CN108921789A - Super-resolution image reconstruction method based on recurrence residual error network - Google Patents
Super-resolution image reconstruction method based on recurrence residual error network Download PDFInfo
- Publication number
- CN108921789A CN108921789A CN201810638253.XA CN201810638253A CN108921789A CN 108921789 A CN108921789 A CN 108921789A CN 201810638253 A CN201810638253 A CN 201810638253A CN 108921789 A CN108921789 A CN 108921789A
- Authority
- CN
- China
- Prior art keywords
- residual
- image
- residual error
- recurrence
- unit
- 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 63
- 238000013528 artificial neural network Methods 0.000 claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 7
- 238000013507 mapping Methods 0.000 claims description 18
- 238000000605 extraction Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 5
- 230000003321 amplification Effects 0.000 claims description 4
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 4
- 230000001052 transient effect Effects 0.000 claims description 4
- 230000005540 biological transmission Effects 0.000 abstract description 3
- 238000013527 convolutional neural network Methods 0.000 description 8
- 230000006870 function Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 6
- 238000011084 recovery Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 230000006399 behavior Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 101100365548 Caenorhabditis elegans set-14 gene Proteins 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000013102 re-test 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
Abstract
The embodiment of the present invention provides a kind of super-resolution image reconstruction method based on recurrence residual error network, including：It will be in the recurrence residual error neural network after low-resolution image input training, obtain super-resolution rebuilding image, wherein, recurrence residual error neural network includes several residual units, for any residual unit, the input information of any residual unit is the output information of a upper residual unit and the high-frequency characteristic image of low resolution input picture.The present invention by local residual error learn rather than VDSR used in the study of overall situation residual error train neural network, it is more conducive to information transmission and gradient flowing, using the low-resolution image of non-interpolation as input, finally network end-point use deconvolute layer be directly upsampled to super-resolution output image, and by introducing recursive structure in residual unit, so that parameter greatly reduces, the computation complexity of recurrence residual error neural network is reduced.
Description
Technical field
The present embodiments relate to Image Reconstruction Technology field more particularly to a kind of super-resolution based on recurrence residual error network
Rate image rebuilding method.
Background technique
Single image super-resolution (Single image super-resolution, abbreviation SISR) is a kind of meter of classics
Calculation machine visual problem, it is intended to restore high-resolution in low resolution (Low-resolution, abbreviation LR) image given from one
Rate (High-resolution, abbreviation HR) image.Since SISR has restored high-frequency information, it is therefore widely used in needs more
The field of more image details, such as imaging of medical, satellite imagery, security monitoring etc..
Existing super-resolution (Super-resolution, abbreviation SR) image rebuilding method is broadly divided into three categories：It is based on
The SR technology of interpolation, the SR technology based on reconstruction and the SR technology based on study.Current SR algorithm is based on study mostly
Method rebuilds SR image by the mapping between study LR image and HR image.
Several super-resolution image reconstruction algorithms commonly used in the prior art have：SRCNN, FSRCNN and VDSR.
SRCNN is consisted of three parts altogether：First part is characterized extraction, the Nonlinear Mapping (volume that second part is characterized
Product neural network), Part III is full resolution pricture reconstruction.SRCNN is using the LR image after interpolation as input, directly output HR
Image, so that Nonlinear Mapping between image can be learnt with mode end to end by demonstrating super-resolution convolutional neural networks.
The it is proposed of FSRCNN is intended to accelerate SRCNN network, and this method has redesigned SRCNN network, and proposition one is compact
Hourglass shape convolutional neural networks (Convolution Neural Network, abbreviation CNN) structure.The model is in network end-point
Introducing is deconvoluted layer, and has reformulated mapping layer, and input feature vector dimension is first reduced before mapping, expands spy after mapping again
Levy dimension.The model has raised speed more than 40 times or even image has better Quality of recovery.
