CN109978165A - A kind of generation confrontation network method merged from attention mechanism - Google Patents
A kind of generation confrontation network method merged from attention mechanism Download PDFInfo
- Publication number
- CN109978165A CN109978165A CN201910272879.8A CN201910272879A CN109978165A CN 109978165 A CN109978165 A CN 109978165A CN 201910272879 A CN201910272879 A CN 201910272879A CN 109978165 A CN109978165 A CN 109978165A
- Authority
- CN
- China
- Prior art keywords
- attention
- arbiter
- characteristic pattern
- generator
- generation
- 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
- 238000000034 method Methods 0.000 title claims abstract description 37
- 230000007246 mechanism Effects 0.000 title claims abstract description 34
- 238000013528 artificial neural network Methods 0.000 claims abstract description 13
- 238000011156 evaluation Methods 0.000 claims abstract description 5
- 230000004927 fusion Effects 0.000 claims abstract description 5
- 238000012549 training Methods 0.000 claims description 35
- 238000009826 distribution Methods 0.000 claims description 22
- 230000008569 process Effects 0.000 claims description 17
- 238000004364 calculation method Methods 0.000 claims description 8
- 238000005457 optimization Methods 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 6
- 230000015556 catabolic process Effects 0.000 claims description 5
- 238000006731 degradation reaction Methods 0.000 claims description 5
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000013473 artificial intelligence Methods 0.000 abstract 1
- 230000006870 function Effects 0.000 description 16
- 238000010586 diagram Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000004519 manufacturing process Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 210000004218 nerve net Anatomy 0.000 description 2
- 238000005303 weighing Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 239000013067 intermediate product Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 239000000047 product Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of fusions to fight network method from the generation of attention mechanism, belongs to computer vision field, especially relates to the generation confrontation network for carrying out image generation.The generation of image is a significant challenge of computer vision field, if it is possible to generate a large amount of high quality graphic sample, rely on the epoch under big data background at this, artificial intelligence field can obtain more rapid development.Therefore, the present invention proposes that a kind of generation merged from attention mechanism fights network, which can be generated the image of high quality, while image and diversity with higher.Specifically, generating confrontation network uses Wasserstein distance instead for the evaluation criteria that generator is distributed with arbiter to measure, loss function is correspondingly improved;It introduces in generator neural network framework corresponding with arbiter from attention mechanism simultaneously, improves the relevance generated between image local pixel region, thus improve the quality for generating image.
Description
Technical field
The invention belongs to computer vision fields, are related to a kind of generation confrontation network method merged from attention mechanism.
Background technique
In recent years, it rises in this world using neural network as the depth learning technology of core in computer vision field, nerve net
Discriminative model in network has been applied to solve such as, the bases such as text description of image recognition, image classification and image
Problem;However, the production model for generating image data but faces modeling process difficulty height, the effect lack of resolution is generated
The problems such as, these reasons cause to be difficult to be applied in image generation field using production model.To solve this problem,
Production model and discriminative model are combined together the training that confrontation type is carried out to image data, solve modeling process difficulty
The quality for generating image data is improved while problem again, this scheme is referred to as to generate confrontation network.Recent years, big number
A large amount of data are relied on according to the deep learning under environment to be trained, and can generate the generation confrontation of a large amount of high image quality degradations
Network is necessary.
Generate arbiter two parts group of the confrontation network by the generator of production neural network, with discriminate neural network
At.Wherein, the training objective of arbiter is to improve its discrimination capabilities to true picture by training, is improved to true picture
Score is reduced to the score for generating image.The training objective of generator is to improve the quality for generating image data by training,
Allow the image data of generation that can obtain higher score in arbiter.Generating confrontation network training process mainly includes two
Stage, first stage carry out the training of true picture discrimination capabilities, second stage is to life to arbiter input image data
It grows up to be a useful person and is trained, improve it and generate score of the image data in arbiter, two step cycles carry out, when arbiter can not
When making accurate judgement to the image data that generator generates, it is flat that we assert that the training for generating confrontation network has reached stable state
Weighing apparatus.
