CN109118428A - A kind of image super-resolution rebuilding method based on feature enhancing - Google Patents
A kind of image super-resolution rebuilding method based on feature enhancing Download PDFInfo
- Publication number
- CN109118428A CN109118428A CN201810581039.5A CN201810581039A CN109118428A CN 109118428 A CN109118428 A CN 109118428A CN 201810581039 A CN201810581039 A CN 201810581039A CN 109118428 A CN109118428 A CN 109118428A
- Authority
- CN
- China
- Prior art keywords
- characteristic pattern
- image
- feature
- resolution
- image super
- 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
- 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
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The present invention relates to a kind of image super-resolution rebuilding methods based on feature enhancing, comprising: construction feature demarcates sub-network;Enhance convolution module according to the features localization sub-network construction feature;Enhance convolution module according to the feature and constructs image super-resolution rebuilding model;Described image Super-resolution reconstruction established model is trained;According to the Super-resolution reconstruction established model and original image acquisition reconstruction image after training.The method of the invention method that the network architecture uses residual error study during the super-resolution rebuilding of image, the feature extracted especially for characteristic pattern has carried out selective enhancing and inhibition, make the image rebuild that there is higher Y-PSNR and structural similarity, it avoids and occurs fake information in reconstruction image, and obtain sharper keen visual effect and details reducing power true to nature.
Description
Technical field
The invention belongs to digital image processing fields, and in particular to a kind of image super-resolution rebuilding based on feature enhancing
Method.
Background technique
Important information form of the image as the human perception world, the abundant and details of content, directly determines this mankind
Experience the level of detail of content.When the pixel density on image as unit scale is higher, then image is more clear, the details of expression
Ability is stronger, and the information of human perception is abundanter, this namely high-definition picture.The super-resolution rebuilding of image is very
Various aspects have corresponding research such as remote sensing images, satellite imagery field, the field of medical imaging etc..
Currently, existing image super-resolution rebuilding method has super resolution ratio reconstruction method based on model and based on study
Super resolution ratio reconstruction method.Wherein, the super resolution ratio reconstruction method based on model mainly has the full calculus of variations, iteration reflective projection
Method, Tikhonov regularization method etc..And the super resolution ratio reconstruction method based on study has SRCNN method, VDSR method etc.,
SRCNN method is to obtain stronger sparse coding ability by study, therefore obtain stronger compared with traditional super-resolution method
Details reducing power;The high-definition picture signal-to-noise ratio that VDSR method obtains is higher, visual effect is sharper keen.
But SRCNN method visual effect is not sharp keen enough, the extracted feature of VDSR method mixes, and easily leads to reconstruction image
In there are fake informations, therefore the super resolution ratio reconstruction method based on model complete can not be retouched by the model that artificially designs
The mapping relations between low fractional diagram picture and high-definition picture are stated, the grain details of image can be lost while reconstruction image
Or generate high-definition picture fake information.
Summary of the invention
In order to solve the above-mentioned problems in the prior art, the present invention provides a kind of image based on feature enhancing is super
Resolution reconstruction method.The technical problem to be solved in the present invention is achieved through the following technical solutions:
The embodiment of the invention provides
A kind of image super-resolution rebuilding method based on feature enhancing, comprising:
Construction feature demarcates sub-network;
Enhance convolution module according to the features localization sub-network construction feature;
Enhance convolution module according to the feature and constructs image super-resolution rebuilding model;
Described image Super-resolution reconstruction established model is trained;
According to the Super-resolution reconstruction established model and original image acquisition reconstruction image after training.
In one embodiment of the invention, the features localization sub-network, comprising: pond layer, the first full articulamentum,
One active coating, the second full articulamentum and the second active coating;Wherein,
The output end of the pond layer connects the input terminal of the described first full articulamentum, for the characteristic pattern progress to input
Compression;
The output end of the first full articulamentum connects the input terminal of first active coating, for defeated to the pond layer
Characteristic pattern out is weighted fusion;
The output end of first active coating connects the input terminal of the described second full articulamentum, complete for increasing described first
The sparsity of articulamentum output feature graph parameter;
The output end of the second full articulamentum connects the input terminal of second active coating, for first activation
The characteristic pattern of layer output is augmented;
Second active coating is used to that the characteristic pattern of the described second full articulamentum output to be normalized.
