CN109543742A - A kind of image local information transfer method based on GAN and Self-Attention - Google Patents
A kind of image local information transfer method based on GAN and Self-Attention Download PDFInfo
- Publication number
- CN109543742A CN109543742A CN201811368715.7A CN201811368715A CN109543742A CN 109543742 A CN109543742 A CN 109543742A CN 201811368715 A CN201811368715 A CN 201811368715A CN 109543742 A CN109543742 A CN 109543742A
- Authority
- CN
- China
- Prior art keywords
- image
- feature
- network
- self
- attention
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The image local information transfer method based on GAN and Self-Attention that the present invention relates to a kind of, comprising the following steps: 1) obtain image and migrate task；2) task is migrated according to described image, carries out image local information transfer using a trained neural network, the image after output migration；The neural network includes image composer and arbiter, wherein, described image generator includes convolutional network and deconvolution network, the convolutional network is by picture convolution at hiding feature space, part is hidden into feature exchange in conjunction with Self-Attention model, the feature after exchange realizes that image local information characteristics are migrated by the picture of the deconvolution network deconvolution Cheng Xin；The arbiter is used to differentiate the true and false relationship for generating picture.Compared with prior art, the present invention has many advantages, such as to avoid the influence of local migration and image overall pixel, solves image local migration to the disturbing factor of general image.
Description
Technical field
The present invention relates to deep learning fields, are related to a kind of image information processing method, are based on GAN more particularly, to one kind
With the image local information transfer method of Self-Attention.
Background technique
As machine learning method continues to develop and develops, big data has become the every profession and trades such as academia and industry
Knowledge precious deposits.However, the case where uneven data category or even shortage of data occur in many fields at present, this is seriously affected
The stability and robustness of machine mould training.How to be had become according to the missing data of existing data set generation high quality
Primarily solve the problems, such as.
Currently, application, which generates the new image data of confrontation network (GAN) generation, has become grinding in deep learning field
Study carefully hot spot.Wherein, generating using the new picture of image-2-image translation technical research is usually to solve picture number
According to the effective way of missing.However, most of in the prior art all concentrate on image Style Transfer field, the content of migration is usual
It is the matching of image whole style and content.Small part image migrating technology is in the migration for solving local feature, however these skills
Art is difficult to control the contradictory relation of the positioning of local feature and general image feature invariant, so the picture robustness after migration
Have to be hoisted.For example, a kind of method that Taihong Xiao et al. proposes ELEGANT (ECCV 2018), ELEGANT are mainly answered
Replace traditional label information with hiding encoded information is exchanged, abundantization markup information is to improve the multiplicity that face character migrates
Property and migration quality.But ELEGANT does not control the local message of image and the disturbing factor of full figure pixel.Therefore it is moving
It is easy to influence the generation of other Pixel Informations in image domains during moving.
Summary of the invention
It is an object of the present invention to overcome the above-mentioned drawbacks of the prior art and provide one kind based on GAN and
The image local information transfer method of Self-Attention.
The purpose of the present invention can be achieved through the following technical solutions:
A kind of image local information transfer method based on GAN and Self-Attention, comprising the following steps:
1) it obtains image and migrates task, including picture pair to be migrated and the feature for needing to migrate；
2) task is migrated according to described image, carries out image local information transfer using a trained neural network, it is defeated
Image after migrating out；
The neural network includes image composer and arbiter, wherein described image generator include convolutional network and
Deconvolution network, the convolutional network by picture convolution at hiding feature space, in conjunction with Self-Attention model by part
Feature exchange is hidden, the feature after exchange realizes that image local is believed by the picture of the deconvolution network deconvolution Cheng Xin
Cease feature migration；The arbiter is used to differentiate the true and false relationship for generating picture.
Further, in described image generator, the company of convolutional network Yu deconvolution network is established by U-net structure
It connects.
Further, the convolutional network include input layer, convolutional layer, non-linear normalizing layer, network activation layer and
Self-Attention layers.
Further, the deconvolution network includes hiding feature input layer, warp lamination, non-linear normalizing layer, net
Network active coating, Tanh active coating and Self-Attention layers.
When further, to the training of described image generator, in each iteration, the output of image composer is put into and is sentenced
Loss error is calculated in other device, carries out backpropagation, is joined using the network that stochastic gradient descent method updates described image generator
Number, until reaching the frequency of training upper limit.
