CN108765290A - A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks - Google Patents
A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks Download PDFInfo
- Publication number
- CN108765290A CN108765290A CN201810535634.5A CN201810535634A CN108765290A CN 108765290 A CN108765290 A CN 108765290A CN 201810535634 A CN201810535634 A CN 201810535634A CN 108765290 A CN108765290 A CN 108765290A
- Authority
- CN
- China
- Prior art keywords
- block structure
- characteristic pattern
- resolution
- sub
- neural networks
- 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
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
- 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
- G06T3/4038—Scaling the whole image or part thereof for image mosaicing, i.e. plane images composed of plane sub-images
Abstract
The invention discloses a kind of super resolution ratio reconstruction methods based on improved dense convolutional neural networks:Including data preparation stage;Network structure builds the stage;Model training stage;The image reconstruction stage.The thought of dense convolutional neural networks structure is applied to the super-resolution rebuilding of single-frame images by the present invention, and improves network structure on the basis of DenseNet structures, has been advanced optimized model framework, has been improved algorithm effect.
Description
Technical field
The invention belongs to technical field of image processing, and more specifically, it relates to one kind based on improved dense convolution god
Super resolution ratio reconstruction method through network.
Background technology
Super-resolution is a kind of method of the raising image definition proposed in recent years.In the hardware for not changing Image Acquisition
Under the conditions of, there are the different sequence of low resolution pictures of space displacement fog-level to rebuild high-resolution for one group of early stage research and utilization
Image, later research direction be partial to be rebuild with the guidance of single low-resolution image additional prior knowledge, referred to as single-frame images is super
Resolution reconstruction technology.Image after super-resolution rebuilding includes abundanter details, and higher details resolving power more meets
Human eye vision demand is applied to the fields such as recognition of face, Car license recognition, Medical Image Processing gradually.
Single-frame images super-resolution rebuilding algorithm can be roughly divided into three classes:Based on interpolation, based on rebuild, based on study
Method.Method based on interpolation assumes that the gray value of pixel is consecutive variations, and is waited for using the calculating of the gray value of neighborhood pixels
The gray value of interpolating pixel, classical image interpolation method have arest neighbors to insert method, bilinear interpolation and bi-cubic interpolation etc..Base
It is simple and practicable in the method for interpolation, but reconstruction image high-frequency information lacks, and image is excessively fuzzy.Therefore, the method based on reconstruction
On the basis of the method based on interpolation, some Image Priori Knowledges are artificially added, reconstruct abundanter detail of the high frequency.
But since Image Priori Knowledge depends on the constraint of practical problem, so the method robustness based on reconstruction is not high.For this
One problem, the method that the method based on study utilizes machine learning, by the elder generation for establishing the current super-resolution problem of model learning
Knowledge is tested, restoring high-frequency information for low sampled data Problems of Reconstruction provides new approaches, becomes research hotspot in recent years.In view of
Successful application of the deep learning in image domains the experiment has found that convolutional neural networks are in single-frame images super-resolution rebuilding problem
On it is also fruitful, general convolutional neural networks structure, such as VGG, ResNet at present all greatly improves single-frame images
The effect of super-resolution rebuilding.
But the method based on VGG and ResNet does not all utilize image local feature fully, it is difficult to by simply
The strategy for increasing network depth improves reconstructed image quality.
Invention content
Purpose of the invention is to overcome the shortcomings in the prior art, provide it is a kind of complete, be used for end to end
The deep learning algorithm of single-frame images super-resolution rebuilding, i.e., a kind of super-resolution based on improved dense convolutional neural networks
Method for reconstructing answers the thought of dense convolutional neural networks structure (Dense Convolutional Network, DenseNet)
The super-resolution rebuilding of single-frame images is used, and network structure is improved on the basis of DenseNet structures, is advanced optimized
Model framework, improves algorithm effect.
The purpose of the present invention is what is be achieved through the following technical solutions.
