CN109543740A - A kind of object detection method based on generation confrontation network - Google Patents
A kind of object detection method based on generation confrontation network Download PDFInfo
- Publication number
- CN109543740A CN109543740A CN201811363392.2A CN201811363392A CN109543740A CN 109543740 A CN109543740 A CN 109543740A CN 201811363392 A CN201811363392 A CN 201811363392A CN 109543740 A CN109543740 A CN 109543740A
- Authority
- CN
- China
- Prior art keywords
- data
- generator
- network
- confrontation
- object detector
- 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.)
- Granted
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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 present invention provides a kind of based on the object detection method for generating confrontation network, design generator, Different categories of samples is generated according to class label, design proxy server, detect the data of generator, pseudo- true value is provided, and the data application for generating proxy server is in the training of object detector, design object detector, judge to generate the promotion whether data are conducive to target detection precision, design confrontation device, in the training stage, differentiate that data are derived from truthful data and still generate data, generator replaces training with arbiter, in test phase, data to be tested directly input object detector, obtain testing result.The present invention generates the sample that network generates can enrich training data in conjunction with authentic specimen, improve detection accuracy, target detection network provides feedback to network is generated, keep the sample generated truer, the data that proxy server generates directly apply to the training of object detector, it is labeled without expending a large amount of manpower and material resources, the configuration of the present invention is simple is easy to dispose.
Description
Technical field
The invention belongs to object detection method fields, and in particular to a kind of based on the target detection side for generating confrontation network
Method.
Background technique
Deep learning is fast-developing in recent years, and the algorithm of target detection based on deep learning yields good result, but
It is that there are still problems to be solved.Firstly, the algorithm of target detection based on deep learning needs a large amount of mark samples, small data set
It is upper to tend to over-fitting using deep learning.It was found that a rough rule of thumb is, deep learning algorithm is supervised every
Class is generally up to acceptable performance in the case of 5000 mark samples to concludeing a contract or treaty, when at least 10,000,000 mark samples
Data set for train when, be up to or be more than the mankind show.In addition, the mark of data set will also expend a large amount of manpower object
Power.Therefore how using do not mark largely sample or generate sample be trained be worth further research.Secondly, at this stage
Target detection research is most of to improve precision dependent on deeper network, but therefore can bring more complicated calculating and more
More memory requirements consume a large amount of hardware, it is difficult to dispose.Generating confrontation network is also that nowadays forward position the most is also in deep learning
One of field the most fascinating, however most research focuses on the quality for improving and generating data, how application generates
The research of data is less.
In view of the above problems, the invention proposes a kind of based on the object detection method for generating confrontation network.Firstly, will give birth to
At confrontation network and the target detection network integration.Training data can be enriched in conjunction with authentic specimen by generating the sample that network generates,
Detection accuracy is improved, while target detection network keeps the sample generated truer also to network offer feedback is generated.The present invention
In refer to proxy server, the sample of generation can generate pseudo- true value by proxy server, make to generate data and directly apply to target detection
The training of device is labeled without expending a large amount of manpower and material resources.Secondly, this method application target during actual test
Network is detected, structure is simple, is easy to dispose.
Summary of the invention
The object of the present invention is to provide a kind of based on the object detection method for generating confrontation network, is suitable for a small amount of mark
Infuse the target detection based on generation confrontation network in the case of sample.
The object of the present invention is achieved like this:
A kind of object detection method based on generation confrontation network, steps are as follows for concrete implementation:
Step 1. designs generator G, generates Different categories of samples according to class label；
Step 2. chooses the high-precision detector trained as proxy server F, examines to the generator G data generated
It surveys, provides pseudo- true value, and the data application that proxy server F is generated is in the training of object detector；
Step 3. design object detector O judges to generate whether data are conducive to target detection precision in the training stage
It is promoted, provides feedback for generator, object detector O is the final output of test phase；
Step 4. design confrontation device A differentiates that data are derived from truthful data and still generate data, is in the training stage
Generator provides feedback；
Step 5. replaces training with arbiter D in training stage, generator G, and wherein the input of generator G standardizes to -1
To between 1 range, the input of arbiter D is truthful data and generates data, and generator uses Adam as optimizer,
Arbiter uses SGD as optimizer；
For step 6. in test phase, data to be tested directly input object detector O, obtain testing result.