In one embodiment of the invention, the pond layer is average pond layer, and the average pond layer exports feature
The size of figure is c × h × w, and output dimension is c × 1 × 1;Wherein, c is the port number that the average pond layer exports characteristic pattern,
H is the height that the average pond layer exports characteristic pattern, and w is the width that the average pond layer exports characteristic pattern.
In one embodiment of the invention, the dimension of the output characteristic pattern of the described first full articulamentum is
Wherein, c is the port number of the output characteristic pattern of the first full articulamentum, when N is the output characteristic pattern fusion of the first full articulamentum
Compression ratio.
In one embodiment of the invention, first active coating is to correct linear unit activating layer, and described second is complete
The dimension that articulamentum exports characteristic pattern is c × 1 × 1;Wherein, c is the port number that the described second full articulamentum exports characteristic pattern.
In one embodiment of the invention, second active coating is S type active coating, and the dimension for exporting characteristic pattern is c
×1×1;Wherein, c is the port number that second active coating exports characteristic pattern.
In one embodiment of the invention, convolution module, packet are enhanced according to the features localization sub-network construction feature
It includes:
Construct convolution sub-network;
Constructing the feature according to the convolution sub-network and the features localization sub-network enhances convolution module.
In one embodiment of the invention, convolution sub-network is constructed, comprising:
Construct convolutional layer and third active coating;
The input terminal that the output end of the convolutional layer is connected to the third active coating constructs the convolution sub-network.
In one embodiment of the invention, convolution module is enhanced according to the feature and constructs image super-resolution rebuilding mould
Type, comprising:
Direct-connected convolutional neural networks are constructed according to multiple feature enhancing convolution modules;
Described image Super-resolution reconstruction established model is constructed according to the direct-connected convolutional neural networks and residual error study bypass,
In, residual error study bypass is for by the input feature vector figure of the direct-connected convolutional neural networks and the output point-to-point phase of characteristic pattern
Add.
In one embodiment of the invention, described image Super-resolution reconstruction established model is trained, comprising:
Choose training sample set and test sample collection;
According to training sample set training described image Super-resolution reconstruction established model.
Compared with prior art, beneficial effects of the present invention:
The method of the invention method that the network architecture uses residual error study during the super-resolution rebuilding of image, especially
The feature extracted for characteristic pattern has carried out selective enhancing and inhibition, and the image rebuild is made to have higher peak value noise
Than and structural similarity, avoid and occur fake information in reconstruction image, and obtain sharper keen visual effect and true to nature
Details reducing power.
Detailed description of the invention
Fig. 1 is a kind of process of image super-resolution rebuilding method based on feature enhancing provided in an embodiment of the present invention
Figure;
Fig. 2 is that a kind of feature of image super-resolution rebuilding method based on feature enhancing provided in an embodiment of the present invention increases
Strong convolution module structure principle chart;
Fig. 3 is that a kind of image of image super-resolution rebuilding method based on feature enhancing provided in an embodiment of the present invention is super
The structure principle chart of resolution reconstruction model;
Fig. 4 is original image of the embodiment of the present invention;
Fig. 5 a-5d is respectively after bicubic interpolation method, SRCNN method, VDSR method and method for reconstructing of the invention
The image of output.
Specific embodiment
Further detailed description is done to the present invention combined with specific embodiments below, but embodiments of the present invention are not limited to
This.
