CN107886162A - A kind of deformable convolution kernel method based on WGAN models - Google Patents
A kind of deformable convolution kernel method based on WGAN models Download PDFInfo
- Publication number
- CN107886162A CN107886162A CN201711123711.8A CN201711123711A CN107886162A CN 107886162 A CN107886162 A CN 107886162A CN 201711123711 A CN201711123711 A CN 201711123711A CN 107886162 A CN107886162 A CN 107886162A
- Authority
- CN
- China
- Prior art keywords
- mrow
- convolution kernel
- maker
- deformable convolution
- msub
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
Abstract
The invention discloses a kind of deformable convolution kernel method based on WGAN models, belong to deep learning field of neural networks, comprise the following steps:S1, construction are originally generated confrontation network model;S2, construction Wo Sesitan distances, the judging quota as confrontation network model;S3, initialization random noise, are inputted in maker;S4, carry out convolution using deformable convolution collecting image in WGAN models;S5, loss function that deformable convolution operation obtains input maker subsequently trained.The deformable convolution kernel method based on WGAN models that the present invention is built, change arbiter, maker receives the convolution mode after picture, arbiter, maker is allowed automatically to change the size of convolution kernel according to the situation of training, so as to adaptively learn to the feature of data images, the robustness of whole network training is improved.
Description
Technical field
The present invention relates to deep learning field of neural networks, and in particular to a kind of deformable convolution kernel based on WGAN models
Method.
Background technology
Production confrontation network (Generative Adversarial Network, abbreviation GAN) is by Goodfellow
In the deep learning framework that 2014 propose, it is based on the thought of " game theory ", construction maker (generator) and arbiter
(discriminator) two kinds of models, the former generates image by the Uniform noise or gaussian random noise for inputting (0,1), after
Person differentiates to the image of input, it is determined that being the image from data set or the image as caused by maker.
In traditional confrontation network model, do not have unified judgment criteria for maker generation picture quality, because
This, it would be highly desirable to a kind of judging quota by the use of Wo Sesitan distances as generation confrontation network is proposed, so that the instruction of whole model
White silk can learn the method for characteristics of image furthermore with deformable convolution, improve whole network toward progress is correctly oriented
Training effectiveness.
The content of the invention
The invention aims to solve drawbacks described above of the prior art, there is provided a kind of based on the variable of WGAN models
Shape convolution kernel method.
The purpose of the present invention can be reached by adopting the following technical scheme that:
A kind of deformable convolution kernel method based on WGAN models, the deformable convolution kernel method comprise the following steps:
S1, construction are originally generated confrontation network model, and generating image by maker inputs to arbiter progress network instruction
Practice;
S2, construction Wo Sesitan distances, the judging quota as confrontation network model;
In the network model that the present invention relates to, the judge by the use of Wo Sesitan distances as generation confrontation network refers to
Mark, so that the training of whole model past can be correctly oriented progress.
S3, initialization random noise, are inputted in maker;
S4, carry out convolution using deformable convolution collecting image in WGAN models;
In original generation confrontation network model, the shape of convolution kernel is generally square, and which has limited neutral net pair
The free degree of characteristics of image study, and in the present invention, for this defect, the shape of convolution kernel is carried out using network training
Adaptively change, so as to the feature of higher efficiency learning image into data set.
S5, loss function that deformable convolution operation obtains input maker subsequently trained.
Further, described step S2 is specific as follows:
Multiple convolution kernels are constructed, different convolution kernels, are represent during study, can be learnt to different images
Feature.
Further, convolution, specific mistake are carried out using deformable convolution collecting image in WGAN in described step S4
Journey is as follows:
S41, the multiple different numerical value of construction but size identical convolution kernel;
S42, using the convolution kernel constructed, convolution is carried out to multiple images of maker generation respectively, it is more so as to obtain
Open characteristic pattern.
Further, in described step S5, the loss function input maker that deformable convolution operation is obtained is carried out
Follow-up training.Detailed process is as follows:
S51, the characteristic pattern after convolution in S4, input arbiter are differentiated;
S52, loss function that deformable convolution operation obtains input maker subsequently trained.
S53, the average of all loss functions is inputted and continues to be trained into maker.