There is mode collapsing, the single lack of diversity of sample mode in previous generation confrontation network technology.Come in detail
It says, generates confrontation network by the training of confrontation type and generator is generated into the data distribution of image to the data distribution of true picture
It draws over to one's side, and the data distribution of true picture is difficult to obtain, then is obtained by using real image data training arbiter close to true
Real image data distribution is originally generated confrontation network technology and describes the standard of the distance between two kinds of image data distributions to hand over
Entropy, that is, JS divergence is pitched, when the two is distributed in trained initial stage there is no when intersection, the training result using JS divergence is poor, can not
Effective gradient is provided for the training of generator so as to cause mode collapsing.In addition to this, previous uncontrollable generation fights net
Network is difficult to generate pixel clearly image, and tracing it to its cause is only to forgive convolutional layer in its network architecture, cannot obtain whole image
Relationship between middle local pixel region, thus the image generated has complete image outline but clarity is not high and details lacks
It loses.Continue a kind of generation confrontation for containing new data distribution and generating image definition mechanism apart from evaluation criteria and raising
Network plan.
Summary of the invention
Network method is fought from the generation of attention mechanism in view of this, the purpose of the present invention is to provide a kind of fusions.
In order to achieve the above objectives, the invention provides the following technical scheme:
A kind of generation confrontation network method merged from attention mechanism, method includes the following steps:
S1: replacing the standard of difference between assessment generator and arbiter data distribution, uses
Wasserstein distance assesses the difference between the two data distribution;Evaluation criteria uses Wasserstein distance instead
Afterwards, loss function also improves, and finally can be improved the diversity for generating image data;
S2: the generation confrontation network from attention mechanism is being merged in its generator neural network corresponding with arbiter
It introduces in framework from attention mechanism;
S3: it in having merged the generation confrontation network from attention mechanism, runs from attention mechanism.
Further, in the step S2, the training process of data includes 4 stages:
Stage 1 loses letter using the difference between Wasserstein distance assessment generator and arbiter data distribution
Number is changed accordingly, and the image data of training set is divided into several batches, sequentially inputs in arbiter and is trained;
Stage 2, using Adam optimization method according to loss function value calculated result, generates after a wheel data training
The weight of device G remains unchanged, and the weight for carrying out arbiter D updates;
Stage 3 samples several random noise variables, carries out the weight in generator G according to loss function and updates;
In the stage 4, the weight θ of G tends to before convergence in generator, the process in 1~stage of cycle stage 3, and training terminates
Generator and arbiter reach the Nash Equilibrium of stable state afterwards, and generator has the ability for generating high image quality degradation at this time.
Further, the S3 specifically includes step:
S31: the characteristic pattern of input carries out the conversion of feature space, and same characteristic pattern is converted respectively to two specifically
In mapping space, while retaining former characteristic pattern;Characteristic pattern x ∈ R from a upper hidden layerC×NIt is first converted to two specifically
Force value is paid attention to calculate in feature space f and g, wherein f (x)=WfX, g (x)=Wgx;
S32: pay attention to according to from each pixel of attention calculation formula to characteristic pattern behind converting characteristic space
Force value calculates, and final result is weighted summation and obtains a corresponding attention characteristic pattern, and calculation formula is as follows:
And sij=f (xi)Tg(xj),
And h (xi)=Whxi,
βj,iWhen value represents j-th of the region in formation zone, model to the attention degree of ith zone, the step for obtain
Attention characteristic pattern be o=(o1,o2,o3,...,oN)∈RC×N;
S33: attention characteristic pattern and former characteristic pattern are added up and is obtained from attention characteristic pattern, yi=γ oi+xi, wherein y be
From the final output of attention layer, γ is initialized as 0;In above-mentioned calculating,Wh∈RC×CBeing can
The weight matrix of study, by 1 × 1 convolution algorithm realization, wherein
The figure for ultimately generating corresponding high quality is trained to input data set the beneficial effects of the present invention are: the present invention
Picture.Image in data set is tailored to same size, several images are that a batch inputs in the correspondence stage of training process
It generates and is trained in generator and arbiter in confrontation network, then terminated when the loss function value of arbiter tends to balance
Training.Finally obtained generator can carry out high quality graphic generation, can improve simultaneously image with problem-solving pattern collapsing problem
The clarity of details.