Embodiment one
Referring to Figure 1, Fig. 2 and Fig. 3, Fig. 1 are a kind of image oversubscription based on feature enhancing provided in an embodiment of the present invention
The flow chart of resolution method for reconstructing;Fig. 2 is a kind of image super-resolution weight based on feature enhancing provided in an embodiment of the present invention
The feature of construction method enhances convolution module structure principle chart;Fig. 3 is provided in an embodiment of the present invention a kind of based on feature enhancing
The structure principle chart of the image super-resolution rebuilding model of image super-resolution rebuilding method.As shown in Figure 1, a kind of be based on feature
The image super-resolution rebuilding method of enhancing, comprising: construction feature demarcates sub-network;It is constructed according to the features localization sub-network
Feature enhances convolution module;Enhance convolution module according to the feature and constructs image super-resolution rebuilding model;To described image
Super-resolution reconstruction established model is trained;According to the Super-resolution reconstruction established model and original image acquisition reconstruction figure after training
Picture.
Preferably, as shown in Fig. 2, the features localization sub-network, comprising: pond layer, the first full articulamentum, the first activation
Layer, the second full articulamentum and the second active coating;Wherein, the output end of the pond layer connects the input of the described first full articulamentum
End, for being compressed to the characteristic pattern of input;The output end of the first full articulamentum connects the defeated of first active coating
Enter end, for being weighted fusion to compressed characteristic pattern;The output end connection described second of first active coating connects entirely
The input terminal for connecing layer accelerates convergence process for increasing the sparsity of feature graph parameter after Weighted Fusion;Described second full connection
The output end of layer connects the input terminal of second active coating, and the characteristic pattern for exporting to first active coating expands
Dimension;Second active coating is used to that the characteristic pattern of the described second full articulamentum output to be normalized.
Preferably, on the one hand pond layer makes characteristic pattern become smaller, simplifies and calculate again for compressing to the characteristic pattern of input
Miscellaneous degree;On the one hand Feature Compression is carried out, main feature is extracted;Wherein, the pond layer that the present embodiment uses is put down for average pond layer
Equal pond layer is exactly the feature that this region is represented using the average value of some special characteristic on one region of characteristic image,
The average value in the region of characteristic pattern is calculated as the value after the pool area.
Preferably, the size of average pond layer output characteristic pattern is c × h × w, and output dimension is c × 1 × 1;Wherein, c is
The port number of characteristic pattern is exported, h is characterized the height of figure, and w is characterized the width of figure.
Preferably, the first full articulamentum is made of the full articulamentum that a compression ratio is N, and the first full articulamentum is to input
Each characteristic value be weighted fusion, and compressed in dimension, reject unnecessary information during recompression, it is defeated
Characteristic value after Weighted Fusion out is weighted fusion to the characteristic pattern of pond layer output.Its export characteristic pattern dimension beWherein, c is the port number of the first full articulamentum input feature vector figure, and N is characterized compression ratio when fusion.
Preferably, the first active coating is to correct linear unit activating layer, i.e. ReLU (Rectified LinearUnit, letter
Claim ReLU) active coating, i.e., operation is executed using ReLU function, wherein ReLU function can indicate are as follows:
F (x)=max (0, x),
Wherein, x is the calibration value that the first full articulamentum exports characteristic pattern.First active coating is for increasing by the first full articulamentum
The sparsity of feature graph parameter is exported, this characteristic can remove the redundant data in characteristic pattern, maximum possible keeping characteristics figure
Feature, to accelerate convergence process.
Preferably, the dimension of the described second full articulamentum output characteristic pattern is c × 1 × 1;Wherein, c connects entirely for described second
Connect the port number of layer output characteristic pattern.Second full articulamentum is augmented the characteristic pattern of input, and exports smooth without activating
Characteristic pattern.
Preferably, the second active coating is S type active coating, i.e. Sigmoid active coating, executes behaviour using sigmoid function
Make, wherein sigmoid function can indicate are as follows:
Wherein, y is the calibration value that the second full articulamentum exports characteristic pattern, and e is the nature truth of a matter.Sigmoid active coating is also referred to as
For S type active coating, the input of this layer is to export without activating smooth characteristic pattern as final characteristic pattern.It exports characteristic pattern
Dimension be c × 1 × 1;Wherein, c is characterized the port number of figure.