When further, to arbiter training, in each iteration, respectively from the defeated of training set and image composer
Out in take out an image, after zooming in and out, while inputting in determining device, obtain loss error, carry out backpropagation, using with
Machine gradient descent method updates the network parameter of the arbiter, until reaching the frequency of training upper limit.
Further, the arbiter is equipped with multiple.
Further, the arbiter is equipped with 4, the original image and the image after migration that two of them differentiate generation respectively,
Other two is true and false to differentiating after the multiple dimensioned scaling of image progress of the original image of generation and migration respectively.
Further, the arbiter includes input layer, convolutional layer, non-linear normalizing layer, network activation layer, full connection
Layer and two classification active coatings.
In order to solve image local migration to the disturbing factor of general image, the present invention is mainly answered when coding is hidden in exchange
With Self-Attention from attention model.Self-Attention is a kind of empty based on feature map from attention
Between autonomous learning method.The present invention mainly passes through convolutional neural networks for picture convolution into hiding feature space, then in conjunction with
Part is hidden feature exchange by Self-Attention model, finally by the picture of the feature deconvolution Cheng Xin after exchange to reality
Existing image local feature migration.Self-Attention can allow the present invention to learn to the pixel inside image from attention model
Related information aimings drill the ability that local feature position can be positioned with boosting algorithm in conjunction with multiple scale arbiters.Therefore, originally
Invention can effectively shift image local feature while reduce the interference that is distributed to image whole pixel information, to be promoted
The effect of the local feature migration of image.
In order to further enhance the effect of image migration, the present invention is also added in a network generates confrontation mechanism, algorithm
It mainly include image composer and arbiter.Generator is used to generate the new images after migration, and arbiter generates figure for differentiating
The true and false relationship of picture.The generation quality of image can be promoted by generating the game mechanism of confrontation in this way.In generator network,
The present invention also applies U-net structure that encoder Encoder and decoder Decoder Fusion Features are improved to the matter that image generates
Amount, reduces interference of local migration's information to the whole pixel of image.In differentiating network, the present invention uses multiple
Discriminators improves picture quality, and generator can be guided further to learn the local message of image, enhancing figure in this way
The migration effect of picture.
To sum up, innovation advantage of the invention embodies in the following areas:
First, the present invention proposes that the exchange of hiding information feature and Self-Attention is combined to do image from attention model
Feature local migration, Self-Attention can learn the direct relationship of pixel in image, effectively avoid local migration and figure
As the influence of global pixel.
Second, present invention uses U-net network layer-across connecting structures, effectively enhance generator generator generation
The quality of the global pixel of picture reduces local message migration bring interference.
Third, the present invention is when application generates confrontation network structure (GAN), using multiple arbiter discriminators
The generation for promoting picture guides e-learning image local feature, enhancing migration effect using multiple scale discriminators
Fruit.
Detailed description of the invention
Fig. 1 is generator structural schematic diagram of the present invention；
Fig. 2 is arbiter structural schematic diagram of the present invention；
Fig. 3 is generator of the present invention training flow diagram；
Fig. 4 is arbiter of the present invention training flow diagram；
Fig. 5 is testing process schematic diagram of the present invention；
Fig. 6 is the effect diagram of migration face fringe of the present invention on celebA public data collection；
Fig. 7 is the effect diagram of migration face beard of the present invention on celebA public data collection.
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with technical solution of the present invention
Premised on implemented, the detailed implementation method and specific operation process are given, but protection scope of the present invention is not limited to
Following embodiments.
The present invention provides a kind of image local information transfer method based on GAN and Self-Attention, this method fortune
For row in GPU, acquisition image first migrates task, including picture pair to be migrated and the feature for needing to migrate；Then according to institute
Image migration task is stated, carries out image local information transfer using a trained neural network；Figure after finally output migration
Picture.
As depicted in figs. 1 and 2, the neural network includes image composer generator and multiple arbiters
Discriminators, wherein described image generator includes convolutional network and deconvolution network, and the convolutional network is by picture
Part is hidden feature exchange at hiding feature space, in conjunction with Self-Attention model by convolution, and the feature after exchange passes through
The picture of the deconvolution network deconvolution Cheng Xin is to realize that image local information characteristics migrate；The arbiter is multiple to sentence
Other device for differentiating the true and false relationship for generating picture, and further promotes the local message migration of image, after promoting image migration
Quality.