Other advantages, target and feature of the invention will be illustrated in the following description to a certain extent, and
And to a certain extent, based on will be apparent to those skilled in the art to investigating hereafter, Huo Zheke
To be instructed from the practice of the present invention.Target of the invention and other advantages can be realized by following specification and
It obtains.
Detailed description of the invention
To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention is made below in conjunction with attached drawing excellent
The detailed description of choosing, in which:
Fig. 1 is the generator G block diagram merged from attention mechanism;
Fig. 2 is the arbiter D block diagram merged from attention mechanism;
Fig. 3 is the system block diagram from attention mechanism.
Specific embodiment
Illustrate embodiments of the present invention below by way of specific specific example, those skilled in the art can be by this specification
Other advantages and efficacy of the present invention can be easily understood for disclosed content.The present invention can also pass through in addition different specific realities
The mode of applying is embodied or practiced, the various details in this specification can also based on different viewpoints and application, without departing from
Various modifications or alterations are carried out under spirit of the invention.It should be noted that diagram provided in following embodiment is only to show
Meaning mode illustrates basic conception of the invention, and in the absence of conflict, the feature in following embodiment and embodiment can phase
Mutually combination.
Wherein, the drawings are for illustrative purposes only and are merely schematic diagrams, rather than pictorial diagram, should not be understood as to this
The limitation of invention;Embodiment in order to better illustrate the present invention, the certain components of attached drawing have omission, zoom in or out, not
Represent the size of actual product;It will be understood by those skilled in the art that certain known features and its explanation may be omitted and be in attached drawing
It is understood that.
The same or similar label correspond to the same or similar components in the attached drawing of the embodiment of the present invention;It is retouched in of the invention
In stating, it is to be understood that if there is the orientation or positional relationship of the instructions such as term " on ", "lower", "left", "right", "front", "rear"
To be based on the orientation or positional relationship shown in the drawings, be merely for convenience of description of the present invention and simplification of the description, rather than indicate or
It implies that signified device or element must have a particular orientation, be constructed and operated in a specific orientation, therefore is described in attached drawing
The term of positional relationship only for illustration, is not considered as limiting the invention, for the ordinary skill of this field
For personnel, the concrete meaning of above-mentioned term can be understood as the case may be.
Before introducing plan content, 7 necessary concepts in invention are first stated.
1st concept: data distribution, the i.e. probability distribution of image data.Since true image data probability distribution is difficult
Often sample distribution possessed by training dataset it is approximately true data distribution in deep learning to obtain, thus instructs
The data sample practiced in data set cannot be very little, and needs representative.
2nd concept: neural network, the present invention in be convolutional neural networks, be the core technology in deep learning, warp
Crossing training can carry out simultaneously feature extraction with pattern classification, more have in terms of image procossing compared to other machines learning art
Advantage.In addition to this, generating the generator neural network G in confrontation network is deconvolution neural network, arbiter neural network D
For convolutional neural networks.
3rd concept: random noise variable is the input z for generating generator in confrontation network, obeys standard normal point
Cloth, i.e. z~N (0,1), the pseudo- image ultimately generated are G (z).
4th concept: convolutional layer is the important composition of convolutional neural networks, if generate confrontation network framework in by
Dry convolution kernel composition, has the local sensing domain for image data, multilayer convolutional layer can extract picture number from low to high
According to feature.
5th concept: characteristic pattern, the data, that is, x ∈ R exported after being handled by upper one layer of convolutional layerC×N, it is convolutional Neural
The intermediate product of network.
6th concept: loss function, in deep learning, loss function be for measure prediction result and actual value it
Between difference size index.Generate the standard loss function of confrontation network are as follows:
Wherein PrThe true distribution of data representing image x, PzRepresent the distribution of random noise variable z, the input of arbiter D
There are real image data x and random noise z simultaneously.