The step 1 generator G uses the GAN network structure of conditional constraint, input condition variable y, i.e. classification mark
Label, generate different classes of sample under the guidance of class label, and generator G uses residual error network structure, including 4 residual blocks
And up-sampling layer, wherein residual block is two BN layers, and ReLU layers, the set of 3*3 convolutional layer, generator G output layer uses Tanh
Activation primitive.
The step 3 object detector O uses full convolution framework, and characteristic extraction part shares 15 layers of convolution, wherein
The convolution kernel size of 2nd, 4,6,8,10 layer of convolution is 3*3, and step-length 2, remaining convolutional layer step-length is 1, and uses multiple features
Integration technology, by low layer convolution feature and high-rise convolution Fusion Features.
It fights device A and introduces conditional-variable y, network structure includes 5 residual blocks and down-sampling layer, is fought residual in device A
Poor block is two LeakyReLU layers, the set of 3*3 convolutional layer, and confrontation device A uses spectrum normalization technology.
The parameter of generator G is θ in the step 1g, by optimizing loss function LDθ can be obtainedg,
Loss function LDL is lost by confrontationD_AWith Detectability loss LD_OTwo parts composition, i.e. LD=LD-A+λLD-O, confrontation loss
LD_AFor
Wherein Z is the input of generator, and y is class label,For the output of generator,Refer to that parameter is θa
Differentiation network confrontation branch,Generation sample is determined as to the probability of authentic specimen for confrontation device；Detection
Lose LD_OFor
WhereinFor object detector output target information,For object detector output class probability,For mesh
The position coordinates of detector output are marked,For proxy server output target information,For proxy server output class probability,
For the position coordinates of proxy server output
The parameter that device A is fought in the step 4 is θa, i.e.,
The parameter of object detector O is θ in the step 3o, i.e.,
The beneficial effects of the present invention are: confrontation network and the target detection network integration will be generated, generates what network generated
Sample can enrich training data in conjunction with authentic specimen, improve detection accuracy, while target detection network also mentions generation network
For feedback, keep the sample generated truer, proxy server is referred in the present invention, the sample of generation can be generated pseudo- by proxy server
True value makes to generate the training that data directly apply to object detector, be labeled without expending a large amount of manpower and material resources,
Secondary, this method application target during actual test detects network, and structure is simple, is easy to dispose.
Detailed description of the invention
Fig. 1 is inventive network architecture diagram.
Fig. 2 is generator G architecture diagram of the invention.
Fig. 3 is residual error block structural diagram in generator G of the invention.
Fig. 4 is object detector O characteristic extraction part architecture diagram of the invention.
Fig. 5 is confrontation device A architecture diagram of the invention.
Fig. 6 is residual error block structural diagram in confrontation device A of the invention.
