted for activation primitive, α is
Trained parameter.
Further, advanced features figure described in step (3.2) refer to primary features figure via residual error resume module after
What is obtained has the characteristic pattern that same size has 64 channels with original image.
Further, step (3.3) is specially：Use the convolution mask pair that 256 sizes are (2n-1) × (2n-1)
Advanced features figure carries out convolution operation, and the channel number of advanced features figure is adjusted to 256, and by the 256 of advanced features figure
Every 4 adjacent channels are divided into 1 group in a channel, amount to 64 groups；Pixel value in every group of characteristic pattern channel same position according to
Channel sequence is rearranged, i.e., 4 width characteristic patterns in every group is fused to 1 characteristic pattern, the length and width point of advanced features figure
Do not become original 2 times；Channel number, which is reduced to previous a quarter, becomes 64, the increase resolution of advanced features figure
One times.
Further, step (3.5) is specially：Use 3 sizes be the convolution mask of (2n-1) × (2n-1) to point
Resolution promotes the advanced features figure after one times and carries out convolution operation, obtains 2 times of super-resolution rebuilding results in the channels Y.
Further, in step (3), the sub-pix convolutional network is carried out using multiple dimensioned label supervised training method
Training, specially：
(1) pass through first reconstruction obtained 2 times of super-resolution rebuilding image of structure, exercised supervision using 2 times of labels
Training, the partial supervised training process mainly act on the first half of network, can assist the training of overall network, Jin Ershi
Existing multiple dimensioned label training；
(2) pass through second reconstruction obtained 4 times of super-resolution rebuilding image of structure, exercised supervision using 4 times of labels
Training, the main supervised training standard of part network as a whole, acts on whole network.
Further, the loss function in training process is to be based on mean square error loss function, by 2 times of loss of amplification
Loss×2(Θ) and the loss Loss of 4 times of amplification×4(Θ) is formed.
Further, the loss function is：L (Θ)=Loss×4(Θ)+λ·Loss×2(Θ)；
The loss of 2 times of amplification and the costing bio disturbance formula for amplifying 4 times are as follows：
Wherein, XHR×2To amplify 2 times of high-definition picture,Indicate the i-th row of high-definition picture of 2 times of amplification
The pixel value of jth row.λ is weight coefficient, λ=1 during hands-on, that is, assigns 2 times of losses and 4 times of losses are identical
Weight；XHR×4To amplify 4 times of high-definition picture,Indicate the picture that high-definition picture the i-th row jth for amplifying 4 times arranges
Element value.YLRFor the low-resolution image of input, F×2(YLR；It is Θ) reconstruction of 2 times of the amplification by the output of super-resolution network
Image, F×4(YLR；It is Θ) reconstruction image of 4 times of the amplification by the output of super-resolution network, Θ is the parameter of network, M, N
Respectively input the length and width of low-resolution image；I, j indicate the i-th row jth row of image respectively.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, can obtain down and show
Beneficial effect：
(1) the method for the present invention completes picture up-sampling process using sub-pix convolutional network, using sub-pix convolutional coding structure,
Periodic arrangement is carried out to feature and obtains high-definition picture, network is greatly simplified, greatly reduces the complexity of calculating,
Algorithm speed is improved significantly.
(2) present invention uses multiple dimensioned label supervised training, exercises supervision training in the different location of network, effectively prevents
Only occur gradient disappearance or gradient disperse in training process.It can be restrained with the accuracy and acceleration model of boosting algorithm prediction
Speed.The loss function that the present invention is built loses two parts loss by 2 times of super-resolution image losses and 4 times of super-resolution images
Composition, considers the prediction result of different scale images, highly shortened the training time of model, obtained high-resolution
Rate image is more in line with human vision characteristics.
Description of the drawings
Fig. 1 is the flow chart of image super-resolution rebuilding method provided in an embodiment of the present invention；
Fig. 2 is low-resolution image provided in an embodiment of the present invention；
Fig. 3 (a) is Y channel decompositions figure provided in an embodiment of the present invention；
Fig. 3 (b) is Cb channel decompositions figure provided in an embodiment of the present invention；
Fig. 3 (c) is Cr channel decompositions figure provided in an embodiment of the present invention；
Fig. 4 (a) is Cb channel images super-resolution rebuilding result figure provided in an embodiment of the present invention；
Fig. 4 (b) is Cr channel images super-resolution rebuilding result figure provided in an embodiment of the present invention；
Fig. 5 is PReLU function curves schematic diagram provided in an embodiment of the present invention；
Fig. 6 is residual error structural schematic diagram provided in an embodiment of the present invention；
Fig. 7 is sub-pix convolution operation schematic diagram provided in an embodiment of the present invention；
Fig. 8 is across scale jump connection diagram provided in an embodiment of the present invention；
Fig. 9 is the channels 4 times of super-resolution Y provided in an embodiment of the present invention reconstructed results；
Figure 10 is that use the method for the present invention provided in an embodiment of the present invention carries out image super-resolution to low-resolution image
The result figure of reconstruction；
Figure 11 is the high-resolution that low-resolution image provided in an embodiment of the present invention is up-sampled using bicubic interpolation
Rate image.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
It does not constitute a conflict with each other and can be combined with each other.
