CN110009568A - The generator construction method of language of the Manchus image super-resolution rebuilding - Google Patents
The generator construction method of language of the Manchus image super-resolution rebuilding Download PDFInfo
- Publication number
- CN110009568A CN110009568A CN201910286781.8A CN201910286781A CN110009568A CN 110009568 A CN110009568 A CN 110009568A CN 201910286781 A CN201910286781 A CN 201910286781A CN 110009568 A CN110009568 A CN 110009568A
- Authority
- CN
- China
- Prior art keywords
- pixels
- resolution
- image
- convolutional layer
- conv
- 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
- 238000010276 construction Methods 0.000 title claims abstract description 8
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims description 27
- 230000009466 transformation Effects 0.000 claims description 19
- 238000001914 filtration Methods 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 9
- 238000013507 mapping Methods 0.000 abstract description 2
- 238000012545 processing Methods 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 24
- 238000000034 method Methods 0.000 description 17
- 238000013461 design Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 210000002569 neuron Anatomy 0.000 description 4
- 239000000203 mixture Substances 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000005215 recombination Methods 0.000 description 3
- 230000006798 recombination Effects 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
- G06T3/4076—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution using the original low-resolution images to iteratively correct the high-resolution images
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The generator construction method of language of the Manchus image super-resolution rebuilding, belong to Computer Image Processing field, in order to solve the generator Construct question that low resolution language of the Manchus image carries out super-resolution rebuilding, use 5 mutually isostructural residual blocks and 2 built-up generators of sub-pix convolutional layer, it can learn mapping relations between the language of the Manchus image of high-low resolution, to carry out super-resolution rebuilding to low resolution language of the Manchus image.
Description
Technical field
The invention belongs to Computer Image Processing fields, and in particular to a kind of generator of language of the Manchus image super-resolution rebuilding
Construction method.
Background technique
Image super-resolution rebuilding (Super Resolution Reconstruction, SRR) technology, refers to and is not mentioning
In the case where rising hardware performance, solved the problems, such as using pure technological means because of reasons such as hardware performance limitation or acquisition targets itself
Caused by the fuzzy problem of pictorial information.In brief, SRR technology can be in the case where not promoting acquisition equipment performance to figure
It is more to obtain from low resolution (Low Resolution) to the reconstruction of high-resolution (High Resolution) as carrying out
Pictorial information.
Traditional image rebuilding method, such as the image rebuilding method based on interpolation, the image rebuilding method based on reconstruction
And it is based partially on the image rebuilding method (SRCNN etc.) of study.Although the result that these methods generate is all with higher
The objective evaluatings index such as PSNR, but its result generated excessively smoothly causes the image detail information for lacking some keys mostly.
Especially in the super resolution task of language of the Manchus file and picture, the language of the Manchus text details information of some keys is ignored to be generated
Very big ambiguity.
Summary of the invention
In order to solve the generator Construct question that low resolution language of the Manchus image carries out super-resolution rebuilding, so that document map
As that can have better details to show, the following technical solutions are proposed: a kind of life of language of the Manchus image super-resolution rebuilding by the present invention
It grows up to be a useful person construction method, uses 5 mutually isostructural residual blocks and 2 built-up generators of sub-pix convolutional layer.
Further, generator structure is:
1st operation is Input input layer, is low resolution RGB triple channel image in training data for input picture;
2nd operation is G-Conv-1 layers, is convolutional layer, and convolution kernel is 9 pixels × 9 pixels, and 1 pixel of step-length includes 64
Filter;
3rd operation is PReLu layers, and G-Conv-1 layers of input signal is carried out nonlinear transformation,;
4-8 operation is Residual block residual block, and five operations are the identical Residual of 5 structures
Block residual block is extracted for the pictorial information feature to low-resolution image;
9th operation includes G-Conv-2 convolutional layer, BN operation, Sum operation, and wherein the convolution kernel of G-Conv-2 convolutional layer is
3 pixels × 3 pixels, 1 pixel of step-length include 64 filters, and BN indicates batch normalization operation, and Sum indicates output summation;
10th operation includes G-Conv-3 convolutional layer, Sub-Pixel CN sub-pix convolutional layer, PReLu layers, wherein G-
The convolution kernel of Conv-3 convolutional layer is 3 pixels × 3 pixels, and 1 pixel of step-length includes 256 filters;The Asia Sub-Pixel CN picture
Plain convolutional layer has 2 layers, generates high-definition picture for being recombinated to the low-resolution image feature extracted,
Upper one layer of input signal is carried out nonlinear transformation by PReLu layers;
It is convolutional layer that 11st operation, which is G-Conv-4, and convolution kernel is 9 pixels × 9 pixels, and 1 pixel of step-length includes 3 filtering
Device;
12nd operation is Output output layer.