7th concept: optimization method, for training algorithm used in deep learning model, the selection pair of optimization method
The performance of model has significant impact.Stochastic gradient descent algorithm is one of the basic skills of trained deep learning model, but
Learning process is slow, cannot automatically adjust learning rate.It thus generates in confrontation network often using the calculation for capableing of autoadapted learning rate
Method, variable learning rate arithmetic have RMSprop and Adam algorithm etc..
A kind of generation confrontation network merged from attention mechanism, using real image data as training set, after training
The neural network of the image of corresponding high quality can be generated.Wherein, the network architecture of generator G is as shown in Figure 1, arbiter D
The network architecture is as shown in Fig. 2, its training process is divided into four-stage, and generator reaches receiving for stable state with arbiter after training
Assorted equilibrium, generator has the ability for generating high image quality degradation at this time.In addition to this, generation of the invention confrontation network exists
It has been merged between generator and the convolutional layer of arbiter neural network from attention mechanism, has generated quality to improve image data,
From running in attention mechanism dependency graph 3 from attention layer, process has forgiven three phases, finally obtained from attention
Input of the characteristic pattern as subsequent convolutional layer.
Specifically, the present invention provides a kind of generation confrontation network concrete schemes merged from attention mechanism such as
Under:
It is replaced, is used firstly, for the standard of difference between assessment generator and arbiter data distribution
Wasserstein distance assesses the difference between the two data distribution.Evaluation criteria uses Wasserstein distance instead
Afterwards, loss function is also improved, and finally can be improved the diversity for generating image data.
Further, a kind of generation confrontation network merged from attention mechanism of the present invention is in its generator and arbiter
It is introduced in corresponding neural network framework from attention mechanism, the training process of data includes 4 stages: the 1st stage, figure
As data are divided into several batches, sequentially input into arbiter and calculate loss function value and be trained;2nd stage, a wheel number
After according to training, using Adam optimization method according to loss function value calculated result, the weight of generator G is remained unchanged, into
The weight of row arbiter D updates;It 3rd stage, several random noise variables of sampling, is carried out in generator G according to loss function
Weight update;4th stage, the weight θ of G tends to before convergence in generator, the process in stage cycle stage 1- 3, instruction
Generator and arbiter reach the Nash Equilibrium of stable state after white silk, and generator has the energy for generating high image quality degradation at this time
Power.
Further, the present invention is a kind of has merged from transporting in the generation confrontation network of attention mechanism from attention mechanism
Include three steps when row: step 1, the characteristic pattern inputted carry out the conversion of feature space, and same characteristic pattern is converted respectively
Into two specific mapping spaces, while retaining former characteristic pattern;Step 2, according to from attention calculation formula to converting characteristic
Behind space each pixel of characteristic pattern carry out pay attention to force value calculate, final result be weighted summation obtain one it is corresponding
Attention characteristic pattern;Attention characteristic pattern is added to obtain from attention characteristic pattern by step 3 with former characteristic pattern.
Firstly, scheme, which uses Wasserstein distance instead, measures the true difference being distributed between model profile, each training in rotation
Practice the value by reducing loss function to a certain extent to draw in the distance between the two distribution, loss function is as follows:
Wherein, the input of generator G is random noise variable z~Pz, c is on true picture x and puppet image G (z) line
The sampling of random difference, arbiter D meets 1-Lipschitz condition, and training process updates nerve net using Adam optimization method
Network weight.
The training process merged from the generation confrontation network of attention mechanism is divided into the following four stage:
1st stage: image data is divided into N batch, and every batch of includes m image patterns, is input in arbiter and calculates
Loss function value, so as to the update of weight in subsequent progress arbiter.Pseudo- image is obtained after random noise variable input generatorPseudo- image and true picture sample are carried outTransformation after input arbiter in, damaged
The calculating of functional value is lost,
After the completion of 2nd stage, a wheel data training, according to obtaining loss function value using Adam optimization method, to sentencing
Weight in other device is updated
It m 3rd stage, sampling noise variation, carries out weighing in generator under the action of arbiter according to loss function
The update of weight.
4th stage, the weight θ in generator tend to before convergence, the process in stage cycle stage 1- 3, most throughout one's life
It grows up to be a useful person and reaches the Nash Equilibrium of stable state with arbiter, generator can generate the higher image data of quality.