To achieve the above object, the sub-pix convolved image super-resolution based on multiple dimensioned label that the present invention provides a kind of
Method for reconstructing, this method can effectively obtain high-definition picture, and described image super resolution ratio reconstruction method includes：
(1) it establishes and trains one by feature extraction structure, residual error module, up-sampling structure, rebuilds structure and across ruler
The sub-pix convolutional network based on multiple dimensioned label of degree jump connection structure composition.The network is to input low-resolution image
Concrete processing procedure is as follows：By the primary features of feature extraction structure extraction low-resolution image, primary features figure is obtained；Through
It crosses by the residual error module of 8 residual error structure compositions, 1 up-sampling structure and reconstruction structure obtain 2 times of super-resolution rebuilding figures
Picture；The characteristic pattern that first up-sampling structure is obtained, is sent in the residual error module by 8 residual error structure compositions, further
Characteristics of image is extracted, advanced residual error characteristic pattern is obtained；Primary features figure by across scale jump connection structure processing after with it is advanced
Residual error characteristic pattern merges, and 4 times of super-resolution rebuilding images are obtained by up-sampling structure and after rebuilding pattern handling.
(1.1) the feature extraction structure of network is used to extract the low-level features of low-resolution image, by convolutional layer and non-thread
Property active coating composition.Wherein, nonlinear activation layer activation primitive uses PReLU functions.
xiIt is inputted for activation primitive, α is to wait for training parameter.When such as input be positive number when, export for itself, when input be negative
When, it exports as direct proportion function, proportionality coefficient α is to be learned.
(1.2) the residual error module of network is for further extracting characteristics of image, each residual error module is by 8 residual error structures
Composition.Residual error structure is further through 2 convolutional layers, 2 batches of dimension normalization layers (Batch Normalization), an activation
Layer (PReLU) and one plus operation are composed.Specifically, the input picture of residual error structure is via convolutional layer, batch scale normalizing
Change layer, activation primitive layer, convolutional layer and the processing of batch dimension normalization layer further extraction characteristics of image, and with the input of the structure
Characteristic pattern is merged, and the output characteristic image of residual error structure is obtained.The structure is further by the Weak characteristic of input feature vector figure
Strengthen, obtains more representational feature in image.Wherein, data of the dimension normalization layer by a batch are criticized, are calculated
The mean μ of dataBAnd variances sigmaB 2, and learn scale parameter γ and offset parameter β, make having the same point of different input data
Cloth.
(1.3) the up-sampling structure of network is made of convolutional layer, sub-pix convolutional layer and nonlinear activation layer.For upper
The input picture of sampling structure, convolutional layer are used to adjust the characteristic dimension of input picture.Sub-pix convolutional layer passes through periodical row
Row operation increases image size and promotes image resolution ratio, by N × r2The low resolution feature of dimension is rearranged into N-dimensional high-resolution
Rate image.R=2 in the method for the present invention, in the characteristic image of obtained N*4 dimensions, the identical bits in 4 adjacent channels
4 pixel values taken are set, the characteristic block of a 2*2 size is rearranged by channel sequence, such 4 width characteristic pattern arranges recombination
Arrange into the characteristic pattern that 1 length and width respectively expands twice.The low-resolution image of whole N*4 dimensions is handled by sub-pix convolutional layer,
The length and width that can obtain N-dimensional degree expand twice of 4 times of high-definition pictures respectively.Low-resolution image is made with conventional method
It is compared with the method that the mode of linear interpolation completes up-sampling, sub-pix process of convolution used by this method can make full use of
Have characteristic pattern, do not need extra computation, effectively reduce the number of parameters of model, the complexity of network obtains letter significantly
Change, while the predictablity rate of model is also obtained for a degree of promotion.Wherein, the input of sub-pix convolutional layer is feature
Dimension is N × r2Characteristic pattern, export characteristic pattern characteristic dimension be N.Nonlinear activation layer uses PReLU activation primitives.
(1.4) the reconstruction structure of network is made of a convolutional layer, not use nonlinear activation layer, convolutional layer it is main
Purpose is to adjust output dimension size, complete high resolution image reconstruction.
(1.5) across the scale jump connection structure of network includes convolutional layer and sub-pix convolutional layer.Across scale jump connection
The primary features figure of image is amplified to correspondingly sized by structure by up-sampling treatment, and the latter half for being sent directly into network is real
Now across scale jump connection.Network synthesis considers primary features and advanced features, and it is pre- to high frequency section to further increase network
The accuracy rate of survey, meanwhile, accelerate the convergence rate of model.
(1.6) training process of network is as follows：Making data set has used 10k pictures, training dataset substantially to cut
250k channel image Y.In training process, batch processing is dimensioned to 16, e-learning rate η=0.0001.
(1.7) network is using multiple dimensioned label supervised training, the multiple dimensioned label supervised training method：In training net
During network, 2 times of resolution reconstruction images can be obtained by rebuilding structure by first, and 4 can be obtained by rebuilding structure for second
Times super-resolution rebuilding image.Finally obtained 4 times of super-resolution rebuilding images of network are carried out using the label of 4 times of sizes
Supervised training.The first half of network is exercised supervision training using 2 times of labels, which can assist the training of overall network,
And then realize multiple dimensioned label training.
(1.8) loss function in training process is based on mean square error loss function, mainly by 2 times of loss of amplification
Loss×2(Θ) and the loss Loss of 4 times of amplification×4(Θ) is formed.Its loss function is：
L (Θ)=Loss×4(Θ)+λ·Loss×2(Θ)
λ is weight coefficient, λ=1 during hands-on, that is, assigns 2 times of losses and the identical weight of 4 times of losses.
