CN109460708A - A kind of Forest fire image sample generating method based on generation confrontation network - Google Patents
A kind of Forest fire image sample generating method based on generation confrontation network Download PDFInfo
- Publication number
- CN109460708A CN109460708A CN201811180582.0A CN201811180582A CN109460708A CN 109460708 A CN109460708 A CN 109460708A CN 201811180582 A CN201811180582 A CN 201811180582A CN 109460708 A CN109460708 A CN 109460708A
- Authority
- CN
- China
- Prior art keywords
- model
- sample
- generating
- training
- image
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the Forest fire image sample generating method for generating confrontation network.Generating model is to be fitted to generate Target Photo by data, and discrimination model then wants to the false sample that will be generated and authentic specimen is distinguished.Model is generated in this method and discrimination model is all made of the sub-network of convolutional neural networks form, while alternately training is carried out to two networks, is finally reached dynamic equilibrium, i.e. Nash Equilibrium by training two sub-networks of optimization.Final purpose is to generate model to learn probability distribution corresponding to image pattern from training sample, to obtain more " generation " samples to realize the expansion of data.The invention belongs to unsupervised learnings, without carrying out sample calibration；Convolutional neural networks can be compared with the traditional method from the acquistion of sample middle school to reliable character representation, and algorithm possesses better robustness and generalization ability.This method has important application value in terms of forest ecology protection, forest rocket.
Description
Technical field
The invention belongs to computer visions and forest rocket video surveillance field, and in particular to one kind is refreshing based on confrontation is generated
Forest fire image sample generating method through network.
Background technique
China is the multiple country of a forest fire, and forest fire protection task is that the primary of National Forest protection of resources is appointed
Business, is related to the safety of people's lives and properties and the forest reserves.In recent years, computer hardware and image processing techniques start to obtain
Compared with much progress, forest fires video monitoring system starts to occupy more importantly position in forest fire detection with its unique advantage
It sets.Forest fires video monitoring system based on computer vision is with embedded device, computer equipment etc. for processing unit, according to
The Image Visual Features such as smog and flame carry out forest fires detection in video image, by Digital Image Processing, signal processing, mode
The technologies such as identification and deep learning implement forest fires monitoring to extensive forest.Forest fires detection method based on computer vision is more
Precise and high efficiency, detection time and pre- call time all greatly shorten, and intelligent monitoring system also facilitates the system at fire prevention direction center
One management.Therefore, as the forest fires recognizer of forest fires video monitoring system core component increasingly by the pass of scientific research personnel
Note.
The rapid emergence of progress and GPU recently as computer performance, computer vision technique obtain considerable hair
Exhibition, obtains using deep neural network as the image processing method of representative in fields such as target detection, action recognition, super-resolutions
The recognition capability of immense success, Forest Fire Monitoring technology based on computer vision is in new Rapid development stage.But it is gloomy
The fast development of woods pyrotechnics video surveillance technology needs to overcome the problems, such as sample size deficiency.It is typically different the forest rocket sample of scene
This is very rare, if therefore the method that directlys adopt general deep learning, in diversified actual scene, algorithm can not reach
To excellent adaptability.
Summary of the invention
Goal of the invention: it is an object of the invention to solve the fast-developing needs of existing forest rocket video surveillance technology
Overcome sample size insufficient, if the method for directlying adopt general deep learning, in diversified actual scene, algorithm can not reach
The problem of to excellent adaptability.
Technical solution: to achieve the above object: the invention adopts the following technical scheme:
The forest rocket image pattern generation method of confrontation neural network is generated based on depth convolution, comprising the following steps:
It is a kind of based on generate confrontation network Forest fire image sample generating method, this method according to include the following steps into
Row:
Step 1: being based on forest rocket video monitoring system, pass through collection or existing forest fires video monitoring system by hand
Automatically the method collected establishes preliminary forest rocket image sample data collection；By all forest rocket image samples being collected into
Notebook data collection is all used as training set；
Step 2: training set design convolution is generated to two sub-networks of confrontation neural network；Two sub-networks are respectively to give birth to
At model and discrimination model, and carry out image generation and genuine/counterfeit discriminating；
Step 3: loading true Forest fire image sample, the weight for generating model fixed first, by true Forest fire image sample
This inputs discrimination model with sample is generated respectively, and training discrimination model reaches certain differentiation accuracy；Then fixed to generate mould
One group of stochastic variable input is generated model, then differentiated the sample input discrimination model that model generates is generated by type；Training
When according to this process, while to model and discrimination model is generated carry out backpropagation, alternately training Optimized model parameter；Reach
After dynamic equilibrium, generating model can be generated and the almost consistent new samples of authentic specimen；
Step 4: generating model, input one by the convolutional neural networks that step 3 training obtains can be used for sample generation
Group stochastic variable carries out forward calculation, available one new Forest fire image sample to model is generated.
