CN109961396A - A kind of image super-resolution rebuilding method based on convolutional neural networks - Google Patents
A kind of image super-resolution rebuilding method based on convolutional neural networks Download PDFInfo
- Publication number
- CN109961396A CN109961396A CN201711417919.0A CN201711417919A CN109961396A CN 109961396 A CN109961396 A CN 109961396A CN 201711417919 A CN201711417919 A CN 201711417919A CN 109961396 A CN109961396 A CN 109961396A
- Authority
- CN
- China
- Prior art keywords
- image
- channel
- layers
- neural networks
- convolutional neural
- 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
Classifications
-
- 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
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
- G06T3/4061—Super resolution, i.e. output image resolution higher than sensor resolution by injecting details from a different spectral band
Abstract
The present invention relates to a kind of image super-resolution rebuilding methods based on convolutional neural networks to obtain shallow-layer texture feature information by data set training SRCNN convolutional neural networks model;Eight layers of end-to-end neural network model based on feature transfer are established, shallow-layer texture feature information is migrated to first four layers of the neural network model, preceding four layers of model parameter is obtained;Four layers after the neural network model of model parameter is obtained, the feature learnt is enhanced;Image data to be reconstructed is inputted, is pre-processed;Obtain the high-definition picture in the channel Y;The high-definition picture in the channel Y, Cb channel image and Cr channel image are merged, the image rebuild.Convolutional neural networks model proposed by the present invention achieves more preferably super-resolution result, it has clear improvement either on subjective vision or on objectively evaluating index, image definition and edge sharpness significantly improve, and convergence rate faster, has higher advantage in terms of fineness.
Description
Technical field
The present invention relates to computer visions and digital image processing field, specifically a kind of to be based on convolutional neural networks
Image super-resolution rebuilding method.
Background technique
Single image super-resolution (Super Resolution, SR) is rebuild, and is a classics of computer vision field
Problem, its object is to obtain a high-definition picture from a low-resolution image.Pass through signal processing and image procossing
Method, reconstruct one with best quality high-resolution (High Resolution, HR) output image the problem of.It is right
In any one given low-resolution pixel image, there are a variety of solutions, this is a typical ill-condition problem, tool
There is serious ill-posedness, solves this problem and need prior information.
The high frequency that convolutional neural networks (Convolutional neural network, CNNs) can obtain image is special
Sign strengthens detailed information, and is especially good at the correlation information for obtaining image pixel in small territory, studies it and is scheming
As the application on rebuilding is highly desirable.By convolutional neural networks apply in image super-resolution rebuilding, and achieve very well
Effect.
Algorithm SRCNN (Super-Resolution Convolutional Neural Network) is by sparse coding side
The each step of method, which integrates, regards a convolutional neural networks as, establishes one and rebuilds network end to end.However SRCNN
Three-decker is unable to satisfy the higher requirement of fineness, preceding several layers of to can only obtain the shallow of image according to convolutional neural networks characteristic
Layer texture information only can just obtain the feature of more details using the convolutional network of deeper, can just reconstruct the figure of high quality
Picture.
Summary of the invention
In view of the deficiencies of the prior art, the present invention provides a kind of image super-resolution rebuilding side based on convolutional neural networks
Method overcomes the shortcomings of current single image convolutional neural networks technology on processing super-resolution rebuilding.
Present invention technical solution used for the above purpose is:
A kind of image super-resolution rebuilding method based on convolutional neural networks, comprising the following steps:
Step 1: by data set training SRCNN convolutional neural networks model, obtaining shallow-layer texture feature information;
Step 2: establishing eight layers of end-to-end neural network model based on feature transfer, shallow-layer texture feature information is migrated
To first four layers of the neural network model, preceding four layers of model parameter is obtained;
Step 3: by eight layer end-to-end neural network model of the data set training based on feature transfer, obtaining the nerve net
Four layers of model parameter after network model enhances the feature learnt;
Step 4: inputting image data to be reconstructed, and the image data is pre-processed, obtain the channel Y, the channel Cb
With Cr channel image data;
Step 5: Y channel image data is input to the end-to-end nerve of eight based on feature transfer layer after step 3 training
Network model obtains the high-definition picture in the channel Y;
Step 6: high-resolution Cb is obtained according to bicubic interpolation algorithm to the image data in the channel Cb and the channel Cr
Channel and Cr channel image;
Step 7: the high-definition picture in the channel Y, Cb channel image and Cr channel image being merged, rebuild
Image.