The present invention is a kind of merged from the generation confrontation network of attention mechanism from attention mechanism operational process packet
Include three steps:
Step 1, input characteristic pattern carry out feature space conversion, same characteristic pattern converted respectively to two it is specific
Mapping space in, while retaining former characteristic pattern;Characteristic pattern x ∈ R from a upper hidden layerC×NBe first converted to two it is specific
Feature space f and g in calculate pay attention to force value, wherein f (x)=WfX, g (x)=Wgx。
Step 2 is infused according to from each pixel of attention calculation formula to characteristic pattern behind converting characteristic space
Force value of anticipating calculates, and final result is weighted summation and obtains a corresponding attention characteristic pattern, and calculation formula is as follows:
And sij=f (xi)Tg(xj),
And h (xi)=Whxi,
βj,iWhen value represents j-th of the region in formation zone, model to the attention degree of ith zone, the step for obtain
Attention characteristic pattern be o=(o1,o2,o3,...,oN)∈RC×N。
Attention characteristic pattern and former characteristic pattern are added up and are obtained from attention characteristic pattern, y by step 3i=γ oi+xi, wherein y
For from the final output of attention layer, γ is initialized as 0.In above-mentioned calculating,Wh∈RC×CIt is
The weight matrix that can learn, by 1 × 1 convolution algorithm realization, wherein
Finally, it is stated that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting, although referring to compared with
Good embodiment describes the invention in detail, those skilled in the art should understand that, it can be to skill of the invention
Art scheme is modified or replaced equivalently, and without departing from the objective and range of the technical program, should all be covered in the present invention
Scope of the claims in.
Claims (3)
1. a kind of fusion fights network method from the generation of attention mechanism, it is characterised in that: method includes the following steps:
S1: the standard of difference between assessment generator and arbiter data distribution is replaced, Wasserstein is used
Distance assesses the difference between the two data distribution;After evaluation criteria uses Wasserstein distance instead, loss function
It improves, finally can be improved the diversity for generating image data;
S2: the generation confrontation network from attention mechanism is being merged in its generator neural network framework corresponding with arbiter
In introduce from attention mechanism;
S3: it in having merged the generation confrontation network from attention mechanism, runs from attention mechanism.
2. a kind of fusion according to claim 1 fights network method from the generation of attention mechanism, it is characterised in that: institute
It states in step S2, the training process of data includes 4 stages:
Stage 1, using Wasserstein distance assessment generator and arbiter data distribution between difference, loss function into
Row is corresponding to be changed, and the image data of training set is divided into several batches, sequentially inputs in arbiter and is trained;
Stage 2, after a wheel data training, using Adam optimization method according to loss function value calculated result, generator G's
Weight remains unchanged, and the weight for carrying out arbiter D updates;
Stage 3 samples several random noise variables, carries out the weight in generator G according to loss function and updates;
In the stage 4, the weight θ of G tends to before convergence in generator, the process in 1~stage of cycle stage 3, raw after training
It grows up to be a useful person and reaches the Nash Equilibrium of stable state with arbiter, generator has the ability for generating high image quality degradation at this time.