Further, forest rocket image sample data collection is established in the step 1, image pattern uses the size of W × H
Size, using unsupervised learning mode, not partition testing collection；The scene of sample collection have different terrain landforms, different distance,
Different illumination and multiple video camera shooting angle.
Further, generation model described in step 2 is needed according to the size that input forest rocket image pattern is W × H
Ultimately generate the image data of such size.
Further, traditional convolutional neural networks can be used in the network architecture of discrimination model described in step 2, such as
VGGNet, Inception structure and Resnet；The input layer of discrimination model is inputted using the triple channel image of W × H, difference
It is in and the classification of the true and false two is done to input picture in the last output probability according to neural network；Model is generated in this method to use
For the stochastic variables of 100 dimensions as input, overall structure is similar with network is differentiated, mainly including input layer, hidden layer and output layer,
Wherein hidden layer is all transposition convolutional layer；
Nash Equilibrium can be preferably converged in order to guarantee that confrontation generates network, when network structure designs, generates model
It is matched substantially with the scale of discrimination model needs；In view of discrimination model only needs to realize that two class of the true and false differentiates, generating model is needed
High dimensional image is generated by low-dimensional data, task is more complicated with respect to discrimination model, therefore the scale of discrimination model is answered
This is slightly less than generation model；But generate model and be unable to too complex, it should be ensured that authentic specimen quantity is much larger than generation model
Parameter amount, network training can just obtain zero-sum game solution；Secondly, in order to guarantee that discrimination model possesses good adaptability and differentiation
Ability, discrimination model uses Dropout and L2 regularization submodel training in this method.
Further, in the step 3, the training process for generating confrontation network can be described as a zero-sum game mistake
Journey, game both sides are to generate model and discrimination model respectively；It is true sample that the corresponding sample of discrimination model output 1 is defined in this method
This, the corresponding sample of output 0 is to generate sample；The following formula of the optimization object function of training process indicates:
Wherein, J (θD, θG) it is loss function, θDFor the parameter to be optimized of discrimination model, θGTo generate the to be optimized of model
Parameter, T are the sample size for participating in training, D (xt, θD) indicate to input picture xtDifferentiate the calculating process of processing, G
(zt, θG) indicate to input random signal zt in parameter θGOn the basis of carry out forward calculation obtain generate image calculating process；
D(G(zt, θG), θD) item indicates to the differentiation for generating image as a result, if being authentic specimen by the image discriminating of generation, in above formula
This will increase, and the direction of model towards Nash Equilibrium is promoted to optimize；Similarly, if model misjudges authentic specimen, log
(D(xt, θD)) penalty values of item can promote model to improve the accuracy of identification to true model；
Specifically comprise the following steps:
Step 301 chooses m image pattern x from authentic speciment, wherein t indicates t-th of image pattern, while random
Sample m random data zt, m generation image pattern G (z is obtained by generating model forward calculation G ()t, θG)；Calculate ladder
Degree:
Wherein, every symbol indicate with it is consistent in optimization object function；It is updated using stochastic gradient climb procedure and differentiates mould
Shape parameter θD；
Step 302, secondly from m random data z of stochastical samplingt, m are obtained by generating model G () forward calculation
Generate image G (zt, θG), prediction result D (G (z is calculated by discrimination modelt, θG), θD) after, utilize stochastic gradient descent
The more newly-generated model parameter θ of methodG:
Wherein, every symbol indicate with it is consistent in optimization object function；Training is ignored when generating model to be calculated to true
The penalty values of dataBecause this not generates model parameter θGFunction, derivative be order；
It needs when step 303, training to parameter (θD, θG) alternative optimization is carried out, select Adam method to carry out in this method
Model optimization；By the way that after more wheel iteration optimizations of step 301 and step 302, model eventually reaches Nash Equilibrium.
