CN106683048A - Image super-resolution method and image super-resolution equipment - Google Patents
Image super-resolution method and image super-resolution equipment Download PDFInfo
- Publication number
- CN106683048A CN106683048A CN201611086392.3A CN201611086392A CN106683048A CN 106683048 A CN106683048 A CN 106683048A CN 201611086392 A CN201611086392 A CN 201611086392A CN 106683048 A CN106683048 A CN 106683048A
- Authority
- CN
- China
- Prior art keywords
- network
- super
- resolution image
- differentiation
- sample
- 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
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000012545 processing Methods 0.000 claims abstract description 19
- 230000004069 differentiation Effects 0.000 claims description 103
- 230000008569 process Effects 0.000 claims description 18
- 230000008485 antagonism Effects 0.000 claims description 17
- 238000013528 artificial neural network Methods 0.000 claims description 10
- 238000012800 visualization Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 abstract description 105
- 230000000694 effects Effects 0.000 abstract description 5
- 238000012549 training Methods 0.000 description 13
- 230000004913 activation Effects 0.000 description 11
- 238000013527 convolutional neural network Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 5
- 238000005070 sampling Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000000644 propagated effect Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000011084 recovery Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000013529 biological neural network Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000003475 lamination Methods 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 230000007935 neutral effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000007634 remodeling Methods 0.000 description 1
- 238000013519 translation 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
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a super-resolution image generation method. The super-resolution image generation method includes inputting a real image sample into a generation network so as to output a super-resolution image sample after the generation network and a distinguishing network are preset, acquiring distinguishing probabilities outputted by the distinguishing network after the real image sample and the super-resolution image sample are outputted, determining generation network loss functions and distinguishing network loss functions according to the real image sample, the super-resolution image sample and the distinguishing probabilities, and adjusting configuration parameters of the generation network and the distinguishing network according to the generation network loss functions and the distinguishing network loss functions; receiving a low-resolution image to be processed after adjustment is completed, generating a super-resolution image of the low-resolution image according to the generation network, and subjecting the super-resolution image to visualized processing. By the super-resolution image generation method, image super-resolution effect and realness of the super-resolution image are both improved remarkably.
Description
Technical field
The present invention relates to communication technical field, more particularly to a kind of image super-resolution method.The present invention is also related to
A kind of image super-resolution equipment.
Background technology
Image super-resolution rebuilding technology is to produce list using one group of low quality, low-resolution image (or motion sequence)
Panel height quality, high-definition picture.Image super-resolution rebuilding application and its broadness, in military affairs, medical science, public safety,
The aspects such as computer vision all have important application prospect.In computer vision field, image super-resolution rebuilding technology
There is a possibility that image is realized from detection level to the conversion of identification level, or further realize to carefully distinguishing horizontal conversion.
Based on the image super-resolution rebuilding technology of the identification ability and accuracy of identification that can improve image, prior art is carried
The single-frame imagess super-resolution that high-definition picture is generated using single frames low resolution, undersampled image is gone out.Single-frame imagess surpass
Resolution reconstruction technology can realize the absorbed analysis of object, such that it is able to obtain area-of-interest more high spatial resolution
Image, without the configuration for directly adopting the huge high spatial resolution images of data volume.
With the continuous development in artificial neural network field, current technical staff can be based on depth convolutional neural networks reality
Existing single-frame imagess super-resolution so that single-frame imagess super-resolution technique has huge progress.Convolutional neural networks are artificial
One kind of neutral net, it has also become current speech analyzes the study hotspot with field of image recognition.The shared network knot of its weights
Structure is allowed to be more closely similar to biological neural network, reduces the complexity of network model, reduces the quantity of weights.The advantage is in net
What is showed when the input of network is multidimensional image becomes apparent from, allow image directly as the input of network, it is to avoid traditional knowledge
Complicated feature extraction and data reconstruction processes in other algorithm.Convolutional network is to recognize one of two-dimensional shapes and particular design
Multilayer perceptron, this network structure has height invariance to translation, proportional zoom, inclination or the deformation of his form common.
At present, in the single-frame imagess super-resolution based on depth convolutional neural networks, the method is by low point of input
Resolution image carries out a series of convolution or deconvolution operation, is output as the high-resolution image of a frame.By calculating output
High-definition picture and the mean square deviation (MSE) of true high-definition picture, believe as the supervision that depth convolutional neural networks are trained
Number.
However, inventor has found during the application is realized, when prior art is carried out at big decimation factor to picture
During reason (such as 4x is up-sampled, i.e. the wide height of image is enlarged into respectively original 4 times), the recovery of grain details yet suffers from problem,
So as to details is not clear enough.Even using the single-frame imagess super-resolution of depth convolutional neural networks, also can be in decimation factor
Result is generally too smoothed when larger, lack detail of the high frequency, lead to not meet demand.
The content of the invention
The invention provides a kind of super-resolution image generation method, believes to the details for preferably retaining pending image
Breath, and solve in prior art to export the unsharp problem of high-definition picture grain details when decimation factor is larger.The party
Method pre-sets generation network and differentiates that the type of network, the generation network and the differentiation network is depth nerve
Network, the method also includes：
Super-resolution image sample after true picture sample is input into into the generation network to export super-resolution processing；
The acquisition differentiation network is respectively after the true picture sample and the super-resolution image sample is input into
The differentiation probability of output, the differentiation probability is the probability of the input picture for true picture of the differentiation network；
Network losses are generated according to the true picture sample, super-resolution image sample and the differentiation determine the probability
Function and differentiation network losses function, and according to the generation network losses function and the differentiation network losses function pair
The configuration parameter of the generation network and the differentiation network is adjusted；
After the adjustment for completing the configuration parameter, pending low-resolution image is received, according to the generation net
Network generates the super-resolution image of the low-resolution image, and carries out visualization processing to the super-resolution image；
Wherein, when the loss function of the generation network is less, the super-resolution image exported with the generation network
Validity it is higher；
When the loss function of the differentiation network is less, the accuracy of the differentiation probability for differentiating network output is higher.
Preferably, the loss function for generating network is according to antagonism loss function, rule constraint loss function and picture
The mean square deviation loss function of plain level is generated, wherein：
The antagonism loss function is according to sentencing that the differentiation network is exported after the super-resolution image sample is input into
Other probability is generated；
The rule constraint function determines according to the Space Consistency of the super-resolution image sample；
The mean square deviation loss function of the Pixel-level is according to the super-resolution image sample and the true picture sample
This determination.
Preferably, the super-resolution after by the true picture sample input generation network to export super-resolution processing
After image pattern, also include：
Respectively described true picture sample and the super-resolution image sample arrange label；
Wherein, the label of the true picture sample puts 1, and the label of the super-resolution image sample sets to 0.
Preferably, label and each described image sample of the loss function for differentiating network according to all image patterns
Generate in the differentiation probability through the differentiation network output.
Preferably, the true picture sample is zoomed in and out at process and normalization before the generation network is input into
Reason；The low-resolution image is normalized before the generation network is input into.