Specifically, it is as follows to amplify 2 times of loss and the costing bio disturbance formula for amplifying 4 times：
Wherein, XHR×2To amplify 2 times of high-definition picture,Indicate the i-th row of high-definition picture of 2 times of amplification
The pixel value of jth row.XHR×4To amplify 4 times of high-definition picture,Indicate the i-th row of high-definition picture of 4 times of amplification
The pixel value of jth row.YLRFor the low-resolution image of input, F×2(YLR；It is Θ) amplification by the output of super-resolution network
2 times of reconstruction image, F×4(YLR；It is Θ) reconstruction image of 4 times of the amplification by the output of super-resolution network, Θ is network
Parameter, M, N are respectively the length and width of input feature vector figure.I, j indicate the i-th row jth row of image respectively.
(2) input picture is transformed into YCbCr color spaces by RGB color, wherein the channel Cb, Cr uses double three
The method of secondary interpolation up-sampling completes image super-resolution rebuilding, and Y channel images are input to trained based on multiple dimensioned label
Sub-pix convolutional network in complete super-resolution rebuilding.The reconstructed results for merging above three channel, obtain final high-resolution
Rate exports image.
(2.1) most of images are to preserve image by storing the RGB information of image, in super-resolution image reconstruction
In the process, YCbCr color spaces, which compare rgb color space, more has advantage.In reconstruction process, first by input picture by
RGB color is transformed into YCbCr color spaces, further processes, and color space conversion formula is as follows：
(2.2) super-resolution rebuilding in the channels Y works by the sub-pix convolution super-resolution network based on multiple dimensioned label
It completes.
The present invention is described in further details by taking low-resolution image shown in Fig. 2 as an example below.Below in conjunction with attached drawing 1~
The present invention is further described for Figure 11 and embodiment.
As shown in Figure 1, a kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label, described image
Super resolution ratio reconstruction method includes：
(1) input low-resolution image is transformed into YCbCr color spaces by RGB color, color space conversion is public
Formula is as follows：
Wherein, R, G, B are respectively the pixel value corresponding to low-resolution image R, G, B triple channel, and Y, Cb, Cr are after converting
The corresponding pixel value of triple channel.Wherein, the channels Y transformation range [16,235], the channel Cb, Cr transformation range [16,240].And it will
Image after color space conversion does channel decomposition operation, carries out super-resolution rebuilding to three channels respectively.Fig. 3 (a) is logical for Y
Road exploded view, Fig. 3 (b) are Cb channel decomposition figures, and Fig. 3 (c) is Cr channel decomposition figures.
(2) for Cb, Cr channel image after decomposition, the method up-sampled using bicubic interpolation completes super-resolution
It rebuilds, Fig. 4 (a) (b) is respectively to rebuild high-definition picture using the channel Cb, Cr that bicubic interpolation top sampling method obtains.
(3) complete using the trained sub-pix convolutional network based on multiple dimensioned label for the Y channel images after decomposition
At image super-resolution rebuilding, detailed process is as follows：
(3.1) primary features figure is obtained after carrying out feature extraction to Y channel images.Primary features figure refers to low resolution figure
There are same size, the feature with 64 channels with input picture as obtained after feature extraction structure extraction feature
Figure.Wherein, feature extraction is carried out by the way of convolution herein, it is the volume of (2n-1) × (2n-1) specifically to use 64 sizes
Product module plate, the method for the present invention n=5.The convolution operation that step-length is 1 is carried out on low-resolution image, to ensure to extract front and back figure
As size does not change, the edge of input picture is filled with 0.Acquired results are carried out non-linear using PReLu functions
Activation.
xiIt is inputted for activation primitive, α is trained parameter.When such as input be positive number when, export for itself, work as input
It when being negative, exports as direct proportion function, proportionality coefficient α determines that Fig. 5 gives the function curve of PReLU in the training process.
(3.2) characteristics of image is further extracted to primary features figure and obtains advanced features figure.Advanced features figure refers to primary
Characteristic pattern is via obtaining having the characteristic pattern that same size has 64 channels with original image after residual error resume module.Wherein, right
Primary features figure carries out convolution operation, and convolution results carry out batch dimension normalization and handle and use the progress of PReLu functions non-linear
Activation, repeats convolution sum batch dimension normalization operation, and acquired results are merged with primary features figure, completed at a residual error structure
Reason.It is the convolution mask of (2n-1) × (2n-1), the method for the present invention n=2 that above-mentioned convolution operation, which uses 64 sizes,.It is primary
Characteristic pattern obtains advanced features figure by 8 residual error pattern handlings.If residual error structure contained by residual error module is very few, spy can be caused
Sign extraction is insufficient；If contained residual error structure is excessive, increase model complexity even over-fitting.Therefore it is determining residual error structure
Number, the present invention design 6 groups of contrast experiments for residual error structure number and super-resolution rebuilding result, and related experiment shows warp
The advanced features plot quality highest obtained by 8 residual error pattern handlings, therefore, each residual error module include 8 residual error structures.Figure
6 give residual error structural schematic diagram.BN layers of introducing can make the data point having the same during training and test
Cloth, network performance are more excellent.