Residual error network (ResNet) is common image classification and object detection method, and main thought is learned according to input
Practise residual error function rather than original function, this makes the training of deep layer network simpler, and can be obtained more by deeper network
Good performance.Assuming that required bottom is mapped as H (x), the non-linear layer of stacking is allowed to be fitted another mapping：F(x):=H
(x)-x.Then original mapping is converted into：F(x)+x.Residual error mapping is easier to optimize than original mappings.In extreme circumstances, such as
Some identical mapping of fruit is optimal, then it is more simpler than being fitted identical mapping with the stacking of non-linear layer that residual error is become 0
It is single, it can be realized by the short connection of feedforward neural network, and identical short connection does not increase additional parameter and meter
Calculate complexity.
“Accurate Image Super-Resolution Using Very Deep Convolutional
Networks (Kim J, Lee J K, Lee K M.) " document proposes a very deep convolutional network (20 layers of convolutional layer), letter
Claim VDSR, so that precision is obviously improved.This method will be by bicubic interpolation by original LR picture up-sampling before CNN study
To required size, cascade convolutional layer and non-linear layer pair are then repeated in deep layer network structure, every layer of convolutional layer uses 64 3
× 3 compact filter study residual error maps to obtain residual image, finally by the LR input picture and residual image phase after interpolation
Add and exports image as super-resolution.
As pioneer's CNN model of SR, SRCNN can learn the Nonlinear Mapping between LR/HR in a manner of end to end,
Performance is significantly better than traditional non-deep learning method.But SRCNN network only haves three layers, and shallow-layer network cannot be acquired more thin
Feature is saved, image Quality of recovery is bad.
FSRCNN introduces the layer that deconvolutes in network end-point, is directly mapped to HR figure from original LR image (no interpolation) study
Picture substantially reduces computation complexity, but it equally uses shallow-layer network directly to learn original mappings function, cannot reconstruct height
Quality SR image.
The it is proposed of ResNet has broken and has deepened the saying that the network number of plies is unable to improving performance.However, original ResNet is mentioned
Be out for solving higher level computer vision problem, it is such as image classification and target detection, ResNet framework is direct
Lower-level vision problem applied to super-resolution image reconstruction may not be best model.
VDSR uses very deep network that image Quality of recovery is promoted, however, deep layer network can generate largely
Parameter occupies excessive memory.Meanwhile VDSR model needs to carry out LR image by bicubic interpolation before CNN study
Sampling, this way increase computation complexity.Excessively high computation complexity usually requires that we use GPU or high performance
CPU carries out operation to neural network.In practical application, much such as mobile device, embedded device are in calculating, volume, power consumption
Etc. it is limited, cause existing high-performance deep neural network effectively can not be calculated and be applied above.
Summary of the invention
In view of the above-mentioned problems, the embodiment of the present invention provides a kind of super-resolution image reconstruction side based on recurrence residual error network
Method.
The embodiment of the present invention provides a kind of super-resolution image reconstruction method based on recurrence residual error network, including：
By in the recurrence residual error neural network after low-resolution image input training, super-resolution rebuilding image is obtained,
In, the recurrence residual error neural network includes several residual units, for any residual unit, any residual unit
Input the output information for the upper residual unit that information is any residual unit and the height of the low resolution input picture
Frequency characteristic image, any residual unit include several convolutional layers and several activation primitive layers.
The embodiment of the present invention provides a kind of non-transient computer readable storage medium, the non-transient computer readable storage
Medium storing computer instruction, the computer instruction make the computer execute the above-mentioned super-resolution based on recurrence residual error network
Rate image rebuilding method.
A kind of super-resolution image reconstruction method based on recurrence residual error network provided in an embodiment of the present invention, passes through part
Residual error study rather than VDSR used in overall situation residual error study to train neural network, be more conducive to information transmission and gradient flowing, by
The low-resolution image of non-interpolation is as input, and finally deconvoluting in network end-point use, to be directly upsampled to super-resolution defeated for layer
Image out, and by introducing recursive structure in residual unit, so that parameter greatly reduces, reduce recurrence residual error neural network
Computation complexity.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is this hair
Bright some embodiments for those of ordinary skill in the art without creative efforts, can be with root
Other attached drawings are obtained according to these attached drawings.