Accordingly, disclosed herein as well is a kind of super-resolution image generates equipment, including：
Preset module, pre-sets generation network and differentiates network, the generation network and the differentiation network
Type is deep neural network；
Input module, the super-resolution after true picture sample is input into into the generation network to export super-resolution processing
Image pattern；
Acquisition module, obtains the differentiation network and is input into the true picture sample and the super-resolution figure respectively
The differentiation probability exported after decent, the differentiation probability is the probability of the input picture for true picture of the differentiation network；
Determining module, gives birth to according to the true picture sample, super-resolution image sample and the differentiation determine the probability
Into network losses function and differentiation network losses function, and according to the generation network losses function and the differentiation network
Loss function is adjusted to the configuration parameter of the generation network and the differentiation network；
Generation module, after the adjustment for completing the configuration parameter, receives pending low-resolution image, according to institute
State and generate the super-resolution image of the network generation low-resolution image, and the super-resolution image is carried out at visualization
Reason；
Wherein, when the loss function of the generation network is less, the super-resolution image exported with the generation network
Validity it is higher；
When the loss function of the differentiation network is less, the accuracy of the differentiation probability for differentiating network output is higher.
Preferably, the loss function for generating network is according to antagonism loss function, rule constraint loss function and picture
The mean square deviation loss function of plain level is generated, wherein：
The antagonism loss function is according to sentencing that the differentiation network is exported after the super-resolution image sample is input into
Other probability is generated；
The rule constraint function determines according to the Space Consistency of the super-resolution image sample；
The mean square deviation loss function of the Pixel-level is according to the super-resolution image sample and the true picture sample
This determination.
Preferably, also include：
Label model, respectively described true picture sample and the super-resolution image sample arrange label；
Wherein, the label of the true picture sample puts 1, and the label of the super-resolution image sample sets to 0.
Preferably, label and each described image sample of the loss function for differentiating network according to all image patterns
Generate in the differentiation probability through the differentiation network output.
Preferably, the true picture sample is zoomed in and out at process and normalization before the generation network is input into
Reason；The low-resolution image is normalized before the generation network is input into.
As can be seen here, by the technical scheme using the application, after pre-setting generation network and differentiating network, will
The super-resolution image sample that true picture sample is input into after generating network to export super-resolution processing, and obtain differentiation network
The differentiation probability for exporting after input true picture sample and super-resolution image sample respectively, finally according to true picture sample
Originally, super-resolution image sample and differentiation determine the probability generate network losses function and differentiate network losses function, and root
Adjusted according to the configuration parameter for generating network losses function and differentiation network losses function pair generation network and differentiation network
It is whole.So when pending low-resolution image is received after the completion of adjustment, low resolution can be generated according to network is generated
The super-resolution image of rate image, and visualization processing is carried out to super-resolution image.So as to significantly improve Image Super-resolution
The verity of rate effect and super resolution image.
Description of the drawings
Fig. 1 is a kind of structural representation of the generation network disclosed in the application specific embodiment；
Fig. 2 is a kind of structural representation of the differentiation network disclosed in the application specific embodiment；
Fig. 3 is a kind of schematic flow sheet of super resolution image generation method that the application is proposed；
The training process workflow schematic diagram that Fig. 4 is proposed by the application specific embodiment；
Fig. 5 is super-resolution image calculation process schematic diagram in the application this specific embodiment；
Fig. 6 is the structural representation that a kind of super resolution image that the application is proposed generates equipment.
Specific embodiment
As stated in the Background Art, the single-frame imagess super-resolution based on depth convolutional neural networks so that single-frame imagess surpass
Resolution technique has huge progress.But for the up-sampling factor than it is larger when (such as 4 times up-sampling) grain details
Recovery yet suffers from problem, so as to cause details not clear enough.
In view of this, present applicant proposes a kind of super resolution image generation method, the method is for single-frame imagess super-resolution
Rate proposes a kind of antagonism network frame, and including two deep neural networks, one is to generate network, and one is to differentiate net
Network, two networks are trained and vied each other simultaneously.The purpose for generating network is to generate high-definition picture from low-resolution image, and
And the high-definition picture for generating is difficult to be distinguished with true picture；The purpose for differentiating network is to discriminate between true picture and super-resolution
The image of generation.By resisting network training, generation network output real high-definition picture as far as possible is made.Just because in this,
Before elaborating the detailed step of technical scheme, the application pre-sets generation network and differentiates network, needs
Illustrate, the type of the generation network and differentiation network is deep neural network, below for both network integrations
Specific embodiment is illustrated：
As shown in figure 1, in the application specific embodiment generate network structural representation, in the generation network structure
Every layer of concrete meaning it is as follows：
1st layer of InputLR represents input low-resolution image；
2nd and the 3rd layer represents that (Rectified linear unit, correct linear unit, are for a convolutional layer and ReLU
One kind of deep learning activation primitive) activation primitive layer, wherein the step-length of convolution operation be 1, convolution kernel size be 3*3, convolution
Nuclear volume is 64；
4th to the 9th layer is a residual error network element function block, has used two groups of convolutional layers to closely follow batch standardization layer, with
ReLU, as activation primitive, is finally that Element-Level is added layer, and the wherein step-length of convolution operation is 1, and convolution kernel size is 3*3, is rolled up
Product nuclear volume is 64；
10th to the 33rd layer is 4 residual error network element function blocks, and each residual error network element function block is ibid；
34th to the 37th layer is two groups of warp product units, for picture up-sampling.The step-length of deconvolution layer operation is 0.5,
Convolution kernel size is 3*3, and convolution nuclear volume is 64；The zoom factor that this example is adopted is for 4, so employing two groups of deconvolution lists
Unit；Zoom factor is such as crossed for 2, then one group of warp product unit can be adopted；
38th layer is a convolutional layer, and convolution operation step-length is 1, and convolution kernel size is 3*3, and convolution nuclear volume is 3, purpose
It is the RGB data for generating 3 passages.
What last layer of the generation network was exported is super-resolution image.
As shown in Fig. 2 in the application specific embodiment differentiation network structural representation, the differentiation network structure
In every layer of concrete meaning it is as follows：
1st layer of Input HR/SR represents the high-resolution true picture or super-resolution image of input；True picture mark
1 is designated as, super-resolution image is labeled as 0；
Layers 2 and 3 represents a convolutional layer and an activation primitive layer, convolution；Wherein convolutional layer step-length is 1, volume
Product core size is 3*3, and convolution nuclear volume is 64；
4th layer to the 6th layer represents convolutional layer, an activation primitive layer and a batch rule layer；Wherein convolution
Layer step-length is 2, and convolution kernel size is 3*3, and convolution nuclear volume is 64；
7th layer to the 9th layer represents convolutional layer, an activation primitive layer and a batch rule layer；Wherein convolution
Layer step-length is 1, and convolution kernel size is 3*3, and convolution nuclear volume is 128；
10th layer to the 12nd layer represents convolutional layer, an activation primitive layer and a batch rule layer；Wherein roll up
Lamination step-length is 2, and convolution kernel size is 3*3, and convolution nuclear volume is 128；
It is that convolution nuclear volume is 256 that 13rd layer is distinguished to the 18th layer similar to the 7th to the 12nd layer, uniquely；
It is that convolution nuclear volume is 512 that 19th layer is distinguished to the 24th layer similar to the 7th to the 12nd layer, uniquely；
25th layer and the 26th layer is a full articulamentum and a ReLU activation primitive layer；
27th layer and the 28th layer is a full articulamentum and Sigmoid (by the use of sigmoid functions as activation letter
Number, is one kind of deep learning activation primitive) activation primitive layer, wherein full articulamentum node number is 1；
Last layer of the differentiation network exports a probit, represents probability of the input picture for true picture.