The utility model has the advantages that
Generator has mainly used 5 mutually isostructural residual blocks (Residual block) and 2 sub-pix convolutional layers
It is built-up.Wherein the main function of residual block is extracted to the pictorial information feature of low-resolution image.And sub-pix
The effect of convolutional layer is to carry out recombination to the low-resolution image feature extracted to generate high-definition picture.
Detailed description of the invention
Fig. 1 language of the Manchus file and picture super-resolution model structure.
Specific embodiment
A kind of language of the Manchus image super-resolution rebuilding method based on generation confrontation network, the method is successively by preparing to instruct
Practice sample, language of the Manchus image super-resolution model is realized based on generation confrontation network establishment, utilizes training sample and loss function pair
Model is adjusted three step compositions.
Make description below for resolution ratio to illustrate: the height of resolution ratio is not stringent boundary and definition, Gao He
Low is all an opposite concept.In experiment of the invention, high resolution graphics seems the figure of 1200 × 800 pixels or so
Picture, low resolution are that high-definition picture is reduced into 1/4 original size, i.e. 300 × 200 pixels or so.While in order to keep
The scalability of model, input picture can be arbitrary size (i.e. resolution ratio), and not to high-definition picture and low
Image in different resolution carries out particularly severe limitation.
Wherein:
The step of preparing training sample are as follows: scanning language of the Manchus document obtains high-resolution language of the Manchus file and picture HR(1~n),
Middle n is the quantity for the high-resolution language of the Manchus image that scanning obtains.Respectively by obtained high-resolution language of the Manchus file and picture using slotting
It is low-resolution image LR that value-based algorithm, which carries out down-sampling,(1~n).By high-resolution language of the Manchus file and picture HR(1~n)With low resolution
Image LR(1~n)Correspond the training dataset of composition model.
Based on the step of generating confrontation network establishment realization language of the Manchus image super-resolution model are as follows: firstly generate the building of device
It has mainly used 5 mutually isostructural residual blocks (Residual block) and 2 sub-pix convolutional layers built-up, has had
Shown in the structural schematic diagram of body such as Fig. 1 (a).If Fig. 1, Input are input layer, input picture is low resolution in training data
RGB triple channel image;G-Conv-1 layers are convolutional layers, and convolution kernel is 9 pixels × 9 pixels, and 1 pixel of step-length includes 64 filtering
Device;Upper one layer of input signal is carried out nonlinear transformation by PReLu layers, followed by the identical Residual of 5 structures
Block is residual block;G-Conv-2 convolutional layer, convolution kernel are 3 pixels × 3 pixels, and 1 pixel of step-length includes 64 filters;BN
Indicate batch normalization operation;Sum indicates summation;G-Conv-3 convolutional layer, convolution kernel be 3 pixels × 3 pixels, 1 pixel of step-length,
Include 256 filters;Sub-Pixel CN is sub-pix convolutional layer, and × 2 indicate 2 sub-pix convolutional layers;PReLu layers will be upper
One layer of input signal carries out nonlinear transformation;G-Conv-4 is convolutional layer, and convolution kernel is 9 pixels × 9 pixels, 1 pixel of step-length,
Include 3 filters;Output is output layer.