In described image generator, the connection of convolutional network Yu deconvolution network is established by U-net structure.Convolutional network
Including input layer, convolutional layer, non-linear normalizing layer, network activation layer and Self-Attention layers；Deconvolution network includes
Hide feature input layer, warp lamination, non-linear normalizing layer, network activation layer and Tanh active coating.
In the present embodiment, building for image composer is as follows:
The level of convolutional network and every layer of characteristic pattern quantity 1a) are preset according to the scale of problem, the present embodiment is adopted
With Input1-Conv64-Norm-LeakyReLu-NonLocalBlock-Conv128-Norm- LeakyReLu-
NonLocalBlock-Conv256-Norm-LeakyReLu-NonLocalBlock-Conv512-Norm-LeakyReLu-
The structure of NonLocalBlock-Conv1024-Norm-LeakyReLu, wherein Input table shows input layer, thereafter the number of institute's band
Word n indicates the feature map quantity of this layer, and Conv indicates that convolutional layer, Norm indicate non-linear normalizing layer, LeakyReLu
Indicate that network activation layer, NonLocalBlock indicate Self-Attention layers；
The structure of deconvolution network 1b) is preset, the present embodiment uses Input1-NonLocalBlock-
ConvTranspose512-Norm-ReLu-NonLocalBlock-ConvTranspose256-Norm-ReLu-
NonLocalBlock-ConvTranspose128-Norm-ReLu-NonLocalBlock-ConvTranspose64-Norm-
The structure of ReLu-ConvTranspose3-Tanh, wherein Input table shows hiding feature input layer, thereafter the digital n table of institute's band
Show the feature map quantity of this layer, ConvTranspose indicates that warp lamination, Norm indicate non-linear normalizing layer, ReLu
Indicate that network activation layer, Tanh indicate Tanh active coating, NonLocalBlock indicates Self-Attention layers；
U-net structure 1c) is finally built, the corresponding Conv and ConvTranspose in Convnet and Deconvnet
Unet model is established, the identical Conv output feature of feature map quantity is input to ConvTranspose computing module
In.
Be illustrated in figure 3 the training process of image composer: input picture has two labeled as A, B, by convolutional network
Available two groups of length is that the convolution in 1024 channels hides characteristic layer, these layers is then divided into 2*n group, wherein n is to need
The feature quantity to be migrated, and hide feature to every group and be numbered.Assuming that i-th of local feature is migrated, here by i-th
The hiding feature of group, which is interchangeable, generates the hiding feature in two groups of new 1024 channels.It is available 4 groups long altogether after exchange feature
Degree hides feature for 1024 channels, hides feature for this four groups and is respectively fed to generate four new pictures in deconvolution network
A',B',C,D.The feature in neural network learning picture finally is allowed with the mode of back-propagation gradient again, optimizes A ', B ', C, D
The quality that four pictures generate.
As shown in Fig. 2, arbiter is equipped with 4, after two of them differentiate the original image and migration of generation respectively in the present embodiment
Image, other two differentiates true and false after zooming in and out respectively to the image of the original image of generation and migration, and then promote image
Local message migration.The build process of arbiter is as follows:
First arbiter discriminator0 2a) is constructed, network structure uses Input1-Conv64-Norm-
LeakyReLu-Conv128-Norm-LeakyReLu-Conv256-Norm-LeakyReLu-Conv512-Norm-
LeakyReLu-Linear-Sigmoid, Input table show input layer, and the digital n of institute's band indicates the feature map of this layer thereafter
Quantity, Conv indicate that convolutional layer, Norm indicate that non-linear normalizing layer, LeakyReLu indicate that network activation layer, Linear indicate
Full articulamentum, Sigmoid indicate two classification active coatings.Training discriminator0 the step of it is as follows: will input picture A, B with
And the input as arbiter discriminator0 of picture A ', B ' original image that generator generates, finally allow arbiter
Discriminator0 training study A, B, A ', four picture of B ' it is true and false, as shown in Figure 4.
Second arbiter discriminator1 2b) is constructed, network structure is consistent with discriminator0.Training
The step of discriminator1, is as follows: will input picture A ', the B ' that picture A, B and generator generate put be twice after conduct
The input of arbiter discriminator1, finally allow arbiter discriminator1 training study A, B, A ', B ' four figure
Piece it is true and false.