The generation network to the application is distributed above in association with specific network structure and function and differentiate that network enters
Go explanation, but it should be noted that above structure is only the one kind proposed in the application specific embodiment is preferable to carry out
Example scheme, is super resolution image deep neural network is based on by the way that input picture is carried out into superresolution processing, and for defeated
The image for entering is carried out on the basis of the judgement of true picture/super resolution image, and other remodeling that those skilled in the art are done are equal
Belong to the protection domain of the application.
Based on the generation network and differentiation network that are pre-created, the super resolution image generation method that the application is proposed
Schematic flow sheet is as shown in figure 3, comprise the following steps：
S301, the super-resolution image after true picture sample is input into into the generation network to export super-resolution processing
Sample.
The defeated of network is generated because the application is intended to be improved by the deep neural network to different functions to optimize
Go out effect, therefore the application is broadly divided into two big flow elements, Part I generates low-resolution image of the network according to input
Output high-definition picture；Part II differentiates that network carries out verity judgement to input picture.Generate the output image of network
As the input picture for differentiating network, differentiate the output of network as the feedback signal for generating network.
Based on described above, the step generates network using existing picture as the input of true picture sample first, defeated
Enter and generate before network, need in the usual course true picture sample to be carried out into some pretreatment works, such as scaling is processed, normalizing
Change process etc..Specific zooming parameter and normalized parameter can be configured by technical staff according to practical situation, these tune
The whole protection domain for belonging to the application.
Because the application needs the verification accuracy rate for differentiating network to test, and true picture sample is raw with follow-up
Into super resolution image there is certain similarity, it is respectively described true therefore in a preferred embodiment of the application
Image pattern and the super-resolution image sample arrange label to make a distinction, and the label of wherein true picture sample puts 1,
The label of super-resolution image sample sets to 0.
S302, obtains the differentiation network and is input into the true picture sample and the super-resolution image sample respectively
The differentiation probability exported after this, the differentiation probability is the probability of the input picture for true picture of the differentiation network.
Based in S301 generate super-resolution image sample and original true picture sample, the application by it simultaneously
Input differentiates network, and the verification accuracy for differentiating network is detected with this.Either it is input into true picture sample or super-resolution
Rate image pattern, differentiates that network can all export a differentiation probability, and the probability is to describe input picture in the inspection for differentiating network
It is the probability of true picture after testing.In specific application scenarios, the probability can be relative with the label value of each sample image
Should, such as probit represents that the probability of true picture is higher closer to 1；Probit represents super-resolution image closer to 0
Probability is higher.
S303, generates network and damages according to true picture sample, super-resolution image sample and the differentiation determine the probability
Lose function and differentiate network losses function, and according to generation network losses function and differentiate that network losses function pair generates net
The configuration parameter of network and differentiation network is adjusted.
Because the application is based on single-frame imagess super-resolution training method come phase in the training stage using two depth networks
Mutually antagonism, therefore based on the differentiation probability in the super-resolution image and S302 in S301, the application will be respectively directed to generate
Network and differentiation network determine its respective loss function, for network is generated, when the loss letter of the generation network
Number is less, and the validity of the super-resolution image exported with the generation network is higher, and for network is differentiated, works as institute
State and differentiate that the loss function of network is less, the accuracy of the differentiation probability for differentiating network output is higher.
As it was previously stated, generate network and differentiate that the type of network is deep neural network, its respectively by convolution unit and
Multiple functions are constituted as functional module, therefore in the concrete application scene of the application, based on generation network losses function pair
The weighted value for generating each network node in network is adjusted, and each in network losses function pair differentiation network based on differentiating
The weighted value of individual network node is adjusted.
It should be noted that the validity of the super-resolution image mentioned in above content is super-resolution image sample
For true picture sample, and accuracy is then to differentiate probability relative to the true picture/super-resolution under truth
For rate image, the specific form of expression can in the light of actual conditions be set by technical staff.Additionally, for generating net
The parameter configuration of network/differentiation network can be adjusted by technical staff according to loss function, and these belong to the guarantor of the application
Shield scope.
In a preferred embodiment of the application, generate network loss function according to antagonism loss function, rule about
The mean square deviation loss function of beam loss function and Pixel-level is generated, and specifically, the generating mode of each function is as follows：
(1) the antagonism loss function is exported according to the differentiation network after the super-resolution image sample is input into
Differentiate that probability is generated.
In specific application scenarios, the determination mode for resisting loss function is as follows：
Wherein, D (ISR) representing that super-resolution output image is input to the output differentiated after network, this value is bigger, represents
Super-resolution output image is closer to true picture, and antagonism loss function value is less.
(2) the rule constraint function determines according to the Space Consistency of the super-resolution image sample.
In the concrete application scene of the application, rule constraint loss function is designated as LGREG, main purpose is to maintain oversubscription
The Space Consistency of resolution image, determination mode is as follows：
(3) the mean square deviation loss function of the Pixel-level is according to the super-resolution image sample and the true picture
Sample determines.
In the concrete application scene of the application, the mean square deviation loss function of Pixel-level is designated as LGMSE, computing formula is such as
Under：
Wherein, IHRRepresent high-resolution true picture, ISRRepresent the image after superresolution processing.
Based on above-mentioned function, the preferred embodiment of the application, can be according to difference when the loss function of network is generated
Function distributes different weights for the accuracy of super-resolution image, and specific weighted value can be by technical staff according to reality
Applicable cases are configured, and these belong to the protection domain of the application.
Above content describes the concrete generating mode of the loss function for generating network, and another in the application is preferable to carry out
In example, the loss function for differentiating network is then based on the mark for distributing for true picture sample and super-resolution image sample in advance
Sign and differentiate the differentiation probability generation that network is drawn after these images are input into respectively.One in the application is embodied as
In example, it is assumed that the loss function for differentiating network is LD, then its corresponding determination mode is as follows：
Wherein, aiInput label is represented, 1 represents true picture, and 0 represents super-resolution image；yiRepresent and differentiate network output
Class probability.
Based on the loss function of the loss function and differentiation network for generating network, the application loses further with these
Function pair generates network and differentiates that the configuration parameter of network is adjusted.Here is it should be noted that S301- in the application
The adjustment process of S303 is not limited in once, but the degree of optimization and technical staff according to network is directed to super-resolution figure
The generation technique required precision of picture repeats necessary number of times, and these belong to the protection domain of the application.