Wherein:
1. the main function of residual block is extracted to the pictorial information feature of low-resolution image, Residual
In block, G-Conv-2 is convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 1 pixel of step-length includes 64 filters;BN is indicated
Batch normalization operation;
2. upper one layer of input signal is carried out nonlinear transformation by PReLu layers, as shown in formula (1):
Wherein xiIt is the input of function, ai changes with trained process;
3. the effect of sub-pix convolutional layer is to carry out recombination to the low-resolution image feature extracted to generate high score
Resolution image, the essence of sub-pix convolutional layer are exactly that low resolution feature according to specific position, is periodically inserted into high score
In resolution image, basic principle is as shown in Figure 1.
" by upper one layer of input signal " addressed makees following understanding, and having sub-pix for Fig. 1, in figure also has commonly
Convolutional layer, naturally it is also possible to be other layer.As long as the output of that layer before PReLu is exactly the defeated of PReLu operation
Enter.
The arbiter part of model is generated in confrontation, using the network structure of the VGG-19 of pre-training, the tool of arbiter
Shown in body structure such as Fig. 1 (b), effect is identified to the image of input.If Fig. 1 (b), Input are input layer, input figure
As being training set middle high-resolution sample;D-Conv-1 is convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 1 pixel of step-length includes
64 filters;Upper one layer of input signal is carried out nonlinear transformation by ReLu layers of Leaky;D-Conv-2 is convolutional layer, convolution
Core is 3 pixels × 3 pixels, and 2 pixel of step-length includes 64 filters;D-Conv-3 is convolutional layer, and convolution kernel is 3 pixels × 3 pictures
Element, 1 pixel of step-length include 128 filters;D-Conv-4 is convolutional layer, and convolution kernel is 3 pixels × 3 pixels, 2 pixel of step-length,
Include 128 filters;D-Conv-5 is convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 1 pixel of step-length includes 256 filtering
Device;D-Conv-6 is convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 2 pixel of step-length includes 256 filters;D-Conv-7 is
Convolutional layer, convolution kernel are 3 pixels × 3 pixels, and 1 pixel of step-length includes 512 filters;D-Conv-8 is convolutional layer, convolution kernel
For 3 pixels × 3 pixels, 2 pixel of step-length includes 512 filters;BN is batch normalization operation;ReLu layers of Leaky will be upper
One layer of input signal carries out nonlinear transformation;Dense layers contain 1024 neuron numbers;ReLu layers of Leaky by upper one layer
Input signal carries out nonlinear transformation;Dense layers contain 1 neuron;Sigmoid function σ (z)=1/ (1+e-z), wherein z table
Show one layer of output;Output output is that arbiter is determined as authentic specimen to input or generates the probability of sample.
Wherein, it as shown in structure in Fig. 1 dotted line frame, in this 6 layers of structures of the D-Conv-n (2≤n≤7) of arbiter D, removes
It also include BN batch normalization operation except convolutional layer D-Conv-n;ReLu layers of Leaky by upper one layer of input signal into
Row nonlinear transformation, as shown in formula (2):
Wherein xiIt is the input of function, it is a variable that ai, which is coefficient, and when fxi is less than 0, (input-input is exhausted by ai=
To value) × 0.5.
Identification result is fed back into generator G, to promote the optimization of generator G, generator G is promoted to generate the height of high quality
Resolution ratio language of the Manchus file and picture.
The step of being adjusted using training sample to model is as follows: by low-resolution image in ready training sample
LR(1~n)Super-resolution reconstruction image SR is generated using generator(1~n), then by the SR of generation(1~n)In image and training sample
High-definition picture HR(1~n)Input arbiter network.The loss function pair of the loss function of generator and arbiter is utilized simultaneously
Model carries out tuning, completes the training to model.
In one embodiment:
Implementation steps of the invention include Generator Design, the design of arbiter, the design of loss function, the training of network
And it is discussed in detail for the use of four using this.