Fig. 1 is a kind of flow chart of super-resolution image reconstruction method of the embodiment of the present invention；
Fig. 2 is the structural schematic diagram of recurrence residual error network in a kind of super-resolution image reconstruction method of the embodiment of the present invention；
Fig. 3 is a kind of structural schematic diagram of residual unit in super-resolution image reconstruction method in the embodiment of the present invention；
Fig. 4 is recursive structure schematic diagram in a kind of super-resolution image reconstruction method of the embodiment of the present invention；
Fig. 5 be a kind of super-resolution image reconstruction method provided in an embodiment of the present invention with other methods respectively to one low point
The reconstruction effect of resolution test image；
Fig. 6 be a kind of super-resolution image reconstruction method provided in an embodiment of the present invention with other methods respectively to another low
The reconstruction effect of resolution test image.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art
Every other embodiment obtained without creative efforts, shall fall within the protection scope of the present invention.
Fig. 1 is a kind of flow chart of super-resolution image reconstruction method of the embodiment of the present invention, as shown in Figure 1, this method packet
It includes：By in the recurrence residual error neural network after low-resolution image input training, super-resolution rebuilding image is obtained, wherein institute
Stating recurrence residual error neural network includes several residual units, for any residual unit, the input of any residual unit
Information is the output information of a upper residual unit for any residual unit and the high frequency spy of the low resolution input picture
Image is levied, any residual unit includes several convolutional layers and several activation primitive layers.
The embodiment of the present invention mainly proposes to realize the reconstruction to low-resolution image by recurrence residual error neural network, is
Acquisition super-resolution rebuilding image passes through first using low-resolution image as the input of entire recurrence residual error neural network
Recurrence residual error neural network after training handles the low resolution input picture, finally obtains super-resolution rebuilding figure
Picture.
The recurrence residual error neural network centainly contains several residual units, connects in sequence between each residual unit
Connect, that is to say, that the output of previous residual unit is connect with the input of the latter residual unit, the latter residual unit it is defeated
It is connect out with the input of next residual unit, on this basis, the input of each residual unit is as follows：
If the residual unit is first residual unit, the input of the residual unit is exactly low resolution input
The high-frequency characteristic image of image, if the residual unit is not first residual unit, illustrate the input of the residual unit with it is upper
The output connection of one residual unit, then another input of the residual unit is the high-frequency characteristic of the low resolution input picture
Image, that is to say, that the input of each residual unit must have an identical branch, be exactly the low resolution input picture
High-frequency characteristic image.This learning method is known as local residual error study by the embodiment of the present invention, due to all perseverances of residual unit
The input of equal branches all keeps identical, and image detail can not only be sent to rear layer by this method, additionally aids gradient flowing, and
And identical branch facilitate training during gradient backpropagation, avoid the occurrence of over-fitting.
On the basis of the above embodiments, it is preferable that it further include feature extraction unit in the recurrence residual error neural network,
The feature extraction unit is used to obtain the high-frequency characteristic image of the low resolution input picture.
Recurrence residual error neural network further includes extraction unit in addition to including several residual units, if extraction unit is located at
Before dry residual unit, extraction unit is used to extract the high-frequency characteristic image of low resolution input picture, and will extract
Input of the high-frequency characteristic image as first residual unit.
Specifically, the structure of extraction unit is a convolutional layer and an activation primitive layer.
Activation primitive refers to Relu function in the embodiment of the present invention.
On the basis of the above embodiments, it is preferable that it further include the layer that deconvolutes in the recurrence residual error neural network, it is described
The layer that deconvolutes obtains the super-resolution rebuilding image, the residual image is described for up-sampling to residual image
The output image of several residual units.