S304, after the adjustment for completing the configuration parameter, receives pending low-resolution image, according to the life
The super-resolution image of the low-resolution image is generated into network, and visualization processing is carried out to the super-resolution image.
After the completion of the step of parameter of network to be generated and differentiation network passes through S301-S303 adjustment, by antagonism
Network training enables generation network to export real high-definition picture as far as possible.Subsequently when technical staff is needed for low resolution
When rate image generates super-resolution image, you can using low-resolution image as input picture to generation network inputs, and with this
Generate super-resolution image.Compared to existing Pixel-level error statistics, can preferably retain image detail information, and solve
Output high-definition picture grain details unsharp problem when decimation factor is larger.
In order to the technological thought of the present invention is expanded on further, in conjunction with super shown in the training process and Fig. 5 shown in Fig. 4
Image in different resolution determines that flow process is illustrated to technical scheme.
Training sample in each training process of this specific embodiment is to go out N from random cropping in training pictures
The picture of Width*Height, such as N are 32, Width=Height=128.These training samples are simultaneously as generation network
With the input for differentiating network.As shown in figure 4, the training process workflow proposed by the application specific embodiment, including such as
Lower step：
Step one, scaling and normalized.Bicubic interpolation down-sampling is carried out to training sample, decimation factor is 4；So
The image after down-sampling is normalized afterwards, i.e., each passage pixel value is divided by 255；As the input data for generating network；
Step 2, generation network propagated forward.According to network structure is generated, forward direction travels through all layers；
Step 3, output super-resolution image, while for differentiating the input of network and generating the costing bio disturbance of network；
Step 4, sample preprocessing.Sample is divided into two sources at this, and one is training sample, each passage pixel value
Divided by 255, as true picture, classification 1 is labeled as；Another is the output of step 3, is program as super-resolution image
Produce, be different from true picture, be labeled as classification 0；Each class sample N, amount to 2N；
Step 5, differentiation network propagated forward.According to network structure is differentiated, forward direction travels through all layers；
Step 6, output class probability, probit represents that the probability of true picture is higher closer to 1；Probit closer to
0, represent that the probability of super-resolution image is higher；
Step 7, generation network losses function are calculated.Generate network losses function and be designated as LG, it is made up of three parts：
1) the mean square deviation loss function of Pixel-level, is designated as LGMSE, determination mode is as follows：
Wherein, x, y represent pixel coordinate position in the picture, IHRRepresent high-resolution true picture, ISRRepresent super
Image after resolution process.
2) loss function is resisted, it is therefore an objective to count the verity of this super-resolution image, be designated as LGADV.Its determination mode
It is as follows：
Wherein, N is quantity N of sample in step 4, and i represents the numbering of current sample, D (ISR) represent that super-resolution is defeated
Go out image and be input to the output differentiated after network, that is, differentiate the differentiation probability of network, this value is bigger, represent super-resolution output
Image is closer to true picture, and antagonism loss function value is less.
3) rule constraint loss function, is designated as LGREG, main purpose is to maintain the Space Consistency of super-resolution image.Its
Determination mode is as follows：
Wherein, ISR X, yRepresent certain pixel in the image after superresolution processing.
Generating network losses function can generate in three partial weightings summation by more than, and concrete mode is as follows：
LG=γ0LGMSE+γ1LGADV+γ2LGREG
Wherein γ0Can be with value 0.9, γ1Can be with value 0.002, γ2Can be with value 2*10-8, can specifically adjust these
Parameter.
Step 8, generation network back propagation.According to the generation network losses function that step 7 is calculated, back propagation is whole
Network is generated, i.e., is generated each network node in network using network losses function pair and is adjusted, in specific applied field
Jing Zhong, technical staff can generate the weighted value of network node according to the loss function adjustment；
Step 9, differentiation network losses function are calculated.Differentiate that network losses function is designated as LD, computing formula equation below 1：
Wherein aiInput label is represented, 1 represents true picture, and 0 represents super-resolution image；yiRepresent and differentiate network output
Differentiation probability；
Step 10, differentiation network back propagation.According to the differentiation network losses function that step 9 is calculated, back propagation is whole
Differentiate network, adjustment differentiates network node weighted value；
In actual application process, technical staff can as needed simultaneously execution step seven, eight and step 9, ten,
Concrete step performs the protection domain that precedence has no effect on the application.
After configuration parameter is have adjusted for the loss function for generating network and differentiation network, in this specific embodiment
Super-resolution image calculation process is as shown in figure 5, comprise the following steps：
Step one, input low-resolution image；
Step 2, normalized, i.e. image each passage pixel value is done to input picture divided by 255；
Step 3, generation network propagated forward.According to network structure is generated, forward direction travels through all layers；
Step 4, output super-resolution image；
Step 5, image viewing, i.e., to the super-resolution image of generation network output, each passage pixel value is multiplied by
255, and fixed point.
By the technical scheme using embodiments above, single-frame imagess super-resolution efect is improve, it is particularly right
In the situation that zoom factor is larger, but also the verity of single-frame imagess super resolution image is improved, and be not limited solely to letter
Make an uproar ratio.
To reach above technical purpose, the invention allows for a kind of super-resolution image generates equipment as shown in fig. 6, bag
Include：
Preset module 610, pre-sets generation network and differentiates network, the generation network and the differentiation network
Type be deep neural network；
Input module 620, the oversubscription after true picture sample is input into into the generation network to export super-resolution processing
Resolution image pattern；
Acquisition module 630, obtains the differentiation network and is input into the true picture sample and the super-resolution respectively
The differentiation probability exported after rate image pattern, it is the general of true picture that the differentiation probability is the input picture of the differentiation network
Rate；
Determining module 640, according to the true picture sample, super-resolution image sample and the differentiation determine the probability
Generate network losses function and differentiate network losses function, and according to the generation network losses function and the differentiation net
Network loss function is adjusted to the configuration parameter of the generation network and the differentiation network；
Generation module 650, after the adjustment for completing the configuration parameter, receives pending low-resolution image, root
The super-resolution image of the low-resolution image is generated according to the generation network, and the super-resolution image is carried out visually
Change is processed；
Wherein, when the loss function of the generation network is less, the super-resolution image exported with the generation network
Validity it is higher；
When the loss function of the differentiation network is less, the accuracy of the differentiation probability for differentiating network output is higher.
In specific application scenarios, the loss function for generating network is damaged according to antagonism loss function, rule constraint
The mean square deviation loss function for losing function and Pixel-level is generated, wherein：
The antagonism loss function is according to sentencing that the differentiation network is exported after the super-resolution image sample is input into
Other probability is generated；
The rule constraint function determines according to the Space Consistency of the super-resolution image sample；
The mean square deviation loss function of the Pixel-level is according to the super-resolution image sample and the true picture sample
This determination.
In specific application scenarios, also include：
Label model, respectively described true picture sample and the super-resolution image sample arrange label；
Wherein, the label of the true picture sample puts 1, and the label of the super-resolution image sample sets to 0.