Preferably, convolution sub-network is constructed, comprising: building convolutional layer and third active coating;By the output of the convolutional layer
End connects the input terminal of the third active coating, constructs the convolution sub-network.Wherein, convolutional layer is by convolution kernel size W × H=
3 × 3, convolution nuclear volume is 64, step value 1, the convolutional layer that edge filling is 1;Wherein, convolution kernel is for doing convolution operation
Template, step value refers to the mobile distance of convolution kernel when each convolution, and edge filling is the figure after convolution in order to prevent
As inconsistent with convolved image size.Third active coating is ReLU active coating, and the output of third active coating is convolution sub-network
Output.
Preferably, after constructing convolution sub-network, institute is constructed according to the convolution sub-network and the features localization sub-network
Enhancing convolution module is stated, i.e., the output end of the convolution sub-network is connected to the input terminal of the features localization sub-network, and will
The output of convolution sub-network is multiplied with the output of features localization sub-network, to obtain feature enhancing convolution module.Wherein,
Features localization sub-network in feature enhancing convolution module can select each characteristic pattern for exporting convolution sub-network
The enhancing and inhibition of selecting property.
Preferably, as shown in figure 3, multiple features enhancing convolution module is sequentially connected, i.e., by first feature enhancing volume
The output end of volume module connects the output end of second feature enhancing convolution module, is sequentially connected, until the M-1 feature increases
The input terminal of the output connection m-th feature enhancing convolution module of strong convolution module, has been built into direct-connected convolutional neural networks,
In, M is the natural number greater than 1, and the input terminal of first feature enhancing convolution module is the input of the direct-connected convolutional neural networks
End, the output end that m-th feature enhances convolution module is the output end of the direct-connected convolutional neural networks.Preferably, the value of M is
10。
Preferably, as shown in figure 3, introducing the bypass for residual error study, shape on the basis of direct-connected convolutional neural networks
At the image super-resolution rebuilding model based on feature enhancing deep learning, wherein the bypass of residual error study is straight for that will input
Even the characteristic pattern of convolutional neural networks carries out point-to-point be added with the characteristic pattern that direct-connected convolutional neural networks export.
Preferably, described image Super-resolution reconstruction established model is trained, comprising: choose training sample set and test specimens
This collection;According to training sample set training described image Super-resolution reconstruction established model;Institute is detected according to the test sample collection
State image super-resolution rebuilding model.
Preferably, training sample set is used for the model parameter of training image Super-resolution reconstruction established model, so that the mould
Type can rebuild figure to band and carry out accurate reconstruction, and test sample collection is used to test the reconstruction model that training is completed, with
Evaluate the performance superiority and inferiority of the reconstruction model.Training sample set chooses BSD400 data set, and test sample collection chooses Set12 data
Collection carries out network training according to the training method of setting to image super-resolution rebuilding model using training sample set.Specifically
Training method are as follows: use existing Adam (A Method for Stochastic Optimization) optimizer, batch is big
It is small to be set as 128, in training with 0.001 learning rate 25 bouts of training, then again with learning rate training 25 times of 0.0001
It closes, amounts to 50 bouts of training.The test sample collection is input in image super-resolution rebuilding model, trained figure is detected
As the performance of Super-resolution reconstruction established model.
It preferably, i.e., will be former according to the Super-resolution reconstruction established model and original image acquisition reconstruction image after training
Super-resolution reconstruction established model after the input training of beginning image exports super-resolution rebuilding figure after the processing of the reconstruction model
Picture.
A kind of image super-resolution rebuilding method based on feature enhancing provided by the invention constructs super-resolution data
Model, and super-resolution model is trained, reconstruction image is obtained finally by super-resolution data model, this method obtains
Reconstruction image use residual error study method, carried out selective enhancing and inhibition especially for the feature of characteristic pattern,
Make the image rebuild that there is higher Y-PSNR and structural similarity, avoids and occur fake information in reconstruction image, and
Obtain sharper keen visual effect and details reducing power true to nature.