The pretreatment includes:
The image of RGB channel is converted to the image in the channel YCbCr, and using the image data in the channel Y as input.
Linear amending unit ReLU is used to replace Sigmoid as activation primitive.
The data set includes Set5, Set14 and B100.
The end-to-end neural network model of eight based on feature transfer layer, including eight layers of convolutional layer and full articulamentum, and
Not comprising pond layer.
The global parameter of eight layers of end-to-end neural network model based on feature transfer is arranged are as follows: learning rate size is fixed as
0.0001, Filling power 1, monmentum=0.3, batch_size=100;
The parameter setting of eight layers of convolutional layer are as follows: first convolutional layer convolution kernel size is 9x9, and convolution nuclear volume is
64;Second convolutional layer convolution kernel size is 3x3, and convolution nuclear volume is 16;Third convolutional layer convolution kernel size is 1x1, volume
Product nuclear volume is 32;4th convolutional layer convolution kernel size is 5x5, and convolution nuclear volume is 1;5th convolutional layer convolution kernel size
For 9x9, convolution nuclear volume is 64;6th convolutional layer convolution kernel size is 3x3, and convolution nuclear volume is 16;7th convolutional layer
Convolution kernel size is 1x1, and convolution nuclear volume is 32;8th convolutional layer convolution kernel size is 5x5, and convolution nuclear volume is 1.
It is described to migrate shallow-layer texture feature information to first four layers of the neural network model are as follows: will be through SRCNN convolution mind
Weight parameter w and offset parameter b after network model training are migrated to eight layers of end-to-end neural network based on feature transfer
First four layers of model.
Weight parameter w and offset parameter b are updated, initialized first, the random value from normal distribution,
It is 0 that mean value is obeyed in normal distribution, and standard deviation is equal toSecondly it completes to update using stochastic gradient descent method.
During eight layers of end-to-end neural network model by data set training based on feature transfer, setting loss
Function, the loss function use root-mean-square error:
Wherein, n indicates the quantity of training sample, Yi' it is original high-resolution image block, Yi' it is prediction high-definition picture
Block.
The invention has the following beneficial effects and advantage:
The present invention can effectively handle the super-resolution rebuilding problem of single width low-resolution image, and network model can be accurate
Ground learns the notable feature to image, and finally provides the reconstruction image of high quality.
Detailed description of the invention
Fig. 1 is the end-to-end Artificial Neural Network Structures figure of eight based on feature transfer layer of the invention;
Specific embodiment
The present invention is described in further detail with reference to the accompanying drawings and embodiments.
The present invention includes following procedure:
The image of RGB channel is converted to the image in the channel YCbCr first, and each channel is rebuild;Then it sets
Eight layers of convolutional neural networks structure are counted, determine the size and quantity of each layer of convolution kernel, and CNN are optimized, including first
The optimization of beginningization parameter weights and biase and the optimization of activation primitive;It is shallow finally by the four-layer network network study of SRCNN
Layer characteristic information, and move in TLSRCNN proposed by the present invention, notable feature is reinforced by rear four-layer network network and completes final figure
As super-resolution rebuilding.
It is as shown in Figure 1 the end-to-end Artificial Neural Network Structures figure of eight based on feature transfer layer of the invention.
Algorithm for reconstructing proposed by the present invention establishes an end to end model TLSRCNN (Transfer Learning
Super-Resolution using Convolutional Neural Networks), it inputs as low-resolution image, output
As full resolution pricture.Due to needing image detail information, the textural characteristics extracted can be then reduced if there is down-sampling,
Therefore without introducing pond layer.Model is full convolutional coding structure, whole to be made of 8 layers of convolutional layer.
8 layers of end-to-end neural network model based on feature transfer are established, first 4 layers of network are by feature transfer and in advance
Training obtains shallow-layer information, and rear 4 layers of realization feature enhancing ultimately forms high-definition picture.Super-resolution, which refers to, passes through hardware
Or the method for software improves the resolution ratio of original image, and a high-resolution figure is obtained by a series of image of low resolution
As process is exactly super-resolution rebuilding.