3. a kind of fusion according to claim 1 fights network method from the generation of attention mechanism, it is characterised in that: institute
It states S3 and specifically includes step:
S31: the characteristic pattern of input carries out the conversion of feature space, and same characteristic pattern is converted respectively to two specific mappings
In space, while retaining former characteristic pattern;Characteristic pattern x ∈ R from a upper hidden layerC×NIt is first converted to two specific features
Force value is paid attention to calculate in space f and g, wherein f (x)=WfX, g (x)=Wgx;
S32: attention force value is carried out according to from each pixel of attention calculation formula to characteristic pattern behind converting characteristic space
It calculates, final result is weighted summation and obtains a corresponding attention characteristic pattern, and calculation formula is as follows:
And sij=f (xi)Tg(xj),
And h (xi)=Whxi,
βj,iWhen value represents j-th of the region in formation zone, model to the attention degree of ith zone, the step for obtained note
Meaning power characteristic pattern is o=(o1,o2,o3,...,oN)∈RC×N;
S33: attention characteristic pattern and former characteristic pattern are added up and is obtained from attention characteristic pattern, yi=γ oi+xi, wherein y is from note
The final output of meaning power layer, γ are initialized as 0;In above-mentioned calculating,Wh∈RC×CIt is that can learn
Weight matrix, realized by 1 × 1 convolution algorithm, wherein
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910272879.8A CN109978165A (en) | 2019-04-04 | 2019-04-04 | A kind of generation confrontation network method merged from attention mechanism |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910272879.8A CN109978165A (en) | 2019-04-04 | 2019-04-04 | A kind of generation confrontation network method merged from attention mechanism |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109978165A true CN109978165A (en) | 2019-07-05 |
Family
ID=67083229
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910272879.8A Pending CN109978165A (en) | 2019-04-04 | 2019-04-04 | A kind of generation confrontation network method merged from attention mechanism |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109978165A (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110399712A (en) * | 2019-07-31 | 2019-11-01 | 网易(杭州)网络有限公司 | Validation-cross method, apparatus, medium and calculating equipment based on identifying code |
CN110458829A (en) * | 2019-08-13 | 2019-11-15 | 腾讯医疗健康(深圳)有限公司 | Image quality control method, device, equipment and storage medium based on artificial intelligence |
CN110555458A (en) * | 2019-07-24 | 2019-12-10 | 中北大学 | Multi-band image feature level fusion method for generating countermeasure network based on attention mechanism |
CN110633655A (en) * | 2019-08-29 | 2019-12-31 | 河南中原大数据研究院有限公司 | Attention-attack face recognition attack algorithm |
CN110659727A (en) * | 2019-09-24 | 2020-01-07 | 中国科学技术大学 | Sketch-based image generation method |
CN110827216A (en) * | 2019-10-23 | 2020-02-21 | 上海理工大学 | Multi-generator generation countermeasure network learning method for image denoising |
CN111079509A (en) * | 2019-10-23 | 2020-04-28 | 西安电子科技大学 | Abnormal behavior detection method based on self-attention mechanism |
CN111091059A (en) * | 2019-11-19 | 2020-05-01 | 佛山市南海区广工大数控装备协同创新研究院 | Data equalization method in household garbage plastic bottle classification |
CN111275613A (en) * | 2020-02-27 | 2020-06-12 | 辽宁工程技术大学 | Editing method for generating confrontation network face attribute by introducing attention mechanism |
CN112001404A (en) * | 2020-08-25 | 2020-11-27 | 华中农业大学 | Image generation model and method for self-adaptive global and local double-layer optimization |
CN112365464A (en) * | 2020-11-09 | 2021-02-12 | 成都信息工程大学 | GAN-based medical image lesion area weak supervision positioning method |
CN113221757A (en) * | 2021-05-14 | 2021-08-06 | 上海交通大学 | Method, terminal and medium for improving accuracy rate of pedestrian attribute identification |
CN113537247A (en) * | 2021-08-13 | 2021-10-22 | 重庆大学 | Data enhancement method for converter transformer vibration signal |
CN113822895A (en) * | 2021-08-29 | 2021-12-21 | 陕西师范大学 | ScanSAR image scallop effect suppression method based on self-attention mechanism and cycleGAN |
CN113935977A (en) * | 2021-10-22 | 2022-01-14 | 河北工业大学 | Solar cell panel defect generation method based on generation countermeasure network |