(3.3) up-sampling treatment is carried out to advanced features figure.Specifically, the use of 256 sizes is (2n-1) × (2n-1)
Convolution mask, convolution operation is carried out to advanced features figure, the channel number of advanced features figure is adjusted to 256, the present invention
Method n=3.By 256 channels of advanced features figure, every 4 adjacent channels are divided into 1 group, amount to 64 groups.Every group of characteristic pattern is logical
Pixel value in road same position is rearranged according to channel sequence, i.e., 4 width characteristic patterns in every group is fused to 1 spy
Sign figure, characteristic pattern length and width become original 2 times respectively.Advanced features figure is by 2 times that become length and width be respectively previous, channel number
To be reduced to previous a quarter become 64, one times of the increase resolution of advanced features figure.Concrete operations as shown in fig. 7,
The input of sub-pix convolutional layer is that characteristic dimension is N × r2Characteristic pattern, export characteristic pattern characteristic dimension be N, picture size
One times is promoted, resolution sizes promote one times.Nonlinear activation is carried out using PReLU activation primitives to advanced features figure.
(3.4) the advanced features figure obtained after up-sampling treatment is rebuild, obtains 2 times of super-resolution rebuilding knots
Fruit.Specifically, to the advanced features figure after one times of increase resolution, it is the convolution mould of (2n-1) × (2n-1) to use 3 sizes
Plate carries out convolution operation, obtains 2 times of super-resolution rebuildings in the channels Y as a result, the method for the present invention n=5.
(3.5) residual error resume module is carried out to the advanced features figure after one times of increase resolution, wherein residual error module is same
It is made of 8 residual error structures, further extracts characteristics of image, gained characteristic pattern is known as advanced residual error characteristic pattern.
(3.6) convolution mask that 256 sizes are (2n-1) × (2n-1) is used to carry out convolution operation to primary features figure,
Obtain the characteristic pattern with 256 channels, n=2 of the embodiment of the present invention.Using aforementioned sub-pix convolution method, it is adjusted to length and width
Respectively expand the twice characteristic pattern with 64 channels, makes one times of its increase resolution.Acquired results and advanced residual error characteristic pattern
Fusion, obtains an abundanter characteristic pattern of information.This feature figure contains the shallow-layer feature initially extracted and depth is special
Reference ceases, above-mentioned that across the scale mixing operation of shallow-layer and depth information is known as across scale jump connection, and Fig. 8 gives across scale
Jump connection diagram.
(3.7) up-sampled and exported reconstruction operation to connecting the characteristic pattern after obtained fusion via jump.Tool
Body, using the convolution mask that 256 sizes are (2n-1) × (2n-1), convolution operation is carried out to the characteristic pattern after fusion, it will
The port number of characteristic pattern is adjusted to 256, the method for the present invention n=2.To the obtained characteristic pattern with 256 channels, per adjacent
4 channels be divided into 1 group, amount to 64 groups.Pixel value in every group in same position is rearranged according to channel sequence, will be deep
Degree information is converted into location information, and each expansion of gained characteristic pattern length and width, which is twice port number, becomes 64, up-sampling is completed, at this point, phase
Than in 4 times of low-resolution image increase resolution.4 times of super-resolution images are used with the convolution mould of 3 (2n-1) × (2n-1)
Plate carries out convolution operation, and takes mean value to three channels of convolution results, as the 4 times of channels super-resolution Y reconstruction images, this hair
Bright embodiment n=5.Fig. 9 gives the 4 times of channels super-resolution Y reconstruction images of sample picture.
(4) reconstructed results in the channel Cb, Cr and the reconstructed results in the channels Y in fusion steps (2), obtain final high score
Resolution image completes super-resolution rebuilding work.
Figure 10 is that use the method for the present invention provided in an embodiment of the present invention carries out super-resolution rebuilding to low-resolution image
Result；By compared with the high-definition picture of Figure 11 obtained using bicubic interpolation：Obtained by the subjective aspect present invention
Super-resolution rebuilding image, image resolution ratio higher can be very good the high-frequency information of prognostic chart picture, meet human vision spy
Point；Objective aspects, as an important indicator for weighing picture quality, usual high quality graphic possesses higher Y-PSNR
4 times of super-resolution rebuilding image Y-PSNRs of Y-PSNR, the embodiment of the present invention are 27.82, use bicubic interpolation
Obtained high-definition picture Y-PSNR is only 22.1, and the method for the present invention rebuilds effect promoting 25.9%.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, all within the spirits and principles of the present invention made by all any modification, equivalent and improvement etc., should all include
Within protection scope of the present invention.
Claims (10)
1. a kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label, which is characterized in that including as follows
Step：
(1) Y channel images, Cb channel images and Cr channel images are obtained after carrying out channel decomposition to input low-resolution image；
(2) channels Cb are obtained after being rebuild respectively to the Cb channel images and the Cr channel images rebuilds high resolution graphics
Picture and the channels Cr rebuild high-definition picture；
(3) using trained, the sub-pix convolutional network based on multiple dimensioned label realizes the super-resolution to the Y channel images
Rate is rebuild；
(4) to the channels Cb rebuild high-definition picture, the channels Cr rebuild high-definition picture and rebuild after Y channel images into
Row fusion treatment obtains the sub-pix convolved image super-resolution image based on multiple dimensioned label.