In specific application scenarios, the loss function for differentiating network is according to the label of all image patterns and each
Described image sample is generated in the differentiation probability through the differentiation network output.
In specific application scenarios, the true picture sample be input into it is described generation network before zoom in and out process with
And normalized；The low-resolution image is normalized before the generation network is input into.
As can be seen here, by the technical scheme using the application, after pre-setting generation network and differentiating network, will
The input of true picture sample generates network to export super-resolution image sample, and obtains differentiation network respectively in the true figure of input
The differentiation probability exported after decent and super-resolution image sample, according to true picture sample, super-resolution image sample
And differentiate determine the probability generate network losses function and differentiate network losses function, according to generate network losses function and
Differentiate that network losses function pair generates network and differentiates that the configuration parameter of network is adjusted.When the tune for completing configuration parameter
After whole, pending low-resolution image is received, can be according to the super-resolution figure for generating network generation low-resolution image
Picture, and visualization processing is carried out to super-resolution image.So as to significantly improve image super-resolution effect and super-resolution figure
The verity of picture.
Through the above description of the embodiments, those skilled in the art can be understood that the present invention can lead to
Cross hardware realization, it is also possible to realize by the mode of software plus necessary general hardware platform.Based on such understanding, this
Bright technical scheme can be embodied in the form of software product, and the software product can be stored in a non-volatile memories
In medium (can be CD-ROM, USB flash disk, portable hard drive etc.), including some instructions are used so that a computer equipment (can be
Personal computer, server, or network equipment etc.) perform method described in each implement scene of the invention.
It will be appreciated by those skilled in the art that accompanying drawing is a schematic diagram for being preferable to carry out scene, module in accompanying drawing or
Flow process is not necessarily implemented necessary to the present invention.
It will be appreciated by those skilled in the art that the module in the device in implement scene can according to implement scene describe into
Row is distributed in the device of implement scene, it is also possible to carry out one or more dresses that respective change is disposed other than this implement scene
In putting.The module of above-mentioned implement scene can merge into a module, it is also possible to be further split into multiple submodule.
The invention described above sequence number is for illustration only, does not represent the quality of implement scene.
Disclosed above is only that the several of the present invention are embodied as scene, but, the present invention is not limited to this, Ren Heben
What the technical staff in field can think change should all fall into protection scope of the present invention.
Claims (10)
1. a kind of super-resolution image generation method, it is characterised in that pre-set generation network and differentiate network, the life
Type into network and the differentiation network is deep neural network, and the method also includes：
Super-resolution image sample after true picture sample is input into into the generation network to export super-resolution processing；
Obtain the differentiation network to export after the true picture sample and the super-resolution image sample is input into respectively
Differentiation probability, the differentiation probability be it is described differentiation network input picture for true picture probability；
Network losses function is generated according to the true picture sample, super-resolution image sample and the differentiation determine the probability
And differentiate network losses function, and according to the generation network losses function and the differentiation network losses function pair
The configuration parameter for generating network and the differentiation network is adjusted；
After the adjustment for completing the configuration parameter, pending low-resolution image is received, according to the generation network life
Into the super-resolution image of the low-resolution image, and visualization processing is carried out to the super-resolution image；
Wherein, it is described to generate the true of the super-resolution image that network is exported when the loss function of the generation network is less
Degree is higher；
When the loss function of the differentiation network is less, the accuracy of the differentiation probability for differentiating network output is higher.
2. the method for claim 1, it is characterised in that the loss function of the generation network is according to antagonism loss letter
The mean square deviation loss function of number, rule constraint loss function and Pixel-level is generated, wherein：
The antagonism loss function is general according to the differentiation that the differentiation network is exported after the super-resolution image sample is input into
Rate is generated；
The rule constraint function determines according to the Space Consistency of the super-resolution image sample；
The mean square deviation loss function of the Pixel-level is true according to the super-resolution image sample and the true picture sample
It is fixed.
3. the method as described in any one of claim 1 or 2, it is characterised in that true picture sample is input into into the generation
Network to export super-resolution processing after super-resolution image sample after, also include：
Respectively described true picture sample and the super-resolution image sample arrange label；
Wherein, the label of the true picture sample puts 1, and the label of the super-resolution image sample sets to 0.
4. method as claimed in claim 3, it is characterised in that
The loss function for differentiating network is according to the label and each described image sample of all image patterns through described
The differentiation probability for differentiating network output is generated.
5. method as claimed in claim 4, it is characterised in that
The true picture sample zooms in and out process and normalized before the generation network is input into；
The low-resolution image is normalized before the generation network is input into.
6. a kind of super-resolution image generates equipment, it is characterised in that include：
Preset module, pre-sets generation network and differentiates the type of network, the generation network and the differentiation network
It is deep neural network；
Input module, the super-resolution image after true picture sample is input into into the generation network to export super-resolution processing
Sample；
Acquisition module, obtains the differentiation network and is input into the true picture sample and the super-resolution image sample respectively
The differentiation probability exported after this, the differentiation probability is the probability of the input picture for true picture of the differentiation network；
Determining module, according to the true picture sample, super-resolution image sample and the differentiation determine the probability net is generated
Network loss function and differentiation network losses function, and according to the generation network losses function and differentiate network losses function
The configuration parameter of the generation network and the differentiation network is adjusted；
Generation module, after the adjustment for completing the configuration parameter, receives pending low-resolution image, according to the life
The super-resolution image of the low-resolution image is generated into network, and visualization processing is carried out to the super-resolution image；
Wherein, when the loss function of the generation network is less, with described the true of the super-resolution image that network is exported is generated
Solidity is higher；
When the loss function of the differentiation network is less, the accuracy of the differentiation probability for differentiating network output is higher.
7. equipment as claimed in claim 6, it is characterised in that the loss function of the generation network is according to antagonism loss letter
The mean square deviation loss function of number, rule constraint loss function and Pixel-level is generated, wherein：
The antagonism loss function is general according to the differentiation that the differentiation network is exported after the super-resolution image sample is input into
Rate is generated；
The rule constraint function determines according to the Space Consistency of the super-resolution image sample；
The mean square deviation loss function of the Pixel-level is true according to the super-resolution image sample and the true picture sample
It is fixed.
8. the equipment as described in any one of claim 6 or 7, it is characterised in that also include：
Label model, respectively described true picture sample and the super-resolution image sample arrange label；
Wherein, the label of the true picture sample puts 1, and the label of the super-resolution image sample sets to 0.
9. equipment as claimed in claim 8, it is characterised in that
The loss function for differentiating network is according to the label and each described image sample of all image patterns through described
The differentiation probability for differentiating network output is generated.