Specific step is as follows:
In general, single image super-resolution rebuilding can all be related to a problem, for given low-resolution image X,
How image Y with higher resolution is reconstructed, and the X observed is the image of the fuzzy downsampled version of Y, and the problem is available
Following formula indicates:
Y=SHX
Wherein, S indicates that down-sampling operator, H indicate fuzzy filter.
Step 1. input image data.
Low resolution image data to be reconstructed is inputted in the image super-resolution rebuilding system based on feature transfer.
Step 2. image preprocessing.
Relative to color, the eyes of the mankind are more sensitive to the brightness of image detail, so when carrying out image preprocessing,
The image of RGB channel is converted to the image in the channel YCbCr first, then the channel Y is carried out using depth convolutional neural networks
Super-resolution rebuilding, and bicubic interpolation algorithm is used to the channel Cb and Cr, directly generate the high-resolution channel Cb, Cr figure
Picture.
Step 3. optimizes convolutional neural networks activation primitive.
Linear amending unit ReLU (Rectified linear unit) replaces Sigmoid to can solve as activation primitive
Gradient disappearance problem, ReLU expression formula are max (0, x+N (0,1)).ReLU function has the spy of unilateral inhibition, sparse activity
Point receives the state of activation of signal closer to brain neuron.The use of activation primitive ReLU, introduces for convolutional neural networks
Sparsity is equivalent to the pretreatment for introducing unsupervised learning.
If X is first layer input, YlIt is exported for l layers, then under ReLU activation primitive, relationship between first time input and output
Are as follows:
Yl=max (0, wl*X+bl)
Thereafter l+1 layers of input/output relation are as follows:
Yl+1=max (0, wl*X+bl)
Wherein wlWith blThe defeated weight and deviation for respectively corresponding first of convolutional layer, " * " indicate convolution operation.
The setting of step 4. model parameter.
Weight and deviation initial value conventional method are the random values in the normal distribution for obeying N (0,1), if inappropriate
Initial value will lead to neural network learning overlong time, seriously affect system convergence speed, it is also possible to fall into local optimum.
Innovatory algorithm is from normal distribution, and mean value is 0, and standard deviation is equal toN is the number of input sample.Or it is false
If neural network, improved algorithm, the new initialization weight method convergence rate of experimental verification is significantly faster than that tradition side
Method, and also improve in accuracy rate.The backpropagation of Web vector graphic standard and stochastic gradient descent method complete instruction
Practice process, makes minimization of loss.Algorithm is restrained using gradient descent method, in gradient descent method, weight w and deviation b all roots
It is updated according to following formula:
Wherein σ is learning rate.The partial derivative of each layer parameter are as follows:
According to obtained updated Δ w(l)With Δ b(l)Update weight parameter.Iterative gradient descent method reduces loss function
Value, solve entire neural network.
The setting of step 5. loss function.
In order to preferably train network, loss function is used as using root-mean-square error (Mean Square Error, MSE).
MSE is closer to human visual perception, and objective experimental result shows that the picture quality obtained using MSE has also been obtained significantly
Promotion.MSE's is expressed mathematically as:
Here n indicates that the quantity of training sample, the physical significance of MSE are to calculate original high-resolution image block YiWith
The prediction high-definition picture block Y of network outputiThe difference of ' center pixel.
Each layer parameter design of step 6. convolutional neural networks.
In training, global parameter has carried out following setting: learning rate size is fixed as 0.0001, Filling power 1,
Monmentum=0.3, batch_size=100.Wherein first convolutional layer convolution kernel size is 9x9, and convolution nuclear volume is
64;Second convolutional layer convolution kernel size is 3x3, and convolution nuclear volume is 16;Third convolutional layer convolution kernel size is 1x1, volume
Product nuclear volume is 32;4th convolutional layer convolution kernel size is 5x5, and convolution nuclear volume is 1;5th convolutional layer convolution kernel size
For 9x9, convolution nuclear volume is 64;6th convolutional layer convolution kernel size is 3x3, and convolution nuclear volume is 16;7th convolutional layer
Convolution kernel size is 1x1, and convolution nuclear volume is 32;8th convolutional layer convolution kernel size is 5x5, and convolution nuclear volume is 1.