2. sub-pix convolved image super resolution ratio reconstruction method as described in claim 1, which is characterized in that step (3) is specific
For：
(3.1) primary features figure is obtained after carrying out feature extraction to Y channel images；
(3.2) characteristics of image is further extracted to primary features figure and obtains advanced features figure；
(3.3) up-sampling treatment is carried out to advanced features figure；
(3.4) the advanced features figure after up-sampling treatment is rebuild, obtains 2 times of super-resolution rebuilding results；
(3.5) advanced residual error characteristic pattern is obtained after carrying out residual error resume module to the advanced features figure after one times of increase resolution；
(3.6) use the primary features figure that convolution mask that 256 sizes are (2n-1) × (2n-1) obtains step (3.1) into
Row convolution operation obtains the characteristic pattern with 256 channels；Characteristic pattern with 256 channels is carried out at sub-pix convolution
Reason carries out feature rearrangement to 256 channels, obtain the characteristic pattern with 64 channels, and this feature figure length and width, which respectively expand, to be twice,
2 times of characteristic pattern increase resolution；
(3.7) the advanced residual error feature for obtaining the characteristic pattern with 64 channels obtained in step (3.6) with step (3.5)
Figure is merged, and the characteristic pattern after fusion is carried out up-sampling treatment, realizes the Super-resolution reconstruction to the Y channel images
It builds.
3. sub-pix convolved image super resolution ratio reconstruction method as claimed in claim 2, which is characterized in that in step (3.1)
The primary features figure refers to that low-resolution image is obtained after feature extraction structure extraction feature to be had with input picture
Same size, the characteristic pattern with 64 channels.
4. sub-pix convolved image super resolution ratio reconstruction method as claimed in claim 2 or claim 3, which is characterized in that in step
(3.1) in, feature extraction is carried out to Y channel images by the way of convolution, is specifically included：
The convolution that step-length is 1 is carried out on low-resolution image using the convolution mask that 64 sizes are (2n-1) × (2n-1) to grasp
Make, and is reused after being filled with 0 to the edge of input picture after PReLu functions carry out nonlinear activation and obtain primary features
Figure；
Wherein, n is convolution kernel size；PReLu functions are：xiIt is inputted for activation primitive, α is
Trained parameter.
5. such as claim 2-4 any one of them sub-pix convolved image super resolution ratio reconstruction methods, which is characterized in that step
(3.2) advanced features figure described in refer to primary features figure has same size via what is obtained after residual error resume module with original image
Characteristic pattern with 64 channels.
6. such as claim 2-5 any one of them sub-pix convolved image super resolution ratio reconstruction methods, which is characterized in that step
(3.3) it is specially：
Convolution operation is carried out to advanced features figure using the convolution mask that 256 sizes are (2n-1) × (2n-1), by advanced spy
The channel number of sign figure is adjusted to 256, and every 4 adjacent channels in 256 channels of advanced features figure are divided into 1 group,
It is 64 groups total；
Pixel value in every group of characteristic pattern channel same position is rearranged according to channel sequence, i.e., by 4 width in every group
Characteristic pattern is fused to 1 characteristic pattern, and the length and width of advanced features figure become original 2 times respectively；Channel number is reduced to previous
A quarter becomes 64, one times of the increase resolution of advanced features figure.
7. such as claim 2-6 any one of them sub-pix convolved image super resolution ratio reconstruction methods, which is characterized in that step
(3.5) it is specially：It is the convolution mask of (2n-1) × (2n-1) to the advanced features after one times of increase resolution to use 3 sizes
Figure carries out convolution operation, obtains 2 times of super-resolution rebuilding results in the channels Y.
8. such as claim 2-7 any one of them sub-pix convolved image super resolution ratio reconstruction methods, which is characterized in that step
(3) in, the sub-pix convolutional network is trained using multiple dimensioned label supervised training method, specially：
(1) pass through first reconstruction obtained 2 times of super-resolution rebuilding image of structure, exercised supervision training using 2 times of labels,
The partial supervised training process mainly acts on the first half of network, can assist the training of overall network, and then realize more
Scale label training；
(2) pass through second reconstruction obtained 4 times of super-resolution rebuilding image of structure, exercised supervision training using 4 times of labels,
The main supervised training standard of part network as a whole, acts on whole network.
9. sub-pix convolved image super resolution ratio reconstruction method as claimed in claim 8, which is characterized in that in training process
Loss function is to be based on mean square error loss function, by 2 times of loss Loss of amplification×2The loss of (Θ) and 4 times of amplification
Loss×4(Θ) is formed.