Specific embodiment
The present invention will be further described with reference to the accompanying drawing:
Embodiment 1
A kind of object detection method based on generation confrontation network, steps are as follows for concrete implementation:
Step 1. designs generator G, generates Different categories of samples according to class label, as shown in Figure 2；
Step 2. chooses the high-precision detector trained as proxy server F, examines to the generator G data generated
It surveys, provides pseudo- true value, and the data application that proxy server F is generated selects YOLO in the training of object detector in this programme
V3 network is acted as agent device, and the sample that generator generates is generated by proxy server and surrounds frame information and classification information, very as agency
Value, the introducing of proxy server realize the direct application for generating data, without additional mark, realize network structure end to end；
Step 3. design object detector O judges to generate whether data are conducive to target detection precision in the training stage
It is promoted, provides feedback for generator, object detector O is the final output of test phase, as shown in figure 4, object detector O makees
For one of arbiter, replace training with generator in the training stage, judges whether the data generated are conducive to mentioning for detection accuracy
It rises, provides feedback for generator, the input of object detector O is the pseudo- true value that the sample generated and proxy server generate, output
For the classification and position coordinates for detecting target, characteristic extraction part uses full convolution framework, altogether by 15 layers of convolution, wherein the
2, the convolution kernel size of 4,6,8,10 layers of convolution is 3*3, and step-length 2, remaining convolutional layer step-length is 1.Here entire in order to guarantee
The stability of network model, all pond layers are replaced by the convolutional layer of involvement step-length, furthermore richer thin in order to obtain
Feature is saved, using multiple features fusion method, low layer convolution feature and high-rise convolution feature are blended, the knot of object detector O
Structure is similar to SSD, YOLO, belongs to stage detection, and therefore, different from RCNN Train detector, bounding box coordinates and classification are general
Rate is predicted to be the output of the last layer simultaneously.Each cell position in the last layer characteristic pattern predicts N number of bounding box, wherein
N is the quantity of anchor frame.The quantity of Feature Mapping in the last layer is arranged to N × (K+5), and wherein K is for predicting that class is general
The quantity of the class of rate, 5 refer to bounding box coordinates and target value (5=4+1)；
Step 4. design confrontation device A differentiates that data are derived from truthful data and still generate data, is in the training stage
Generator provides feedback, as shown in figure 5, confrontation device A is one of arbiter, inputs as truthful data and generates the defeated of model
Out, while input is also introduced into classification information, exports as two classes 0/1, wherein 1 is true, i.e. truthful data, 0 is puppet, i.e. generation number
According to confrontation device A uses residual error network structure, and residual block includes two ReLU layers of Leaky, 3*3 convolutional layers.Share 5 residual errors
Block, each residual block are followed by a down-sampling layer, and fight device A and normalize this weight normalization technology, the party using spectrum
Method calculation amount is few, and can generate or quality comparable picture higher compared with other technologies quality；
Step 5. replaces training with arbiter D in training stage, generator G, and wherein the input of generator G standardizes to -1
To between 1 range, the input of arbiter D is truthful data and generates data, and generator uses Adam as optimizer,
Arbiter uses SGD as optimizer, and the training process of whole network is that generator is allowed to vie each other with arbiter, and the two exists
Ability is constantly promoted during alternating training, is ultimately generated device by the intrinsic propesties of study truthful data, is generated and true
The similar new data of data.Discrimination model consists of two parts, and confrontation device A judges the true and false of input data, and object detector O sentences
Can the data of disconnected input improve the precision of target detection, and discrimination model is played the role of being that guidance generates how model adjusts
The generation data being desirably to obtain are closer to truthful data, prevent repetition training process in divergent state, in the training process
Arbiter is first trained, generator is then trained, the two is alternately trained；
For step 6. in test phase, data to be tested directly input object detector O, obtain testing result.
The step 1 generator G uses the GAN network structure of conditional constraint, and G is from a noise pzStart, passes through
Network propagated forward generates piece image, and it is desirable that the picture generated tends to true picture, the deception that can mix the spurious with the genuine is sentenced
Other device, input condition variable y, i.e. class label generate different classes of sample under the guidance of class label, and generator G makes
With residual error network structure, including 4 residual blocks and up-sampling layer, wherein residual block is two BN layers, ReLU layers, 3*3 convolutional layer
Set, generator G output layer use Tanh activation primitive.
The step 3 object detector O uses full convolution framework, and characteristic extraction part shares 15 layers of convolution, convolution
Core size is 3*3,1*1, and uses multiple features fusion technology, by low layer convolution feature and high-rise convolution Fusion Features.
The confrontation device A introduces conditional-variable y, and network structure includes 5 residual blocks and down-sampling layer, fights device A
In residual block be two LeakyReLU layer, the set of 3*3 convolutional layer, confrontation device A using compose normalization technology.