Further, in the step 4, stochastic variable dimension is 100 dimensions；It can be tested when realistic model training different
Training hyper parameter setting, chooses and shows most excellent model.
The utility model has the advantages that compared with prior art, the invention has the following advantages that
(1) enhancing of traditional sample generallys use the image processing methods such as mirror image, overturning, rotation and increases to sample data
By force, but the effect of data enhancing is extremely limited, easily causes the over-fitting of model.
To solve the above-mentioned problems, confrontation neural network is generated the present invention is based on depth convolution generate new sample data,
The texture and distribution characteristics of forest rocket image pattern can effectively be learnt, data set can be advised in high quality by generating model
Mould expands.
(2) traditional generation confrontation network relies solely on full articulamentum learning sample feature, and it is simple can only to handle some comparisons
Single image.The forest rocket image pattern not being suitable in this scene differentiates and generates.
To solve the above-mentioned problems, the present invention devises the discrimination model based on deep layer convolutional neural networks and generates mould
Type.
By to discrimination model and generating the alternative optimization of model, may finally generate model and discrimination model jointly into
In the case where step, reach the Nash Equilibrium for generating model and discrimination model.Finally obtained generation model can be generated with really
The closely similar generation sample of sample.
Detailed description of the invention
Fig. 1 is model overall structure figure of the invention；
Fig. 2 is discrimination model neural network structure simplification figure of the invention；
Fig. 3 is generation Model Neural structure simplification figure of the invention；
Fig. 4 is the flow chart of the method for the present invention.
Specific embodiment
The invention will be further described with attached drawing is illustrated combined with specific embodiments below.
The forest rocket image pattern generation method that confrontation neural network is generated based on depth convolution, such as Fig. 4 flow chart institute
Show, comprising the following steps:
Step 1: being based on forest rocket video monitoring system, pass through collection or existing forest fires video monitoring system by hand
The methods of automatic collection, establishes preliminary forest rocket image sample data collection.Because it is nothing that convolution, which generates confrontation neural network,
Supervised learning, therefore be not required to divide training set and test set, by all forest rocket image patterns being collected into all as instruction
Practice collection；
Step 2: as shown in Figure 1, needing to design two sub-networks that convolution generates confrontation neural network.Assuming that input forest
Pyrotechnics image pattern is the size of W × H, generates model and needs to ultimately generate the image data of such size, and differentiates mould
Type then inputs the image data of such size, and output is that the true and false two is classified.
Step 3: loading true Forest fire image sample, the weight for generating model fixed first, by true Forest fire image sample
This inputs discrimination model with sample is generated respectively, and training discrimination model reaches certain differentiation accuracy.Then fixed to generate mould
One group of stochastic variable input is generated model, then differentiated the sample input discrimination model that model generates is generated by type.Training
When according to this process, while to model and discrimination model is generated carry out backpropagation, alternately training Optimized model parameter.Reach
After dynamic equilibrium, generating model can be generated and the almost consistent new samples of authentic specimen.
Step 4: generating model by the convolutional neural networks that step 3 training obtains can be used for sample generation, pass through this
The kind available a large amount of Forest fire image sample of mode, solves the sample scarcity problem in the fields such as forest fires video monitoring.
2. the method according to claim 1, wherein establishing forest rocket image sample data in step 1
Collection.Image pattern uses the size of W × H, and this method uses unsupervised learning mode, therefore is not required to partition testing collection.Together
When, to guarantee that training sample has enough representativenesses, when sample collection, needs to consider different terrain landforms, different distance, difference
The covering of the several scenes such as illumination, video camera shooting angle.
3. the method according to claim 1, wherein the confrontation network of generation described in step 2 includes generating
Two sub-networks of model and discrimination model.As shown in Fig. 2, traditional convolutional Neural net can be used in the network architecture of discrimination model
Network, such as VGGNet, Inception structure, the networks such as Resnet.The input layer of discrimination model is defeated using the triple channel image of W × H
Enter, the difference is that finally use Softmax layers as two classifier of the true and false.It is adopted as shown in figure 3, generating model in this method
Use the stochastic variable of 100 dimensions as input, overall structure is similar with network is differentiated, mainly includes input layer, hidden layer and output
Layer, wherein hidden layer is all transposition convolutional layer.Transposition convolutional layer is traditional convolutional layer " reverse " process, and main function is will be defeated
The low-dimensional data entered is mapped to higher dimensional space.