10. sub-pix convolved image super resolution ratio reconstruction method as claimed in claim 9, which is characterized in that the loss letter
Number is：L (Θ)=Loss×4(Θ)+λ·Loss×2(Θ)；
The loss of 2 times of amplification and the costing bio disturbance formula for amplifying 4 times are as follows：
Wherein, XHR×2To amplify 2 times of high-definition picture,Indicate that high-definition picture the i-th row jth for amplifying 2 times arranges
Pixel value；λ is weight coefficient, λ=1 during hands-on, that is, assigns 2 times of losses and the identical weight of 4 times of losses；
XHR×4To amplify 4 times of high-definition picture,Indicate the pixel value that high-definition picture the i-th row jth for amplifying 4 times arranges；
YLRFor the low-resolution image of input, F×2(YLR；It is Θ) reconstruction image of 2 times of the amplification by the output of super-resolution network,
F×4(YLR；It is Θ) reconstruction image of 4 times of the amplification by the output of super-resolution network, Θ is the parameter of network, M, N difference
To input the length and width of low-resolution image；I, j indicate the i-th row jth row of image respectively.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810473375.8A CN108734659B (en) | 2018-05-17 | 2018-05-17 | Sub-pixel convolution image super-resolution reconstruction method based on multi-scale label |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810473375.8A CN108734659B (en) | 2018-05-17 | 2018-05-17 | Sub-pixel convolution image super-resolution reconstruction method based on multi-scale label |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108734659A true CN108734659A (en) | 2018-11-02 |
| CN108734659B CN108734659B (en) | 2021-08-20 |
Family
ID=63937545
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810473375.8A Expired - Fee Related CN108734659B (en) | 2018-05-17 | 2018-05-17 | Sub-pixel convolution image super-resolution reconstruction method based on multi-scale label |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108734659B (en) |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109522958A (en) * | 2018-11-16 | 2019-03-26 | 中山大学 | Based on the depth convolutional neural networks object detection method merged across scale feature |
| CN109544451A (en) * | 2018-11-14 | 2019-03-29 | 武汉大学 | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection |
| CN109934771A (en) * | 2019-03-08 | 2019-06-25 | 北京航空航天大学 | Unsupervised Remote sensed image super-resolution reconstruction method based on Recognition with Recurrent Neural Network |
| CN109977942A (en) * | 2019-02-02 | 2019-07-05 | 浙江工业大学 | A kind of scene character recognition method based on scene classification and super-resolution |
| CN109978763A (en) * | 2019-03-01 | 2019-07-05 | 昆明理工大学 | A kind of image super-resolution rebuilding algorithm based on jump connection residual error network |
| CN110009568A (en) * | 2019-04-10 | 2019-07-12 | 大连民族大学 | The generator construction method of language of the Manchus image super-resolution rebuilding |
| CN110020989A (en) * | 2019-05-23 | 2019-07-16 | 西华大学 | A kind of depth image super resolution ratio reconstruction method based on deep learning |
| CN110097505A (en) * | 2019-05-16 | 2019-08-06 | 中国人民解放军海军工程大学 | A kind of Law of DEM Data processing method and processing device |
| CN110163803A (en) * | 2019-05-24 | 2019-08-23 | 厦门大学 | A kind of image super-resolution rebuilding method and system based on convolutional neural networks |
| CN110245683A (en) * | 2019-05-13 | 2019-09-17 | 华中科技大学 | The residual error relational network construction method that sample object identifies a kind of less and application |
| CN110675321A (en) * | 2019-09-26 | 2020-01-10 | 兰州理工大学 | Super-resolution image reconstruction method based on progressive depth residual error network |
| CN111062867A (en) * | 2019-11-21 | 2020-04-24 | 浙江大华技术股份有限公司 | Video super-resolution reconstruction method |
| CN111160413A (en) * | 2019-12-12 | 2020-05-15 | 天津大学 | Thyroid nodule classification method based on multi-scale feature fusion |
| CN111161166A (en) * | 2019-12-16 | 2020-05-15 | 西安交通大学 | Image moire eliminating method based on depth multi-resolution network |
| CN111402138A (en) * | 2020-03-24 | 2020-07-10 | 天津城建大学 | Image super-resolution reconstruction method of supervised convolutional neural network based on multi-scale feature extraction fusion |
| CN111461983A (en) * | 2020-03-31 | 2020-07-28 | 华中科技大学鄂州工业技术研究院 | Image super-resolution reconstruction model and method based on different frequency information |
| CN111476719A (en) * | 2020-05-06 | 2020-07-31 | Oppo广东移动通信有限公司 | Image processing method, image processing device, computer equipment and storage medium |
| CN111640067A (en) * | 2020-06-10 | 2020-09-08 | 华侨大学 | Single image super-resolution reconstruction method based on three-channel convolutional neural network |
| CN111754400A (en) * | 2020-06-01 | 2020-10-09 | 杭州电子科技大学 | Efficient picture super-resolution reconstruction method |
| WO2020220797A1 (en) * | 2019-04-30 | 2020-11-05 | 北京市商汤科技开发有限公司 | Method, device and apparatus for amplifying feature map, and computer readable storage medium |
| CN112381839A (en) * | 2020-11-14 | 2021-02-19 | 四川大学华西医院 | Breast cancer pathological image HE cancer nest segmentation method based on deep learning |
| CN113222822A (en) * | 2021-06-02 | 2021-08-06 | 西安电子科技大学 | Hyperspectral image super-resolution reconstruction method based on multi-scale transformation |
| CN113298716A (en) * | 2021-05-31 | 2021-08-24 | 重庆师范大学 | Image super-resolution reconstruction method based on convolutional neural network |