CN114372554A (en) * | 2021-11-28 | 2022-04-19 | 国网辽宁省电力有限公司鞍山供电公司 | Flexible load scene generation and reduction method based on generation countermeasure network |
CN114549608A (en) * | 2022-04-22 | 2022-05-27 | 季华实验室 | Point cloud fusion method and device, electronic equipment and storage medium |
CN114936168A (en) * | 2022-07-21 | 2022-08-23 | 南京争锋信息科技有限公司 | Automatic test case generation method in real user intelligent perception system |
CN115035353A (en) * | 2022-08-11 | 2022-09-09 | 粤港澳大湾区数字经济研究院(福田) | Image classification method, image classification model, intelligent terminal and storage medium |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109544442A (en) * | 2018-11-12 | 2019-03-29 | 南京邮电大学 | The image local Style Transfer method of production confrontation network based on dual confrontation |
-
2019
- 2019-04-04 CN CN201910272879.8A patent/CN109978165A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109544442A (en) * | 2018-11-12 | 2019-03-29 | 南京邮电大学 | The image local Style Transfer method of production confrontation network based on dual confrontation |
Non-Patent Citations (3)
Title |
---|
HAN ZHANG等: "Self-Attention Generative Adversarial Networks", 《ARXIV》 * |
ISHAAN GULRAJANI等: "Improved Training of Wasserstein GANs20", 《ARXIV》 * |
罗佳等: "生成式对抗网络研究综述", 《仪器仪表学报》 * |
Cited By (31)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110555458A (en) * | 2019-07-24 | 2019-12-10 | 中北大学 | Multi-band image feature level fusion method for generating countermeasure network based on attention mechanism |
CN110555458B (en) * | 2019-07-24 | 2022-04-19 | 中北大学 | Multi-band image feature level fusion method for generating countermeasure network based on attention mechanism |
CN110399712B (en) * | 2019-07-31 | 2022-03-22 | 杭州网易智企科技有限公司 | Interactive verification method, device, medium and computing equipment based on verification code |
CN110399712A (en) * | 2019-07-31 | 2019-11-01 | 网易(杭州)网络有限公司 | Validation-cross method, apparatus, medium and calculating equipment based on identifying code |
CN110458829A (en) * | 2019-08-13 | 2019-11-15 | 腾讯医疗健康(深圳)有限公司 | Image quality control method, device, equipment and storage medium based on artificial intelligence |
CN110458829B (en) * | 2019-08-13 | 2024-01-30 | 腾讯医疗健康(深圳)有限公司 | Image quality control method, device, equipment and storage medium based on artificial intelligence |
CN110633655A (en) * | 2019-08-29 | 2019-12-31 | 河南中原大数据研究院有限公司 | Attention-attack face recognition attack algorithm |
CN110659727B (en) * | 2019-09-24 | 2022-05-13 | 中国科学技术大学 | Sketch-based image generation method |
CN110659727A (en) * | 2019-09-24 | 2020-01-07 | 中国科学技术大学 | Sketch-based image generation method |
CN111079509A (en) * | 2019-10-23 | 2020-04-28 | 西安电子科技大学 | Abnormal behavior detection method based on self-attention mechanism |
CN111079509B (en) * | 2019-10-23 | 2023-03-24 | 西安电子科技大学 | Abnormal behavior detection method based on self-attention mechanism |
WO2021077997A1 (en) * | 2019-10-23 | 2021-04-29 | 上海理工大学 | Multi-generator generative adversarial network learning method for image denoising |
CN110827216B (en) * | 2019-10-23 | 2023-07-14 | 上海理工大学 | Multi-generator generation countermeasure network learning method for image denoising |
US11308587B2 (en) | 2019-10-23 | 2022-04-19 | University Of Shanghai For Science And Technology | Learning method of generative adversarial network with multiple generators for image denoising |
CN110827216A (en) * | 2019-10-23 | 2020-02-21 | 上海理工大学 | Multi-generator generation countermeasure network learning method for image denoising |
CN111091059A (en) * | 2019-11-19 | 2020-05-01 | 佛山市南海区广工大数控装备协同创新研究院 | Data equalization method in household garbage plastic bottle classification |
CN111275613A (en) * | 2020-02-27 | 2020-06-12 | 辽宁工程技术大学 | Editing method for generating confrontation network face attribute by introducing attention mechanism |
CN112001404A (en) * | 2020-08-25 | 2020-11-27 | 华中农业大学 | Image generation model and method for self-adaptive global and local double-layer optimization |
CN112365464A (en) * | 2020-11-09 | 2021-02-12 | 成都信息工程大学 | GAN-based medical image lesion area weak supervision positioning method |
CN112365464B (en) * | 2020-11-09 | 2021-08-10 | 成都信息工程大学 | GAN-based medical image lesion area weak supervision positioning method |
CN113221757A (en) * | 2021-05-14 | 2021-08-06 | 上海交通大学 | Method, terminal and medium for improving accuracy rate of pedestrian attribute identification |