The super resolution ratio reconstruction method based on improved dense convolutional neural networks of the present invention, includes the following steps:
Step 1, data preparation
1) data set is divided:Using public data collection DIV2K, which includes 800 training figures, 100 proof diagrams
With 100 test charts, wherein 100 proof diagrams rebuild effect for testing, 800 training figures and 100 proof diagrams be all by
High-definition picture and its corresponding low-resolution image composition, low-resolution image is high-definition picture via the model that degrades
It generates;
2) the non-overlapping image block for being divided into 96 × 96 sizes is schemed into 800 training, as network inputs;
Step 2, network structure are built
1) the low-resolution image block of 96 × 96 sizes is inputted to one 7 × 7 convolutional layer, and uses ReLU
(Rectified Linear Units) is used as activation primitive;
2) the output characteristic pattern of 7 × 7 convolutional layers inputs multi-connection block structure, wherein each multi-connection block structure includes
One sub- block structure and an identical mapping;
3) multi-connection block structure is stacked;
4) the output characteristic pattern of the last one multi-connection block structure is inputted into 1 × 1 convolutional layer, reduces the dimension of characteristic pattern
Number;
5) skip floor is used, the characteristic pattern after the characteristic pattern and 1 × 1 convolutional layer dimensionality reduction of the output of 7 × 7 convolutional layers is stacked
Get up, collectively as the input of amplification module;
6) method for using low resolution characteristic pattern to be spliced into high-resolution features figure in order, by low resolution characteristic pattern
It is amplified to certain multiple;
7) it uses 1 × 1 convolutional layer that characteristic pattern is limited to tri- channels RGB, and uses Tanh activation primitives;
Step 3, model training
1) learning rate is set as 10-3, and every 75000 iterative learning rates reduce half, no longer reduce until error or
After 250000 iteration, deconditioning;
2) use L1 norms as loss function;
3) ADAM optimization methods, β are used1=0.9, β2=0.999;
Step 4, image reconstruction
The low-resolution image for inputting arbitrary size, loads trained model, exports the high-definition picture of reconstruction.
The building process of sub-block structure described in step 2:First, the thought based on dense convolutional neural networks structure, is adopted
The structure that 3 × 3 convolutional layers are connect with 1 × 1 convolutional layer, as the benchmark architecture of sub-block, and each convolutional layer is made using ReLU
For activation primitive;Then, on the basis of benchmark architecture, increase by 3 × 3 convolutional layers, the improved structure as sub-block structure;
Finally, the benchmark architecture of sub-block structure and improved structure are stacked in characteristic pattern channel dimension, as melting for sub-block structure
Close structure.
The definition of the identical mapping:In each multi-connection block structure, all comprising one from the connection for being input to output, directly
Connect the input as next multi-connection block structure by the output of a upper multi-connection block structure.
Compared with prior art, advantageous effect caused by technical scheme of the present invention is:
(1) since single-frame images super-resolution rebuilding problem is to the shallow-layers feature-sensitive such as edge, texture, it is proposed that using thick
Close convolutional neural networks make network structure extract the reconstruction that abundanter detail of the high frequency is used for image.
(2) since the feature under big receptive field can include the information between abundanter neighbor pixel.And it is small
The concatenated network structure of convolution kernel can increase the non-thread sexuality of network under the premise of ensureing identical receptive field size.Cause
This, it is proposed that the dense convolutional neural networks based on small convolution kernel of connecting improve reconstructed image quality.
(3) feature for being directed to different feeling open country size includes different image abstraction information, it is proposed that dense piece of multi-connection
Network structure, further improve the index and visual effect of reconstruction image.
(4) invention increases the recycling to shallow-layer local feature, strengthen local feature under different feeling open country and
The fusion of global characteristics optimizes the existing super resolution ratio reconstruction method based on convolutional neural networks.Secondly, the present invention devises
One network structure end to end, eliminates cumbersome pretreatment and post-treating and other steps, method is simple and practicable.Moreover, because
The present invention can amplify the low-resolution image of arbitrary size, have only with the structure of convolutional layer so in phase of regeneration
Preferable generalization ability.