1, the design of generator
Generator has mainly used 5 mutually isostructural residual blocks (Residual block) and 2 sub-pix convolutional layers
It is built-up.Wherein the main function of residual block is extracted to the pictorial information feature of low-resolution image.And sub-pix
The effect of convolutional layer is to carry out recombination to the low-resolution image feature extracted to generate high-definition picture.Specific knot
Structure is as shown in Fig. 1 (a), wherein the detail parameters of each convolutional layer are as shown in table 1:
Each layer parameter in 1 generator G of table
2, the design of arbiter
The effect of arbiter is identified to the image of input.Identification result is fed back into generator G, to promote to generate
The optimization of device G.Shown in its detailed construction such as Fig. 1 (b), wherein the detail parameters of each convolutional layer are as shown in table 2
Convolution layer parameter in 2 arbiter D of table
3, the design of loss function
The loss function expression formula for the language of the Manchus file and picture Super-resolution reconstruction established model that the present invention realizes is as shown in Equation 1.
WhereinIndicate content loss,Indicate confrontation loss, the total losses function of model is the weighted sum of the two.
Content lossThe same layer in VGG network using calculating source high-definition picture and the high-definition picture of generation
The method of characteristic spectrum Euclidean distance calculates loss bring negative effect so as to avoid in pixel level, so that generating figure
As there is better details to show.The expression formula of content loss is as shown in Equation 2.
Wherein φi,jIndicate the feature that j-th of convolutional layer in 19 layers of VGG network model after i-th layer of maximum pond layer obtains
Map, Wi,jAnd Hi,jRespectively indicate the dimension of characteristic pattern.It implies that and generates image by VGG feature extraction, in pair with original image
Answer layer feature as close as possible, to guarantee the consistency of picture material.VGG model is Oxonian Oxford Visual
Geometry Group was in ILSVRC (ImageNet Large Scale Visual Recognition in 2014
Challenge) the model that contest proposes, because having good transportable property, at present in deep learning field as classical mould
Type is used by many algorithms.lSR VGG/i,jIndicate that the loss function of super-resolution model, subscript SR indicate super-resolution, subscript VGG
It indicating to use VGG model ,/i, j indicate j-th of convolutional layer after i-th layer of maximum pond layer of VGG model, soTable
What is shown is the loss function of j-th of convolutional layer after i-th layer of VGG model maximum pond layer.IHRIndicate high-definition picture.
Wi,jIndicate the width of characteristic pattern, unit is pixel;Hi,jIndicate the height of characteristic pattern, unit is pixel.
And fight lossAddition, generator can be motivated to be generated as far as possible with source high-definition picture details
It generates as a result, the definition of confrontation loss is performance based on all training samples on arbiter, fights the expression of loss function
Formula is as shown in Equation 3.
WhereinIt indicates to generate resultIt is considered as true high-definition picture in arbiter
Probability.Indicate the loss function of the generation model in super-resolution model, ILRIndicate low-resolution image,Indicate ginseng
Number is θGGenerator, thenIt indicates by generating modelAccording to low-resolution image ILRThe high resolution graphics of generation
Picture;Indicate arbiter network,Indicate that the high-definition picture for generating above-mentioned generator inputs arbiter
Output afterwards;N=1~N indicates the quantity of the low resolution of input.
4, the training and use of model
When being trained to model, training dataset, which is used, obtains HR image using interpolation algorithm 4 times of down-samplings of progress
LR image.Respectively using LR image and HR image as the input of generator and desired output, the SR image that generator is generated and
HR image inputs arbiter, and the output of arbiter is fed back to generator.It is right under the loss function constraint of model during this
The network weight of generator and arbiter optimizes, after the completion of model training.By low resolution language of the Manchus document map to be reconstructed
As input model can be obtained rebuild after high-resolution version,
The disclosed language of the Manchus file and picture super resolution ratio reconstruction method for focusing on image reconstruction details of the present embodiment, so that the language of the Manchus
File and picture Super-resolution reconstruction established model can be by 4 times of increase resolution of source document image.Compared to traditional reconstructing method, originally
Method can reconstruct many Key details lacked in low resolution language of the Manchus image to a certain extent, so that language of the Manchus document map
As having higher readability.