2c) construct third arbiter discriminator2, network structure and discriminator0 and
Discriminator1 is consistent.The step of training discriminator2, is as follows: input picture A, B and generator are generated
Input of picture C, D original image as arbiter discriminator2 finally allows arbiter discriminator2 training to learn
A, tetra- picture of B, C, D is true and false；
The 4th arbiter discriminator3 2d) is constructed, network structure is consistent with discriminator2.Training
The step of discriminator3, is as follows: will input after picture C, D that picture A, B and generator generate reduce one times to be used as and sentence
The input of other device discriminator3 finally allows arbiter discriminator3 training to learn tetra- picture of A, B, C, D
It is true and false.
The training of above-mentioned image composer and determining device executes respectively.When to the training of described image generator, changing every time
The output of image composer is put into arbiter and calculates loss error, carries out backpropagation, utilize stochastic gradient descent by Dai Zhong
Method updates the network parameter of described image generator, until reaching the frequency of training upper limit.When to arbiter training, each
In iteration, an image is taken out from the output of training set and image composer respectively, after zooming in and out, while inputting determining device
In, loss error is obtained, backpropagation is carried out, the network parameter of the arbiter is updated using stochastic gradient descent method, until
Reach the frequency of training upper limit.
After obtaining trained neural network, the test of image local information transfer is carried out using image measurement collection, is surveyed
It is similar with the training process of image composer to try process, as shown in figure 5, specific steps are as follows: obtain classification 1 and classification in test set
Any one image in 2 inputs in generator generator, obtains four images after generator generator operation
A',B',C,D.Wherein C, D are the result after input picture A, B local feature exchange migration respectively.
In order to verify performance of the invention, the present embodiment is on the public data collection (celebA) of a face attribute migration
It is tested.Good training set test set is divided first, in accordance with celebA, training is concentrated with 202000 picture input models
It is trained, test set has 599.After picture training one time all in training set, a model is obtained for testing.
Fig. 6 and Fig. 7 is the test effect result schematic diagram that face Liu Haihe beard is migrated in test set respectively.The Liu Haihe beard of face
The local message migration of face characteristic is belonged to, wherein face is opened on the left side Fig. 6 and Fig. 7 two is taken out from celebA test set
The output that face is training pattern of the present invention is opened on face original image, the right two.It can be seen that the fringe of Ms has been in Fig. 6
Man's head portrait is successfully migrated to suffer, the upper corners of the mouth beard of young man is successfully migrated in middle-aged and the old's head portrait in Fig. 7, and
And the face feature at other positions remains intact in their head portraits, is not interfered.
Above two picture is model time later test result of training on training set, and it is several all over can be with to continue more training
Obtain better test result.
The preferred embodiment of the present invention has been described in detail above.It should be appreciated that those skilled in the art without
It needs creative work according to the present invention can conceive and makes many modifications and variations.Therefore, all technologies in the art
Personnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea
Technical solution, all should be within the scope of protection determined by the claims.
Claims (9)
1. a kind of image local information transfer method based on GAN and Self-Attention, which is characterized in that including following step
It is rapid:
1) it obtains image and migrates task, including picture pair to be migrated and the feature for needing to migrate；
2) task is migrated according to described image, carries out image local information transfer using a trained neural network, output is moved
Image after shifting；
The neural network includes image composer and arbiter, wherein described image generator includes convolutional network and warp
Part at hiding feature space, is hidden picture convolution by product network, the convolutional network in conjunction with Self-Attention model
Feature exchange, the feature after exchange realize that image local information is special by the picture of the deconvolution network deconvolution Cheng Xin
It relocates residents from locations to be used for construction of new buildings or factories shifting；The arbiter is used to differentiate the true and false relationship for generating picture.
2. the image local information transfer method according to claim 1 based on GAN and Self-Attention, feature
It is, in described image generator, the connection of convolutional network Yu deconvolution network is established by U-net structure.
3. the image local information transfer method according to claim 1 based on GAN and Self-Attention, feature
It is, the convolutional network includes input layer, convolutional layer, non-linear normalizing layer, network activation layer and Self-Attention
Layer.
4. the image local information transfer method according to claim 1 based on GAN and Self-Attention, feature
It is, the deconvolution network includes hiding feature input layer, warp lamination, non-linear normalizing layer, network activation layer, Tanh
Active coating and Self-Attention layers.
5. the image local information transfer method according to claim 1 based on GAN and Self-Attention, feature
It is, when to the training of described image generator, in each iteration, the output of image composer is put into calculate in arbiter and is damaged
It is poor to make mistakes, and carries out backpropagation, and the network parameter of described image generator is updated using stochastic gradient descent method, until reaching instruction
Practice maximum number of times.