The migration of step 7. feature and enhancing.
After four layers of neural network extract characteristics of image before SRCNN, transfer parameter is carried out, is moved to proposed by the present invention
TLSRCNN algorithm, then the feature arrived by rear four layers of Neural-Network-Based Reinforcement Learning of TLSRCNN, it is thin can to reinforce image texture
Section etc. improves and rebuilds effect.
Step 8. super-resolution rebuilding.
After having trained eight layers of end-to-end neural network model based on feature transfer, input as low-resolution image, output
As high-definition picture.
Claims (10)
1. a kind of image super-resolution rebuilding method based on convolutional neural networks, it is characterised in that: the following steps are included:
Step 1: by data set training SRCNN convolutional neural networks model, obtaining shallow-layer texture feature information;
Step 2: establishing eight layers of end-to-end neural network model based on feature transfer, shallow-layer texture feature information is migrated to this
First four layers of neural network model, obtain preceding four layers of model parameter;
Step 3: by eight layer end-to-end neural network model of the data set training based on feature transfer, obtaining the neural network mould
Four layers of model parameter after type enhances the feature learnt;
Step 4: inputting image data to be reconstructed, and the image data is pre-processed, obtain the channel Y, the channel Cb and Cr
Channel image data;
Step 5: Y channel image data is input to the end-to-end neural network of eight based on feature transfer layer after step 3 training
Model obtains the high-definition picture in the channel Y;
Step 6: the high-resolution channel Cb is obtained according to bicubic interpolation algorithm to the image data in the channel Cb and the channel Cr
With Cr channel image;
Step 7: the high-definition picture in the channel Y, Cb channel image and Cr channel image being merged, the figure rebuild
Picture.
2. the image super-resolution rebuilding method according to claim 1 based on convolutional neural networks, it is characterised in that: institute
Stating pretreatment includes:
The image of RGB channel is converted to the image in the channel YCbCr, and using the image data in the channel Y as input.
3. the image super-resolution rebuilding method according to claim 1 based on convolutional neural networks, it is characterised in that: make
Linear amending unit ReLU is used to replace Sigmoid as activation primitive.
4. the image super-resolution rebuilding method according to claim 1 based on convolutional neural networks, it is characterised in that: institute
Stating data set includes Set5, Set14 and B100.
5. the image super-resolution rebuilding method according to claim 1 based on convolutional neural networks, it is characterised in that: institute
Eight layers of end-to-end neural network model based on feature transfer, including eight layers of convolutional layer and full articulamentum are stated, and do not include pond
Layer.
6. the image super-resolution rebuilding method according to claim 5 based on convolutional neural networks, it is characterised in that: base
In the global parameter setting of eight layers of end-to-end neural network model of feature transfer are as follows: learning rate size is fixed as 0.0001, fills out
It supplements with money as 1, monmentum=0.3, batch_size=100.
7. the image super-resolution rebuilding method according to claim 5 based on convolutional neural networks, it is characterised in that: institute
State the parameter setting of eight layers of convolutional layer are as follows: first convolutional layer convolution kernel size is 9x9, and convolution nuclear volume is 64;Second volume
Lamination convolution kernel size is 3x3, and convolution nuclear volume is 16;Third convolutional layer convolution kernel size is 1x1, and convolution nuclear volume is
32;4th convolutional layer convolution kernel size is 5x5, and convolution nuclear volume is 1;5th convolutional layer convolution kernel size is 9x9, convolution
Nuclear volume is 64;6th convolutional layer convolution kernel size is 3x3, and convolution nuclear volume is 16;7th convolutional layer convolution kernel size
For 1x1, convolution nuclear volume is 32;8th convolutional layer convolution kernel size is 5x5, and convolution nuclear volume is 1.
8. the image super-resolution rebuilding method according to claim 1 based on convolutional neural networks, it is characterised in that: institute
It states and migrates shallow-layer texture feature information to first four layers of the neural network model are as follows: will be through SRCNN convolutional neural networks model
Weight parameter w and offset parameter b after training migrate to eight layers of end-to-end neural network model based on feature transfer before four
Layer.