10. equipment as claimed in claim 9, it is characterised in that
The true picture sample zooms in and out process and normalized before the generation network is input into；
The low-resolution image is normalized before the generation network is input into.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611086392.3A CN106683048B (en) | 2016-11-30 | 2016-11-30 | Image super-resolution method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611086392.3A CN106683048B (en) | 2016-11-30 | 2016-11-30 | Image super-resolution method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106683048A true CN106683048A (en) | 2017-05-17 |
| CN106683048B CN106683048B (en) | 2020-09-01 |
Family
ID=58867637
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611086392.3A Active CN106683048B (en) | 2016-11-30 | 2016-11-30 | Image super-resolution method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106683048B (en) |
Cited By (36)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107133934A (en) * | 2017-05-18 | 2017-09-05 | 北京小米移动软件有限公司 | Image completion method and device |
| CN107274358A (en) * | 2017-05-23 | 2017-10-20 | 广东工业大学 | Image Super-resolution recovery technology based on cGAN algorithms |
| CN107423700A (en) * | 2017-07-17 | 2017-12-01 | 广州广电卓识智能科技有限公司 | The method and device of testimony verification |
| CN107578377A (en) * | 2017-08-31 | 2018-01-12 | 北京飞搜科技有限公司 | A kind of super-resolution image reconstruction method and system based on deep learning |
| CN107590774A (en) * | 2017-09-18 | 2018-01-16 | 北京邮电大学 | A kind of car plate clarification method and device based on generation confrontation network |
| CN107633218A (en) * | 2017-09-08 | 2018-01-26 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating image |
| CN107862785A (en) * | 2017-10-16 | 2018-03-30 | 深圳市中钞信达金融科技有限公司 | Bill authentication method and device |
| CN107945204A (en) * | 2017-10-27 | 2018-04-20 | 西安电子科技大学 | A kind of Pixel-level portrait based on generation confrontation network scratches drawing method |
| CN107977511A (en) * | 2017-11-30 | 2018-05-01 | 浙江传媒学院 | A kind of industrial design material high-fidelity real-time emulation algorithm based on deep learning |
| CN108122249A (en) * | 2017-12-20 | 2018-06-05 | 长沙全度影像科技有限公司 | A kind of light stream method of estimation based on GAN network depth learning models |
| CN108122209A (en) * | 2017-12-14 | 2018-06-05 | 浙江捷尚视觉科技股份有限公司 | A kind of car plate deblurring method based on confrontation generation network |
| CN108182669A (en) * | 2018-01-02 | 2018-06-19 | 华南理工大学 | A kind of Super-Resolution method of the generation confrontation network based on multiple dimension of pictures |
| CN108235058A (en) * | 2018-01-12 | 2018-06-29 | 广州华多网络科技有限公司 | Video quality processing method, storage medium and terminal |
| CN108805188A (en) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | A kind of feature based recalibration generates the image classification method of confrontation network |
| CN108960014A (en) * | 2017-05-23 | 2018-12-07 | 北京旷视科技有限公司 | Image processing method, device and system and storage medium |
| CN108961161A (en) * | 2018-05-24 | 2018-12-07 | 上海商汤智能科技有限公司 | A kind of image processing method, device and computer storage medium |
| CN109003253A (en) * | 2017-05-24 | 2018-12-14 | 通用电气公司 | Neural network point cloud generates system |
| CN109118470A (en) * | 2018-06-26 | 2019-01-01 | 腾讯科技（深圳）有限公司 | A kind of image quality evaluating method, device, terminal and server |
| CN109377459A (en) * | 2018-09-30 | 2019-02-22 | 国网山东省电力公司电力科学研究院 | A kind of super-resolution deblurring method of production confrontation network |
| CN109525859A (en) * | 2018-10-10 | 2019-03-26 | 腾讯科技（深圳）有限公司 | Model training, image transmission, image processing method and relevant apparatus equipment |
| CN109636721A (en) * | 2018-11-29 | 2019-04-16 | 武汉大学 | Video super-resolution method based on confrontation study and attention mechanism |
| CN109697470A (en) * | 2018-12-27 | 2019-04-30 | 百度在线网络技术（北京）有限公司 | A kind of training method of identification model, device, electronic equipment and storage medium |
| CN109801215A (en) * | 2018-12-12 | 2019-05-24 | 天津津航技术物理研究所 | The infrared super-resolution imaging method of network is generated based on confrontation |
| CN109816764A (en) * | 2019-02-02 | 2019-05-28 | 深圳市商汤科技有限公司 | Image generating method and device, electronic equipment and storage medium |
| WO2019104705A1 (en) * | 2017-12-01 | 2019-06-06 | 华为技术有限公司 | Image processing method and device |
| CN109859113A (en) * | 2018-12-25 | 2019-06-07 | 北京奇艺世纪科技有限公司 | Model generating method, image enchancing method, device and computer readable storage medium |
| WO2019114523A1 (en) * | 2017-12-12 | 2019-06-20 | 腾讯科技（深圳）有限公司 | Classification training method, server and storage medium |
| CN109934102A (en) * | 2019-01-28 | 2019-06-25 | 浙江理工大学 | A kind of finger vein identification method based on image super-resolution |
| CN109949219A (en) * | 2019-01-12 | 2019-06-28 | 深圳先进技术研究院 | A kind of reconstructing method of super-resolution image, device and equipment |
| CN109993694A (en) * | 2017-12-29 | 2019-07-09 | Tcl集团股份有限公司 | A kind of method and device generating super-resolution image |
| CN110472457A (en) * | 2018-05-10 | 2019-11-19 | 成都视观天下科技有限公司 | Low-resolution face image identification, restoring method, equipment and storage medium |
| CN110580682A (en) * | 2019-09-16 | 2019-12-17 | 电子科技大学 | Countermeasure network seismic data super-resolution reconstruction method based on optimization generation |
| CN111402133A (en) * | 2020-03-13 | 2020-07-10 | 北京字节跳动网络技术有限公司 | Image processing method, image processing device, electronic equipment and computer readable medium |
| CN113628121A (en) * | 2020-05-06 | 2021-11-09 | 阿里巴巴集团控股有限公司 | Method and device for processing data and training multimedia data |
| KR20220131808A (en) * | 2021-03-22 | 2022-09-29 | 재단법인대구경북과학기술원 | Method and apparatus for generating imaga classification model |
| US11823349B2 (en) | 2020-11-19 | 2023-11-21 | Samsung Electronics Co., Ltd. | Image generators with conditionally-independent pixel synthesis |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102693419A (en) * | 2012-05-24 | 2012-09-26 | 武汉大学 | Super-resolution face recognition method based on multi-manifold discrimination and analysis |
| CN103985107A (en) * | 2014-05-27 | 2014-08-13 | 武汉科技大学 | Self-adaptive image super-resolution reconstruction method based on visual perception |
| CN104778659A (en) * | 2015-04-15 | 2015-07-15 | 杭州电子科技大学 | Single-frame image super-resolution reconstruction method on basis of deep learning |
| CN104992410A (en) * | 2015-02-10 | 2015-10-21 | 国网重庆市电力公司电力科学研究院 | Monocular visual pattern processing method |