6. the image local information transfer method according to claim 1 based on GAN and Self-Attention, feature
It is, when to arbiter training, in each iteration, takes out one from the output of training set and image composer respectively
Image after zooming in and out, while inputting in determining device, obtains loss error, carries out backpropagation, utilize stochastic gradient descent method
The network parameter of the arbiter is updated, until reaching the frequency of training upper limit.
7. the image local information transfer method according to claim 1 based on GAN and Self-Attention, feature
It is, the arbiter is equipped with multiple.
8. the image local information transfer method according to claim 7 based on GAN and Self-Attention, feature
It is, the arbiter is equipped with 4, and the original image and the image after migration that two of them differentiate generation respectively, other two is distinguished
Differentiate after the multiple dimensioned scaling of image progress of original image and migration to generation true and false.
9. the image local information transfer method according to claim 1 based on GAN and Self-Attention, feature
It is, the arbiter includes that input layer, convolutional layer, non-linear normalizing layer, network activation layer, full articulamentum and two classification swash
Layer living.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811368715.7A CN109543742A (en) | 2018-11-16 | 2018-11-16 | A kind of image local information transfer method based on GAN and Self-Attention |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811368715.7A CN109543742A (en) | 2018-11-16 | 2018-11-16 | A kind of image local information transfer method based on GAN and Self-Attention |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109543742A true CN109543742A (en) | 2019-03-29 |
Family
ID=65847856
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811368715.7A Pending CN109543742A (en) | 2018-11-16 | 2018-11-16 | A kind of image local information transfer method based on GAN and Self-Attention |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109543742A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110033034A (en) * | 2019-04-01 | 2019-07-19 | 深圳大学 | A kind of image processing method, device and the computer equipment of non-homogeneous texture |
| CN110288537A (en) * | 2019-05-20 | 2019-09-27 | 湖南大学 | Facial image complementing method based on the depth production confrontation network from attention |
| CN111037365A (en) * | 2019-12-26 | 2020-04-21 | 大连理工大学 | Cutter state monitoring data set enhancing method based on generative countermeasure network |
| CN111127309A (en) * | 2019-12-12 | 2020-05-08 | 杭州格像科技有限公司 | Portrait style transfer model training method, portrait style transfer method and device |
| CN111198964A (en) * | 2020-01-10 | 2020-05-26 | 中国科学院自动化研究所 | Image retrieval method and system |
| CN112132133A (en) * | 2020-06-16 | 2020-12-25 | 杭州中科睿鉴科技有限公司 | Identification image data enhancement method and authenticity intelligent identification method |
| CN112699726A (en) * | 2020-11-11 | 2021-04-23 | 中国科学院计算技术研究所数字经济产业研究院 | Image enhancement method, genuine-fake commodity identification method and equipment |
| WO2021097429A1 (en) * | 2019-11-15 | 2021-05-20 | Waymo Llc | Multi object tracking using memory attention |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107577651A (en) * | 2017-08-25 | 2018-01-12 | 上海交通大学 | Chinese character style migratory system based on confrontation network |
| CN108564119A (en) * | 2018-04-04 | 2018-09-21 | 华中科技大学 | A kind of any attitude pedestrian Picture Generation Method |
| CN108647560A (en) * | 2018-03-22 | 2018-10-12 | 中山大学 | A kind of face transfer method of the holding expression information based on CNN |
-
2018
- 2018-11-16 CN CN201811368715.7A patent/CN109543742A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107577651A (en) * | 2017-08-25 | 2018-01-12 | 上海交通大学 | Chinese character style migratory system based on confrontation network |
| CN108647560A (en) * | 2018-03-22 | 2018-10-12 | 中山大学 | A kind of face transfer method of the holding expression information based on CNN |
| CN108564119A (en) * | 2018-04-04 | 2018-09-21 | 华中科技大学 | A kind of any attitude pedestrian Picture Generation Method |
Non-Patent Citations (1)
| Title |
|---|
| WUGUANGBIN1230: "SAGAN—Self-Attention Generative Adversarial Networks", 《HTTPS://BLOG.CSDN.NET/WUGUANGBIN1230/ARTICLE/DETAILS/83829879》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110033034B (en) * | 2019-04-01 | 2023-09-12 | 深圳大学 | Picture processing method and device for non-uniform texture and computer equipment |