The parameter of generator G is θ in the step 1g, by optimizing loss function LDθ can be obtainedg,
Loss function LDL is lost by confrontationD_AWith Detectability loss LD_OTwo parts composition, i.e. LD=LD-A+λLD-O, confrontation loss
LD_AFor
Wherein Z is the input of generator, and y is class label,For the output of generator, Detectability loss LD_OFor
WhereinFor object detector output target information,For object detector output class probability,For mesh
The position coordinates of detector output are marked,For proxy server output target information,For proxy server output class probability,
For the position coordinates of proxy server output
The parameter that device A is fought in the step 4 is θa, i.e.,
The parameter of object detector O is θ in the step 3o, i.e.,
Based on the object detection method for generating confrontation network, the present invention relates to a kind of object detection methods, are specifically based on
Generate the object detection method in the case of a small amount of mark sample of confrontation network.
Deep learning is fast-developing in recent years, and the algorithm of target detection based on deep learning yields good result, but
It is that there are still problems to be solved.Firstly, the algorithm of target detection based on deep learning needs a large amount of mark samples, small data set
It is upper to tend to over-fitting using deep learning.It was found that a rough rule of thumb is, deep learning algorithm is supervised every
Class is generally up to acceptable performance in the case of 5000 mark samples to concludeing a contract or treaty, when at least 10,000,000 mark samples
Data set for train when, be up to or be more than the mankind show.In addition, the mark of data set will also expend a large amount of manpower object
Power.Therefore how using do not mark largely sample or generate sample be trained be worth further research.Secondly, at this stage
Target detection research is most of to improve precision dependent on deeper network, but therefore can bring more complicated calculating and more
More memory requirements consume a large amount of hardware, it is difficult to dispose.Generating confrontation network is also that nowadays forward position the most is also in deep learning
One of field the most fascinating, however most research focuses on the quality for improving and generating data, how application generates
The research of data is less.
In view of the above problems, the invention proposes a kind of based on the object detection method for generating confrontation network.Firstly, will give birth to
At confrontation network and the target detection network integration.Training data can be enriched in conjunction with authentic specimen by generating the sample that network generates,
Detection accuracy is improved, while target detection network keeps the sample generated truer also to network offer feedback is generated.The present invention
In refer to proxy server, the sample of generation can generate pseudo- true value by proxy server, make to generate data and directly apply to target detection
The training of device is labeled without expending a large amount of manpower and material resources.Secondly, this method application target during actual test
Network is detected, structure is simple, is easy to dispose.
It is an object of the invention to propose it is a kind of suitable for mark on a small quantity in the case of sample based on generating confrontation network
Object detection method, this method, which introduces, generates confrontation network, will generate confrontation network and target detection network integration, and design suitable
Generator and arbiter for this structure.Overall network structure is as shown in Fig. 1, is mainly made of three parts: generator
G, proxy server F, arbiter D, wherein arbiter includes object detector O and confrontation device A.This method comprises the following steps:
S1. generator G is designed, Different categories of samples is generated according to class label, structure chart is as shown in Fig. 2.
S2. proxy server F is introduced, selects in this programme YOLO v3 detector to act as agent device, the data generated to generator G
It is detected, pseudo- true value is provided.Proxy server F can make to generate the training that data directly apply to object detector, without additional mark
Note realizes network structure end to end.
S3. design object detector O, structure chart are as shown in Fig. 4.In the training stage, judge to generate whether data have
Conducive to the promotion of target detection precision, feedback is provided for generator.Object detector O is also the final output of test phase.
S4. design confrontation device A, structure chart are as shown in Fig. 5.In the training stage, differentiate that data are derived from true number
According to data are still generated, feedback is provided for generator.
S5. in the training stage, generator G replaces training with arbiter D, and wherein the input of generator G standardizes to -1 to 1
Range between, the input of arbiter D is truthful data and generates data, and generator uses Adam as optimizer, sentences
Other device uses SGD as optimizer.Ability is constantly promoted both in training process, is kept the data generated truer, is examined simultaneously
Precision is surveyed also to be continuously improved.
S6. in test phase, data to be tested directly input object detector O and obtain test result.
1 couple of present invention is more fully described with reference to the accompanying drawing.
S1. as shown in Fig. 2, constructing generator a G, G from a noise pzStart, generates one by network propagated forward
Width image, and it is desirable that the true picture of picture trend generated, can mix the spurious with the genuine and cheat arbiter.Meanwhile classification being believed
Breath, which incorporates, generates network, generates Different categories of samples under the guidance of classification information.Generator uses residual error network structure, and residual block is such as
It include two BN layers, RELU layers and 3*3 convolutional layer shown in Fig. 3.Altogether there are four residual block, each residual block is followed by adopting on one
Sample layer.And the output layer of generator uses Tanh activation primitive.