| CN113313632A (en) * | 2021-06-11 | 2021-08-27 | 展讯通信(天津)有限公司 | Image reconstruction method, system and related equipment |
| CN113596576A (en) * | 2021-07-21 | 2021-11-02 | 杭州网易智企科技有限公司 | Video super-resolution method and device |
| CN113837015A (en) * | 2021-08-31 | 2021-12-24 | 艾普工华科技（武汉）有限公司 | Face detection method and system based on feature pyramid |
| CN113902658A (en) * | 2021-09-01 | 2022-01-07 | 西安电子科技大学 | RGB image-to-hyperspectral image reconstruction method based on dense multiscale network |
| CN116702876A (en) * | 2023-04-27 | 2023-09-05 | 贵州大学 | Image countermeasure defense method based on preprocessing |
| CN118229532A (en) * | 2024-05-21 | 2024-06-21 | 华侨大学 | Image super-resolution method and system based on composite transducer |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130170767A1 (en) * | 2012-01-04 | 2013-07-04 | Anustup Kumar CHOUDHURY | Image content enhancement using a dictionary technique |
| CN106683067A (en) * | 2017-01-20 | 2017-05-17 | 福建帝视信息科技有限公司 | Deep learning super-resolution reconstruction method based on residual sub-images |
| CN107154023A (en) * | 2017-05-17 | 2017-09-12 | 电子科技大学 | Face super-resolution reconstruction method based on generation confrontation network and sub-pix convolution |
| CN107240066A (en) * | 2017-04-28 | 2017-10-10 | 天津大学 | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks |
-
2018
- 2018-05-17 CN CN201810473375.8A patent/CN108734659B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130170767A1 (en) * | 2012-01-04 | 2013-07-04 | Anustup Kumar CHOUDHURY | Image content enhancement using a dictionary technique |
| CN106683067A (en) * | 2017-01-20 | 2017-05-17 | 福建帝视信息科技有限公司 | Deep learning super-resolution reconstruction method based on residual sub-images |
| CN107240066A (en) * | 2017-04-28 | 2017-10-10 | 天津大学 | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks |
| CN107154023A (en) * | 2017-05-17 | 2017-09-12 | 电子科技大学 | Face super-resolution reconstruction method based on generation confrontation network and sub-pix convolution |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109544451A (en) * | 2018-11-14 | 2019-03-29 | 武汉大学 | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection |
| CN109522958A (en) * | 2018-11-16 | 2019-03-26 | 中山大学 | Based on the depth convolutional neural networks object detection method merged across scale feature |
| CN109977942A (en) * | 2019-02-02 | 2019-07-05 | 浙江工业大学 | A kind of scene character recognition method based on scene classification and super-resolution |
| CN109978763A (en) * | 2019-03-01 | 2019-07-05 | 昆明理工大学 | A kind of image super-resolution rebuilding algorithm based on jump connection residual error network |
| CN109934771A (en) * | 2019-03-08 | 2019-06-25 | 北京航空航天大学 | Unsupervised Remote sensed image super-resolution reconstruction method based on Recognition with Recurrent Neural Network |
| CN109934771B (en) * | 2019-03-08 | 2021-03-12 | 北京航空航天大学 | Unsupervised remote sensing image super-resolution reconstruction method based on recurrent neural network |
| CN110009568A (en) * | 2019-04-10 | 2019-07-12 | 大连民族大学 | The generator construction method of language of the Manchus image super-resolution rebuilding |
| WO2020220797A1 (en) * | 2019-04-30 | 2020-11-05 | 北京市商汤科技开发有限公司 | Method, device and apparatus for amplifying feature map, and computer readable storage medium |
| US11049217B2 (en) | 2019-04-30 | 2021-06-29 | Beijing Sensetime Technology Development Co., Ltd. | Magnifying feature map |
| CN110245683A (en) * | 2019-05-13 | 2019-09-17 | 华中科技大学 | The residual error relational network construction method that sample object identifies a kind of less and application |
| CN110245683B (en) * | 2019-05-13 | 2021-07-27 | 华中科技大学 | Residual error relation network construction method for less-sample target identification and application |
| CN110097505A (en) * | 2019-05-16 | 2019-08-06 | 中国人民解放军海军工程大学 | A kind of Law of DEM Data processing method and processing device |
| CN110020989A (en) * | 2019-05-23 | 2019-07-16 | 西华大学 | A kind of depth image super resolution ratio reconstruction method based on deep learning |
| CN110163803A (en) * | 2019-05-24 | 2019-08-23 | 厦门大学 | A kind of image super-resolution rebuilding method and system based on convolutional neural networks |
| CN110675321A (en) * | 2019-09-26 | 2020-01-10 | 兰州理工大学 | Super-resolution image reconstruction method based on progressive depth residual error network |
| CN111062867A (en) * | 2019-11-21 | 2020-04-24 | 浙江大华技术股份有限公司 | Video super-resolution reconstruction method |
| CN111160413A (en) * | 2019-12-12 | 2020-05-15 | 天津大学 | Thyroid nodule classification method based on multi-scale feature fusion |
| CN111160413B (en) * | 2019-12-12 | 2023-11-17 | 天津大学 | Thyroid nodule classification method based on multi-scale feature fusion |
| CN111161166A (en) * | 2019-12-16 | 2020-05-15 | 西安交通大学 | Image moire eliminating method based on depth multi-resolution network |
| CN111402138A (en) * | 2020-03-24 | 2020-07-10 | 天津城建大学 | Image super-resolution reconstruction method of supervised convolutional neural network based on multi-scale feature extraction fusion |
| CN111461983A (en) * | 2020-03-31 | 2020-07-28 | 华中科技大学鄂州工业技术研究院 | Image super-resolution reconstruction model and method based on different frequency information |
| CN111461983B (en) * | 2020-03-31 | 2023-09-19 | 华中科技大学鄂州工业技术研究院 | Image super-resolution reconstruction model and method based on different frequency information |
| CN111476719A (en) * | 2020-05-06 | 2020-07-31 | Oppo广东移动通信有限公司 | Image processing method, image processing device, computer equipment and storage medium |