Embodiment two
Fig. 4, Fig. 5 a-5d are referred to, Fig. 4 is original image of the embodiment of the present invention;Fig. 5 a-5d is respectively through bicubic interpolation
The image exported after method, SRCNN method, VDSR method and method for reconstructing of the invention.The present embodiment is in above-described embodiment
On the basis of, citing description is carried out to image reconstruction to method for reconstructing provided by the invention and existing method for reconstructing.Specifically, respectively
Use bicubic interpolation method, SRCNN method, VDSR method and the method for the present invention to the low-resolution image in Fig. 4 to scale
The factor 4 carries out image super-resolution rebuilding.Wherein Fig. 5 a is the image exported after being rebuild using bicubic interpolation method;Fig. 5 b is
The image exported using SRCNN method;Fig. 5 c is the image exported after being rebuild using VDSR method;Fig. 5 d is using of the invention
The image that method exports after rebuilding.
Preferably, by the comparison of Fig. 5 a to Fig. 5 d as can be seen that image is compared with other after the method for the present invention is rebuild
Reconstructed results, details is richer, and edge is apparent.
Preferably, Y-PSNR (PSNR) and structural similarity (SSIM) is respectively adopted to quantify the control assessment present invention
The image super-resolution rebuilding method and existing bicubic interpolation method, SRCNN method and VDSR based on feature enhancing proposed
The performance of method, control assessment result are as follows:
As seen from the above table:
(1) Y-PSNR (PSNR) of the image after method for reconstructing proposed by the present invention reconstruction is apparently higher than bicubic
Interpolation method, SRCNN method and VDSR method, i.e. image of the explanation after the method for the present invention is rebuild remain more image details
Information.
(2) the structural similarity coefficient (SSIM) of the image after method Super-resolution Reconstruction proposed by the present invention is apparently higher than
Bicubic interpolation method, SRCNN method and VDSR method as a result, i.e. image of the explanation after the method for the present invention super-resolution rebuilding
Remain the more architectural characteristics of original image.
Method for reconstructing reconstruction effect provided by the invention is preferable, has the edge being more clear, the detailed information in image
Also it is more clear, the structural informations such as edge and the details of original image can be retained to a greater extent.
The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that
Specific implementation of the invention is only limited to these instructions.For those of ordinary skill in the art to which the present invention belongs, exist
Under the premise of not departing from present inventive concept, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to of the invention
Protection scope.
Claims (10)
1. a kind of image super-resolution rebuilding method based on feature enhancing characterized by comprising
Construction feature demarcates sub-network;
Enhance convolution module according to the features localization sub-network construction feature;
Enhance convolution module according to the feature and constructs image super-resolution rebuilding model;
Described image Super-resolution reconstruction established model is trained;
According to the Super-resolution reconstruction established model and original image acquisition reconstruction image after training.
2. image super-resolution rebuilding method according to claim 1, which is characterized in that the features localization sub-network,
It include: pond layer, the first full articulamentum, the first active coating, the second full articulamentum and the second active coating;Wherein,
The output end of the pond layer connects the input terminal of the described first full articulamentum, for pressing the characteristic pattern of input
Contracting;
The output end of the first full articulamentum connects the input terminal of first active coating, for what is exported to the pond layer
Characteristic pattern is weighted fusion;
The output end of first active coating connects the input terminal of the described second full articulamentum, for increasing by the described first full connection
The sparsity of layer output feature graph parameter;
The output end of the second full articulamentum connects the input terminal of second active coating, for defeated to first active coating
Characteristic pattern out is augmented;
Second active coating is used to that the characteristic pattern of the described second full articulamentum output to be normalized.
3. image super-resolution rebuilding method according to claim 2, which is characterized in that the pond layer is average pond
The size of layer, the average pond layer output characteristic pattern is c × h × w, and output dimension is c × 1 × 1;Wherein, c is described average
Pond layer exports the port number of characteristic pattern, and h is the height that the average pond layer exports characteristic pattern, and w is that the average pond layer is defeated
The width of characteristic pattern out.
4. image super-resolution rebuilding method according to claim 2, which is characterized in that the first full articulamentum it is defeated
The dimension of characteristic pattern is outWherein, c is the port number of the output characteristic pattern of the first full articulamentum, and N connects entirely for first
Connect the compression ratio when output characteristic pattern fusion of layer.