Description of the drawings
Fig. 1 is the network overall structure figure of the present invention;
Fig. 2 is the different structure figure of sub-block in multi-connection block structure in the present invention;
Fig. 3 is amplification module structure chart in the present invention.
Specific implementation mode
The invention will be further described below in conjunction with the accompanying drawings.
The super resolution ratio reconstruction method based on improved dense convolutional neural networks of the present invention, includes the following steps:
One, data preparation
1) data set is divided:Using public data collection DIV2K, which includes 800 training figures, 100 proof diagrams
With 100 test charts.Since the high-definition picture of test chart is not open, 100 proof diagrams rebuild effect for testing
Fruit.Wherein, for trained 800 training figure, be all by high-definition picture for 100 proof diagrams of test and it is corresponding
Low-resolution image composition, low-resolution image is high-definition picture to be generated via the certain specific model that degrades.
2) the non-overlapping image block for being divided into 96 × 96 sizes is schemed into 800 training, as network inputs.
Two, network structure is built
1) the low-resolution image block of 96 × 96 sizes is inputted to one 7 × 7 convolutional layer, and uses ReLU
(Rectified Linear Units) is used as activation primitive.
2) the output characteristic pattern of 7 × 7 convolutional layers inputs multi-connection block structure.Wherein, each multi-connection block structure includes
One sub- block structure and an identical mapping.
The building process of sub-block structure:First, the thought based on dense convolutional neural networks structure, the present invention use 1 × 1
Convolutional layer connects the structure of 3 × 3 convolutional layers, and as the benchmark architecture of sub-block, and each convolutional layer is used as activation using ReLU
Function;Then, on the basis of benchmark architecture, the present invention increases by 3 × 3 convolutional layers, increases the receptive field of sub-block structure,
Improved structure as sub-block structure;Finally, the present invention by the benchmark architecture of sub-block structure and improved structure in characteristic pattern channel
Dimension is stacked, the fusion structure as sub-block structure.
Identical mapping:In each multi-connection block structure, all comprising one from the connection for being input to output, directly by upper one
Input of the output of multi-connection block structure as next multi-connection block structure increases to strengthen the recycling of shallow-layer feature
The complexity and diversity of extraction feature are added.
3) multi-connection block structure is stacked, network depth, the characteristic information that extraction is more abundant, complicated are increased.
4) the output characteristic pattern of the last one multi-connection block structure is inputted into 1 × 1 convolutional layer, reduces the dimension of characteristic pattern
Number increases non-linear.
5) skip floor is used, the characteristic pattern after the characteristic pattern and 1 × 1 convolutional layer dimensionality reduction of the output of 7 × 7 convolutional layers is stacked
Get up, collectively as the input of amplification module.
6) method for using low resolution characteristic pattern to be spliced into high-resolution features figure in order, by low resolution characteristic pattern
It is amplified to certain multiple.For example, being H by height, width W, port number are C × r2Low resolution characteristic pattern (H × W × C
×r2) amplification r times, obtain highly be rH, the high-resolution characteristic pattern (rH × rW × C) that width rW, port number are C.
7) it uses 1 × 1 convolutional layer that characteristic pattern is limited to tri- channels RGB, and uses Tanh activation primitives.
Three, model training
1) learning rate is set as 10-3, and every 75000 iterative learning rates reduce half, no longer reduce until error or
After 250000 iteration, deconditioning.
2) use L1 norms as loss function.
3) ADAM optimization methods, β are used1=0.9, β2=0.999.
Four, image reconstruction
The low-resolution image for inputting arbitrary size, loads trained model, exports the high-definition picture of reconstruction.
Embodiment one
One, data preparation:
1) data set is divided.Using public data collection DIV2K, which includes 800 training figures, 100 proof diagrams
With 100 test charts.Since the high-definition picture of test chart is not open, 100 proof diagrams are rebuild for testing
Effect.Wherein, it is all made of high-definition picture and its corresponding low-resolution image for training and the image tested, it is low
Resolution chart seems that high-definition picture degrades what model generated via certain specific.
2) the non-overlapping image block for being divided into 96 × 96 sizes is schemed into 800 training, as network inputs.