Nash Equilibrium can be preferably converged in order to guarantee that confrontation generates network, when network structure designs, generates model
It is matched substantially with the scale of discrimination model needs.In view of discrimination model only needs to realize that two class of the true and false differentiates, generating model is needed
High dimensional image is generated by low-dimensional data, task is more complicated with respect to discrimination model, therefore the scale of discrimination model is answered
This is slightly less than generation model.But generate model and be unable to too complex, it should be ensured that authentic specimen quantity is much larger than generation model
Parameter amount, network training can just obtain zero-sum game solution.Secondly, in order to guarantee that discrimination model possesses good adaptability and differentiation
Ability, discrimination model uses Dropout and L2 regularization submodel training in this method.
4. the method according to claim 1, wherein the training process for generating confrontation network can in step 3
To be described as a zero-sum game process, game both sides are to generate model and discrimination model respectively.Definition differentiates mould in this method
The corresponding sample of type output 1 is authentic specimen, and the corresponding sample of output 0 is to generate sample.The optimization object function of training process is as follows
Formula indicates:
Wherein, J (θD, θG) it is loss function, θDFor the parameter to be optimized of discrimination model, θGTo generate the to be optimized of model
Parameter, T are the sample size for participating in training, D (xt, θD) indicate the calculating process for carrying out differentiating processing to input picture xt, G
(zt, θG) indicate to input random signal zt in parameter θGOn the basis of carry out forward calculation obtain generate image calculating process.
D(G(zt, θG), θD) item indicate to generate image differentiation as a result, if by the image discriminating of generation be authentic specimen (prediction result
Closer to 1), then this in above formula will increase, and the direction of model towards Nash Equilibrium is promoted to optimize.Similarly, if model pair
Authentic specimen is misjudged, then log (D (xt, θD)) penalty values of item can promote model to improve the accuracy of identification to true model.
It is as follows that process is specifically disposed in step 3:
Step 301 chooses m image pattern x from authentic speciment(wherein t indicates t-th of image pattern), while with
Machine samples m random data zt, m generation image pattern G (z is obtained by generating model forward calculation G ()t, θG).It calculates
Gradient:
Wherein, every symbol indicate with it is consistent in optimization object function.It is updated using stochastic gradient climb procedure and differentiates mould
Shape parameter θD。
Step 302, secondly from m random data z of stochastical samplingt, m are obtained by generating model G () forward calculation
Generate image G (zt, θG), prediction result D (G (z is calculated by discrimination modelt, θG), θD) after, utilize stochastic gradient descent
The more newly-generated model parameter θ of methodG:
Wherein, every symbol indicate with it is consistent in optimization object function.Training is ignored when generating model to be calculated to true
The penalty values of dataBecause this not generates model parameter θGFunction, derivative be order.
It needs when step 303, training to parameter (θD, θG) alternative optimization is carried out, select Adam method to carry out in this method
Model optimization.By the way that after more wheel iteration optimizations of step 301 and step 302, model eventually reaches Nash Equilibrium.
5. the method according to claim 1, wherein in step 4, the stochastic variable dimension in this method is
100 dimensions.Different training hyper parameter settings can be tested when realistic model training, chosen and showed most excellent model.
It should be understood that above-described embodiment is merely to illustrate the specific embodiment of technical solution of the present invention, rather than limitation is originally
The range of invention.After the present invention has been read, those skilled in the art to the modifications of various equivalent forms of the invention and replace
It changes and falls within protection scope defined by the claim of this application.
