CN109359667A - A kind of feature recalibration convolution method based on WGAN model - Google Patents
A kind of feature recalibration convolution method based on WGAN model Download PDFInfo
- Publication number
- CN109359667A CN109359667A CN201811041215.2A CN201811041215A CN109359667A CN 109359667 A CN109359667 A CN 109359667A CN 201811041215 A CN201811041215 A CN 201811041215A CN 109359667 A CN109359667 A CN 109359667A
- Authority
- CN
- China
- Prior art keywords
- convolution
- weight
- training
- image
- feature recalibration
- 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
- G06F18/20—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
-
- 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
Abstract
The invention discloses a kind of feature recalibration convolution method based on WGAN model, belongs to deep learning field of neural networks, comprising the following steps: S1, construction are originally generated confrontation network model；S2, construction Wo Sesitan distance, the judging quota as confrontation network model；S3, initialization random noise, input in generator；S4, corresponding weight is set for each layer of convolution, convolution operation is carried out to image using feature recalibration convolution；The situation of change of loss function, the weight for dynamically updating each layer of convolution carry out subsequent training after training every time for S5, basis.In traditional model, image can obtain multiple characteristic patterns by multilayer convolution, but every characteristic pattern weight is equal, this method can allow generate confrontation every characteristic pattern of e-learning weight, that is significance level, efficiency is improved for subsequent training, provides the direction of network training simultaneously also by Wo Sesitan distance.
Description
Technical field
The present invention relates to deep learning nerual network technique fields, and in particular to a kind of feature based on WGAN model is marked again
Determine convolution method.
Background technique
It is by Goodfellow that production, which fights network (Generative Adversarial Network, abbreviation GAN),
In the deep learning frame that 2014 propose, it is based on the thought of " game theory ", constructs generator (generator) and arbiter
(discriminator) two kinds of models, the former generates image by the Uniform noise or gaussian random noise of input (0,1), after
Person differentiates that determination is the image from data set or the image by generator generation to the image of input.
In traditional confrontation network model, for the characteristics of image after convolution do not have significance level point, i.e., all one
Treat as benevolence.And in the present invention, using Wo Sesitan distance as the judging quota for generating confrontation network, to make entire model
Training can learn the method for characteristics of image furthermore with feature recalibration convolution, so that network toward being correctly oriented progresss
It can be subject to different weights to different features during training, that is, know the secondary from the primary, to improve whole network
Training effectiveness.
Summary of the invention
The purpose of the present invention is to solve drawbacks described above in the prior art, construct a kind of spy based on WGAN model
Levy recalibration convolution method.
The purpose of the present invention can be reached by adopting the following technical scheme that:
A kind of feature recalibration convolution method based on WGAN model, the feature recalibration convolution method includes following
Step:
S1, construction are originally generated confrontation network model, and generator is input to arbiter progress network instruction by generating image
Practice；
S2, the multiple convolution kernels of construction, meanwhile, Wo Sesitan distance is constructed, the judging quota as confrontation network model；
It is involved in the present invention to network model in, using Wo Sesitan distance as generate fight network judge refer to
Mark, to enable the training of entire model is past to be correctly oriented progress.
S3, initialization random noise, input in generator；
S4, corresponding weight is set for each layer of convolution, convolution operation is carried out to image using feature recalibration convolution；
The situation of change of each loss function after training of S5, basis, after the weight progress for dynamically updating each layer of convolution
Continuous training.
Further, in the step S2, multiple convolution kernels are constructed, different convolution kernels represents the mistake in study
Cheng Zhong can learn to different characteristics of image.
Further, Wo Sesitan distance is as follows:
Wherein, two data distribution psrAnd pgThe set of all Joint Distributions be denoted as ∏ (pr,pg)。
Further, it is that each layer of convolution sets corresponding weight in the step S4, utilizes feature recalibration convolution
Convolution operation is carried out to image, detailed process is as follows:
S41, corresponding weight is set for different convolutional layers；
S42, weight is multiplied with the characteristic pattern after convolution, carries out subsequent training.
Further, in the step S5, according to the situation of change of loss function, dynamic update every after training every time
The weight of one layer of convolution carries out subsequent training.Detailed process is as follows:
S51, to the characteristic pattern after convolution in S4, input arbiter and differentiated；
The situation of change of each loss function after training of S52, basis, after the weight progress for dynamically updating each layer of convolution
Continuous training.
The present invention has the following advantages and effects with respect to the prior art:
High efficiency: the present invention is arranged according to the operating process of feature recalibration convolution and constructs multiple feature recalibrations volumes
Product core improves efficiency by generating the weight of confrontation every characteristic pattern of e-learning, i.e. significance level for subsequent training,
The direction of network training is provided simultaneously also by Wo Sesitan distance.