S2. YOLO v3 network is selected to act as agent device in this programme, the input of proxy server is the output of generator.It generates
The sample that device generates generates encirclement frame information and classification information by proxy server, as acting on behalf of true value.The introducing of proxy server is realized
The direct application for generating data realizes network structure end to end without additional mark.
S3. as shown in figure 4, building object detector O, object detector O is as one of arbiter, in training stage and life
It grows up to be a useful person and alternately trains, judge whether the data generated are conducive to the promotion of detection accuracy, provide feedback for generator.Target detection
The input of device O is the pseudo- true value that the sample generated and proxy server generate, and exports classification and position to detect target and sits
Mark.Its characteristic extraction part uses full convolution framework, altogether by 15 layers of convolution, wherein the convolution kernel of the 2nd, 4,6,8,10 layer of convolution is big
Small is 3*3, and step-length 2, remaining convolutional layer step-length is 1.Here in order to guarantee the stability of whole network model, all ponds
Layer is replaced by the convolutional layer of involvement step-length.It furthermore, will using multiple features fusion method in order to obtain richer minutia
Low layer convolution feature is blended with high-rise convolution feature.The structure of object detector O is similar to SSD, YOLO, belongs to stage inspection
It surveys, therefore, different from RCNN Train detector, bounding box coordinates and class probability while the output for being predicted to be the last layer.
Each cell position in the last layer characteristic pattern predicts N number of bounding box, and wherein N is the quantity of anchor frame.Spy in the last layer
The quantity of sign mapping is arranged to N × (K+5), and wherein K is the quantity for predicting the class of class probability, and 5 refer to bounding box seat
Mark and target value (5=4+1).
S4. it as shown in figure 5, building confrontation device A, confrontation device A are one of arbiter, inputs as truthful data and generates mould
The output (generating data) of type, while input is also introduced into classification information.Output is two classes 0/1, wherein 1 is true, i.e. truthful data,
0 is puppet, i.e. generation data.It fights device A and uses residual error network structure, residual block includes two ReLU layers of Leaky, 3*3 convolution
Layer.5 residual blocks are shared, each residual block is followed by a down-sampling layer.And fighting device A, this weight is returned using spectrum normalization
One changes technology, and this method calculation amount is few, and can generate or quality comparable picture higher compared with other technologies quality.
S5. the training process of whole network is that generator is allowed to vie each other with arbiter, mistake of the two in alternately training
Ability is constantly promoted in journey, is ultimately generated device by the intrinsic propesties of study truthful data, is generated similar with truthful data new
Data.Discrimination model consists of two parts, and confrontation device A judges the true and false of input data, the data of object detector O judgement input
The precision of target detection can be improved.Discrimination model is played the role of being that guidance generates how model adjusts the life being desirably to obtain
It is closer to truthful data at data, prevents repetition training process in divergent state.Arbiter is first trained in the training process,
Then generator is trained, the two is alternately trained, uses Adam as optimizer for generator, made for arbiter using SGD
For optimizer, loss function is as follows:
θgFor the parameter of generator G, generator wishes that the data generated can cheat arbiter, makes arbiter will
It is mistaken for very, by optimizing loss function LDθ can be obtainedg, it is specific as follows shown,
Wherein loss function LDL is lost by confrontationD_AWith Detectability loss LD_OTwo parts composition.I.e.
LD=LD-A+λLD-O, (2)
Wherein fight loss function LD_AAs follows, Z is the input of generator, and y is class label,For generator
Output,Refer to that parameter is θaDifferentiation network confrontation branch,Sample will be generated for confrontation device to differentiate
For the probability of authentic specimen.
Detectability loss LD_OThree parts can be divided into: returning loss, target loss and Classification Loss.Wherein,For object detector output target information, class probability and position coordinates,For proxy server
The class label and position coordinates of output.