In another embodiment:
A kind of language of the Manchus image super-resolution rebuilding method based on generation confrontation network, includes the following steps:
S1. prepare training sample;
S2. language of the Manchus image super-resolution model is realized based on generation confrontation network establishment.
Further, described based on the language of the Manchus image super-resolution rebuilding method for generating confrontation network, further include
S3. model is adjusted by training sample and loss function.
Further, the step of preparing training sample are as follows:
Language of the Manchus document is scanned, and obtains high-resolution language of the Manchus file and picture HR(1~n), wherein n is the high score that scanning obtains
Obtained high-resolution language of the Manchus file and picture is carried out down-sampling using interpolation algorithm, and obtained by the quantity of resolution language of the Manchus image
Low-resolution image LR(1~n), wherein n is the quantity of low resolution language of the Manchus image, by high-resolution language of the Manchus file and picture HR(1 ~n)With low-resolution image LR(1~n)Correspond the training dataset of composition model.
Further, based on confrontation network establishment realization language of the Manchus image super-resolution model is generated the step of are as follows: building life
It grows up to be a useful person and arbiter.
Further, 5 mutually isostructural residual blocks and 2 sub-pix convolution layer buildings the building generator: are used
Into generator, generator structure is:
1st operation is Input input layer, is low resolution RGB triple channel image in training data for input picture;
2nd operation is G-Conv-1 layers, is convolutional layer, and convolution kernel is 9 pixels × 9 pixels, and 1 pixel of step-length includes 64
Filter;
3rd operation is PReLu layers, and G-Conv-1 layers of input signal is carried out nonlinear transformation,;
4-8 operation is Residual block residual block, and five operations are the identical Residual of 5 structures
Block residual block is extracted for the pictorial information feature to low-resolution image;
9th operation includes G-Conv-2 convolutional layer, BN operation, Sum operation, and wherein the convolution kernel of G-Conv-2 convolutional layer is
3 pixels × 3 pixels, 1 pixel of step-length include 64 filters, and BN indicates batch normalization operation, and Sum indicates output summation;
10th operation includes G-Conv-3 convolutional layer, Sub-Pixel CN sub-pix convolutional layer, PReLu layers, wherein G-
The convolution kernel of Conv-3 convolutional layer is 3 pixels × 3 pixels, and 1 pixel of step-length includes 256 filters;The Asia Sub-Pixel CN picture
Plain convolutional layer has 2 layers, generates high-definition picture for being recombinated to the low-resolution image feature extracted,
Upper one layer of input signal is carried out nonlinear transformation by PReLu layers;
It is convolutional layer that 11st operation, which is G-Conv-4, and convolution kernel is 9 pixels × 9 pixels, and 1 pixel of step-length includes 3 filtering
Device;
12nd operation is Output output layer.
Further, upper one layer of input signal is subjected to nonlinear transformation PReLu layers, by shown in formula (1):
Wherein xiIt is the input of function, ai is coefficient, is changed with trained process.
Further, the arbiter of building, structure are:
1st operation is Input input layer, and input picture is training set middle high-resolution sample;
It is convolutional layer that 2nd operation, which is D-Conv-1, and convolution kernel is 3 pixels × 3 pixels, and 1 pixel of step-length includes 64 filtering
Device;
3rd operation is ReLu layers of Leaky, and upper one layer of input signal is carried out nonlinear transformation;
4th operation is D-Conv-2 convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 2 pixel of step-length includes 64 filtering
Device;
5th operation is D-Conv-3 convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 1 pixel of step-length includes 128 filtering
Device;
6th operation is D-Conv-4 convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 2 pixel of step-length includes 128 filtering
Device;
7th operation is D-Conv-5 convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 1 pixel of step-length includes 256 filtering
Device;
8th operation is D-Conv-6 convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 2 pixel of step-length includes 256 filtering
Device;
9th operation is D-Conv-7 convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 1 pixel of step-length includes 512 filtering
Device;
10th operation is D-Conv-8 convolutional layer, and convolution kernel is 3 pixels × 3 pixels, and 2 pixel of step-length includes 512 filtering
Device;
It is batch normalization operation that 11st operation, which is BN,;
12nd operation is ReLu layers of Leaky and upper one layer of input signal is carried out nonlinear transformation;
13rd operation is Dense layers and contains 1024 neuron numbers;ReLu layers of Leaky by upper one layer of input signal into
Row nonlinear transformation;
14th operation is Dense layers, contains 1 neuron;
15th operation is Sigmoid function:
σ (z)=1/ (1+e-z)
Wherein z indicates one layer of output;
16th operation is output output, and the output is that arbiter is determined as authentic specimen to input or generates sample
Probability, be identification result.