Two, network structure is built:
The network structure of the present invention is mainly made of N number of multi-connection block structure, an amplification module and a skip floor.Below
Explanation will be described in detail in conjunction with Fig. 1, the network structure built to the present invention.Fig. 1 is is proposed based on improved thick
The network overall structure figure of the super resolution ratio reconstruction method of close convolutional neural networks, " C " indicates characteristic pattern in channel dimension in figure
Stacking.
1) each iteration inputs the low-resolution image block of 16 96 × 96 sizes.Low-resolution image block first passes around
One 7 × 7 convolutional layer extracts 64 characteristic patterns, and using ReLU (Rectified Linear Units) as activation letter
Number.
2) 64 characteristic patterns input multi-connection block structure.Wherein, multi-connection block structure is by a sub- block structure (shown in Fig. 2)
It is formed with an identical mapping.Based on the thought of dense convolutional neural networks structure, using whole shallow-layer features as further feature
The shallow-layers feature such as texture, profile is recycled in input when extraction in network training.The benchmark sub-block knot as shown in Fig. 2 (a)
Structure, the present invention connects the structure of 3 × 3 convolutional layers using 1 × 1 convolutional layer, as the benchmark architecture of sub-block, and each convolutional layer
Using ReLU as activation primitive.On the basis of benchmark architecture, the present invention increases by 3 × 3 convolutional layers, increases sub-block knot
The receptive field of structure and non-linear improves sub-block structure as the improved structure of sub-block as shown in Fig. 2 (b).Finally, such as Fig. 2 (c)
Shown fusant block structure, the present invention stack the benchmark architecture of sub-block and improved structure in characteristic pattern channel dimension, make
For the fusion structure of sub-block.Wherein, in order to keep output characteristic pattern number identical, each branch characteristic pattern in sub-block fusion structure
Number be benchmark architecture and improved structure characteristic pattern number half.All include an input in each multi-connection block structure
To the identical mapping of output, the input directly by the output of a upper multi-connection block structure as next multi-connection block structure,
To strengthen the recycling of shallow-layer feature, the complexity and diversity of feature are increased.In the present invention, multi-connection characteristic block heap
16 are folded.
3) it is to reduce the characteristic pattern number for stacking multi-connection block structure, by the output feature of the last one multi-connection block structure
Figure one 1 × 1 convolutional layer of input obtains 256 output characteristic patterns.
4) it is further to utilize shallow-layer feature, local feature and global characteristics is fully combined, present invention employs
One skip floor stacks 256 characteristic patterns exported after 64 characteristic patterns and 1 × 1 convolutional layer dimensionality reduction of the output of 7 × 7 convolutional layers
Get up, then by this 320 characteristic patterns collectively as the input of amplification module.
5) method that the present invention is spliced into high-resolution features figure in order using low resolution characteristic pattern, by low resolution
Characteristic pattern is amplified to corresponding multiple.As shown in figure 3, being H by height, width W, port number are C × r2Low resolution it is special
Sign figure (H × W × C × r2) r times of amplification, to obtain be highly rH, high-resolution features figure that width rW, port number are C (rH ×
rW×C).The case where exemplary in Fig. 3 is r=2, C=1, Fig. 3 (a) are low resolution characteristic patterns, and Fig. 3 (b) is high-resolution spy
Sign figure.
6) finally, characteristic pattern is limited to tri- channels RGB by the present invention using 1 × 1 convolutional layer, and is swashed using Tanh
Function living, output are the high-definition picture that super-resolution rebuilding comes out.
Three, model training:
The learning rate of the present invention is set as 10-3, and every 75000 iterative learning rates reduce half, until error no longer subtracts
After small or 250000 iteration, deconditioning.Using L1 norms as loss function.Using ADAM optimization methods, β1=
0.9, β2=0.999.
Four, image reconstruction:
The low-resolution image for inputting arbitrary size, loads trained model, exports the high-definition picture of reconstruction.