θaFor the parameter for fighting device A, truthful data is judged as very by the expectation of confrontation device, generates data and is judged as
Parameter θ can be obtained by following formula in vacationa。
θoFor the parameter of object detector O, object detector be desirable to the classification for correctly detecting target with
And location information, parameter θ is waited until by following formulao。
S6. trained target detection O is used in test phase, input data is tested, exported as corresponding classification
Information and encirclement frame information.
Claims (7)
1. a kind of based on the object detection method for generating confrontation network, which is characterized in that steps are as follows for concrete implementation:
Step 1. designs generator G, generates Different categories of samples according to class label；
Step 2. chooses the high-precision detector trained as proxy server F, detects, mentions to the generator G data generated
For pseudo- true value, and the data application that proxy server F is generated is in the training of object detector；
Step 3. design object detector O judges to generate whether data are conducive to mentioning for target detection precision in the training stage
It rises, provides feedback for generator, object detector O is the final output of test phase；
Step 4. design confrontation device A differentiates that data are derived from truthful data and still generate data, in the training stage to generate
Device provides feedback；
Step 5. replaces training with arbiter D in training stage, generator G, and wherein the input of generator G standardizes to -1 to 1
Range between, the input of arbiter D is truthful data and generates data, and generator uses Adam as optimizer, sentences
Other device uses SGD as optimizer；
For step 6. in test phase, data to be tested directly input object detector O, obtain testing result.
2. according to claim 1 a kind of based on the object detection method for generating confrontation network, it is characterised in that: described
Step 1 generator G uses the GAN network structure of conditional constraint, input condition variable y, i.e. class label, in class label
Guidance is lower to generate different classes of sample, and generator G uses residual error network structure, including 4 residual blocks and up-sampling layer, wherein
Residual block is two BN layers, ReLU layers with the set of 3*3 convolutional layer, generator G output layer uses Tanh activation primitive.
3. according to claim 1 a kind of based on the object detection method for generating confrontation network, it is characterised in that: described
Step 3 object detector O uses full convolution framework, and characteristic extraction part shares 15 layers of convolution, wherein the 2nd, 4,6,8,10 layer
The convolution kernel size of convolution is 3*3, and step-length 2, remaining convolutional layer step-length is 1, and uses multiple features fusion technology, by low layer
Convolution feature and high-rise convolution Fusion Features.
4. according to claim 1 a kind of based on the object detection method for generating confrontation network, it is characterised in that: confrontation device
A introduces conditional-variable y, and network structure includes 5 residual blocks and down-sampling layer, and fighting the residual block in device A is two
The LeakyReLU layers of set with 3*3 convolutional layer, confrontation device A use spectrum normalization technology.
5. according to claim 1 a kind of based on the object detection method for generating confrontation network, it is characterised in that: described
The parameter of generator G is θ in step 1g, by optimizing loss function LDθ can be obtainedg,
Loss function LDL is lost by confrontationD_AWith Detectability loss LD_OTwo parts composition, i.e. LD=LD-A+λLD-O, confrontation loss LD_A
For
Wherein Z is the input of generator, and y is class label,For the output of generator,Refer to that parameter is θaSentence
The confrontation branch of other network,Generation sample is determined as to the probability of authentic specimen for confrontation device；Detectability loss
LD_OFor
WhereinFor object detector output target information,For object detector output class probability,For target detection
The position coordinates of device output,For proxy server output target information,For proxy server output class probability,For agency
The position coordinates of device output.