The present invention is had the following advantages relative to prior art and effect:
Robustness:The present invention sets according to the operating process of deformable convolution and constructs multiple deformable convolution kernels, pass through
The mode of convolution kernel size is dynamically changed in the training process, is applied and is being served as maker with sentencing with depth convolutional neural networks
In the confrontation network model of other device, while the judging quota by the use of Wo Sesitan distances as generation confrontation network, so that whole
The training of individual model past can be correctly oriented progress.
Brief description of the drawings
Fig. 1 is the training flow chart of the deformable convolution kernel method based on WGAN models disclosed in the present invention;
Fig. 2 is the schematic diagram for being transformed into deformable convolution kernel in the present invention to original convolution core.
Embodiment
To make the purpose, technical scheme and advantage of the embodiment of the present invention clearer, below in conjunction with the embodiment of the present invention
In accompanying drawing, the technical scheme in the embodiment of the present invention is clearly and completely described, it is clear that described embodiment is
Part of the embodiment of the present invention, rather than whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art
The every other embodiment obtained under the premise of creative work is not made, belongs to the scope of protection of the invention.
Embodiment
Present embodiment discloses a kind of deformable convolution kernel method based on WGAN models, the following steps are specifically included:
Step S1, construction is originally generated confrontation network model, and maker is inputted to arbiter by generating image and carries out net
Network training.
Step S2, Wo Sesitan distances are constructed, the judging quota as confrontation network model;
Different convolution kernels, it is embodied in difference, the difference of ranks number of matrix numerical value.
Multiple convolution kernels are constructed, during image is handled, different convolution kernels is meant in network training
Different characteristic of the process learning to generation image.
In the network model that the present invention relates to, the judge by the use of Wo Sesitan distances as generation confrontation network refers to
Mark, so that the training of whole model past can be correctly oriented progress.
In the model of tradition confrontation network, the convolution kernel used in arbiter and maker is all fixed size and numerical value
Consistent, training effectiveness in this case is relatively low, and the characteristics of image scope learnt is relatively small.And at this
In invention, using deformable convolution, the operation of " 0 " is interleave in being carried out to original convolution core, can be learned so as to increase convolution kernel
The characteristic range practised, further increase the efficiency of whole network study.
In actual applications, it should which according to the complexity of data images feature, the number of convolution kernel is set.
Step S3, random noise is initialized, is inputted in maker.
Step S4, convolution is carried out in WGAN models using deformable convolution collecting image.
In original generation confrontation network model, the shape of convolution kernel is generally square, and which has limited neutral net pair
The free degree of characteristics of image study, and in the present invention, for this defect, the shape of convolution kernel is carried out using network training
Adaptively change, so as to the feature of higher efficiency learning image into data set.
Specific method is as follows:
S41, the multiple different numerical value of construction but size identical convolution kernel;
S42, the error by anti-pass in network training process, adaptive change is carried out to the shape of convolution kernel.
Step S5, the loss function input maker that deformable convolution operation obtains subsequently is trained.Detailed process
It is as follows:
S51, by the characteristic pattern after convolution in step S4, input arbiter is differentiated;
S52, loss function that deformable convolution operation obtains input maker subsequently trained.
S53, the average of all loss functions is inputted and continues to be trained into maker.
The effect of loss function is to weigh the ability that arbiter is judged generation image.The value of loss function is smaller, explanation
In current iteration, arbiter can have the generation image of preferable performance discrimination maker;Property that is on the contrary then illustrating arbiter
Can be poor.
The expression formula of loss function is:
Wherein, D (x) represents differentiation of the arbiter to image, and pr represents the distribution of data images, and pg represents generation image
Distribution, λ is hyper parameter,For gradient, E is the functional symbol for taking average.
In summary, present embodiment discloses a kind of deformable convolution kernel method based on WGAN models, compared to tradition
Original confrontation network model, change arbiter receive picture after the mode learnt to characteristics of image.It is right in tradition
In the model of anti-network, arbiter is all fixed size with the convolution kernel used in maker and numerical value is consistent, in this feelings
Training effectiveness under condition is relatively low, and the characteristics of image scope learnt is relatively small.And in the present invention, utilization is variable
Shape convolution, the effect learnt according to network in the training process to characteristics of image, the size of convolution kernel is dynamically changed, so as to increase
Big convolution kernel can learn the adaptivity of scope, further increase the efficiency that whole network learns.