Low resolution input picture first passes around extraction unit in recurrence residual error neural network, and it is defeated to get the low resolution
Enter the high-frequency characteristic image of image, then, using high-frequency characteristic image as the input of first residual unit, the company of residual unit
It connects in a manner previously described, the output of the last one residual unit is the residual image learnt, which passes through
Deconvolute layer, and the layer that deconvolutes directly up-samples the residual image acquired, and reconstructs super-resolution output image.
Fig. 2 is the structural schematic diagram of recurrence residual error network in a kind of super-resolution image reconstruction method of the embodiment of the present invention,
As shown in Fig. 2, low resolution input picture first passes around feature extraction unit in recurrence residual error neural network, this low point is got
The high-frequency characteristic image of resolution input picture, then, using high-frequency characteristic image as the input of first residual unit, by
After one residual unit, the high-frequency characteristic image of the output of first residual unit and low resolution input picture is input to
In two residual units, such recurrence is gone down, and the output of last n-th residual unit is residual image, which passes through
Deconvolute layer, obtains super-resolution rebuilding image.
Fig. 3 is a kind of structural schematic diagram of residual unit in super-resolution image reconstruction method in the embodiment of the present invention, such as
Shown in Fig. 3, each residual unit includes 2 convolutional layers and 2 activation primitive layers, and convolutional layer and activation primitive layer alternately connect.
On the basis of the above embodiments, it is preferable that the output result of any residual unit is：
Hu=R (Hu-1)=F (Hu-1,W)+H0, (1)
Wherein, HuIndicate the output of any residual unit as a result, R indicates the residual unit of any residual unit
Function, F (Hu-1, W) and indicate residual error mapping to be learned, Hu-1Indicate the output of a upper residual unit for any residual unit
As a result, W indicates weight set, H0Indicate the high-frequency characteristic image of the low resolution input picture.
Since the identical mapping of each residual unit is H0, recursive structure is consequently formed, constitutes recurrence residual error network.Figure
4 be recursive structure schematic diagram in a kind of super-resolution image reconstruction method of the embodiment of the present invention, as shown in figure 4, B indicates recurrence block
Number, H0For identical mapping, HuFor by the output result of u-th of recurrence block.(a) network structure of only one recurrence block is indicated；
(b) network structure there are two recurrence block is indicated；(c) network structure of u recurrence block is indicated.
By recursive learning, weight set W shares residual unit between the residual unit in recurrence block, can increase net
Controlling model parameter saves memory space to greatly reduce number of parameters while network depth.
According to formula (1), the result that can obtain u-th of residual unit is：
Hu=Ru(H0)=R (R (... (R (H0)) ...)), (2)
Recurrence residual error neural network end is the layer that deconvolutes in the embodiment of the present invention, which is deconvoluted using one group
Filter up-samples the residual image of output.Different from traditional interpolation method, deconvoluting is that may learn image
The up-sampling kernel of feature, can be considered as the inverse operation of convolution.
Convolution operation is exported when the Jump step of filter in convolutional layer is k as inputConversely, going to roll up
The output of product operation is then k times of input.
When k is equal to amplification factor n, by deconvoluting, layer can directly export super-resolution image, this way reduces
Network query function complexity.
On the basis of the above embodiments, it is preferable that the computation complexity of the recurrence residual error network is：
Wherein, fi(i=1,2 ..., L) indicates the size of i-th layer of filter in the recurrence residual error network, ni(i=1,
2 ..., L) indicate the number of i-th layer of filter in the recurrence residual error network, SinputIndicate the big of low resolution input picture
It is small.
It is 3 × 3 filter, the layer that deconvolutes of the invention that VDSR model and convolutional layer of the invention, which use 64 sizes,
The filter for the use of 1 size being 5 × 5.
As shown in formula (3), network query function complexity is directly proportional to the size of input picture, the size of VDSR input picture
The n of input picture about of the present invention2(n is amplification factor herein) again.In addition, VDSR network shares 20 layers, and the present invention only has
12 layers, i.e. the computation complexity of VDSR is about 2n of the invention2Times.