| US20160217552A1 (en) * | 2015-01-22 | 2016-07-28 | Samsung Electronics Co., Ltd. | Video super-resolution by fast video segmentation for boundary accuracy control |
| CN106127684A (en) * | 2016-06-22 | 2016-11-16 | 中国科学院自动化研究所 | Image super-resolution Enhancement Method based on forward-backward recutrnce convolutional neural networks |
-
2016
- 2016-11-30 CN CN201611086392.3A patent/CN106683048B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102693419A (en) * | 2012-05-24 | 2012-09-26 | 武汉大学 | Super-resolution face recognition method based on multi-manifold discrimination and analysis |
| CN103985107A (en) * | 2014-05-27 | 2014-08-13 | 武汉科技大学 | Self-adaptive image super-resolution reconstruction method based on visual perception |
| US20160217552A1 (en) * | 2015-01-22 | 2016-07-28 | Samsung Electronics Co., Ltd. | Video super-resolution by fast video segmentation for boundary accuracy control |
| CN104992410A (en) * | 2015-02-10 | 2015-10-21 | 国网重庆市电力公司电力科学研究院 | Monocular visual pattern processing method |
| CN104778659A (en) * | 2015-04-15 | 2015-07-15 | 杭州电子科技大学 | Single-frame image super-resolution reconstruction method on basis of deep learning |
| CN106127684A (en) * | 2016-06-22 | 2016-11-16 | 中国科学院自动化研究所 | Image super-resolution Enhancement Method based on forward-backward recutrnce convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| 杨宇翔: ""图像超分辨率重建算法研究"", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (60)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107133934A (en) * | 2017-05-18 | 2017-09-05 | 北京小米移动软件有限公司 | Image completion method and device |
| CN107133934B (en) * | 2017-05-18 | 2020-03-17 | 北京小米移动软件有限公司 | Image completion method and device |
| CN108960014A (en) * | 2017-05-23 | 2018-12-07 | 北京旷视科技有限公司 | Image processing method, device and system and storage medium |
| CN107274358A (en) * | 2017-05-23 | 2017-10-20 | 广东工业大学 | Image Super-resolution recovery technology based on cGAN algorithms |
| CN108960014B (en) * | 2017-05-23 | 2021-05-11 | 北京旷视科技有限公司 | Image processing method, device and system and storage medium |
| CN109003253B (en) * | 2017-05-24 | 2021-10-22 | 通用电气公司 | Neural network point cloud generating system |
| CN109003253A (en) * | 2017-05-24 | 2018-12-14 | 通用电气公司 | Neural network point cloud generates system |
| CN107423700A (en) * | 2017-07-17 | 2017-12-01 | 广州广电卓识智能科技有限公司 | The method and device of testimony verification |
| CN107423700B (en) * | 2017-07-17 | 2020-10-20 | 广州广电卓识智能科技有限公司 | Method and device for verifying testimony of a witness |
| WO2019015466A1 (en) * | 2017-07-17 | 2019-01-24 | 广州广电运通金融电子股份有限公司 | Method and apparatus for verifying person and certificate |
| CN107578377A (en) * | 2017-08-31 | 2018-01-12 | 北京飞搜科技有限公司 | A kind of super-resolution image reconstruction method and system based on deep learning |
| CN107633218B (en) * | 2017-09-08 | 2021-06-08 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating image |
| CN107633218A (en) * | 2017-09-08 | 2018-01-26 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating image |
| US10650492B2 (en) | 2017-09-08 | 2020-05-12 | Baidu Online Network Technology (Beijing) Co., Ltd. | Method and apparatus for generating image |
| CN107590774A (en) * | 2017-09-18 | 2018-01-16 | 北京邮电大学 | A kind of car plate clarification method and device based on generation confrontation network |
| CN107862785A (en) * | 2017-10-16 | 2018-03-30 | 深圳市中钞信达金融科技有限公司 | Bill authentication method and device |
| CN107945204B (en) * | 2017-10-27 | 2021-06-25 | 西安电子科技大学 | Pixel-level image matting method based on generation countermeasure network |
| CN107945204A (en) * | 2017-10-27 | 2018-04-20 | 西安电子科技大学 | A kind of Pixel-level portrait based on generation confrontation network scratches drawing method |
| CN107977511A (en) * | 2017-11-30 | 2018-05-01 | 浙江传媒学院 | A kind of industrial design material high-fidelity real-time emulation algorithm based on deep learning |
| US11704771B2 (en) | 2017-12-01 | 2023-07-18 | Huawei Technologies Co., Ltd. | Training super-resolution convolutional neural network model using a high-definition training image, a low-definition training image, and a mask image |
| WO2019104705A1 (en) * | 2017-12-01 | 2019-06-06 | 华为技术有限公司 | Image processing method and device |
| WO2019114523A1 (en) * | 2017-12-12 | 2019-06-20 | 腾讯科技（深圳）有限公司 | Classification training method, server and storage medium |
| US11017220B2 (en) | 2017-12-12 | 2021-05-25 | Tencent Technology (Shenzhen) Company Limited | Classification model training method, server, and storage medium |
| CN108122209B (en) * | 2017-12-14 | 2020-05-15 | 浙江捷尚视觉科技股份有限公司 | License plate deblurring method based on countermeasure generation network |
| CN108122209A (en) * | 2017-12-14 | 2018-06-05 | 浙江捷尚视觉科技股份有限公司 | A kind of car plate deblurring method based on confrontation generation network |
| CN108122249A (en) * | 2017-12-20 | 2018-06-05 | 长沙全度影像科技有限公司 | A kind of light stream method of estimation based on GAN network depth learning models |
| CN109993694A (en) * | 2017-12-29 | 2019-07-09 | Tcl集团股份有限公司 | A kind of method and device generating super-resolution image |
| CN108182669A (en) * | 2018-01-02 | 2018-06-19 | 华南理工大学 | A kind of Super-Resolution method of the generation confrontation network based on multiple dimension of pictures |
| CN108235058A (en) * | 2018-01-12 | 2018-06-29 | 广州华多网络科技有限公司 | Video quality processing method, storage medium and terminal |
| CN110472457A (en) * | 2018-05-10 | 2019-11-19 | 成都视观天下科技有限公司 | Low-resolution face image identification, restoring method, equipment and storage medium |
| CN108961161B (en) * | 2018-05-24 | 2023-09-22 | 上海商汤智能科技有限公司 | Image data processing method, device and computer storage medium |
| CN108961161A (en) * | 2018-05-24 | 2018-12-07 | 上海商汤智能科技有限公司 | A kind of image processing method, device and computer storage medium |
| CN108805188A (en) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | A kind of feature based recalibration generates the image classification method of confrontation network |
| CN108805188B (en) * | 2018-05-29 | 2020-08-21 | 徐州工程学院 | Image classification method for generating countermeasure network based on feature recalibration |
| CN109118470A (en) * | 2018-06-26 | 2019-01-01 | 腾讯科技（深圳）有限公司 | A kind of image quality evaluating method, device, terminal and server |
| CN109377459B (en) * | 2018-09-30 | 2022-03-18 | 国网智能科技股份有限公司 | Super-resolution deblurring method of generative confrontation network |
| CN109377459A (en) * | 2018-09-30 | 2019-02-22 | 国网山东省电力公司电力科学研究院 | A kind of super-resolution deblurring method of production confrontation network |