5. image super-resolution rebuilding method according to claim 2, which is characterized in that first active coating is amendment
The dimension of linear unit active coating, the second full articulamentum output characteristic pattern is c × 1 × 1;Wherein, c connects entirely for described second
Connect the port number of layer output characteristic pattern.
6. image super-resolution rebuilding method according to claim 2, which is characterized in that second active coating is S type
Active coating, the dimension for exporting characteristic pattern is c × 1 × 1;Wherein, c is the port number that second active coating exports characteristic pattern.
7. image super-resolution rebuilding method according to claim 1, which is characterized in that according to the features localization subnet
Network construction feature enhances convolution module, comprising:
Construct convolution sub-network;
Constructing the feature according to the convolution sub-network and the features localization sub-network enhances convolution module.
8. image super-resolution rebuilding method according to claim 7, which is characterized in that building convolution sub-network, comprising:
Construct convolutional layer and third active coating;
The input terminal that the output end of the convolutional layer is connected to the third active coating constructs the convolution sub-network.
9. image super-resolution rebuilding method according to claim 1 or claim 7, which is characterized in that enhanced according to the feature
Convolution module constructs image super-resolution rebuilding model, comprising:
Direct-connected convolutional neural networks are constructed according to multiple feature enhancing convolution modules;
According to the direct-connected convolutional neural networks and residual error study bypass building described image Super-resolution reconstruction established model, wherein
Residual error study bypass is for by the input feature vector figure of the direct-connected convolutional neural networks and exporting that characteristic pattern is point-to-point to be added.
10. image super-resolution rebuilding method according to claim 1, which is characterized in that described image super-resolution
Reconstruction model is trained, comprising:
Choose training sample set and test sample collection;
According to training sample set training described image Super-resolution reconstruction established model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810581039.5A CN109118428B (en) | 2018-06-07 | 2018-06-07 | Image super-resolution reconstruction method based on feature enhancement |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810581039.5A CN109118428B (en) | 2018-06-07 | 2018-06-07 | Image super-resolution reconstruction method based on feature enhancement |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109118428A true CN109118428A (en) | 2019-01-01 |
CN109118428B CN109118428B (en) | 2023-05-19 |
Family
ID=64822962
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810581039.5A Active CN109118428B (en) | 2018-06-07 | 2018-06-07 | Image super-resolution reconstruction method based on feature enhancement |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109118428B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110120019A (en) * | 2019-04-26 | 2019-08-13 | 电子科技大学 | A kind of residual error neural network and image deblocking effect method based on feature enhancing |
CN112508780A (en) * | 2019-09-16 | 2021-03-16 | 中移(苏州)软件技术有限公司 | Training method and device of image processing model and storage medium |
CN114245126A (en) * | 2021-11-26 | 2022-03-25 | 电子科技大学 | Depth feature map compression method based on texture cooperation |
CN115082317A (en) * | 2022-07-11 | 2022-09-20 | 四川轻化工大学 | Image super-resolution reconstruction method for attention mechanism enhancement |
Citations (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103077511A (en) * | 2013-01-25 | 2013-05-01 | 西安电子科技大学 | Image super-resolution reconstruction method based on dictionary learning and structure similarity |
CN104991241A (en) * | 2015-06-30 | 2015-10-21 | 西安电子科技大学 | Target signal extraction and super-resolution enhancement processing method in strong clutter condition |
US20160042537A1 (en) * | 2013-03-15 | 2016-02-11 | Real Time Tomography, Llc | Enhancements for displaying and viewing tomosynthesis images |