Above-described embodiment is the preferable embodiment of the present invention, but embodiments of the present invention are not by above-described embodiment
Limitation, other any Spirit Essences without departing from the present invention with made under principle change, modification, replacement, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (4)
- A kind of 1. deformable convolution kernel method based on WGAN models, it is characterised in that described deformable convolution kernel method bag Include the following steps:S1, construction are originally generated confrontation network model, and generating image by maker inputs to arbiter progress network training;S2, construction Wo Sesitan distances, the judging quota as confrontation network model;S3, initialization random noise, are inputted in maker;S4, carry out convolution using deformable convolution collecting image in WGAN models;S5, loss function that deformable convolution operation obtains input maker subsequently trained.
- 2. a kind of deformable convolution kernel method based on WGAN models according to claim 1, it is characterised in that described Step S4 detailed processes are as follows:S41, the multiple different numerical value of construction but size identical convolution kernel;S42, the error by anti-pass in network training process, adaptive change is carried out to the shape of convolution kernel.
- 3. a kind of deformable convolution kernel method based on WGAN models according to claim 1, it is characterised in that described Step S5 detailed processes are as follows:S51, the characteristics of image figure that will be obtained after deformable convolution, input in arbiter and are differentiated;S52, loss function that deformable convolution operation obtains input maker subsequently trained;S53, the average of all loss functions is inputted and continues to be trained into maker.
- 4. a kind of deformable convolution kernel method based on WGAN models according to claim 1, it is characterised in that described The expression formula of loss function is:<mrow> <mi>L</mi> <mrow> <mo>(</mo> <mi>D</mi> <mo>)</mo> </mrow> <mo>=</mo> <mo>-</mo> <msub> <mi>E</mi> <mrow> <mi>x</mi> <mo>~</mo> <mi>p</mi> <mi>r</mi> </mrow> </msub> <mo>&lsqb;</mo> <mi>D</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>&rsqb;</mo> <mo>+</mo> <msub> <mi>E</mi> <mrow> <mi>x</mi> <mo>~</mo> <mi>p</mi> <mi>g</mi> </mrow> </msub> <mo>&lsqb;</mo> <mi>D</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>&rsqb;</mo> <mo>+</mo> <msub> <mi>&lambda;E</mi> <mrow> <mi>x</mi> <mo>~</mo> <mi>X</mi> </mrow> </msub> <msub> <mo>&dtri;</mo> <mi>x</mi> </msub> </mrow>Wherein, D (x) represents differentiation of the arbiter to image, and pr represents the distribution of data images, and pg represents point of generation image Cloth, λ are hyper parameter,For gradient, E is the functional symbol for taking average.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711123711.8A CN107886162A (en) | 2017-11-14 | 2017-11-14 | A kind of deformable convolution kernel method based on WGAN models |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711123711.8A CN107886162A (en) | 2017-11-14 | 2017-11-14 | A kind of deformable convolution kernel method based on WGAN models |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107886162A true CN107886162A (en) | 2018-04-06 |
Family
ID=61776659
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711123711.8A Pending CN107886162A (en) | 2017-11-14 | 2017-11-14 | A kind of deformable convolution kernel method based on WGAN models |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107886162A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109359667A (en) * | 2018-09-07 | 2019-02-19 | 华南理工大学 | A kind of feature recalibration convolution method based on WGAN model |
CN109933677A (en) * | 2019-02-14 | 2019-06-25 | 厦门一品威客网络科技股份有限公司 | Image generating method and image generation system |
CN110728303A (en) * | 2019-09-12 | 2020-01-24 | 东南大学 | Dynamic self-adaptive computing array based on convolutional neural network data complexity |
CN112102306A (en) * | 2020-09-25 | 2020-12-18 | 西安交通大学 | Dual-GAN-based defect detection method for edge repair feature fusion |
TWI714397B (en) * | 2019-03-19 | 2020-12-21 | 大陸商深圳市商湯科技有限公司 | Method, device for video processing and computer storage medium thereof |
CN112292694A (en) * | 2018-04-19 | 2021-01-29 | 智动科技有限公司 | Method for accelerating operation and accelerator device |
CN114239814A (en) * | 2022-02-25 | 2022-03-25 | 杭州研极微电子有限公司 | Training method of convolution neural network model for image processing |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107103590A (en) * | 2017-03-22 | 2017-08-29 | 华南理工大学 | A kind of image for resisting generation network based on depth convolution reflects minimizing technology |