| CN111476719B (en) * | 2020-05-06 | 2023-05-12 | Oppo广东移动通信有限公司 | Image processing method, device, computer equipment and storage medium |
| CN111754400B (en) * | 2020-06-01 | 2023-12-26 | 杭州电子科技大学 | Efficient picture super-resolution reconstruction method |
| CN111754400A (en) * | 2020-06-01 | 2020-10-09 | 杭州电子科技大学 | Efficient picture super-resolution reconstruction method |
| CN111640067B (en) * | 2020-06-10 | 2022-08-09 | 华侨大学 | Single image super-resolution reconstruction method based on three-channel convolutional neural network |
| CN111640067A (en) * | 2020-06-10 | 2020-09-08 | 华侨大学 | Single image super-resolution reconstruction method based on three-channel convolutional neural network |
| CN112381839A (en) * | 2020-11-14 | 2021-02-19 | 四川大学华西医院 | Breast cancer pathological image HE cancer nest segmentation method based on deep learning |
| CN113298716A (en) * | 2021-05-31 | 2021-08-24 | 重庆师范大学 | Image super-resolution reconstruction method based on convolutional neural network |
| CN113298716B (en) * | 2021-05-31 | 2023-09-12 | 重庆师范大学 | Image super-resolution reconstruction method based on convolutional neural network |
| CN113222822A (en) * | 2021-06-02 | 2021-08-06 | 西安电子科技大学 | Hyperspectral image super-resolution reconstruction method based on multi-scale transformation |
| CN113222822B (en) * | 2021-06-02 | 2023-01-24 | 西安电子科技大学 | Hyperspectral image super-resolution reconstruction method based on multi-scale transformation |
| CN113313632A (en) * | 2021-06-11 | 2021-08-27 | 展讯通信(天津)有限公司 | Image reconstruction method, system and related equipment |
| CN113596576A (en) * | 2021-07-21 | 2021-11-02 | 杭州网易智企科技有限公司 | Video super-resolution method and device |
| CN113837015A (en) * | 2021-08-31 | 2021-12-24 | 艾普工华科技（武汉）有限公司 | Face detection method and system based on feature pyramid |
| CN113902658B (en) * | 2021-09-01 | 2023-02-10 | 西安电子科技大学 | RGB image-to-hyperspectral image reconstruction method based on dense multiscale network |
| CN113902658A (en) * | 2021-09-01 | 2022-01-07 | 西安电子科技大学 | RGB image-to-hyperspectral image reconstruction method based on dense multiscale network |
| CN116702876A (en) * | 2023-04-27 | 2023-09-05 | 贵州大学 | Image countermeasure defense method based on preprocessing |
| CN116702876B (en) * | 2023-04-27 | 2024-04-12 | 贵州大学 | Image countermeasure defense method based on preprocessing |
| CN118229532A (en) * | 2024-05-21 | 2024-06-21 | 华侨大学 | Image super-resolution method and system based on composite transducer |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108734659B (en) | 2021-08-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108734659A (en) | A kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label | |
| Du et al. | Super-resolution reconstruction of single anisotropic 3D MR images using residual convolutional neural network | |
| WO2021022929A1 (en) | Single-frame image super-resolution reconstruction method | |
| CN105046672B (en) | A kind of image super-resolution rebuilding method | |
| CN108734660A (en) | A kind of image super-resolution rebuilding method and device based on deep learning | |
| CN110458906B (en) | Medical image coloring method based on depth color migration | |
| CN108460726A (en) | A kind of magnetic resonance image super-resolution reconstruction method based on enhancing recurrence residual error network | |
| CN106157244A (en) | A kind of QR Code Image Super-resolution Reconstruction method based on rarefaction representation | |
| CN107609638A (en) | A kind of method based on line decoder and interpolation sampling optimization convolutional neural networks | |
| CN108492248A (en) | Depth map super-resolution method based on deep learning | |
| CN109903292A (en) | A kind of three-dimensional image segmentation method and system based on full convolutional neural networks | |
| CN107369189A (en) | The medical image super resolution ratio reconstruction method of feature based loss | |
| CN107464216A (en) | A kind of medical image ultra-resolution ratio reconstructing method based on multilayer convolutional neural networks | |
| Chen et al. | Single image super-resolution using deep CNN with dense skip connections and inception-resnet | |
| CN109685716A (en) | A kind of image super-resolution rebuilding method of the generation confrontation network based on Gauss encoder feedback | |
| CN109544457A (en) | Image super-resolution method, storage medium and terminal based on fine and close link neural network | |
| CN109961396A (en) | A kind of image super-resolution rebuilding method based on convolutional neural networks | |
| CN110322402A (en) | Medical image super resolution ratio reconstruction method based on dense mixing attention network | |
| CN106709872B (en) | A kind of rapid image ultra-resolution ratio reconstructing method | |
| CN109214989A (en) | Single image super resolution ratio reconstruction method based on Orientation Features prediction priori | |
| CN105631807A (en) | Single-frame image super resolution reconstruction method based on sparse domain selection | |
| CN109559278B (en) | Super resolution image reconstruction method and system based on multiple features study | |
| CN112926607B (en) | Two-branch network image steganography framework and method based on convolutional neural network | |
| CN108470350A (en) | Broken line dividing method in line chart and device | |
| CN109410114A (en) | Compressed sensing image reconstruction algorithm based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| 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: 20210820 |