CN106228124A (en) * | 2016-07-17 | 2016-12-14 | 西安电子科技大学 | SAR image object detection method based on convolutional neural networks |
CN106709875A (en) * | 2016-12-30 | 2017-05-24 | 北京工业大学 | Compressed low-resolution image restoration method based on combined deep network |
US20170193680A1 (en) * | 2016-01-04 | 2017-07-06 | Kla-Tencor Corporation | Generating high resolution images from low resolution images for semiconductor applications |
CN106952229A (en) * | 2017-03-15 | 2017-07-14 | 桂林电子科技大学 | Image super-resolution rebuilding method based on the enhanced modified convolutional network of data |
CN107123089A (en) * | 2017-04-24 | 2017-09-01 | 中国科学院遥感与数字地球研究所 | Remote sensing images super-resolution reconstruction method and system based on depth convolutional network |
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 |
CN107358576A (en) * | 2017-06-24 | 2017-11-17 | 天津大学 | Depth map super resolution ratio reconstruction method based on convolutional neural networks |
CN107464216A (en) * | 2017-08-03 | 2017-12-12 | 济南大学 | A kind of medical image ultra-resolution ratio reconstructing method based on multilayer convolutional neural networks |
CN107563965A (en) * | 2017-09-04 | 2018-01-09 | 四川大学 | Jpeg compressed image super resolution ratio reconstruction method based on convolutional neural networks |
CN107633482A (en) * | 2017-07-24 | 2018-01-26 | 西安电子科技大学 | A kind of super resolution ratio reconstruction method based on sequence image |
CN107633520A (en) * | 2017-09-28 | 2018-01-26 | 福建帝视信息科技有限公司 | A kind of super-resolution image method for evaluating quality based on depth residual error network |
CN107977932A (en) * | 2017-12-28 | 2018-05-01 | 北京工业大学 | It is a kind of based on can differentiate attribute constraint generation confrontation network face image super-resolution reconstruction method |
CN108122197A (en) * | 2017-10-27 | 2018-06-05 | 江西高创保安服务技术有限公司 | A kind of image super-resolution rebuilding method based on deep learning |
-
2018
- 2018-06-07 CN CN201810581039.5A patent/CN109118428B/en active Active
Patent Citations (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103077511A (en) * | 2013-01-25 | 2013-05-01 | 西安电子科技大学 | Image super-resolution reconstruction method based on dictionary learning and structure similarity |
US20160042537A1 (en) * | 2013-03-15 | 2016-02-11 | Real Time Tomography, Llc | Enhancements for displaying and viewing tomosynthesis images |
CN104991241A (en) * | 2015-06-30 | 2015-10-21 | 西安电子科技大学 | Target signal extraction and super-resolution enhancement processing method in strong clutter condition |
US20170193680A1 (en) * | 2016-01-04 | 2017-07-06 | Kla-Tencor Corporation | Generating high resolution images from low resolution images for semiconductor applications |
CN106228124A (en) * | 2016-07-17 | 2016-12-14 | 西安电子科技大学 | SAR image object detection method based on convolutional neural networks |
CN106709875A (en) * | 2016-12-30 | 2017-05-24 | 北京工业大学 | Compressed low-resolution image restoration method based on combined deep network |
CN106952229A (en) * | 2017-03-15 | 2017-07-14 | 桂林电子科技大学 | Image super-resolution rebuilding method based on the enhanced modified convolutional network of data |
CN107123089A (en) * | 2017-04-24 | 2017-09-01 | 中国科学院遥感与数字地球研究所 | Remote sensing images super-resolution reconstruction method and system based on depth convolutional network |
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 |
CN107358576A (en) * | 2017-06-24 | 2017-11-17 | 天津大学 | Depth map super resolution ratio reconstruction method based on convolutional neural networks |
CN107633482A (en) * | 2017-07-24 | 2018-01-26 | 西安电子科技大学 | A kind of super resolution ratio reconstruction method based on sequence image |
CN107464216A (en) * | 2017-08-03 | 2017-12-12 | 济南大学 | A kind of medical image ultra-resolution ratio reconstructing method based on multilayer convolutional neural networks |
CN107563965A (en) * | 2017-09-04 | 2018-01-09 | 四川大学 | Jpeg compressed image super resolution ratio reconstruction method based on convolutional neural networks |
CN107633520A (en) * | 2017-09-28 | 2018-01-26 | 福建帝视信息科技有限公司 | A kind of super-resolution image method for evaluating quality based on depth residual error network |
CN108122197A (en) * | 2017-10-27 | 2018-06-05 | 江西高创保安服务技术有限公司 | A kind of image super-resolution rebuilding method based on deep learning |