9. the image super-resolution rebuilding method according to claim 8 based on convolutional neural networks, it is characterised in that: right
Weight parameter w and offset parameter b are updated, and are initialized first, the random value from normal distribution, normal distribution clothes
It is 0 from mean value, standard deviation is equal toSecondly it completes to update using stochastic gradient descent method.
10. the image super-resolution rebuilding method according to claim 1 based on convolutional neural networks, it is characterised in that:
During eight layers of end-to-end neural network model by data set training based on feature transfer, loss function is set, it should
Loss function uses root-mean-square error:
Wherein, n indicates the quantity of training sample, Yi' it is original high-resolution image block, Yi' it is prediction high-definition picture block.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711417919.0A CN109961396B (en) | 2017-12-25 | 2017-12-25 | Image super-resolution reconstruction method based on convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711417919.0A CN109961396B (en) | 2017-12-25 | 2017-12-25 | Image super-resolution reconstruction method based on convolutional neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109961396A true CN109961396A (en) | 2019-07-02 |
CN109961396B CN109961396B (en) | 2020-12-22 |
Family
ID=67020556
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711417919.0A Active CN109961396B (en) | 2017-12-25 | 2017-12-25 | Image super-resolution reconstruction method based on convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109961396B (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110554991A (en) * | 2019-09-03 | 2019-12-10 | 浙江传媒学院 | Method for correcting and managing text picture |
CN110827200A (en) * | 2019-11-04 | 2020-02-21 | Oppo广东移动通信有限公司 | Image super-resolution reconstruction method, image super-resolution reconstruction device and mobile terminal |
CN110880160A (en) * | 2019-11-14 | 2020-03-13 | Oppo广东移动通信有限公司 | Picture frame super-division method and device, terminal equipment and computer readable storage medium |
CN111091521A (en) * | 2019-12-05 | 2020-05-01 | 腾讯科技(深圳)有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
CN111182254A (en) * | 2020-01-03 | 2020-05-19 | 北京百度网讯科技有限公司 | Video processing method, device, equipment and storage medium |
CN111340696A (en) * | 2020-02-10 | 2020-06-26 | 南京理工大学 | Convolutional neural network image super-resolution reconstruction method fused with bionic visual mechanism |
CN111654621A (en) * | 2020-05-26 | 2020-09-11 | 浙江大学 | Dual-focus camera continuous digital zooming method based on convolutional neural network model |
CN111861961A (en) * | 2020-07-25 | 2020-10-30 | 安徽理工大学 | Multi-scale residual error fusion model for single image super-resolution and restoration method thereof |
CN112927136A (en) * | 2021-03-05 | 2021-06-08 | 江苏实达迪美数据处理有限公司 | Image reduction method and system based on convolutional neural network domain adaptation |
CN113221842A (en) * | 2021-06-04 | 2021-08-06 | 第六镜科技(北京)有限公司 | Model training method, image recognition method, device, equipment and medium |
CN113365107A (en) * | 2020-03-05 | 2021-09-07 | 阿里巴巴集团控股有限公司 | Video processing method, movie video processing method and device |
CN113409195A (en) * | 2021-07-06 | 2021-09-17 | 中国标准化研究院 | Image super-resolution reconstruction method based on improved deep convolutional neural network |
CN113747099A (en) * | 2020-05-29 | 2021-12-03 | 华为技术有限公司 | Video transmission method and device |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105069825A (en) * | 2015-08-14 | 2015-11-18 | 厦门大学 | Image super resolution reconstruction method based on deep belief network |
WO2016132147A1 (en) * | 2015-02-19 | 2016-08-25 | Magic Pony Technology Limited | Enhancement of visual data |
CN106204449A (en) * | 2016-07-06 | 2016-12-07 | 安徽工业大学 | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network |
CN106251289A (en) * | 2016-07-21 | 2016-12-21 | 北京邮电大学 | A kind of based on degree of depth study and the video super-resolution method for reconstructing of self-similarity |
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 |