In order to verify a kind of effect of super-resolution image reconstruction method provided in an embodiment of the present invention, the present invention will be with
The comparison result of Bicubic, SelfEx, SRCNN, FSRCNN and VDSR these types SISR method.Table 1 provides several data sets
Qualitative assessment summarize.Wherein, the result of Bicubic, SelfEx, SRCNN and VDSR are quoted from VDSR, the result of FSRCNN
From our retest, code is the source code that author publishes.
Table 1
Table 1 is different SR methods when amplification factor is 2,3 and 4, flat on test set Set5, Set14, B100 respectively
Equal PSNR value and SSIM value.
Fig. 5 and Fig. 6 illustrates context of methods compared with advanced method generates image.Fig. 5 provides for the embodiment of the present invention
A kind of super-resolution image reconstruction method and other methods respectively to the reconstruction effect of a low resolution sample image, such as Fig. 5
Shown, with " img_092 " in B100 test set for reconstructed object, the first row picture is different CNN methods to " img_092 " figure
The reconstructed results of picture；Second row shows the enlarged drawing of red block tab area in corresponding method reconstruction image；Third behavior pair
The PSNR value and SSIM value of induction method reconstruction image.
Reconstruction for plank on bridge, in addition to VDSR other methods seriously obscure or even striped generates distortion, and the side this paper
Plank gap is clear in the recovery image of method, and striped is parallel, and mutually the Quality of recovery of method image, context of methods have very than before
Big raising.
Fig. 6 be a kind of super-resolution image reconstruction method provided in an embodiment of the present invention and other methods respectively to anotherly
The reconstruction effect of resolution ratio sample image, as shown in fig. 6, using the butterfly_GT in Set5 test set as reconstructed object, the
A line picture is reconstructed results of the different CNN methods to butterfly_GT image；Second row shows corresponding method reconstruction image
The enlarged drawing of middle red block tab area；The PSNR value and SSIM value of third behavior corresponding method reconstruction image.The present invention is real
The method for applying example offer has ideally rebuild the decorative pattern on butterfly's wing, and the image that other methods generate has apparent ring existing
As, and edge is relatively fuzzyyer.
To sum up, a kind of super-resolution image reconstruction method provided in an embodiment of the present invention, specifically there is following three points tribute
It offers：
(1) present invention introduces the study of local residual error.In VDSR, residual image is to output and input estimation from network
, referred to as global residual error study.In addition, very deep network may as observed in visual identity and image recovery
Performance degradation problem can be encountered, reason is that image after multilayer transmission, can lose a large amount of detailed information.In order to solve
This problem, the present invention proposes the enhanced residual unit structure of local residual error study, wherein identical branch is not only by deep layer figure
As details is transmitted to rear layer, gradient flowing is additionally aided.
(2) present invention uses recursive structure, reduces number of parameters.Recursive structure is also introduced into residual unit by the present invention
In, a recurrence residual error network is constituted, and weight is integrated between these residual units and shares, and greatly reduces the number of parameter
Amount.
(3) computation complexity of the invention is low.Network query function complexity is directly proportional to the size of input picture, this paper model
Use the LR image of original non-interpolation as input, and VDSR is then using the LR image after interpolation as input, size is about
The n of RRSR2Times.Meanwhile the present invention uses the less network number of plies, is computed, the computation complexity of VDSR is about of the invention
2n2Times.
Through the above description of the embodiments, those skilled in the art can be understood that each embodiment can
It realizes by means of software and necessary general hardware platform, naturally it is also possible to pass through hardware.Based on this understanding, on
Stating technical solution, substantially the part that contributes to existing technology can be embodied in the form of software products in other words, should
Computer software product may be stored in a computer readable storage medium, such as ROM/RAM, magnetic disk, CD, including several fingers
It enables and using so that a computer equipment (can be personal computer, server or the network equipment etc.) executes each implementation
Method described in certain parts of example or embodiment.