CN107154023A (en) * | 2017-05-17 | 2017-09-12 | 电子科技大学 | Face super-resolution reconstruction method based on generation confrontation network and sub-pix convolution |
-
2017
- 2017-11-14 CN CN201711123711.8A patent/CN107886162A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107103590A (en) * | 2017-03-22 | 2017-08-29 | 华南理工大学 | A kind of image for resisting generation network based on depth convolution reflects minimizing technology |
CN107154023A (en) * | 2017-05-17 | 2017-09-12 | 电子科技大学 | Face super-resolution reconstruction method based on generation confrontation network and sub-pix convolution |
Non-Patent Citations (3)
Title |
---|
IAN J. GOODFELLOW 等: "Generative Adversarial Nets", 《ARXIV》 * |
ISHAAN GULRAJANI 等: "Improved Training of Wasserstein GANs", 《ARXIV》 * |
MARTIN ARJOVSKY 等: "Wasserstein GAN", 《ARXIV》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112292694A (en) * | 2018-04-19 | 2021-01-29 | 智动科技有限公司 | Method for accelerating operation and accelerator device |
CN109359667A (en) * | 2018-09-07 | 2019-02-19 | 华南理工大学 | A kind of feature recalibration convolution method based on WGAN model |
CN109933677A (en) * | 2019-02-14 | 2019-06-25 | 厦门一品威客网络科技股份有限公司 | Image generating method and image generation system |
TWI714397B (en) * | 2019-03-19 | 2020-12-21 | 大陸商深圳市商湯科技有限公司 | Method, device for video processing and computer storage medium thereof |
CN110728303A (en) * | 2019-09-12 | 2020-01-24 | 东南大学 | Dynamic self-adaptive computing array based on convolutional neural network data complexity |
CN110728303B (en) * | 2019-09-12 | 2022-03-11 | 东南大学 | Dynamic self-adaptive computing array based on convolutional neural network data complexity |
CN112102306A (en) * | 2020-09-25 | 2020-12-18 | 西安交通大学 | Dual-GAN-based defect detection method for edge repair feature fusion |
CN112102306B (en) * | 2020-09-25 | 2022-10-25 | 西安交通大学 | Dual-GAN-based defect detection method for edge repair feature fusion |
CN114239814A (en) * | 2022-02-25 | 2022-03-25 | 杭州研极微电子有限公司 | Training method of convolution neural network model for image processing |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107886162A (en) | A kind of deformable convolution kernel method based on WGAN models | |
CN107871142A (en) | A kind of empty convolution method based on depth convolution confrontation network model | |
CN107590518A (en) | A kind of confrontation network training method of multiple features study | |
CN107886169A (en) | A kind of multiple dimensioned convolution kernel method that confrontation network model is generated based on text image | |
Zhang et al. | Hierarchical feature fusion with mixed convolution attention for single image dehazing | |
CN105657402B (en) | A kind of depth map restoration methods | |
CN108009568A (en) | A kind of pedestrian detection method based on WGAN models | |
CN105825484B (en) | A kind of depth image denoising and Enhancement Method based on deep learning | |
CN106991408A (en) | The generation method and method for detecting human face of a kind of candidate frame generation network | |
CN107945118A (en) | A kind of facial image restorative procedure based on production confrontation network | |
CN108961245A (en) | Picture quality classification method based on binary channels depth parallel-convolution network | |
CN107767413A (en) | A kind of image depth estimation method based on convolutional neural networks | |
CN104217404B (en) | Haze sky video image clearness processing method and its device | |
CN107506722A (en) | One kind is based on depth sparse convolution neutral net face emotion identification method | |
CN107944546A (en) | It is a kind of based on be originally generated confrontation network model residual error network method | |
CN106447626A (en) | Blurred kernel dimension estimation method and system based on deep learning | |
CN107408211A (en) | Method for distinguishing is known again for object | |
CN107577985A (en) | The implementation method of the face head portrait cartooning of confrontation network is generated based on circulation | |
CN107992944A (en) | It is a kind of based on be originally generated confrontation network model multiple dimensioned convolution method | |
CN106339984B (en) | Distributed image ultra-resolution method based on K mean value driving convolutional neural networks | |
CN108615228A (en) | Facial image complementing method based on hybrid neural networks | |
CN106022392A (en) | Deep neural network sample automatic accepting and rejecting training method | |
CN106934455A (en) | Remote sensing image optics adapter structure choosing method and system based on CNN | |
CN109360159A (en) | A kind of image completion method based on generation confrontation network model | |
CN107943750A (en) | A kind of decomposition convolution method based on WGAN models |
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: 20180406 |