Evaluation index:Present invention experiment is weighed using Y-PSNR (Peak Signal to Noise Ratio, PSNR)
Quantity algorithm effect, PSNR are defined as,
Wherein, f (x, y) and f'(x, y) artwork and reconstruction figure are respectively represented, M × N represents spatial resolution.Higher
PSNR values represent better reconstruction effect.On DIV2K data sets, it is as shown in table 1 that the present invention tests specific PSNR values.
The PSNR values of the different sub-block structures of 1 present invention of table compare (unit:dB)
Although the function and the course of work of the present invention are described above in conjunction with attached drawing, the invention is not limited in
Above-mentioned concrete function and the course of work, the above mentioned embodiment is only schematical, rather than restrictive, ability
The those of ordinary skill in domain under the inspiration of the present invention, is not departing from present inventive concept and scope of the claimed protection situation
Under, many forms can also be made, all of these belong to the protection of the present invention.
Claims (2)
1. a kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks, which is characterized in that including following step
Suddenly:
Step 1, data preparation
1) data set is divided:Using public data collection DIV2K, which includes 800 training figures, 100 proof diagrams and 100
Open test chart, wherein 100 proof diagrams rebuild effect for testing, and 800 training scheme and 100 proof diagrams are all by high score
Resolution image and its corresponding low-resolution image composition, low-resolution image are that high-definition picture is generated via the model that degrades
's;
2) the non-overlapping image block for being divided into 96 × 96 sizes is schemed into 800 training, as network inputs;
Step 2, network structure are built
1) the low-resolution image block of 96 × 96 sizes is inputted to one 7 × 7 convolutional layer, and using ReLU (Rectified
Linear Units) it is used as activation primitive;
2) the output characteristic pattern of 7 × 7 convolutional layers inputs multi-connection block structure, wherein each multi-connection block structure includes one
Sub-block structure and an identical mapping;
3) multi-connection block structure is stacked;
4) the output characteristic pattern of the last one multi-connection block structure is inputted into 1 × 1 convolutional layer, reduces the dimension of characteristic pattern;
5) skip floor is used, the characteristic pattern heap after the characteristic pattern and 1 × 1 convolutional layer dimensionality reduction of the output of 7 × 7 convolutional layers is stacked
Come, collectively as the input of amplification module;
6) method for using low resolution characteristic pattern to be spliced into high-resolution features figure in order, low resolution characteristic pattern is amplified
To certain multiple;
7) it uses 1 × 1 convolutional layer that characteristic pattern is limited to tri- channels RGB, and uses Tanh activation primitives;
Step 3, model training
1) learning rate is set as 10-3, and every 75000 iterative learning rates reduce half, no longer reduce until error or
After 250000 iteration, deconditioning;
2) use L1 norms as loss function;
3) ADAM optimization methods, β are used1=0.9, β2=0.999;
Step 4, image reconstruction
The low-resolution image for inputting arbitrary size, loads trained model, exports the high-definition picture of reconstruction.
2. the super resolution ratio reconstruction method according to claim 1 based on improved dense convolutional neural networks, feature
It is, the building process of sub-block structure described in step 2:First, the thought based on dense convolutional neural networks structure, using 1
× 1 convolutional layer connects the structure of 3 × 3 convolutional layers, and as the benchmark architecture of sub-block, and each convolutional layer is used as using ReLU and is swashed
Function living;Then, on the basis of benchmark architecture, increase by 3 × 3 convolutional layers, the improved structure as sub-block structure;Most
Afterwards, the benchmark architecture of sub-block structure and improved structure are stacked in characteristic pattern channel dimension, the fusion as sub-block structure
Structure.