Finally it should be noted that：The above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that：It still may be used
To modify the technical solutions described in the foregoing embodiments or equivalent replacement of some of the technical features；
And these are modified or replaceed, technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution spirit and
Range.
Claims (10)
1. a kind of super-resolution image reconstruction method, which is characterized in that including：
Using low resolution input picture as the input of the recurrence residual error neural network after training, super-resolution rebuilding figure is obtained
Picture, wherein the recurrence residual error neural network includes several residual units, for any residual unit, any residual error
The input information of unit is the output information and low resolution input figure of a upper residual unit for any residual unit
The high-frequency characteristic image of picture, any residual unit include several convolutional layers and several activation primitive layers.
2. method according to claim 1, which is characterized in that further include feature extraction list in the recurrence residual error neural network
Member, the feature extraction unit are used to obtain the high-frequency characteristic image of the low resolution input picture.
3. method according to claim 1, which is characterized in that it further include the layer that deconvolutes in the recurrence residual error neural network,
The layer that deconvolutes obtains the super-resolution rebuilding image, the residual image is for up-sampling to residual image
The output image of several residual units.
4. method according to claim 2, which is characterized in that the feature extraction unit includes that a convolutional layer and one swash
Function layer living.
5. method according to claim 1, which is characterized in that the output result of any residual unit is：
Hu=R (Hu-1)=F (Hu-1,W)+H0,
Wherein, HuIndicate the output of any residual unit as a result, R indicates the residual unit function of any residual unit,
F(Hu-1, W) and indicate residual error mapping to be learned, Hu-1Indicate the output of a upper residual unit for any residual unit as a result,
W indicates weight set, H0Indicate the high-frequency characteristic image of the low resolution input picture.
6. method according to claim 5, which is characterized in that the weight is integrated between several residual units altogether
It enjoys.
7. method according to claim 1, which is characterized in that the size of the super-resolution rebuilding image is the low resolution
N times of rate input picture, n indicate amplification factor.
8. method according to claim 1, which is characterized in that the computation complexity of the recurrence residual error network is：
Wherein, fi(i=1,2 ..., L) indicates the size of i-th layer of filter in the recurrence residual error network, ni(i=1,2 ...,
L the number of i-th layer of filter in the recurrence residual error network, S) are indicatedinputIndicate the size of low resolution input picture.
9. method according to claim 1, which is characterized in that each convolutional layer in the recurrence residual error network includes 64
3 × 3 filter, the layer that deconvolutes in the recurrence residual error network include 15 × 5 filter.
10. a kind of non-transient computer readable storage medium, which is characterized in that the non-transient computer readable storage medium is deposited
Computer instruction is stored up, the computer instruction makes the computer execute method as described in any one of claim 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810638253.XA CN108921789A (en) | 2018-06-20 | 2018-06-20 | Super-resolution image reconstruction method based on recurrence residual error network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810638253.XA CN108921789A (en) | 2018-06-20 | 2018-06-20 | Super-resolution image reconstruction method based on recurrence residual error network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108921789A true CN108921789A (en) | 2018-11-30 |
Family
ID=64421540
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810638253.XA Pending CN108921789A (en) | 2018-06-20 | 2018-06-20 | Super-resolution image reconstruction method based on recurrence residual error network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921789A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109686425A (en) * | 2019-01-17 | 2019-04-26 | 南京晓庄学院 | A method of accelerating global reconstruction human brain neuro images technology |