CN106930770A (en) * | 2017-02-06 | 2017-07-07 | 西安科技大学 | Shield machine shield gap method of estimation based on convolutional neural networks |
CN107203999A (en) * | 2017-04-28 | 2017-09-26 | 北京航空航天大学 | A kind of skin lens image automatic division method based on full convolutional neural networks |
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 |
CN107507134A (en) * | 2017-09-21 | 2017-12-22 | 大连理工大学 | Super-resolution method based on convolutional neural networks |
-
2017
- 2017-12-25 CN CN201711417919.0A patent/CN109961396B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2016132147A1 (en) * | 2015-02-19 | 2016-08-25 | Magic Pony Technology Limited | Enhancement of visual data |
CN105069825A (en) * | 2015-08-14 | 2015-11-18 | 厦门大学 | Image super resolution reconstruction method based on deep belief network |
CN106204449A (en) * | 2016-07-06 | 2016-12-07 | 安徽工业大学 | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network |
CN106251289A (en) * | 2016-07-21 | 2016-12-21 | 北京邮电大学 | A kind of based on degree of depth study and the video super-resolution method for reconstructing of self-similarity |
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 |
CN106930770A (en) * | 2017-02-06 | 2017-07-07 | 西安科技大学 | Shield machine shield gap method of estimation based on convolutional neural networks |
CN107203999A (en) * | 2017-04-28 | 2017-09-26 | 北京航空航天大学 | A kind of skin lens image automatic division method based on full convolutional neural networks |
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 |
CN107507134A (en) * | 2017-09-21 | 2017-12-22 | 大连理工大学 | Super-resolution method based on convolutional neural networks |
Non-Patent Citations (2)
Title |
---|
孙旭 等: "基于深度学习的图像超分辨率复原研究进展", 《自动化学报》 * |
孙超 等: "基于迁移学习的红外图像超分辨率方法研究", 《激光与红外》 * |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110554991A (en) * | 2019-09-03 | 2019-12-10 | 浙江传媒学院 | Method for correcting and managing text picture |
US11151690B2 (en) | 2019-11-04 | 2021-10-19 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Image super-resolution reconstruction method, mobile terminal, and computer-readable storage medium |
CN110827200A (en) * | 2019-11-04 | 2020-02-21 | Oppo广东移动通信有限公司 | Image super-resolution reconstruction method, image super-resolution reconstruction device and mobile terminal |
WO2021088473A1 (en) * | 2019-11-04 | 2021-05-14 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Image super-resolution reconstruction method, image super-resolution reconstruction apparatus, and computer-readable storage medium |
CN110827200B (en) * | 2019-11-04 | 2023-04-07 | Oppo广东移动通信有限公司 | Image super-resolution reconstruction method, image super-resolution reconstruction device and mobile terminal |
CN110880160B (en) * | 2019-11-14 | 2023-04-18 | Oppo广东移动通信有限公司 | Picture frame super-division method and device, terminal equipment and computer readable storage medium |
CN110880160A (en) * | 2019-11-14 | 2020-03-13 | Oppo广东移动通信有限公司 | Picture frame super-division method and device, terminal equipment and computer readable storage medium |
CN111091521A (en) * | 2019-12-05 | 2020-05-01 | 腾讯科技(深圳)有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
CN111182254A (en) * | 2020-01-03 | 2020-05-19 | 北京百度网讯科技有限公司 | Video processing method, device, equipment and storage medium |
US11416967B2 (en) | 2020-01-03 | 2022-08-16 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Video processing method, apparatus, device and storage medium |
CN111182254B (en) * | 2020-01-03 | 2022-06-24 | 北京百度网讯科技有限公司 | Video processing method, device, equipment and storage medium |
CN111340696A (en) * | 2020-02-10 | 2020-06-26 | 南京理工大学 | Convolutional neural network image super-resolution reconstruction method fused with bionic visual mechanism |
CN111340696B (en) * | 2020-02-10 | 2022-11-04 | 南京理工大学 | Convolutional neural network image super-resolution reconstruction method fused with bionic visual mechanism |
CN113365107A (en) * | 2020-03-05 | 2021-09-07 | 阿里巴巴集团控股有限公司 | Video processing method, movie video processing method and device |
CN111654621B (en) * | 2020-05-26 | 2021-04-16 | 浙江大学 | Dual-focus camera continuous digital zooming method based on convolutional neural network model |
CN111654621A (en) * | 2020-05-26 | 2020-09-11 | 浙江大学 | Dual-focus camera continuous digital zooming method based on convolutional neural network model |