6. according to claim 5 a kind of based on the object detection method for generating confrontation network, it is characterised in that: described
The parameter that device A is fought in step 4 is θa, i.e.,
7. according to claim 5 a kind of based on the object detection method for generating confrontation network, it is characterised in that: described
The parameter of object detector O is θ in step 3o, i.e.,
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811363392.2A CN109543740B (en) | 2018-11-14 | 2018-11-14 | Target detection method based on generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811363392.2A CN109543740B (en) | 2018-11-14 | 2018-11-14 | Target detection method based on generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109543740A true CN109543740A (en) | 2019-03-29 |
| CN109543740B CN109543740B (en) | 2022-07-15 |
Family
ID=65847906
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811363392.2A Active CN109543740B (en) | 2018-11-14 | 2018-11-14 | Target detection method based on generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109543740B (en) |
Cited By (12)
| 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 |
| CN110458185A (en) * | 2019-06-26 | 2019-11-15 | 平安科技（深圳）有限公司 | Image-recognizing method and device, storage medium, computer equipment |
| CN110503974A (en) * | 2019-08-29 | 2019-11-26 | 泰康保险集团股份有限公司 | Fight audio recognition method, device, equipment and computer readable storage medium |
| CN110795892A (en) * | 2019-10-23 | 2020-02-14 | 北京邮电大学 | Channel simulation method and device based on generation countermeasure network |
| CN110969200A (en) * | 2019-11-27 | 2020-04-07 | 中国科学院自动化研究所 | Image target detection model training method and device based on consistency negative sample |
| CN111651561A (en) * | 2020-06-05 | 2020-09-11 | 拾音智能科技有限公司 | High-quality difficult sample generation method |
| CN111860593A (en) * | 2020-06-15 | 2020-10-30 | 北京华电天仁电力控制技术有限公司 | Fan blade fault detection method based on generation countermeasure network |
| CN111899251A (en) * | 2020-08-06 | 2020-11-06 | 中国科学院深圳先进技术研究院 | Copy-move type forged image detection method for distinguishing forged source and target area |
| CN112465037A (en) * | 2020-11-30 | 2021-03-09 | 深圳供电局有限公司 | Target detection method and device, computer equipment and storage medium |
| WO2021169157A1 (en) * | 2020-02-25 | 2021-09-02 | 浙江工业大学 | Feature remapping-based adversarial sample defense method and application |
| CN114648724A (en) * | 2022-05-18 | 2022-06-21 | 成都航空职业技术学院 | Lightweight efficient target segmentation and counting method based on generation countermeasure network |
| WO2023050746A1 (en) * | 2021-09-30 | 2023-04-06 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Method for enhancing sar image data for ship target detection |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108564109A (en) * | 2018-03-21 | 2018-09-21 | 天津大学 | A kind of Remote Sensing Target detection method based on deep learning |
| US20180293721A1 (en) * | 2017-04-07 | 2018-10-11 | Kla-Tencor Corporation | Contour based defect detection |
-
2018
- 2018-11-14 CN CN201811363392.2A patent/CN109543740B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180293721A1 (en) * | 2017-04-07 | 2018-10-11 | Kla-Tencor Corporation | Contour based defect detection |
| CN108564109A (en) * | 2018-03-21 | 2018-09-21 | 天津大学 | A kind of Remote Sensing Target detection method based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| FERNANDO CAMARO NOGUES ET AL.: "Object Detection using Domain Randomization and Generative Adversarial Refinement of Synthetic Images", 《HTTPS://ARXIV.ORG/ABS/1805.11778》 * |
| 杨沛: "双判别器生成对抗网络及其在接触网鸟巢检测的应用研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (18)
| 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 |
| WO2020258667A1 (en) * | 2019-06-26 | 2020-12-30 | 平安科技（深圳）有限公司 | Image recognition method and apparatus, and non-volatile readable storage medium and computer device |
| CN110458185A (en) * | 2019-06-26 | 2019-11-15 | 平安科技（深圳）有限公司 | Image-recognizing method and device, storage medium, computer equipment |
| CN110503974B (en) * | 2019-08-29 | 2022-02-22 | 泰康保险集团股份有限公司 | Confrontation voice recognition method, device, equipment and computer readable storage medium |
| CN110503974A (en) * | 2019-08-29 | 2019-11-26 | 泰康保险集团股份有限公司 | Fight audio recognition method, device, equipment and computer readable storage medium |
| CN110795892A (en) * | 2019-10-23 | 2020-02-14 | 北京邮电大学 | Channel simulation method and device based on generation countermeasure network |