| CN109525859A (en) * | 2018-10-10 | 2019-03-26 | 腾讯科技（深圳）有限公司 | Model training, image transmission, image processing method and relevant apparatus equipment |
| CN109525859B (en) * | 2018-10-10 | 2021-01-15 | 腾讯科技（深圳）有限公司 | Model training method, image sending method, image processing method and related device equipment |
| CN109636721A (en) * | 2018-11-29 | 2019-04-16 | 武汉大学 | Video super-resolution method based on confrontation study and attention mechanism |
| CN109636721B (en) * | 2018-11-29 | 2023-06-23 | 武汉大学 | Video super-resolution method based on countermeasure learning and attention mechanism |
| CN109801215A (en) * | 2018-12-12 | 2019-05-24 | 天津津航技术物理研究所 | The infrared super-resolution imaging method of network is generated based on confrontation |
| CN109801215B (en) * | 2018-12-12 | 2020-04-28 | 天津津航技术物理研究所 | Infrared super-resolution imaging method based on countermeasure generation network |
| CN109859113A (en) * | 2018-12-25 | 2019-06-07 | 北京奇艺世纪科技有限公司 | Model generating method, image enchancing method, device and computer readable storage medium |
| CN109697470B (en) * | 2018-12-27 | 2021-02-09 | 百度在线网络技术（北京）有限公司 | Recognition model training method and device, electronic equipment and storage medium |
| CN109697470A (en) * | 2018-12-27 | 2019-04-30 | 百度在线网络技术（北京）有限公司 | A kind of training method of identification model, device, electronic equipment and storage medium |
| CN109949219B (en) * | 2019-01-12 | 2021-03-26 | 深圳先进技术研究院 | Reconstruction method, device and equipment of super-resolution image |
| CN109949219A (en) * | 2019-01-12 | 2019-06-28 | 深圳先进技术研究院 | A kind of reconstructing method of super-resolution image, device and equipment |
| CN109934102B (en) * | 2019-01-28 | 2020-11-27 | 浙江理工大学 | Finger vein identification method based on image super-resolution |
| CN109934102A (en) * | 2019-01-28 | 2019-06-25 | 浙江理工大学 | A kind of finger vein identification method based on image super-resolution |
| CN109816764A (en) * | 2019-02-02 | 2019-05-28 | 深圳市商汤科技有限公司 | Image generating method and device, electronic equipment and storage medium |
| US11410344B2 (en) | 2019-02-02 | 2022-08-09 | Shenzhen Sensetime Technology Co., Ltd. | Method for image generation, electronic device, and storage medium |
| JP2021518961A (en) * | 2019-02-02 | 2021-08-05 | 深▲せん▼市商▲湯▼科技有限公司Ｓｈｅｎｚｈｅｎ Ｓｅｎｓｅｔｉｍｅ Ｔｅｃｈｎｏｌｏｇｙ Ｃｏ．， Ｌｔｄ． | Image generation methods and devices, electronic devices and storage media |
| CN110580682A (en) * | 2019-09-16 | 2019-12-17 | 电子科技大学 | Countermeasure network seismic data super-resolution reconstruction method based on optimization generation |
| CN111402133A (en) * | 2020-03-13 | 2020-07-10 | 北京字节跳动网络技术有限公司 | Image processing method, image processing device, electronic equipment and computer readable medium |
| CN113628121A (en) * | 2020-05-06 | 2021-11-09 | 阿里巴巴集团控股有限公司 | Method and device for processing data and training multimedia data |
| CN113628121B (en) * | 2020-05-06 | 2023-11-14 | 阿里巴巴集团控股有限公司 | Method and device for processing and training multimedia data |
| US11823349B2 (en) | 2020-11-19 | 2023-11-21 | Samsung Electronics Co., Ltd. | Image generators with conditionally-independent pixel synthesis |
| KR20220131808A (en) * | 2021-03-22 | 2022-09-29 | 재단법인대구경북과학기술원 | Method and apparatus for generating imaga classification model |
| KR102611121B1 (en) | 2021-03-22 | 2023-12-07 | 재단법인대구경북과학기술원 | Method and apparatus for generating imaga classification model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106683048B (en) | 2020-09-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106683048A (en) | Image super-resolution method and image super-resolution equipment | |
| Xiao et al. | Satellite video super-resolution via multiscale deformable convolution alignment and temporal grouping projection | |
| EP3540637A1 (en) | Neural network model training method, device and storage medium for image processing | |
| CN105069746B (en) | Video real-time face replacement method and its system based on local affine invariant and color transfer technology | |
| CN108334847A (en) | A kind of face identification method based on deep learning under real scene | |
| CN103824272B (en) | The face super-resolution reconstruction method heavily identified based on k nearest neighbor | |
| CN111209811B (en) | Method and system for detecting eyeball attention position in real time | |
| CN114898284B (en) | Crowd counting method based on feature pyramid local difference attention mechanism | |
| CN109191511A (en) | A kind of binocular solid matching process based on convolutional neural networks | |
| CN114022506B (en) | Image restoration method for edge prior fusion multi-head attention mechanism | |
| CN114463759A (en) | Lightweight character detection method and device based on anchor-frame-free algorithm | |
| CN111666852A (en) | Micro-expression double-flow network identification method based on convolutional neural network | |
| CN115661459A (en) | 2D mean teacher model using difference information | |
| CN116563916A (en) | Attention fusion-based cyclic face super-resolution method and system | |
| Ma et al. | Multi-task interaction learning for spatiospectral image super-resolution | |
| CN115661530A (en) | Image data enhancement method and system and electronic equipment | |
| CN111753670A (en) | Human face overdividing method based on iterative cooperation of attention restoration and key point detection | |
| CN104091364B (en) | Single-image super-resolution reconstruction method | |
| CN111882516A (en) | Image quality evaluation method based on visual saliency and deep neural network | |
| CN114943646A (en) | Gradient weight loss and attention mechanism super-resolution method based on texture guidance | |
| CN117291803B (en) | PAMGAN lightweight facial super-resolution reconstruction method | |
| CN117456330A (en) | MSFAF-Net-based low-illumination target detection method | |
| CN111401209B (en) | Action recognition method based on deep learning | |
| CN115578260B (en) | Attention method and system for directional decoupling of image super-resolution | |
| CN111368831A (en) | System and method for positioning vertically arranged characters |
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 | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20221116 Address after: 15 / F, Zhongchuang Plaza, 385 Hangtian Middle Road, national civil aerospace industrial base, Xi'an City, Shaanxi Province 710100 Patentee after: Xi'an Yu vision Mdt InfoTech Ltd. Address before: 2 / F, South Block, building 10, wanlun Science Park, 88 Jiangling Road, Binjiang District, Hangzhou City, Zhejiang Province, 310051 Patentee before: ZHEJIANG UNIVIEW TECHNOLOGIES Co.,Ltd. |