Further, in the D-Conv-n structure of arbiter D, 2≤n≤7, including convolutional layer D-Conv-n, BN batch
ReLu layers of normalization operation, Leaky, ReLu layers of Leaky are that upper one layer of input signal is carried out nonlinear transformation;
ReLu layers of Leaky are that upper one layer of input signal is carried out nonlinear transformation, as shown in formula (2):
Wherein xiIt is the input of function, wherein a=0.2.
Further, identification result is fed back into generator G, promotes the optimization of generator G, promote generator G to generate high
The high-resolution language of the Manchus file and picture of quality.
Further, the step of being adjusted by training sample to model is as follows: by low point in ready training sample
Resolution image LR(1~n), super-resolution reconstruction image SR is generated using generator(1~n), n is the quantity of reconstruction image, then will be given birth to
At figure rebuild as SR(1~n)With training sample middle high-resolution image HR(1~n)Arbiter network is inputted, while utilizing generator
Loss function and the loss function of arbiter tuning is carried out to model, complete training to model.
The present invention carries out the super-resolution rebuilding of language of the Manchus file and picture using the method for deep learning, can learn height point
Mapping relations between the language of the Manchus image of resolution, to carry out super-resolution rebuilding to low resolution language of the Manchus image.
Present invention utilizes the characteristics of production confrontation this generation model of network, so that the high-resolution language of the Manchus text rebuild
Shelves image is showed with better details.
The preferable specific embodiment of the above, only the invention, but the protection scope of the invention is not
It is confined to this, anyone skilled in the art is in the technical scope that the invention discloses, according to the present invention
The technical solution of creation and its inventive concept are subject to equivalent substitution or change, should all cover the invention protection scope it
It is interior.
Claims (2)
1. a kind of generator construction method of language of the Manchus image super-resolution rebuilding, it is characterised in that: mutually isostructural residual using 5
Poor block and 2 built-up generators of sub-pix convolutional layer.
2. the generator construction method of language of the Manchus image super-resolution rebuilding as described in claim 1, it is characterised in that:
Generator structure is:
1st operation is Input input layer, is low resolution RGB triple channel image in training data for input picture;
2nd operation is G-Conv-1 layers, is convolutional layer, and convolution kernel is 9 pixels × 9 pixels, and 1 pixel of step-length includes 64 filtering
Device;
3rd operation is PReLu layers, and G-Conv-1 layers of input signal is carried out nonlinear transformation,;
4-8 operation is Residual block residual block, and five operations are the identical Residual block of 5 structures
Residual block is extracted for the pictorial information feature to low-resolution image;
9th operation includes G-Conv-2 convolutional layer, BN operation, Sum operation, and wherein the convolution kernel of G-Conv-2 convolutional layer is 3 pictures
Element × 3 pixels, 1 pixel of step-length include 64 filters, and BN indicates batch normalization operation, and Sum indicates output summation;
10th operation includes G-Conv-3 convolutional layer, Sub-Pixel CN sub-pix convolutional layer, PReLu layers, wherein G-Conv-3
The convolution kernel of convolutional layer is 3 pixels × 3 pixels, and 1 pixel of step-length includes 256 filters;Sub-Pixel CN sub-pix convolution
Layer has 2 layers, generates high-definition picture for being recombinated to the low-resolution image feature extracted, and PReLu layers will
Upper one layer of input signal carries out nonlinear transformation;
It is convolutional layer that 11st operation, which is G-Conv-4, and convolution kernel is 9 pixels × 9 pixels, and 1 pixel of step-length includes 3 filters;