Claims (6)
1. it is a kind of based on generate confrontation network Forest fire image sample generating method, which is characterized in that this method according to include with
Lower step carries out:
Step 1: it is based on forest rocket video monitoring system, it is automatic by collection or existing forest fires video monitoring system by hand
The method of collection establishes preliminary forest rocket image sample data collection；By all forest rocket image pattern numbers being collected into
According to collection all as training set；
Step 2: training set design convolution is generated to two sub-networks of confrontation neural network；Two sub-networks are respectively to generate mould
Type and discrimination model, and carry out image generation and genuine/counterfeit discriminating；
Step 3: load true Forest fire image sample, the weight for generating model fixed first, by true Forest fire image sample with
It generates sample and inputs discrimination model respectively, training discrimination model reaches certain differentiation accuracy；Then fixed to generate model, it will
One group of stochastic variable input generates model, then differentiates the sample input discrimination model that model generates is generated；It is pressed when training
According to this process, while to model and discrimination model progress backpropagation is generated, Optimized model parameter is alternately trained；Reach dynamic
After balance, generating model can be generated and the almost consistent new samples of authentic specimen；
Step 4: generate model by the convolutional neural networks that step 3 training obtains can be used for sample generation, one group of input with
Machine variable carries out forward calculation, available one new Forest fire image sample to model is generated.
2. according to claim 1 based on the Forest fire image sample generating method for generating confrontation network, which is characterized in that institute
It states and establishes forest rocket image sample data collection in step 1, image pattern uses the size of W × H, using unsupervised learning
Mode, not partition testing collection；The scene of sample collection has different terrain landforms, different distance, different illumination and multiple camera shootings
Machine shooting angle.
3. according to claim 1 based on the Forest fire image sample generating method for generating confrontation network, which is characterized in that step
The rapid 2 generation model needs to ultimately generate such size according to the size that input forest rocket image pattern is W × H
Image data.
4. according to claim 1 based on the Forest fire image sample generating method for generating confrontation network, which is characterized in that step
The network architecture of rapid 2 discrimination model can be used traditional convolutional neural networks, such as VGGNet, Inception structure and
Resnet；The input layer of discrimination model is inputted using the triple channel image of W × H, the difference is that last according to neural network
Output probability to input picture do the true and false two classify；Model is generated in this method, and input is used as using the stochastic variable of 100 dimensions,
Overall structure is similar with network is differentiated, includes mainly input layer, hidden layer and output layer, wherein hidden layer is all transposition convolution
Layer；
Nash Equilibrium can be preferably converged in order to guarantee that confrontation generates network, when network structure designs, model is generated and sentences
The scale of other model needs basic matching；In view of discrimination model only needs to realize that two class of the true and false differentiates, generates model and need to lead to
It crosses low-dimensional data and generates high dimensional image, task is more complicated with respect to discrimination model, therefore the scale of discrimination model should be omited
Less than generation model；But generate model and be unable to too complex, it should be ensured that authentic specimen quantity is much larger than the parameter for generating model
Amount, network training can just obtain zero-sum game solution；Secondly, in order to guarantee that discrimination model possesses good adaptability and differentiation energy
Power, discrimination model uses Dropout and L2 regularization submodel training in this method.
5. according to claim 1 based on the Forest fire image sample generating method for generating confrontation network, which is characterized in that institute
It states in step 3, the training process for generating confrontation network can be described as a zero-sum game process, and game both sides are to generate respectively
Model and discrimination model；It is authentic specimen that the corresponding sample of discrimination model output 1 is defined in this method, and the corresponding sample of output 0 is made a living
At sample；The following formula of the optimization object function of training process indicates:
Wherein, J (θD, θG) it is loss function, θDFor the parameter to be optimized of discrimination model, θGFor the parameter to be optimized for generating model, T
For the sample size for participating in training, D (xt, θD) indicate to input picture xtDifferentiate the calculating process of processing, G (zt, θG) table
Show to input random signal ztIn parameter θGOn the basis of carry out forward calculation obtain generate image calculating process；D(G(zt,
θG), θD) item indicates to the differentiation for generating image as a result, if being authentic specimen, this meeting in above formula by the image discriminating of generation
Increase, the direction of model towards Nash Equilibrium is promoted to optimize；Similarly, if model misjudges authentic specimen, log (D (xt,
θD)) penalty values of item can promote model to improve the accuracy of identification to true model；
Specifically comprise the following steps:
Step 301 chooses m image pattern x from authentic speciment, wherein t indicates t-th of image pattern, while stochastical sampling m
A random data zt, m generation image pattern G (z is obtained by generating model forward calculation G ()t, θG)；Calculate gradient:
Wherein, every symbol indicate with it is consistent in optimization object function；Discrimination model is updated using stochastic gradient climb procedure to join
Number θD；
Step 302, secondly from m random data z of stochastical samplingt, m generation figure is obtained by generating model G () forward calculation
As G (zt, θG), prediction result D (G (z is calculated by discrimination modelt, θG), θD) after, more using stochastic gradient descent method
Newly-generated model parameter θG:
Wherein, every symbol indicate with it is consistent in optimization object function；Training is ignored when generating model to be calculated to truthful data
Penalty valuesBecause this not generates model parameter θGFunction, derivative be order；
It needs when step 303, training to parameter (θD, θG) alternative optimization is carried out, select Adam method to carry out model in this method
Optimization；By the way that after more wheel iteration optimizations of step 301 and step 302, model eventually reaches Nash Equilibrium.