| CN110795892B (en) * | 2019-10-23 | 2021-10-01 | 北京邮电大学 | Channel simulation method and device based on generation countermeasure network |
| CN110969200B (en) * | 2019-11-27 | 2023-05-12 | 中国科学院自动化研究所 | Image target detection model training method and device based on consistency negative sample |
| CN110969200A (en) * | 2019-11-27 | 2020-04-07 | 中国科学院自动化研究所 | Image target detection model training method and device based on consistency negative sample |
| US11921819B2 (en) | 2020-02-25 | 2024-03-05 | Zhejiang University Of Technology | Defense method and an application against adversarial examples based on feature remapping |
| WO2021169157A1 (en) * | 2020-02-25 | 2021-09-02 | 浙江工业大学 | Feature remapping-based adversarial sample defense method and application |
| CN111651561A (en) * | 2020-06-05 | 2020-09-11 | 拾音智能科技有限公司 | High-quality difficult sample generation method |
| CN111860593A (en) * | 2020-06-15 | 2020-10-30 | 北京华电天仁电力控制技术有限公司 | Fan blade fault detection method based on generation countermeasure network |
| CN111899251A (en) * | 2020-08-06 | 2020-11-06 | 中国科学院深圳先进技术研究院 | Copy-move type forged image detection method for distinguishing forged source and target area |
| CN112465037B (en) * | 2020-11-30 | 2023-12-19 | 深圳供电局有限公司 | Target detection method, device, computer equipment and storage medium |
| CN112465037A (en) * | 2020-11-30 | 2021-03-09 | 深圳供电局有限公司 | Target detection method and device, computer equipment and storage medium |
| WO2023050746A1 (en) * | 2021-09-30 | 2023-04-06 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Method for enhancing sar image data for ship target detection |
| CN114648724A (en) * | 2022-05-18 | 2022-06-21 | 成都航空职业技术学院 | Lightweight efficient target segmentation and counting method based on generation countermeasure network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109543740B (en) | 2022-07-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109543740A (en) | A kind of object detection method based on generation confrontation network | |
| Bivand | R packages for analyzing spatial data: A comparative case study with areal data | |
| Puduppully et al. | Data-to-text generation with content selection and planning | |
| Odenbaugh et al. | Buyer beware: Robustness analyses in economics and biology | |
| CN103650033B (en) | Use the speech recognition of variable-length context | |
| Warrier et al. | Risk factors associated with circumferential resection margin positivity in rectal cancer: a binational registry study | |
| CN106933806A (en) | The determination method and apparatus of medical synonym | |
| CN110110754B (en) | Method for classifying imbalance problems based on cost local generalization errors | |
| Bois | Bayesian inference | |
| CN103473416B (en) | The method for establishing model of protein interaction and device | |
| CN112990343A (en) | Water environment quality evaluation method based on artificial intelligence algorithm | |
| Ambalavanan et al. | Using contextual representations for suicide risk assessment from Internet forums | |
| Huang et al. | Can Language Models Be Specific? How? | |
| Lasky et al. | A hierarchical model of whole assemblage island biogeography | |
| Chen et al. | Diagnostic evaluation model of English learning based on machine learning | |
| Isoda et al. | Strengths of exaggerated tsunami-originated placenames: Disaster subculture in Sanriku Coast, Japan | |
| CN116597916A (en) | Prediction method of antitumor compound prognosis efficacy based on organ chip and deep learning | |
| Harris | Conceptualizing uncertainty: the IPCC, model robustness and the weight of evidence | |
| Zeileis et al. | Parties, models, mobsters: a new implementation of model-based recursive partitioning in R | |
| Rojas‐Cerda et al. | Presence of southeast pacific blue whales (balaenoptera musculus) off south georgia in the south atlantic ocean | |
| Meysami et al. | Estimating the optimal population upper bound for scan methods in retrospective disease surveillance | |
| Köllner | Automatic loanword identification using tree reconciliation | |
| Debastiani et al. | Phylogenetic uncertainty and the inference of patterns in community ecology and comparative studies | |
| Tian | Do You Want to Foresee Your Future? The Best Model Predicting the Success of Kickstarter Campaigns | |
| CN116307792B (en) | Urban physical examination subject scene-oriented evaluation method and device |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |