CN109544451A - A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection - Google Patents
A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection Download PDFInfo
- Publication number
- CN109544451A CN109544451A CN201811352508.2A CN201811352508A CN109544451A CN 109544451 A CN109544451 A CN 109544451A CN 201811352508 A CN201811352508 A CN 201811352508A CN 109544451 A CN109544451 A CN 109544451A
- Authority
- CN
- China
- Prior art keywords
- layer
- image
- gradual
- characteristic pattern
- resolution
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000005070 sampling Methods 0.000 claims abstract description 49
- 230000003321 amplification Effects 0.000 claims abstract description 14
- 238000003199 nucleic acid amplification method Methods 0.000 claims abstract description 14
- 101000911390 Homo sapiens Coagulation factor VIII Proteins 0.000 claims description 10
- 102000057593 human F8 Human genes 0.000 claims description 10
- 229940047431 recombinate Drugs 0.000 claims description 10
- 238000005215 recombination Methods 0.000 claims description 8
- 230000006798 recombination Effects 0.000 claims description 8
- 230000006835 compression Effects 0.000 claims description 6
- 238000007906 compression Methods 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 239000000203 mixture Substances 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims 1
- 238000013507 mapping Methods 0.000 abstract description 5
- 238000007796 conventional method Methods 0.000 abstract description 4
- 230000000694 effects Effects 0.000 abstract description 4
- 238000013527 convolutional neural network Methods 0.000 abstract description 3
- 238000012549 training Methods 0.000 abstract description 2
- 238000013135 deep learning Methods 0.000 description 4
- 230000008713 feedback mechanism Effects 0.000 description 3
- 238000003384 imaging method Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004883 computer application Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000002360 preparation method 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
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (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)
Abstract
The invention discloses a kind of image super-resolution rebuilding method and system based on gradual iterative backprojection, it is intended to explore a kind of method for adapting to big multiplying power (such as 8x) image resolution ratio amplification task using depth convolutional neural networks.The system includes characteristic extracting module, gradual iterative backprojection module, image reconstruction module.Wherein, characteristic extracting module is used to extract the characteristic information of input picture, generates multichannel image characteristic pattern；Gradual iterative backprojection module is used to carry out details prediction to input picture, sampling is iterated to characteristics of image figure using pyramid back projection unit, gradual cascade sampling avoids the disposable mapping in conventional method in big multiplying power factor super-resolution rebuilding, improves the training effectiveness of big multiplying power factor model and rebuilds effect；Multi-channel feature of the image reconstruction module for image compresses, and recovers the amplified triple channel RGB image of resolution ratio.
Description
Technical field
The present invention relates to machine learning, computer application technology, anti-based on gradual iteration more particularly to one kind
The image super-resolution rebuilding method and system of projection.
Background technique
Image resolution ratio indicates that image is capable of providing the level of detail of information, it is a weight for measuring picture quality quality
Index is wanted, can be defined with the number of pixel in unit area.What the higher expression image of resolution ratio was capable of providing recognizes
Details it is abundanter, entrained useful information is also more.Image super-resolution rebuilding, which refers to, utilizes a frame low resolution figure
As or image sequence, finally reconstruct a vertical frame dimension image in different resolution) image processing techniques.High-definition picture tool after reconstruction
There is higher details resolution capability, plays a key effect in engineering practice and practical application.
Super-resolution rebuilding difficult point essentially consists in the picture reconstruction from Poor information to the picture of high information quantity, lacks necessity
Minutia.The super-resolution rebuilding in one's early years, which represents method, adaptive approach, projections onto convex sets and regularization method.These
Although the processing based on airspace can make full use of the prior information of image, limited by technology, they are to low-resolution image
Increased quality it is fairly limited, many minutias of missing image are generated, far from meeting people to high-definition picture
Demand.
As machine learning is using more and more extensive, the super-resolution rebuilding research based on rarefaction representation becomes this field
Hot topic, and a collection of super resolution ratio reconstruction method based on rarefaction representation that has been born has his own strong points in computational accuracy and speed.
However, the quality of image reconstruction is highly dependent on each image block in super complete dictionary.If desired the low resolution figure rebuild
As block and the image block similarity in dictionary be not high, it is bad to will lead to reconstruction quality；In addition, the super-resolution based on sparse coding
Reconstruction is limited for the assessment of high-definition picture details, optimizes not comprehensive enough.Based on the above reasons, researchers begin look for energy
Enough learn more details feature, compatibility is better, optimizes more comprehensive method.
With the fast development of deep learning in recent years, the researchers in super-resolution rebuilding field also start attention
Deep learning is invested, and proposes a series of deep learning network frames applied to image super-resolution rebuilding: SRCNN,
VDSR, ESPCN, SRGAN, LapSRN, EDSR, compared to the super resolution ratio reconstruction method based on sparse coding, these are based on deep
The super-resolution network model of degree study achieves higher accuracy.
Achieved many breakthroughs using the super-resolution rebuilding technology of deep learning, however these networks usually just for
The small multiplying power super-resolution rebuilding task of 2x, 3x, 4x, for big multiplying power (such as 8x) super-resolution rebuilding task treatment effect not
It is good.The innovation of the invention consists in that introducing gradual sampling thought using iterative backprojection method, proposing a kind of pyramid
Iterative backprojection depth convolutional neural networks model, overcome conventional method handle big multiplying power factored sampling when deficiency,
Improve the quality of reconstruction image.
Summary of the invention
The present invention proposes a kind of based on gradual iterative backprojection to meet the needs of big multiplying power super-resolution rebuilding
Image super-resolution rebuilding method.This method carries out input picture by cascading a series of gradual back projection units
Circulation up/down sampling with feedback, makes full use of the correlation between low resolution-high-definition picture block；Gradual classification
Sampling, disposable mapping when avoiding big multiplying power factored sampling in conventional method improve training effectiveness and generate picture quality.
To achieve the above object, it is as follows to provide technical solution by the present invention:
A kind of image super-resolution rebuilding method based on gradual iterative backprojection, includes the following steps:
Step S1 carries out feature extraction to input picture first, obtains the characteristic pattern of multiple input pictures；
Step S2, using the detailed information of the further predicted characteristics figure of gradual iterative backprojection method, after being amplified
Image, the gradual iterative backprojection method is that successively cascade up/down samples back projection unit by constructing several grades
It realizes, in every level-one back projection unit, up-samples the quantity of projecting cell always 1 more than downsampling unit；
Step S3 carries out channel compressions to the image of amplification, finally exports high-resolution triple channel image.
Further, gradual iterative backprojection method is that the size of the multiplying power factor r amplified as needed adopts up/down
Sample back projection unit is divided into n grades, in which:
N=log2r。
Further, every level-one up/down sampling back projection unit executes 2 to characteristic pattern in gradual iterative backprojection method
Resolution ratio amplification again, concrete processing procedure is as follows,
Step S21, input feature vector figure recombinate layer L by one layer of sub-pix1Carry out 2 times of up-samplings, wherein sub-pix recombination
Layer is made of one layer of convolutional layer and a shuffler, and convolutional layer will be originally inputted feature M1The port number of (H x W x C) is widened
Extremely original 4 times (H x W x 4C), then each channel is shuffled by shuffler and merges the feature of a Src Chan number size out
Scheme M2(2H x2W X C)；
Step S22, amplified characteristic pattern M2The contrary operation L recombinated using a sub-pix2(referred to as against sub-pix
Recombinate layer) Lai Jinhang down-sampling, M at this time2(2H x 2W x C), which first passes through a shuffler and retrieve, is originally inputted size
Characteristic pattern (H x W x 4C), then the port number size M for being originally inputted port number boil down to via one layer of convolutional layer3(H x W
x C)；
Step S23, the characteristic pattern M that step S22 is obtained3Be originally inputted M1Subtract each other to obtain feedback error e1, by this error
e1Layer L is recombinated by the same level the last layer sub-pix3It is once up-sampled, obtains amplified feedback error e2；
Step S24, by amplified feedback error e2The characteristic pattern M obtained with step S212It is added to obtain resolution ratio
The same level that twice of amplification exports O1。
The present invention also provides a kind of image super-resolution rebuilding systems based on gradual iterative backprojection, including such as lower die
Block:
Characteristic extracting module obtains the characteristic pattern of multiple input pictures for carrying out feature extraction to input picture；
Gradual iterative backprojection module obtains amplified image for the detailed information of further predicted characteristics figure,
By several grades, successively cascade up/down sampling back projection unit is formed, and in every level-one back projection unit, up-samples projecting cell
Quantity is always 1 more than downsampling unit；
Image reconstruction module finally exports high-resolution triple channel figure for carrying out channel compressions to the image of amplification
Picture.
Further, the characteristic extracting module includes that two layers of convolutional layer and one layer of ReLU activate unit；Described image weight
Structure module includes one layer of convolutional layer.
Further, the size of the multiplying power factor r amplified as needed in gradual iterative backprojection module adopts up/down
Sample back projection unit is divided into n grades, in which:
N=log2r。
Further, the resolution ratio that every level-one up/down sampling back projection unit executes 2 times to characteristic pattern is amplified, specific to locate
Reason process is as follows,
Step S21, input feature vector figure recombinate layer L by one layer of sub-pix1Carry out 2 times of up-samplings, wherein sub-pix recombination
Layer is made of one layer of convolutional layer and a shuffler, and convolutional layer will be originally inputted feature M1The port number of (H x W x C) is widened
Extremely original 4 times (H x W x 4C), then each channel is shuffled by shuffler and merges the feature of a Src Chan number size out
Scheme M2(2H x2W X C)；
Step S22, amplified characteristic pattern M2The contrary operation L recombinated using a sub-pix2(referred to as against sub-pix
Recombinate layer) Lai Jinhang down-sampling, M at this time2(2H x 2W x C), which first passes through a shuffler and retrieve, is originally inputted size
Characteristic pattern (H x W x 4C), then the port number size M for being originally inputted port number boil down to via one layer of convolutional layer3(H x W
x C)；
Step S23, the characteristic pattern M that step S22 is obtained3Be originally inputted M1Subtract each other to obtain feedback error e1, by this error
e1Layer L is recombinated by the same level the last layer sub-pix3It is once up-sampled, obtains amplified feedback error e2；
Step S24, by amplified feedback error e2The characteristic pattern M obtained with step S212It is added to obtain resolution ratio
The same level that twice of amplification exports O1.Further, the quantity for up-sampling projecting cell is more than or equal to 5.
Compared with prior art, the present invention has the following advantages and beneficial effect:
(1) according to scientific investigations showed that, eyes imaging system have iterative feedback mechanism come guide imaging, traditional oversubscription
Resolution rebuilds network model and lacks feedback mechanism, does not efficiently use related between low resolution and high-definition picture block
Property.Items are research shows that the sample mode of iterative backprojection can greatly improve generation pattern accuracy.
(2) the existing image super-resolution rebuilding method based on iterative back projection is facing big multiplying power decimation factor
When, sampling is still realized using disposable mapping, increases trained difficulty.Gradual up/down proposed by the present invention is anti-
Projecting cell avoids disposable big multiplying power mapping by multistage operations, improves final imaging effect.
(3) the existing image super-resolution rebuilding method based on iterative back projection is all made of convolution/deconvolution side
Formula realizes mapping.This method replaces original convolution/deconvolution operation using sub-pix recombination and its inverse operation, is keeping convolution
Under conditions of core size is certain, the receptive field of each convolution can be expanded, to make to train more efficient.
Detailed description of the invention
Fig. 1 is the general frame of technical solution of the present invention；
Fig. 2 is model structure of the invention；
Fig. 3 is the structure chart of the up-sampling back projection unit in gradual iterative backprojection module；
Fig. 4 is the inverse process schematic diagram of sub-pix recombination.
Specific embodiment
Below in conjunction with the attached drawing in the embodiment of the present application, technical solutions in the embodiments of the present application carries out clear, complete
Site preparation description, it should be apparent that, drawings in the following description are only some embodiments of the invention, skill common for this field
For art personnel, without creative efforts, it is also possible to obtain other drawings based on these drawings:
A kind of image super-resolution rebuilding method packet based on gradual iterative backprojection provided by the embodiment of the present invention
Include following steps:
Step S1 carries out feature extraction to input picture first, obtains the characteristic pattern of multiple input pictures；
Step S2, using the detailed information of the further predicted characteristics figure of gradual iterative backprojection method, after being amplified
Image, the gradual iterative backprojection method is that successively cascade up/down samples back projection unit by constructing several grades
It realizes, in every level-one back projection unit, up-samples the quantity of projecting cell always 1 more than downsampling unit；
Step S3 carries out channel compressions to the image of amplification, finally exports high-resolution triple channel image.
When it is implemented, computer software technology, which can be used, in the above process realizes automatic running process.It is shown in Figure 1,
A kind of image super-resolution rebuilding system based on gradual iterative backprojection contains altogether three modules: characteristic extracting module, gradually
Into formula iterative backprojection module, image reconstruction module.Wherein, characteristic extracting module is used to extract the characteristic information of input picture,
Generate multichannel image characteristic pattern；Gradual iterative backprojection module is used to carry out details prediction to input picture, using golden word
Tower back projection unit is iterated formula sampling to characteristics of image figure；Image reconstruction module is used for the multi-channel feature pressure of image
Contracting, recovers the amplified triple channel RGB image of resolution ratio.
The overall structure of the embodiment of the present invention is as shown in Figure 2.Characteristic extracting module is by two layers of 3x3 convolutional layer and one layer of ReLU
Unit composition is activated, inputs triple channel RGB image, first layer convolution extracts 256 characteristic patterns, then compresses via second layer convolution
To 64 characteristic patterns；Gradual iterative backprojection module is made of a series of gradual back projection units.Wherein up/down sampling is single
Member is arranged successively, and up-sampling unit quantity is always 1 more than down-sampling.In unit by the way of intensively connecting, guarantee on each/
The output cascade of the input of lower unit all lower/upper sampling units all for before；Image reconstruction module is by one layer of 3x3 convolutional layer group
At, it is therefore an objective to pressure channel number generates triple channel RGB image.
Input picture (being defaulted as triple channel RGB image) it is defeated to be obtained multiple by characteristic extracting module first by step S1
Enter the characteristic pattern of image.
Step S2, it is further pre- that the characteristics of image figure that characteristic extracting module is obtained is sent into gradual iterative backprojection module
Survey detailed information.In the module, it is single that input feature vector figure passes sequentially through the sampling of the gradual up/down with feedback mechanism back projection
Member, the number of up-sampling is always than down-sampling often 1 to ensure the amplification of last image resolution ratio.Element number determines network
Depth, and then determine image quality, therefore up-sampling unit quantity is traditionally arranged to be more than or equal to 5, when it is implemented, this field
Technical staff can voluntarily adjust according to actual demand.
The amplified feature of gradual iterative backprojection module output is sent into image reconstruction unit and led to by step S3
Road compression, finally exports high-resolution RGB triple channel image.