6. according to claim 1 based on the Forest fire image sample generating method for generating confrontation network, which is characterized in that institute
It states in step 4, stochastic variable dimension is 100 dimensions；Different training hyper parameter settings can be tested when realistic model training, chosen
Show most excellent model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811180582.0A CN109460708A (en) | 2018-10-09 | 2018-10-09 | A kind of Forest fire image sample generating method based on generation confrontation network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811180582.0A CN109460708A (en) | 2018-10-09 | 2018-10-09 | A kind of Forest fire image sample generating method based on generation confrontation network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109460708A true CN109460708A (en) | 2019-03-12 |
Family
ID=65607497
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811180582.0A Pending CN109460708A (en) | 2018-10-09 | 2018-10-09 | A kind of Forest fire image sample generating method based on generation confrontation network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109460708A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110147758A (en) * | 2019-05-17 | 2019-08-20 | 电子科技大学成都学院 | A kind of forest fire protection method based on deep learning |
| CN110246093A (en) * | 2019-05-05 | 2019-09-17 | 北京大学 | A kind of decoding image enchancing method |
| CN110351536A (en) * | 2019-08-20 | 2019-10-18 | 深圳江行联加智能科技有限公司 | A kind of substation abnormality detection system, method and device |
| CN110390347A (en) * | 2019-06-10 | 2019-10-29 | 河海大学 | Conditions leading formula confrontation for deep neural network generates test method and system |
| CN110472528A (en) * | 2019-07-29 | 2019-11-19 | 江苏必得科技股份有限公司 | A kind of metro environment target training set creation method and system |
| CN110598530A (en) * | 2019-07-30 | 2019-12-20 | 浙江工业大学 | Small sample radio signal enhanced identification method based on ACGAN |
| CN110765937A (en) * | 2019-10-22 | 2020-02-07 | 新疆天业（集团）有限公司 | Coal yard spontaneous combustion detection method based on transfer learning |
| CN111127447A (en) * | 2019-12-26 | 2020-05-08 | 河南工业大学 | Blood vessel segmentation network and method based on generative confrontation network |
| CN111242216A (en) * | 2020-01-13 | 2020-06-05 | 北京工业大学 | Image generation method for generating anti-convolution neural network based on conditions |
| CN111275640A (en) * | 2020-01-17 | 2020-06-12 | 天津大学 | Image enhancement method for fusing two-dimensional discrete wavelet transform and generating countermeasure network |
| CN111476965A (en) * | 2020-03-13 | 2020-07-31 | 深圳信息职业技术学院 | Method for constructing fire detection model, fire detection method and related equipment |
| CN111563482A (en) * | 2020-06-18 | 2020-08-21 | 深圳天海宸光科技有限公司 | Gas station dangerous scene picture generation method based on GAN |
| CN111814689A (en) * | 2020-07-09 | 2020-10-23 | 浙江大华技术股份有限公司 | Fire recognition network model training method, fire recognition method and related equipment |
| CN112309068A (en) * | 2020-10-29 | 2021-02-02 | 电子科技大学中山学院 | Forest fire early warning method based on deep learning |
| CN112560575A (en) * | 2020-11-09 | 2021-03-26 | 北京物资学院 | Red Fuji apple shape data enhancement device and method |
| CN113128422A (en) * | 2021-04-23 | 2021-07-16 | 重庆市海普软件产业有限公司 | Image smoke and fire detection method and system of deep neural network |
| CN116822589A (en) * | 2023-06-14 | 2023-09-29 | 西南交通大学 | Pyrotechnic image generation method and system |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
| CN107392255A (en) * | 2017-07-31 | 2017-11-24 | 深圳先进技术研究院 | Generation method, device, computing device and the storage medium of minority class picture sample |
| CN107958216A (en) * | 2017-11-27 | 2018-04-24 | 沈阳航空航天大学 | Based on semi-supervised multi-modal deep learning sorting technique |
| CN108229381A (en) * | 2017-12-29 | 2018-06-29 | 湖南视觉伟业智能科技有限公司 | Face image synthesis method, apparatus, storage medium and computer equipment |