12nd operation is Output output layer.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910286781.8A CN110009568A (en) | 2019-04-10 | 2019-04-10 | The generator construction method of language of the Manchus image super-resolution rebuilding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910286781.8A CN110009568A (en) | 2019-04-10 | 2019-04-10 | The generator construction method of language of the Manchus image super-resolution rebuilding |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110009568A true CN110009568A (en) | 2019-07-12 |
Family
ID=67170956
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910286781.8A Pending CN110009568A (en) | 2019-04-10 | 2019-04-10 | The generator construction method of language of the Manchus image super-resolution rebuilding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110009568A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111275108A (en) * | 2020-01-20 | 2020-06-12 | 国网山东省电力公司枣庄供电公司 | Method for performing sample expansion on partial discharge data based on generation countermeasure network |
CN112381720A (en) * | 2020-11-30 | 2021-02-19 | 黑龙江大学 | Construction method of super-resolution convolutional neural network model |
Citations (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107154023A (en) * | 2017-05-17 | 2017-09-12 | 电子科技大学 | Face super-resolution reconstruction method based on generation confrontation network and sub-pix convolution |
US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial 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 |
US20180268284A1 (en) * | 2017-03-15 | 2018-09-20 | Samsung Electronics Co., Ltd. | System and method for designing efficient super resolution deep convolutional neural networks by cascade network training, cascade network trimming, and dilated convolutions |
CN108734659A (en) * | 2018-05-17 | 2018-11-02 | 华中科技大学 | A kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label |
CN108921783A (en) * | 2018-06-01 | 2018-11-30 | 武汉大学 | A kind of satellite image super resolution ratio reconstruction method based on losses by mixture function constraint |
CN109035142A (en) * | 2018-07-16 | 2018-12-18 | 西安交通大学 | A kind of satellite image ultra-resolution method fighting network integration Aerial Images priori |
CN109064407A (en) * | 2018-09-13 | 2018-12-21 | 武汉大学 | Intensive connection network image super-resolution method based on multi-layer perception (MLP) layer |
CN109087243A (en) * | 2018-06-29 | 2018-12-25 | 中山大学 | A kind of video super-resolution generation method generating confrontation network based on depth convolution |
CN109190722A (en) * | 2018-08-06 | 2019-01-11 | 大连民族大学 | Font style based on language of the Manchus character picture migrates transform method |
CN109509152A (en) * | 2018-12-29 | 2019-03-22 | 大连海事大学 | A kind of image super-resolution rebuilding method of the generation confrontation network based on Fusion Features |
US20190095795A1 (en) * | 2017-03-15 | 2019-03-28 | Samsung Electronics Co., Ltd. | System and method for designing efficient super resolution deep convolutional neural networks by cascade network training, cascade network trimming, and dilated convolutions |
-
2019
- 2019-04-10 CN CN201910286781.8A patent/CN110009568A/en active Pending
Patent Citations (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
US20180268284A1 (en) * | 2017-03-15 | 2018-09-20 | Samsung Electronics Co., Ltd. | System and method for designing efficient super resolution deep convolutional neural networks by cascade network training, cascade network trimming, and dilated convolutions |
US20190095795A1 (en) * | 2017-03-15 | 2019-03-28 | Samsung Electronics Co., Ltd. | System and method for designing efficient super resolution deep convolutional neural networks by cascade network training, cascade network trimming, and dilated convolutions |
CN107154023A (en) * | 2017-05-17 | 2017-09-12 | 电子科技大学 | Face super-resolution reconstruction method based on generation confrontation network and sub-pix convolution |
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 |
CN108734659A (en) * | 2018-05-17 | 2018-11-02 | 华中科技大学 | A kind of sub-pix convolved image super resolution ratio reconstruction method based on multiple dimensioned label |
CN108921783A (en) * | 2018-06-01 | 2018-11-30 | 武汉大学 | A kind of satellite image super resolution ratio reconstruction method based on losses by mixture function constraint |
CN109087243A (en) * | 2018-06-29 | 2018-12-25 | 中山大学 | A kind of video super-resolution generation method generating confrontation network based on depth convolution |