| CN110033034A (en) * | 2019-04-01 | 2019-07-19 | 深圳大学 | A kind of image processing method, device and the computer equipment of non-homogeneous texture |
| CN110288537A (en) * | 2019-05-20 | 2019-09-27 | 湖南大学 | Facial image complementing method based on the depth production confrontation network from attention |
| WO2021097429A1 (en) * | 2019-11-15 | 2021-05-20 | Waymo Llc | Multi object tracking using memory attention |
| CN111127309A (en) * | 2019-12-12 | 2020-05-08 | 杭州格像科技有限公司 | Portrait style transfer model training method, portrait style transfer method and device |
| CN111127309B (en) * | 2019-12-12 | 2023-08-11 | 杭州格像科技有限公司 | Portrait style migration model training method, portrait style migration method and device |
| CN111037365B (en) * | 2019-12-26 | 2021-08-20 | 大连理工大学 | Cutter state monitoring data set enhancing method based on generative countermeasure network |
| CN111037365A (en) * | 2019-12-26 | 2020-04-21 | 大连理工大学 | Cutter state monitoring data set enhancing method based on generative countermeasure network |
| CN111198964B (en) * | 2020-01-10 | 2023-04-25 | 中国科学院自动化研究所 | Image retrieval method and system |
| CN111198964A (en) * | 2020-01-10 | 2020-05-26 | 中国科学院自动化研究所 | Image retrieval method and system |
| CN112132133A (en) * | 2020-06-16 | 2020-12-25 | 杭州中科睿鉴科技有限公司 | Identification image data enhancement method and authenticity intelligent identification method |
| CN112132133B (en) * | 2020-06-16 | 2023-11-17 | 中国科学院计算技术研究所数字经济产业研究院 | Identification image data enhancement method and true-false intelligent identification method |
| CN112699726A (en) * | 2020-11-11 | 2021-04-23 | 中国科学院计算技术研究所数字经济产业研究院 | Image enhancement method, genuine-fake commodity identification method and equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109543742A (en) | A kind of image local information transfer method based on GAN and Self-Attention | |
| Huang et al. | Architectural drawings recognition and generation through machine learning | |
| CN107833183A (en) | A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring | |
| CN110427875A (en) | Infrared image object detection method based on depth migration study and extreme learning machine | |
| CN107358648A (en) | Real-time full-automatic high quality three-dimensional facial reconstruction method based on individual facial image | |
| CN110046572A (en) | A kind of identification of landmark object and detection method based on deep learning | |
| CN107563439A (en) | A kind of model for identifying cleaning food materials picture and identification food materials class method for distinguishing | |
| CN108960499A (en) | A kind of Fashion trend predicting system merging vision and non-vision feature | |
| CN109492229A (en) | A kind of cross-cutting sensibility classification method and relevant apparatus | |
| CN110008961A (en) | Text real-time identification method, device, computer equipment and storage medium | |
| CN114092697B (en) | Building facade semantic segmentation method with attention fused with global and local depth features | |
| CN109147772A (en) | A kind of DNN-HMM acoustic model parameters migration structure | |
| CN112069993B (en) | Dense face detection method and system based on five-sense organ mask constraint and storage medium | |
| Liu et al. | Dunhuang murals contour generation network based on convolution and self-attention fusion | |
| Jiang et al. | [Retracted] Application of Intelligent Education Using Interactive Modeling in Higher Ideological and Political Education Platform | |
| Gui et al. | DLP-GAN: learning to draw modern Chinese landscape photos with generative adversarial network | |
| CN110264551A (en) | A kind of motion retargeting method and system | |
| Song et al. | Graphdecoder: Recovering diverse network graphs from visualization images via attention-aware learning | |
| CN110472108A (en) | Garment fabric sample retrieving method based on text profile matching | |
| CN107977372A (en) | A kind of mask method and device of face key element annotation | |
| Maietti et al. | Documentation, Analysis and Representation of Modernist Heritage Through Building Information Modeling | |
| CN110222693A (en) | The method and apparatus for constructing character recognition model and identifying character | |
| CN113704490B (en) | New industrial and scientific education knowledge graph construction method for heterogeneous data | |
| Akboy-İlk | Architectural Documentation: Built Environment, Modernization, and Turkish Nationalism | |
| Yang et al. | Development And Challenges of Generative Artificial Intelligence in Education and Art |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190329 |