CN107977932A (en) * | 2017-12-28 | 2018-05-01 | 北京工业大学 | It is a kind of based on can differentiate attribute constraint generation confrontation network face image super-resolution reconstruction method |
Non-Patent Citations (4)
Title |
---|
JIWON KIM: ""Deeply-Recursive Convolutional Network for Image Super-Resolution"", 《PROCESS OF THE IEEE CONFERENCE OF COMPUTER VISION AND PATTERN RECOGINITION》 * |
刘颖等: "基于边缘增强的正则化超分辨率图像重建", 《西安邮电大学学报》 * |
秦翰林等: "基于改进的分块压缩感知红外图像重建", 《强激光与粒子束》 * |
罗鸣威等: "变焦序列图像超分辨率重建算法研究", 《南京大学学报(自然科学)》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110120019A (en) * | 2019-04-26 | 2019-08-13 | 电子科技大学 | A kind of residual error neural network and image deblocking effect method based on feature enhancing |
CN110120019B (en) * | 2019-04-26 | 2023-03-28 | 电子科技大学 | Residual error neural network based on feature enhancement and image deblocking method |
CN112508780A (en) * | 2019-09-16 | 2021-03-16 | 中移(苏州)软件技术有限公司 | Training method and device of image processing model and storage medium |
CN114245126A (en) * | 2021-11-26 | 2022-03-25 | 电子科技大学 | Depth feature map compression method based on texture cooperation |
CN115082317A (en) * | 2022-07-11 | 2022-09-20 | 四川轻化工大学 | Image super-resolution reconstruction method for attention mechanism enhancement |
Also Published As
Publication number | Publication date |
---|---|
CN109118428B (en) | 2023-05-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109118428A (en) | A kind of image super-resolution rebuilding method based on feature enhancing | |
CN111861961B (en) | Single image super-resolution multi-scale residual error fusion model and restoration method thereof | |
CN110197468A (en) | A kind of single image Super-resolution Reconstruction algorithm based on multiple dimensioned residual error learning network | |
CN109389556A (en) | The multiple dimensioned empty convolutional neural networks ultra-resolution ratio reconstructing method of one kind and device | |
CN107240066A (en) | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks | |
CN109064396A (en) | A kind of single image super resolution ratio reconstruction method based on depth ingredient learning network | |
Zhu et al. | Efficient single image super-resolution via hybrid residual feature learning with compact back-projection network | |
CN114092330A (en) | Lightweight multi-scale infrared image super-resolution reconstruction method | |
CN108830790A (en) | It is a kind of based on the fast video super resolution ratio reconstruction method for simplifying convolutional neural networks | |
CN108805808A (en) | A method of improving video resolution using convolutional neural networks | |
CN111932461B (en) | Self-learning image super-resolution reconstruction method and system based on convolutional neural network | |
Xu et al. | Dense bynet: Residual dense network for image super resolution | |
CN106157244A (en) | A kind of QR Code Image Super-resolution Reconstruction method based on rarefaction representation | |
CN110111256A (en) | Image Super-resolution Reconstruction method based on residual error distillation network | |
CN110322402A (en) | Medical image super resolution ratio reconstruction method based on dense mixing attention network | |
CN110084745A (en) | Image super-resolution rebuilding method based on dense convolutional neural networks in parallel | |
Yang et al. | Image super-resolution based on deep neural network of multiple attention mechanism | |
CN108765290A (en) | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks | |
CN114331831A (en) | Light-weight single-image super-resolution reconstruction method | |
CN112365554A (en) | Compressed sensing image reconstruction method based on multi-scale residual error neural network | |
CN112767243B (en) | Method and system for realizing super-resolution of hyperspectral image | |
CN115829876A (en) | Real degraded image blind restoration method based on cross attention mechanism | |
CN112767252A (en) | Image super-resolution reconstruction method based on convolutional neural network | |
CN111951203A (en) | Viewpoint synthesis method, apparatus, device and computer readable storage medium | |
CN108416736A (en) | A kind of image super-resolution rebuilding method returned based on secondary anchor point neighborhood |
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 |