Detailed description of the invention
Fig. 1 is the training flow chart of the feature recalibration convolution method in the present invention based on WGAN model；
Fig. 2 is the schematic diagram for carrying out weight update in the present invention for each layer of convolution.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art
Every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
Embodiment
Present embodiment discloses a kind of feature recalibration convolution method based on WGAN model, specifically includes the following steps:
Step S1, construction is originally generated confrontation network model, and generator is input to arbiter by generation image and carries out net
Network training.
Step S2, multiple convolution kernels are constructed, meanwhile, Wo Sesitan distance is constructed, the judge as confrontation network model refers to
Mark；
Different convolution kernels is embodied in the difference of matrix numerical value, the difference of ranks number.
Multiple convolution kernels are constructed, during handling image, different convolution kernels is meant in network training
Different characteristic of the study to generation image in the process.
It is involved in the present invention to network model in, using Wo Sesitan distance as generate fight network judge refer to
Mark, to enable the training of entire model is past to be correctly oriented progress.
The Wo Sesitan distance is as follows:
Wherein, two data distribution psrAnd pgThe set of all Joint Distributions be denoted as ∏ (pr,pg)。
In practical 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 generator.
Step S4, corresponding weight is set for each layer of convolution, convolution behaviour is carried out to image using feature recalibration convolution
Make.
The specific method is as follows:
S41, corresponding weight is set for different convolutional layers；
S42, the characteristic pattern after each layer of convolution is multiplied with weight, carries out subsequent training.
Step S5, according to the situation of change of loss function after training every time, dynamically update the weight of each layer of convolution into
The subsequent training of row.Detailed process is as follows:
S51, by the characteristic pattern after convolution in step S4, input arbiter and differentiated；
The situation of change of each loss function after training of S52, basis, after the weight progress for dynamically updating each layer of convolution
Continuous training.
The effect of loss function is to measure arbiter to the ability for generating image judgement.The value of loss function is smaller, explanation
In current iteration, arbiter can have the generation image of preferable performance discrimination generator；Property that is on the contrary then illustrating arbiter
It can be poor.
The expression formula of loss function are as follows:
Wherein, D (x) indicates differentiation of the arbiter to image, and pr indicates the distribution of data images, and pg indicates to generate image
Distribution, λ is hyper parameter,For gradient.
In conclusion present embodiment discloses a kind of feature recalibration convolution method based on WGAN model, compared to biography
The original confrontation network model of system changes arbiter and receives the mode learnt to characteristics of image after picture.In tradition
Model in, image can obtain multiple characteristic patterns by multilayer convolution, but every characteristic pattern weight is equal, and this method can allow
The weight of confrontation every characteristic pattern of e-learning, i.e. significance level are generated, improves efficiency for subsequent training, simultaneously also by
Wo Sesitan distance provides the direction of network training.
The above embodiment is a preferred embodiment of the present invention, but embodiments of the present invention are not by above-described embodiment
Limitation, other any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present invention,
It should be equivalent substitute mode, be included within the scope of the present invention.
Claims (5)
1. a kind of feature recalibration convolution method based on WGAN model, which is characterized in that the feature recalibration convolution side
Method includes the following steps:
S1, construction are originally generated confrontation network model；
S2, construction Wo Sesitan distance, the judging quota as confrontation network model；
S3, initialization random noise, input in generator；
S4, corresponding weight is set for each layer of convolution, convolution operation is carried out to image using feature recalibration convolution；
The situation of change of loss function, the weight for dynamically updating each layer of convolution carry out subsequent instruction after training every time for S5, basis
Practice.
2. a kind of feature recalibration convolution method based on WGAN model according to claim 1, which is characterized in that described
Wo Sesitan distance it is as follows:
Wherein, two data distribution psrAnd pgThe set of all Joint Distributions be denoted as ∏ (pr,pg)。
3. a kind of feature recalibration convolution method based on WGAN model according to claim 1, which is characterized in that described
Step S4 detailed process is as follows:
S41, corresponding weight is set for each layer of convolution；
S42, weight is multiplied with the characteristic pattern after convolution, carries out subsequent training.
4. a kind of feature recalibration convolution method based on WGAN model according to claim 1, which is characterized in that described
Step S5 process it is as follows:
S51, the characteristics of image figure that will be obtained after feature recalibration convolution, input in arbiter and are differentiated；
The situation of change of loss function, the weight for dynamically updating each layer of convolution carry out subsequent instruction after training every time for S52, basis
Practice.
5. a kind of feature recalibration convolution method based on WGAN model according to claim 1, which is characterized in that described
Loss function expression formula are as follows:
Wherein, D (x) indicates differentiation of the arbiter to image, and pr indicates the distribution of data images, and pg indicates to generate point of image
Cloth, λ are hyper parameter,For gradient.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811041215.2A CN109359667A (en) | 2018-09-07 | 2018-09-07 | A kind of feature recalibration convolution method based on WGAN model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811041215.2A CN109359667A (en) | 2018-09-07 | 2018-09-07 | A kind of feature recalibration convolution method based on WGAN model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109359667A true CN109359667A (en) | 2019-02-19 |
Family
ID=65350496
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811041215.2A Pending CN109359667A (en) | 2018-09-07 | 2018-09-07 | A kind of feature recalibration convolution method based on WGAN model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109359667A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563155A (en) * | 2017-08-08 | 2018-01-09 | 中国科学院信息工程研究所 | A kind of safe steganography method and device based on generation confrontation network |
| CN107886162A (en) * | 2017-11-14 | 2018-04-06 | 华南理工大学 | A kind of deformable convolution kernel method based on WGAN models |
| CN107943750A (en) * | 2017-11-14 | 2018-04-20 | 华南理工大学 | A kind of decomposition convolution method based on WGAN models |
| CN108021979A (en) * | 2017-11-14 | 2018-05-11 | 华南理工大学 | It is a kind of based on be originally generated confrontation network model feature recalibration convolution method |
-
2018
- 2018-09-07 CN CN201811041215.2A patent/CN109359667A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563155A (en) * | 2017-08-08 | 2018-01-09 | 中国科学院信息工程研究所 | A kind of safe steganography method and device based on generation confrontation network |
| CN107886162A (en) * | 2017-11-14 | 2018-04-06 | 华南理工大学 | A kind of deformable convolution kernel method based on WGAN models |
| CN107943750A (en) * | 2017-11-14 | 2018-04-20 | 华南理工大学 | A kind of decomposition convolution method based on WGAN models |
| CN108021979A (en) * | 2017-11-14 | 2018-05-11 | 华南理工大学 | It is a kind of based on be originally generated confrontation network model feature recalibration convolution method |
Non-Patent Citations (1)
| Title |
|---|
| 林懿伦 等: "人工智能研究的新前线: 生成式对抗网络", 《自动化学报》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107590518A (en) | A kind of confrontation network training method of multiple features study | |
| CN110097178A (en) | It is a kind of paid attention to based on entropy neural network model compression and accelerated method | |
| Pardos et al. | Using HMMs and bagged decision trees to leverage rich features of user and skill from an intelligent tutoring system dataset | |
| CN108021979A (en) | It is a kind of based on be originally generated confrontation network model feature recalibration convolution method | |
| CN107886169A (en) | A kind of multiple dimensioned convolution kernel method that confrontation network model is generated based on text image | |
| CN108175402A (en) | The intelligent identification Method of electrocardiogram (ECG) data based on residual error network | |
| CN107886162A (en) | A kind of deformable convolution kernel method based on WGAN models | |
| CN106411896A (en) | APDE-RBF neural network based network security situation prediction method | |
| CN107689034A (en) | A kind of training method of neutral net, denoising method and device | |
| Grappiolo et al. | Towards player adaptivity in a serious game for conflict resolution | |
| CN108230278A (en) | A kind of image based on generation confrontation network goes raindrop method | |
| CN107944546A (en) | It is a kind of based on be originally generated confrontation network model residual error network method | |
| CN107563509A (en) | A kind of dynamic adjustment algorithm for the condition DCGAN models that feature based returns | |
| CN108009568A (en) | A kind of pedestrian detection method based on WGAN models | |
| CN108970119A (en) | The adaptive game system strategic planning method of difficulty | |
| CN109344879A (en) | A kind of decomposition convolution method fighting network model based on text-image | |
| WO2020259504A1 (en) | Efficient exploration method for reinforcement learning | |
| CN107992944A (en) | It is a kind of based on be originally generated confrontation network model multiple dimensioned convolution method | |
| CN108776835A (en) | A kind of deep neural network training method | |
| CN107590139A (en) | A kind of knowledge mapping based on circular matrix translation represents learning method | |
| CN107943750A (en) | A kind of decomposition convolution method based on WGAN models | |
| CN113033822A (en) | Antagonistic attack and defense method and system based on prediction correction and random step length optimization | |
| CN108470208A (en) | It is a kind of based on be originally generated confrontation network model grouping convolution method | |
| Lv | Martial arts competitive decision-making algorithm based on improved BP neural network | |
| CN109359667A (en) | A kind of feature recalibration convolution method based on WGAN model |
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: 20190219 |