The step S2, gradual iterative backprojection module, comprising the following steps:
Gradual iterative backprojection module is successively cascaded by a series of up/down sampling back projection unit and is formed, total unit number
Amount can be optionally customized.The quantity of up-sampling projecting cell is always 1 more than downsampling unit.The image oversubscription amplified with 8 times
For resolution reconstruction tasks, a up-sampling back projection unit is divided into 3 grades, every level-one executes 2 times of resolution ratio amplification.It is each
In grade:
Step S21, input feature vector recombinate layer L by one layer of sub-pix1Carry out 2 times of up-samplings.Wherein, sub-pix recombinates layer
It is made of one layer of convolutional layer and a shuffler.Convolutional layer will be originally inputted feature M1The port number of (H x W x C) widen to
4 times original (H x W x 4C), then each channel is shuffled by shuffler and merges the characteristic pattern of a Src Chan number size out
M2(2H x2W X C)。
Step S22, amplified characteristic pattern M2The contrary operation L recombinated using a sub-pix2(referred to as against sub-pix
Recombinate layer) Lai Jinhang down-sampling.M at this time2(2H x 2W x C), which first passes through a shuffler and retrieve, is originally inputted size
Characteristic pattern (H x W x 4C), then the port number size M for being originally inputted port number boil down to via one layer of convolutional layer3(H x W
x C)。
Step S23, the characteristic pattern M that step S22 is obtained3Be originally inputted M1Subtract each other to obtain feedback error e1, by this error
e1Layer L is recombinated by the same level the last layer sub-pix3It is once up-sampled, obtains amplified feedback error e2。
Step S24, by amplified feedback error e2The characteristic pattern M obtained with step S212It is added to obtain resolution ratio
The same level that twice of amplification exports O1。
Hereafter, every level-one output repeats the operation of S21-S24 as the input of next stage, until final enlargement ratio meets
Setting demand.Gradual down-sampling back projection unit with up-sampling back projection unit structure it is similar, by step S21-S24 it is upper/
Down-sampling operation replaces all with lower/upper sampling and can be realized.
Gradual back projection unit structure chart is as shown in Figure 3.Size by taking up-sampling unit as an example, according to multiplying power factor r
Unit is divided into n-layer, in which:
N=log2r
Every level-one only carries out the map operation that multiplying power is 2, avoids the disposable big multiplying power factor map in conventional method: to
Fixed n-th grade of input Ln, h is obtained by once up-sampling0, then to h0It carries out down-sampling and obtains l1；Calculate l1L is inputted with the same leveln
Between feedback error el, to feedback error elUp-sampling operation again obtains the error e of big resolution ratioh, by ehWith h0By adding
Musical instruments used in a Buddhist or Taoist mass obtains the same level output Hn.The output H of every level-onenBy the input L as next stagen+1, the above-mentioned iteration map throwing of circulation execution
The operation of shadow.
Sampling each time is completed using sub-pix recombination and its inverse operation in unit, and the inverse operation process of sub-pix recombination is such as
Shown in Fig. 4.Input feature vector is decomposed into port number more small figure by shuffler, then via one layer of 3x3 convolutional layer pressure channel
Number.
The embodiment of the present invention is the image super-resolution rebuilding method realized using depth convolutional neural networks.Input picture
It is respectively set to 16x16 and 128x128 (the multiplying power factor is 8) with the patch size of target image, batch size is set as
16, optimizer uses Adam optimizer, and up-sampling unit quantity is set as 5,7,10 3 kind.
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention
The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method
In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (7)
1. a kind of image super-resolution rebuilding method based on gradual iterative backprojection, which comprises the steps of:
Step S1 carries out feature extraction to input picture first, obtains the characteristic pattern of multiple input pictures；
Step S2 obtains amplified figure using the detailed information of the further predicted characteristics figure of gradual iterative backprojection method
Picture, the gradual iterative backprojection method are that successively cascade up/down samples back projection unit come real by constructing several grades
It is existing, in every level-one back projection unit, the quantity of projecting cell is up-sampled always 1 more than downsampling unit；
Step S3 carries out channel compressions to the image of amplification, finally exports high-resolution triple channel image.
2. a kind of image super-resolution rebuilding method based on gradual iterative backprojection as described in claim 1, feature
Be: gradual iterative backprojection method is that up/down is sampled back projection unit by the size of the multiplying power factor r amplified as needed
It is divided into n grades, in which:
N=log2r。
3. a kind of image super-resolution rebuilding method based on gradual iterative backprojection as described in claim 1, feature
Be: the resolution ratio that every level-one up/down sampling back projection unit executes 2 times to characteristic pattern in gradual iterative backprojection method is put
Greatly, concrete processing procedure is as follows,
Step S21, input feature vector figure recombinate layer L by one layer of sub-pix1Carry out 2 times up-sampling, wherein sub-pix recombinate layer by
One layer of convolutional layer and a shuffler composition, convolutional layer will be originally inputted feature M1The port number of (H x W x C) is widened to original
4 times (the H x W x 4C) to begin, then each channel is shuffled by shuffler and merges the characteristic pattern M of a Src Chan number size out2
(2H x 2W X C)；
Step S22, amplified characteristic pattern M2The contrary operation L recombinated using a sub-pix2It (is referred to as recombinated against sub-pix
Layer) Lai Jinhang down-sampling, M at this time2(2H x 2W x C) first passes through a shuffler and retrieves the feature for being originally inputted size
Scheme (H x W x 4C), then the port number size M for being originally inputted port number boil down to via one layer of convolutional layer3(H x W x
C)；
Step S23, the characteristic pattern M that step S22 is obtained3Be originally inputted M1Subtract each other to obtain feedback error e1, by this error e1It is logical
Cross the same level the last layer sub-pix recombination layer L3It is once up-sampled, obtains amplified feedback error e2；
Step S24, by amplified feedback error e2The characteristic pattern M obtained with step S212It is added to obtain resolution ratio amplification
Twice the same level exports O1。
4. a kind of image super-resolution rebuilding system based on gradual iterative backprojection, which is characterized in that including following module:
Characteristic extracting module obtains the characteristic pattern of multiple input pictures for carrying out feature extraction to input picture；
Gradual iterative backprojection module obtains amplified image for the detailed information of further predicted characteristics figure, if by
Successively cascade up/down sampling back projection unit forms dry grade, in every level-one back projection unit, up-samples the quantity of projecting cell
Always 1 more than downsampling unit；
Image reconstruction module finally exports high-resolution triple channel image for carrying out channel compressions to the image of amplification.
5. a kind of image super-resolution rebuilding system based on gradual iterative backprojection as claimed in claim 4, feature
Be: the characteristic extracting module includes two layers of convolutional layer and one layer of ReLU activation unit；Described image reconstructed module includes one
Layer convolutional layer.
6. a kind of image super-resolution rebuilding system based on gradual iterative backprojection as described in claim 4 or 5, special
Sign is: the size of the multiplying power factor r amplified as needed in gradual iterative backprojection module is single by up/down sampling back projection
Member is divided into n grades, in which:
N=log2r。
7. a kind of image super-resolution rebuilding system based on gradual iterative backprojection as described in claim 4 or 5, special
Sign is: the resolution ratio that every level-one up/down sampling back projection unit executes 2 times to characteristic pattern is amplified, and concrete processing procedure is as follows,
Step S21, input feature vector figure recombinate layer L by one layer of sub-pix1Carry out 2 times up-sampling, wherein sub-pix recombinate layer by
One layer of convolutional layer and a shuffler composition, convolutional layer will be originally inputted feature M1The port number of (H x W x C) is widened to original
4 times (the H x W x 4C) to begin, then each channel is shuffled by shuffler and merges the characteristic pattern M of a Src Chan number size out2
(2H x 2W X C)；
Step S22, amplified characteristic pattern M2The contrary operation L recombinated using a sub-pix2It (is referred to as recombinated against sub-pix
Layer) Lai Jinhang down-sampling, M at this time2(2H x 2W x C) first passes through a shuffler and retrieves the feature for being originally inputted size
Scheme (H x W x 4C), then the port number size M for being originally inputted port number boil down to via one layer of convolutional layer3(H x W x
C)；
Step S23, the characteristic pattern M that step S22 is obtained3Be originally inputted M1Subtract each other to obtain feedback error e1, by this error e1It is logical
Cross the same level the last layer sub-pix recombination layer L3It is once up-sampled, obtains amplified feedback error e2；
Step S24, by amplified feedback error e2The characteristic pattern M obtained with step S212It is added to obtain resolution ratio amplification
Twice the same level exports O1。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811352508.2A CN109544451A (en) | 2018-11-14 | 2018-11-14 | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811352508.2A CN109544451A (en) | 2018-11-14 | 2018-11-14 | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109544451A true CN109544451A (en) | 2019-03-29 |
Family
ID=65847262
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811352508.2A Pending CN109544451A (en) | 2018-11-14 | 2018-11-14 | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109544451A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110717856A (en) * | 2019-09-03 | 2020-01-21 | 天津大学 | Super-resolution reconstruction algorithm for medical imaging |
| CN110853040A (en) * | 2019-11-12 | 2020-02-28 | 北京深境智能科技有限公司 | Image collaborative segmentation method based on super-resolution reconstruction |
| CN111353940A (en) * | 2020-03-31 | 2020-06-30 | 成都信息工程大学 | Image super-resolution reconstruction method based on deep learning iterative up-down sampling |
| CN111579506A (en) * | 2020-04-20 | 2020-08-25 | 湖南大学 | Multi-camera hyperspectral imaging method, system and medium based on deep learning |
| CN111815546A (en) * | 2020-06-23 | 2020-10-23 | 浙江大华技术股份有限公司 | Image reconstruction method and related equipment and device |
| CN112215755A (en) * | 2020-10-28 | 2021-01-12 | 南京信息工程大学 | Image super-resolution reconstruction method based on back projection attention network |
| CN112785506A (en) * | 2021-02-25 | 2021-05-11 | 北京中科深智科技有限公司 | Image super-resolution reconstruction method and device for real-time video stream |
| CN112950692A (en) * | 2019-11-26 | 2021-06-11 | 福建天晴数码有限公司 | Image depth of field processing method and system based on mobile game platform |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107358575A (en) * | 2017-06-08 | 2017-11-17 | 清华大学 | A kind of single image super resolution ratio reconstruction method based on depth residual error network |
| CN108596330A (en) * | 2018-05-16 | 2018-09-28 | 中国人民解放军陆军工程大学 | A kind of full convolutional neural networks of Concurrent Feature and its construction method |
| CN108734659A (en) * | 2018-05-17 | 2018-11-02 | 华中科技大学 | A kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label |
-
2018
- 2018-11-14 CN CN201811352508.2A patent/CN109544451A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107358575A (en) * | 2017-06-08 | 2017-11-17 | 清华大学 | A kind of single image super resolution ratio reconstruction method based on depth residual error network |
| CN108596330A (en) * | 2018-05-16 | 2018-09-28 | 中国人民解放军陆军工程大学 | A kind of full convolutional neural networks of Concurrent Feature and its construction method |
| CN108734659A (en) * | 2018-05-17 | 2018-11-02 | 华中科技大学 | A kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label |
Non-Patent Citations (5)
| Title |
|---|
| KAI ZHANG等: "Learning a Single Convolutional Super-Resolution Network for Multiple Degradations", 《ARXIV》 * |
| MUHAMMAD HARIS等: "Deep Back-Projection Networks For Super-Resolution", 《ARXIV》 * |
| WEI-SHENG LAI等: "Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 王一宁等: "基于残差神经网络的图像超分辨率改进算法", 《计算机应用》 * |
| 王晓晖等: "基于深度学习的深度图超分辨率采样", 《计算机工程》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110717856A (en) * | 2019-09-03 | 2020-01-21 | 天津大学 | Super-resolution reconstruction algorithm for medical imaging |
| CN110853040A (en) * | 2019-11-12 | 2020-02-28 | 北京深境智能科技有限公司 | Image collaborative segmentation method based on super-resolution reconstruction |
| CN110853040B (en) * | 2019-11-12 | 2023-04-28 | 北京深境智能科技有限公司 | Image collaborative segmentation method based on super-resolution reconstruction |
| CN112950692A (en) * | 2019-11-26 | 2021-06-11 | 福建天晴数码有限公司 | Image depth of field processing method and system based on mobile game platform |
| CN112950692B (en) * | 2019-11-26 | 2023-07-14 | 福建天晴数码有限公司 | Image depth of field processing method and system based on mobile game platform |
| CN111353940A (en) * | 2020-03-31 | 2020-06-30 | 成都信息工程大学 | Image super-resolution reconstruction method based on deep learning iterative up-down sampling |
| CN111579506A (en) * | 2020-04-20 | 2020-08-25 | 湖南大学 | Multi-camera hyperspectral imaging method, system and medium based on deep learning |
| CN111579506B (en) * | 2020-04-20 | 2021-04-09 | 湖南大学 | Multi-camera hyperspectral imaging method, system and medium based on deep learning |
| CN111815546A (en) * | 2020-06-23 | 2020-10-23 | 浙江大华技术股份有限公司 | Image reconstruction method and related equipment and device |
| CN112215755A (en) * | 2020-10-28 | 2021-01-12 | 南京信息工程大学 | Image super-resolution reconstruction method based on back projection attention network |
| CN112785506A (en) * | 2021-02-25 | 2021-05-11 | 北京中科深智科技有限公司 | Image super-resolution reconstruction method and device for real-time video stream |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109544451A (en) | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection | |
| AU2020100200A4 (en) | Content-guide Residual Network for Image Super-Resolution | |
| CN111861961A (en) | Multi-scale residual error fusion model for single image super-resolution and restoration method thereof | |
| CN106204449A (en) | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network | |
| CN109671022A (en) | A kind of picture texture enhancing super-resolution method based on depth characteristic translation network | |
| CN109636721A (en) | Video super-resolution method based on confrontation study and attention mechanism | |
| CN104301728B (en) | Compression video acquisition and reconfiguration system based on structural sparse dictionary learning | |
| CN114821100B (en) | Image compressed sensing reconstruction method based on structural group sparse network | |
| CN115546032B (en) | Single-frame image super-resolution method based on feature fusion and attention mechanism | |
| CN110349087A (en) | RGB-D image superior quality grid generation method based on adaptability convolution | |
| CN112288630A (en) | Super-resolution image reconstruction method and system based on improved wide-depth neural network | |
| CN114333074A (en) | Human body posture estimation method based on dynamic lightweight high-resolution network | |
| CN115100039B (en) | Lightweight image super-resolution reconstruction method based on deep learning | |
| CN106097253A (en) | A kind of based on block rotation and the single image super resolution ratio reconstruction method of definition | |
| CN106934458A (en) | Multilayer automatic coding and system based on deep learning | |
| CN111461978A (en) | Attention mechanism-based resolution-by-resolution enhanced image super-resolution restoration method | |
| CN111626296B (en) | Medical image segmentation system and method based on deep neural network and terminal | |
| CN109993702A (en) | Based on the language of the Manchus image super-resolution rebuilding method for generating confrontation network | |
| CN116167920A (en) | Image compression and reconstruction method based on super-resolution and priori knowledge | |
| CN113837290A (en) | Unsupervised unpaired image translation method based on attention generator network | |
| CN115829876A (en) | Real degraded image blind restoration method based on cross attention mechanism | |
| CN109118428A (en) | A kind of image super-resolution rebuilding method based on feature enhancing | |
| CN116188274A (en) | Image super-resolution reconstruction method | |
| CN110288529A (en) | A kind of single image super resolution ratio reconstruction method being locally synthesized network based on recurrence | |
| CN117078941B (en) | Cardiac MRI segmentation method based on context cascade attention |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190329 |
|
| RJ01 | Rejection of invention patent application after publication |