CN113747099A (en) * | 2020-05-29 | 2021-12-03 | 华为技术有限公司 | Video transmission method and device |
CN113747099B (en) * | 2020-05-29 | 2022-12-06 | 华为技术有限公司 | Video transmission method and device |
CN111861961A (en) * | 2020-07-25 | 2020-10-30 | 安徽理工大学 | Multi-scale residual error fusion model for single image super-resolution and restoration method thereof |
CN111861961B (en) * | 2020-07-25 | 2023-09-22 | 安徽理工大学 | Single image super-resolution multi-scale residual error fusion model and restoration method thereof |
CN112927136B (en) * | 2021-03-05 | 2022-05-10 | 江苏实达迪美数据处理有限公司 | Image reduction method and system based on convolutional neural network domain adaptation |
CN112927136A (en) * | 2021-03-05 | 2021-06-08 | 江苏实达迪美数据处理有限公司 | Image reduction method and system based on convolutional neural network domain adaptation |
CN113221842A (en) * | 2021-06-04 | 2021-08-06 | 第六镜科技(北京)有限公司 | Model training method, image recognition method, device, equipment and medium |
CN113221842B (en) * | 2021-06-04 | 2023-12-29 | 第六镜科技(北京)集团有限责任公司 | Model training method, image recognition method, device, equipment and medium |
CN113409195A (en) * | 2021-07-06 | 2021-09-17 | 中国标准化研究院 | Image super-resolution reconstruction method based on improved deep convolutional neural network |
Also Published As
Publication number | Publication date |
---|---|
CN109961396B (en) | 2020-12-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109961396A (en) | A kind of image super-resolution rebuilding method based on convolutional neural networks | |
CN111563902B (en) | Lung lobe segmentation method and system based on three-dimensional convolutional neural network | |
Xu et al. | EMFusion: An unsupervised enhanced medical image fusion network | |
Li et al. | Luminance-aware pyramid network for low-light image enhancement | |
CN105046672B (en) | A kind of image super-resolution rebuilding method | |
CN107977932A (en) | It is a kind of based on can differentiate attribute constraint generation confrontation network face image super-resolution reconstruction method | |
CN107369189A (en) | The medical image super resolution ratio reconstruction method of feature based loss | |
CN109214250A (en) | A kind of static gesture identification method based on multiple dimensioned convolutional neural networks | |
CN108460726A (en) | A kind of magnetic resonance image super-resolution reconstruction method based on enhancing recurrence residual error network | |
CN109544457A (en) | Image super-resolution method, storage medium and terminal based on fine and close link neural network | |
CN107424145A (en) | The dividing method of nuclear magnetic resonance image based on three-dimensional full convolutional neural networks | |
CN110458906A (en) | A kind of medical image color method based on depth color transfer | |
CN110288597A (en) | Wireless capsule endoscope saliency detection method based on attention mechanism | |
CN109033945A (en) | A kind of human body contour outline extracting method based on deep learning | |
CN107633522A (en) | Brain image dividing method and system based on local similarity movable contour model | |
CN106023098B (en) | Image mending method based on the more dictionary learnings of tensor structure and sparse coding | |
Yang et al. | Image super-resolution based on deep neural network of multiple attention mechanism | |
CN110322530A (en) | It is a kind of based on depth residual error network can interaction figure picture coloring | |
CN110322402A (en) | Medical image super resolution ratio reconstruction method based on dense mixing attention network | |
CN111161203A (en) | Multi-focus image fusion method based on memristor pulse coupling neural network | |
CN114581356B (en) | Image enhancement model generalization method based on style migration data augmentation | |
CN110503614A (en) | A kind of Magnetic Resonance Image Denoising based on sparse dictionary study | |
CN114820579A (en) | Semantic segmentation based image composite defect detection method and system | |
CN109785279B (en) | Image fusion reconstruction method based on deep learning | |
CN113256494B (en) | Text image super-resolution method |
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 |