The definition of the identical mapping:In each multi-connection block structure, all comprising one from the connection for being input to output, directly will
Input of the output of a upper multi-connection block structure as next multi-connection block structure.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810535634.5A CN108765290A (en) | 2018-05-29 | 2018-05-29 | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810535634.5A CN108765290A (en) | 2018-05-29 | 2018-05-29 | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108765290A true CN108765290A (en) | 2018-11-06 |
Family
ID=64003877
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810535634.5A Pending CN108765290A (en) | 2018-05-29 | 2018-05-29 | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108765290A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109726657A (en) * | 2018-12-21 | 2019-05-07 | 万达信息股份有限公司 | A kind of deep learning scene text recognition sequence method |
CN109903223A (en) * | 2019-01-14 | 2019-06-18 | 北京工商大学 | A kind of image super-resolution method based on dense connection network and production confrontation network |
CN110163802A (en) * | 2019-05-20 | 2019-08-23 | 电子科技大学 | A kind of SAR image ultra-resolution method neural network based |
CN110189308A (en) * | 2019-05-17 | 2019-08-30 | 山东财经大学 | A kind of lesion detection approach and device based on BM3D and the fusion of dense convolutional network |
CN110322402A (en) * | 2019-04-30 | 2019-10-11 | 武汉理工大学 | Medical image super resolution ratio reconstruction method based on dense mixing attention network |
CN110738231A (en) * | 2019-07-25 | 2020-01-31 | 太原理工大学 | Method for classifying mammary gland X-ray images by improving S-DNet neural network model |
CN111353940A (en) * | 2020-03-31 | 2020-06-30 | 成都信息工程大学 | Image super-resolution reconstruction method based on deep learning iterative up-down sampling |
CN112308119A (en) * | 2020-10-15 | 2021-02-02 | 中国医学科学院北京协和医院 | Immunofluorescence classification method and device for glomerulonephritis |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106228512A (en) * | 2016-07-19 | 2016-12-14 | 北京工业大学 | Based on learning rate adaptive convolutional neural networks image super-resolution rebuilding method |
CN106991646A (en) * | 2017-03-28 | 2017-07-28 | 福建帝视信息科技有限公司 | A kind of image super-resolution method based on intensive connection network |
CN107221013A (en) * | 2017-05-16 | 2017-09-29 | 山东财经大学 | One kind is based on variation light stream estimation lung 4D CT Image Super Resolution Processing methods |
CN107358575A (en) * | 2017-06-08 | 2017-11-17 | 清华大学 | A kind of single image super resolution ratio reconstruction method based on depth residual error network |
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 |
EP3319039A1 (en) * | 2016-11-07 | 2018-05-09 | UMBO CV Inc. | A method and system for providing high resolution image through super-resolution reconstruction |
-
2018
- 2018-05-29 CN CN201810535634.5A patent/CN108765290A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106228512A (en) * | 2016-07-19 | 2016-12-14 | 北京工业大学 | Based on learning rate adaptive convolutional neural networks image super-resolution rebuilding method |
EP3319039A1 (en) * | 2016-11-07 | 2018-05-09 | UMBO CV Inc. | A method and system for providing high resolution image through super-resolution reconstruction |
CN106991646A (en) * | 2017-03-28 | 2017-07-28 | 福建帝视信息科技有限公司 | A kind of image super-resolution method based on intensive connection network |
CN107221013A (en) * | 2017-05-16 | 2017-09-29 | 山东财经大学 | One kind is based on variation light stream estimation lung 4D CT Image Super Resolution Processing methods |
CN107358575A (en) * | 2017-06-08 | 2017-11-17 | 清华大学 | A kind of single image super resolution ratio reconstruction method based on depth residual error network |
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 |
Non-Patent Citations (2)
Title |
---|
JINGHUI CHU等: "A Novel Multi-Connected Convolutional Network for Super-Resolution", 《IEEE SIGNAL PROCESSING LETTERS》 * |
LIM B等: "Enhanced deep residual networks for single image super-resolution", 《IN THE IEEE CONFERENCE ON COMPUTER VISION AND PATTERN》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109726657A (en) * | 2018-12-21 | 2019-05-07 | 万达信息股份有限公司 | A kind of deep learning scene text recognition sequence method |
CN109726657B (en) * | 2018-12-21 | 2023-06-09 | 万达信息股份有限公司 | Deep learning scene text sequence recognition method |
CN109903223A (en) * | 2019-01-14 | 2019-06-18 | 北京工商大学 | A kind of image super-resolution method based on dense connection network and production confrontation network |
CN109903223B (en) * | 2019-01-14 | 2023-08-25 | 北京工商大学 | Image super-resolution method based on dense connection network and generation type countermeasure network |
CN110322402A (en) * | 2019-04-30 | 2019-10-11 | 武汉理工大学 | Medical image super resolution ratio reconstruction method based on dense mixing attention network |
CN110189308A (en) * | 2019-05-17 | 2019-08-30 | 山东财经大学 | A kind of lesion detection approach and device based on BM3D and the fusion of dense convolutional network |
CN110163802A (en) * | 2019-05-20 | 2019-08-23 | 电子科技大学 | A kind of SAR image ultra-resolution method neural network based |
CN110163802B (en) * | 2019-05-20 | 2020-09-01 | 电子科技大学 | SAR image super-resolution method based on neural network |
CN110738231B (en) * | 2019-07-25 | 2022-12-27 | 太原理工大学 | Method for classifying mammary gland X-ray images by improving S-DNet neural network model |
CN110738231A (en) * | 2019-07-25 | 2020-01-31 | 太原理工大学 | Method for classifying mammary gland X-ray images by improving S-DNet neural network model |
CN111353940A (en) * | 2020-03-31 | 2020-06-30 | 成都信息工程大学 | Image super-resolution reconstruction method based on deep learning iterative up-down sampling |
CN112308119B (en) * | 2020-10-15 | 2021-11-05 | 中国医学科学院北京协和医院 | Immunofluorescence classification method and device for glomerulonephritis |
CN112308119A (en) * | 2020-10-15 | 2021-02-02 | 中国医学科学院北京协和医院 | Immunofluorescence classification method and device for glomerulonephritis |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108765290A (en) | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks | |
CN108765291A (en) | Super resolution ratio reconstruction method based on dense neural network and two-parameter loss function | |
CN107240066A (en) | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks | |
Zhong et al. | Joint sub-bands learning with clique structures for wavelet domain super-resolution | |
CN109903223B (en) | Image super-resolution method based on dense connection network and generation type countermeasure network | |
CN110197468A (en) | A kind of single image Super-resolution Reconstruction algorithm based on multiple dimensioned residual error learning network | |
CN109903228A (en) | A kind of image super-resolution rebuilding method based on convolutional neural networks | |
CN109064396A (en) | A kind of single image super resolution ratio reconstruction method based on depth ingredient learning network | |
CN110751597B (en) | Video super-resolution method based on coding damage repair | |
CN110084745A (en) | Image super-resolution rebuilding method based on dense convolutional neural networks in parallel | |
CN106709875A (en) | Compressed low-resolution image restoration method based on combined deep network | |
CN111861961A (en) | Multi-scale residual error fusion model for single image super-resolution and restoration method thereof | |
KR102289045B1 (en) | Method and Apparatus for Providing Multi-scale Object Image Restoration | |
CN108805808A (en) | A method of improving video resolution using convolutional neural networks | |
CN109934771B (en) | Unsupervised remote sensing image super-resolution reconstruction method based on recurrent neural network | |
CN106920214A (en) | Spatial target images super resolution ratio reconstruction method | |
CN111951164B (en) | Image super-resolution reconstruction network structure and image reconstruction effect analysis method | |
CN111340696B (en) | Convolutional neural network image super-resolution reconstruction method fused with bionic visual mechanism | |
CN107590775B (en) | Image super-resolution amplification method using regression tree field | |
CN112288632A (en) | Single image super-resolution method and system based on simplified ESRGAN | |
CN115100039B (en) | Lightweight image super-resolution reconstruction method based on deep learning | |
CN111402138A (en) | Image super-resolution reconstruction method of supervised convolutional neural network based on multi-scale feature extraction fusion | |
CN110163803A (en) | A kind of image super-resolution rebuilding method and system based on convolutional neural networks | |
CN112017116A (en) | Image super-resolution reconstruction network based on asymmetric convolution and construction method thereof | |
CN115953294A (en) | Single-image super-resolution reconstruction method based on shallow channel separation and aggregation |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181106 |