CN113221757B (en) * | 2021-05-14 | 2022-09-02 | 上海交通大学 | Method, terminal and medium for improving accuracy rate of pedestrian attribute identification |
CN113537247A (en) * | 2021-08-13 | 2021-10-22 | 重庆大学 | Data enhancement method for converter transformer vibration signal |
CN113822895A (en) * | 2021-08-29 | 2021-12-21 | 陕西师范大学 | ScanSAR image scallop effect suppression method based on self-attention mechanism and cycleGAN |
CN113935977A (en) * | 2021-10-22 | 2022-01-14 | 河北工业大学 | Solar cell panel defect generation method based on generation countermeasure network |
CN114372554A (en) * | 2021-11-28 | 2022-04-19 | 国网辽宁省电力有限公司鞍山供电公司 | Flexible load scene generation and reduction method based on generation countermeasure network |
CN114372554B (en) * | 2021-11-28 | 2024-07-12 | 国网辽宁省电力有限公司鞍山供电公司 | Flexible load scene generation and reduction method based on generation countermeasure network |
CN114549608A (en) * | 2022-04-22 | 2022-05-27 | 季华实验室 | Point cloud fusion method and device, electronic equipment and storage medium |
CN114936168A (en) * | 2022-07-21 | 2022-08-23 | 南京争锋信息科技有限公司 | Automatic test case generation method in real user intelligent perception system |
CN115035353B (en) * | 2022-08-11 | 2022-12-23 | 粤港澳大湾区数字经济研究院(福田) | Image classification method, image classification model, intelligent terminal and storage medium |
CN115035353A (en) * | 2022-08-11 | 2022-09-09 | 粤港澳大湾区数字经济研究院(福田) | Image classification method, image classification model, intelligent terminal and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109978165A (en) | A kind of generation confrontation network method merged from attention mechanism | |
CN113469356B (en) | Improved VGG16 network pig identity recognition method based on transfer learning | |
CN110097609B (en) | Sample domain-based refined embroidery texture migration method | |
CN110097185A (en) | A kind of Optimized model method and application based on generation confrontation network | |
CN107292813A (en) | A kind of multi-pose Face generation method based on generation confrontation network | |
CN109584337A (en) | A kind of image generating method generating confrontation network based on condition capsule | |
CN108062551A (en) | A kind of figure Feature Extraction System based on adjacency matrix, figure categorizing system and method | |
CN107679522A (en) | Action identification method based on multithread LSTM | |
CN106599789A (en) | Video class identification method and device, data processing device and electronic device | |
CN110197514A (en) | A kind of mushroom phenotype image generating method based on production confrontation network | |
CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
CN106570522A (en) | Object recognition model establishment method and object recognition method | |
CN106203625A (en) | A kind of deep-neural-network training method based on multiple pre-training | |
CN108830913A (en) | Semantic level line original text painting methods based on User Colors guidance | |
CN108932517A (en) | A kind of multi-tag clothes analytic method based on fining network model | |
CN106485259A (en) | A kind of image classification method based on high constraint high dispersive principal component analysiss network | |
CN110070116A (en) | Segmented based on the tree-shaped Training strategy of depth selects integrated image classification method | |
CN110188449A (en) | Interpretable clothing information recommended method, system, medium and equipment based on attribute | |
CN108596256A (en) | One kind being based on RGB-D object identification grader building methods | |
Khanum et al. | Two new improved variants of grey wolf optimizer for unconstrained optimization | |
CN110210371A (en) | A kind of aerial hand-written inertia sensing signal creating method based on depth confrontation study | |
CN112906888A (en) | Task execution method and device, electronic equipment and storage medium | |
CN110188621B (en) | Three-dimensional facial expression recognition method based on SSF-IL-CNN | |
CN104376585B (en) | Non-convex compressed sensing image reconstructing method based on picture block structure property policy | |
CN109598771A (en) | A kind of landform synthetic method of more geomorphic feature constraints |
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: 20190705 |
|
RJ01 | Rejection of invention patent application after publication |