| CN108256238A (en) * | 2018-01-22 | 2018-07-06 | 武汉理工大学 | A kind of optic fiber grating wavelength demodulation method and device based on deep learning |
| CN108268638A (en) * | 2018-01-18 | 2018-07-10 | 浙江工业大学 | A kind of generation confrontation network distribution type implementation method based on Spark frames |
| CN108334936A (en) * | 2018-01-30 | 2018-07-27 | 华中科技大学 | Failure prediction method based on migration convolutional neural networks |
-
2018
- 2018-10-09 CN CN201811180582.0A patent/CN109460708A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
| CN107392255A (en) * | 2017-07-31 | 2017-11-24 | 深圳先进技术研究院 | Generation method, device, computing device and the storage medium of minority class picture sample |
| CN107958216A (en) * | 2017-11-27 | 2018-04-24 | 沈阳航空航天大学 | Based on semi-supervised multi-modal deep learning sorting technique |
| CN108229381A (en) * | 2017-12-29 | 2018-06-29 | 湖南视觉伟业智能科技有限公司 | Face image synthesis method, apparatus, storage medium and computer equipment |
| CN108268638A (en) * | 2018-01-18 | 2018-07-10 | 浙江工业大学 | A kind of generation confrontation network distribution type implementation method based on Spark frames |
| CN108256238A (en) * | 2018-01-22 | 2018-07-06 | 武汉理工大学 | A kind of optic fiber grating wavelength demodulation method and device based on deep learning |
| CN108334936A (en) * | 2018-01-30 | 2018-07-27 | 华中科技大学 | Failure prediction method based on migration convolutional neural networks |
Non-Patent Citations (3)
| Title |
|---|
| ALEC RADFORD 等: "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks", 《HTTPS://ARXIV.ORG/ABS/1511.06434》 * |
| IAN J.GOODFELLOW 等: "Generative Adversarial Nets", 《28TH CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS》 * |
| 洪洋 等: "深度卷积对抗生成网络综述", 《第18届中国系统仿真技术及其应用学术年会论文集（18TH CCSSTA 2017）》 * |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110246093B (en) * | 2019-05-05 | 2021-05-04 | 北京大学 | Method for enhancing decoded image |
| CN110246093A (en) * | 2019-05-05 | 2019-09-17 | 北京大学 | A kind of decoding image enchancing method |
| CN110147758A (en) * | 2019-05-17 | 2019-08-20 | 电子科技大学成都学院 | A kind of forest fire protection method based on deep learning |
| CN110390347A (en) * | 2019-06-10 | 2019-10-29 | 河海大学 | Conditions leading formula confrontation for deep neural network generates test method and system |
| CN110472528A (en) * | 2019-07-29 | 2019-11-19 | 江苏必得科技股份有限公司 | A kind of metro environment target training set creation method and system |
| CN110598530A (en) * | 2019-07-30 | 2019-12-20 | 浙江工业大学 | Small sample radio signal enhanced identification method based on ACGAN |
| CN110351536A (en) * | 2019-08-20 | 2019-10-18 | 深圳江行联加智能科技有限公司 | A kind of substation abnormality detection system, method and device |
| CN110765937A (en) * | 2019-10-22 | 2020-02-07 | 新疆天业（集团）有限公司 | Coal yard spontaneous combustion detection method based on transfer learning |
| CN111127447A (en) * | 2019-12-26 | 2020-05-08 | 河南工业大学 | Blood vessel segmentation network and method based on generative confrontation network |
| CN111127447B (en) * | 2019-12-26 | 2023-03-31 | 河南工业大学 | Blood vessel segmentation network and method based on generative confrontation network |
| CN111242216A (en) * | 2020-01-13 | 2020-06-05 | 北京工业大学 | Image generation method for generating anti-convolution neural network based on conditions |
| CN111275640B (en) * | 2020-01-17 | 2022-12-09 | 天津大学 | Image enhancement method for fusing two-dimensional discrete wavelet transform and generation of countermeasure network |
| CN111275640A (en) * | 2020-01-17 | 2020-06-12 | 天津大学 | Image enhancement method for fusing two-dimensional discrete wavelet transform and generating countermeasure network |
| CN111476965B (en) * | 2020-03-13 | 2021-08-03 | 深圳信息职业技术学院 | Method for constructing fire detection model, fire detection method and related equipment |
| CN111476965A (en) * | 2020-03-13 | 2020-07-31 | 深圳信息职业技术学院 | Method for constructing fire detection model, fire detection method and related equipment |
| CN111563482A (en) * | 2020-06-18 | 2020-08-21 | 深圳天海宸光科技有限公司 | Gas station dangerous scene picture generation method based on GAN |
| CN111814689A (en) * | 2020-07-09 | 2020-10-23 | 浙江大华技术股份有限公司 | Fire recognition network model training method, fire recognition method and related equipment |
| CN111814689B (en) * | 2020-07-09 | 2024-08-20 | 浙江大华技术股份有限公司 | Training method of fire disaster identification network model, fire disaster identification method and related equipment |
| CN112309068A (en) * | 2020-10-29 | 2021-02-02 | 电子科技大学中山学院 | Forest fire early warning method based on deep learning |
| CN112560575A (en) * | 2020-11-09 | 2021-03-26 | 北京物资学院 | Red Fuji apple shape data enhancement device and method |
| CN112560575B (en) * | 2020-11-09 | 2023-07-18 | 北京物资学院 | Red Fuji apple shape data enhancement device and method |
| CN113128422A (en) * | 2021-04-23 | 2021-07-16 | 重庆市海普软件产业有限公司 | Image smoke and fire detection method and system of deep neural network |
| CN113128422B (en) * | 2021-04-23 | 2024-03-29 | 重庆市海普软件产业有限公司 | Image smoke and fire detection method and system for deep neural network |
| CN116822589A (en) * | 2023-06-14 | 2023-09-29 | 西南交通大学 | Pyrotechnic image generation method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109460708A (en) | A kind of Forest fire image sample generating method based on generation confrontation network | |
| CN109522819B (en) | Fire image identification method based on deep learning | |
| CN108537743B (en) | Face image enhancement method based on generation countermeasure network | |
| Prema et al. | A novel efficient video smoke detection algorithm using co-occurrence of local binary pattern variants | |
| CN109670528B (en) | Data expansion method facing pedestrian re-identification task and based on paired sample random occlusion strategy | |
| CN103886344B (en) | A kind of Image Fire Flame recognition methods | |
| CN110008842A (en) | A kind of pedestrian's recognition methods again for more losing Fusion Model based on depth | |
| CN110472627A (en) | One kind SAR image recognition methods end to end, device and storage medium | |
| Li et al. | Adversarial open-world person re-identification | |
| CN113221655B (en) | Face spoofing detection method based on feature space constraint | |
| CN104504362A (en) | Face detection method based on convolutional neural network | |
| CN113642621B (en) | Zero sample image classification method based on generation countermeasure network | |
| CN109886227A (en) | Inside fire video frequency identifying method based on multichannel convolutive neural network | |
| CN108805078A (en) | Video pedestrian based on pedestrian's average state recognition methods and system again | |
| CN110652726B (en) | Game auxiliary system based on image recognition and audio recognition | |
| CN109344856A (en) | A kind of off-line signature verification method based on multilayer discriminate feature learning | |
| Capece et al. | Implementation of a coin recognition system for mobile devices with deep learning | |
| CN109919921A (en) | Based on the influence degree modeling method for generating confrontation network | |
| CN113947579A (en) | Confrontation sample detection method for image target detection neural network | |
| Liu et al. | Weighted sequence loss based spatial-temporal deep learning framework for human body orientation estimation | |
| Zhao et al. | Cross-modal knowledge distillation for fine-grained one-shot classification | |
| CN108960013A (en) | A kind of pedestrian recognition methods and device again | |
| CN117036875A (en) | Infrared weak and small moving target generation algorithm based on fusion attention GAN | |
| Yin et al. | An improved algorithm based on convolutional neural network for smoke detection | |
| CN110516573A (en) | The two stages smog identification convolutional neural networks of color combining, textural characteristics |
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: 20190312 |