CN109035142A (en) * | 2018-07-16 | 2018-12-18 | 西安交通大学 | A kind of satellite image ultra-resolution method fighting network integration Aerial Images priori |
CN109190722A (en) * | 2018-08-06 | 2019-01-11 | 大连民族大学 | Font style based on language of the Manchus character picture migrates transform method |
CN109064407A (en) * | 2018-09-13 | 2018-12-21 | 武汉大学 | Intensive connection network image super-resolution method based on multi-layer perception (MLP) layer |
CN109509152A (en) * | 2018-12-29 | 2019-03-22 | 大连海事大学 | A kind of image super-resolution rebuilding method of the generation confrontation network based on Fusion Features |
Non-Patent Citations (3)
Title |
---|
吴洋洋等: "生成对抗网络的血管内超声图像超分辨率重建", 《南方医科大学学报》 * |
张晓阳等: "基于深度卷积网络的红外遥感图像超分辨率重建", 《黑龙江大学自然科学学报》 * |
韩森森: "深度学习在超分辨率图像重建中的应用", 《计算机时代》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111275108A (en) * | 2020-01-20 | 2020-06-12 | 国网山东省电力公司枣庄供电公司 | Method for performing sample expansion on partial discharge data based on generation countermeasure network |
CN112381720A (en) * | 2020-11-30 | 2021-02-19 | 黑龙江大学 | Construction method of super-resolution convolutional neural network model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109993702A (en) | Based on the language of the Manchus image super-resolution rebuilding method for generating confrontation network | |
CN110119780A (en) | Based on the hyperspectral image super-resolution reconstruction method for generating confrontation network | |
CN106683067A (en) | Deep learning super-resolution reconstruction method based on residual sub-images | |
CN106600538A (en) | Human face super-resolution algorithm based on regional depth convolution neural network | |
CN106127684A (en) | Image super-resolution Enhancement Method based on forward-backward recutrnce convolutional neural networks | |
CN109671022A (en) | A kind of picture texture enhancing super-resolution method based on depth characteristic translation network | |
CN109685716A (en) | A kind of image super-resolution rebuilding method of the generation confrontation network based on Gauss encoder feedback | |
CN109903236A (en) | Facial image restorative procedure and device based on VAE-GAN to similar block search | |
CN107341765A (en) | A kind of image super-resolution rebuilding method decomposed based on cartoon texture | |
CN109035146A (en) | A kind of low-quality image oversubscription method based on deep learning | |
CN115546032B (en) | Single-frame image super-resolution method based on feature fusion and attention mechanism | |
CN110349087A (en) | RGB-D image superior quality grid generation method based on adaptability convolution | |
CN114331830B (en) | Super-resolution reconstruction method based on multi-scale residual error attention | |
CN110009568A (en) | The generator construction method of language of the Manchus image super-resolution rebuilding | |
CN115100039B (en) | Lightweight image super-resolution reconstruction method based on deep learning | |
CN114066871A (en) | Method for training new coronary pneumonia focus region segmentation model | |
CN111461978A (en) | Attention mechanism-based resolution-by-resolution enhanced image super-resolution restoration method | |
CN114897694A (en) | Image super-resolution reconstruction method based on mixed attention and double-layer supervision | |
CN114841859A (en) | Single-image super-resolution reconstruction method based on lightweight neural network and Transformer | |
CN114943646A (en) | Gradient weight loss and attention mechanism super-resolution method based on texture guidance | |
CN111414988A (en) | Remote sensing image super-resolution method based on multi-scale feature self-adaptive fusion network | |
CN113096015B (en) | Image super-resolution reconstruction method based on progressive perception and ultra-lightweight network | |
Sun et al. | ESinGAN: Enhanced single-image GAN using pixel attention mechanism for image super-resolution | |
Wang | Single image super-resolution with u-net generative adversarial networks | |
CN103198495B (en) | The texture compression method that importance degree drives |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190712 |
|
RJ01 | Rejection of invention patent application after publication |