| CN109903226A (en) * | 2019-01-30 | 2019-06-18 | 天津城建大学 | Image super-resolution rebuilding method based on symmetrical residual error convolutional neural networks |
| CN109903219A (en) * | 2019-02-28 | 2019-06-18 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment, computer readable storage medium |
| CN110197468A (en) * | 2019-06-06 | 2019-09-03 | 天津工业大学 | A kind of single image Super-resolution Reconstruction algorithm based on multiple dimensioned residual error learning network |
| CN110288529A (en) * | 2019-06-28 | 2019-09-27 | 闽江学院 | A kind of single image super resolution ratio reconstruction method being locally synthesized network based on recurrence |
| CN110796623A (en) * | 2019-10-31 | 2020-02-14 | 上海融军科技有限公司 | Infrared image rain removing method and device based on progressive residual error network |
| CN111405283A (en) * | 2020-02-20 | 2020-07-10 | 北京大学 | End-to-end video compression method, system and storage medium based on deep learning |
| CN111583107A (en) * | 2020-04-03 | 2020-08-25 | 长沙理工大学 | Image super-resolution reconstruction method and system based on attention mechanism |
| CN111861886A (en) * | 2020-07-15 | 2020-10-30 | 南京信息工程大学 | Image super-resolution reconstruction method based on multi-scale feedback network |
| WO2021042270A1 (en) * | 2019-09-03 | 2021-03-11 | 中山大学 | Compression artifacts reduction method based on dual-stream multi-path recursive residual network |
| CN113378704A (en) * | 2021-06-09 | 2021-09-10 | 武汉理工大学 | Multi-target detection method, equipment and storage medium |
| CN113902618A (en) * | 2021-10-09 | 2022-01-07 | 普达迪泰（天津）智能装备科技有限公司 | Image super-resolution algorithm based on multi-mode spatial filtering |
| CN115131612A (en) * | 2022-07-02 | 2022-09-30 | 哈尔滨理工大学 | Retina OCT image classification method based on recursive residual error network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN106960415A (en) * | 2017-03-17 | 2017-07-18 | 深圳市唯特视科技有限公司 | A kind of method for recovering image based on pixel-recursive super-resolution model |
| US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
-
2018
- 2018-06-20 CN CN201810638253.XA patent/CN108921789A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
| 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 |
| CN106960415A (en) * | 2017-03-17 | 2017-07-18 | 深圳市唯特视科技有限公司 | A kind of method for recovering image based on pixel-recursive super-resolution model |
Non-Patent Citations (3)
| Title |
|---|
| DONG C ET AL: "Accelerating the Super-Resolution Convolutional Neural Network", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
| SHI J ET AL: "MR image super-resolution via wide residual networks with fixed skip connection", 《IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS》 * |
| TAI Y ET AL: "Image Super-Resolution via Deep Recursive Residual Network", 《IEEE COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2017)》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109686425B (en) * | 2019-01-17 | 2020-08-11 | 南京晓庄学院 | System and method for accelerating global reconstruction technology of human brain nerve image |
| CN109686425A (en) * | 2019-01-17 | 2019-04-26 | 南京晓庄学院 | A method of accelerating global reconstruction human brain neuro images technology |
| CN109903226A (en) * | 2019-01-30 | 2019-06-18 | 天津城建大学 | Image super-resolution rebuilding method based on symmetrical residual error convolutional neural networks |
| CN109903226B (en) * | 2019-01-30 | 2023-08-15 | 天津城建大学 | Image super-resolution reconstruction method based on symmetric residual convolution neural network |
| CN109903219A (en) * | 2019-02-28 | 2019-06-18 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment, computer readable storage medium |
| CN110197468A (en) * | 2019-06-06 | 2019-09-03 | 天津工业大学 | A kind of single image Super-resolution Reconstruction algorithm based on multiple dimensioned residual error learning network |
| CN110288529B (en) * | 2019-06-28 | 2022-06-07 | 闽江学院 | Single image super-resolution reconstruction method based on recursive local synthesis network |
| CN110288529A (en) * | 2019-06-28 | 2019-09-27 | 闽江学院 | A kind of single image super resolution ratio reconstruction method being locally synthesized network based on recurrence |
| WO2021042270A1 (en) * | 2019-09-03 | 2021-03-11 | 中山大学 | Compression artifacts reduction method based on dual-stream multi-path recursive residual network |
| CN110796623A (en) * | 2019-10-31 | 2020-02-14 | 上海融军科技有限公司 | Infrared image rain removing method and device based on progressive residual error network |
| CN111405283A (en) * | 2020-02-20 | 2020-07-10 | 北京大学 | End-to-end video compression method, system and storage medium based on deep learning |
| CN111583107A (en) * | 2020-04-03 | 2020-08-25 | 长沙理工大学 | Image super-resolution reconstruction method and system based on attention mechanism |
| CN111861886B (en) * | 2020-07-15 | 2023-08-08 | 南京信息工程大学 | Image super-resolution reconstruction method based on multi-scale feedback network |
| CN111861886A (en) * | 2020-07-15 | 2020-10-30 | 南京信息工程大学 | Image super-resolution reconstruction method based on multi-scale feedback network |
| CN113378704A (en) * | 2021-06-09 | 2021-09-10 | 武汉理工大学 | Multi-target detection method, equipment and storage medium |
| CN113902618A (en) * | 2021-10-09 | 2022-01-07 | 普达迪泰（天津）智能装备科技有限公司 | Image super-resolution algorithm based on multi-mode spatial filtering |
| CN113902618B (en) * | 2021-10-09 | 2024-03-29 | 普达迪泰（天津）智能装备科技有限公司 | Image super-resolution algorithm based on multi-modal spatial filtering |
| CN115131612A (en) * | 2022-07-02 | 2022-09-30 | 哈尔滨理工大学 | Retina OCT image classification method based on recursive residual error network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921789A (en) | Super-resolution image reconstruction method based on recurrence residual error network | |
| CN111369440B (en) | Model training and image super-resolution processing method, device, terminal and storage medium | |
| CN109389556A (en) | The multiple dimensioned empty convolutional neural networks ultra-resolution ratio reconstructing method of one kind and device | |
| CN108475415A (en) | Method and system for image procossing | |
| CN107369189A (en) | The medical image super resolution ratio reconstruction method of feature based loss | |
| CN108229497A (en) | Image processing method, device, storage medium, computer program and electronic equipment | |
| CN106683067A (en) | Deep learning super-resolution reconstruction method based on residual sub-images | |
| CN110232661A (en) | Low illumination colour-image reinforcing method based on Retinex and convolutional neural networks | |
| CN109801221A (en) | Generate training method, image processing method, device and the storage medium of confrontation network | |
| Chen et al. | Single image super-resolution using deep CNN with dense skip connections and inception-resnet | |
| CN109544457A (en) | Image super-resolution method, storage medium and terminal based on fine and close link neural network | |
| CN111242846A (en) | Fine-grained scale image super-resolution method based on non-local enhancement network | |
| CN111340696B (en) | Convolutional neural network image super-resolution reconstruction method fused with bionic visual mechanism | |
| CN105513033B (en) | A kind of super resolution ratio reconstruction method that non local joint sparse indicates | |
| Yang et al. | Image super-resolution based on deep neural network of multiple attention mechanism | |
| CN110322402A (en) | Medical image super resolution ratio reconstruction method based on dense mixing attention network | |
| CN113421187B (en) | Super-resolution reconstruction method, system, storage medium and equipment | |
| CN113781308A (en) | Image super-resolution reconstruction method and device, storage medium and electronic equipment | |
| CN113744136A (en) | Image super-resolution reconstruction method and system based on channel constraint multi-feature fusion | |
| CN115797176A (en) | Image super-resolution reconstruction method | |
| Li et al. | Lightweight adaptive weighted network for single image super-resolution | |
| Yang et al. | An image super-resolution network based on multi-scale convolution fusion | |
| Wang et al. | 3D dense convolutional neural network for fast and accurate single MR image super-resolution | |
| CN106981046A (en) | Single image super resolution ratio reconstruction method based on multi-gradient constrained regression | |
| CN110223224A (en) | A kind of Image Super-resolution realization algorithm based on information filtering network |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181130 |