CN109389136A - Classifier training method - Google Patents
Classifier training method Download PDFInfo
- Publication number
- CN109389136A CN109389136A CN201710670289.1A CN201710670289A CN109389136A CN 109389136 A CN109389136 A CN 109389136A CN 201710670289 A CN201710670289 A CN 201710670289A CN 109389136 A CN109389136 A CN 109389136A
- Authority
- CN
- China
- Prior art keywords
- training
- classifier
- scene type
- neural network
- new samples
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of classifier training methods, comprising the following steps: generates data set by computer graphics model, data set includes the sample of several scenes classification；Training pattern is obtained using deep neural network training dataset；Training pattern is tested to obtain the test result of each scene type；The information of the corresponding scene type of accuracy minimum in test result is fed back into computer graphics model；Computer graphics model modifies parameter corresponding with the scene type, generates the new samples about the scene type；And new samples are added in data set.Classifier training method through the invention can be improved the accuracy of classifier training and reduce the time of classifier training.
Description
Technical field
The present invention relates to a kind of classifier training methods, more particularly to one kind to be applied to ADAS (advanced driver
Assistance system, driver assistance system) classifier training method.
Background technique
ADAS may be defined as in driving procedure by the perception to vehicle-surroundings environment and to the reason of driving behavior
Solution judges danger that may be present, gives driver's warning, by the way of interactive so as to reducing traffic accident
Incidence extends the safe driving time.ADAS system based on image has at low cost, easy for installation and high accuracy rate excellent
Point.ADAS system needs to participate in object to the traffic in driving conditions to identify, to judge its shadow to current driving environment
It rings.Usually, the traffic in driving conditions, which participates in object, to be divided into: static traffic participates in object, such as trees, fence, traffic mark
Board, traffic lights, lane line, fixed obstacle etc.；Mobile traffic participate in object, as vehicle, pedestrian, bicycle, power assist vehicle, animal,
Other objects with displacement characteristic.
ADAS system needs to participate in object to above-mentioned movement/static traffic to analyze with the motion conditions from vehicle, to reach
The effect driven to auxiliary, primary goal is the detection and identification to above-mentioned target.Detection and identification for specific objective
The machine learning algorithm that supervision can be used solves, and is broadly divided into two processes of training and test.It is needed in training process
The sample marked is inputed into machine, machine is using specific algorithm and goes out optimal parameter according to sample training, provides point
Class model；Machine examines the position of target and size according to trained model and the image of input during the test
It surveys.Detection effect, which is directly undergone training, the influence of sample, the especially complexity of ADAS application environment and participates in the diversity of object,
Influence of the Many times training sample to performance is even more than algorithm itself.
By taking vehicle as an example, the type of vehicle for needing to identify is numerous, comprising: car, SUV, MPV, sport car, microcaloire, pick up, big
Bar, mini-bus, trailer, tank car, lorry etc..Vehicle running surface type can be divided into again: super expressway, urban road, suburb road
Road and natural environment etc..Vehicle running environment is influenced by various illumination and weather condition again.Therefore, it collects in various weather
And under natural environment, a variety of vehicle samples of various information of road surface have very big difficulty, in addition in view of collecting sample regards
The sample size of the demand at angle, demand will be very huge, it will usually reach the sample of millions of or even more than one hundred million number of levels.Such as
Fruit needs to collect the video data of substantial amounts by the way of manually marking, and therefrom calibrates the actual size of target by hand
And position, it will bring huge workload and data integration time.
It is therefore desirable to be able to provide Massive Sample and using these sample training classifiers quickly and easily with effectively reliable
Ground is applied to the method for ADAS.
Summary of the invention
The purpose of the present invention is to provide a kind of classifier training methods that can be applied to ADAS, use and pass through CG
The magnanimity sample for the different scenes condition for selected target that (computer graphics, computer graphical) technology generates
This.
According to an embodiment of the invention, providing a kind of classifier training method, comprising the following steps: pass through computer graphic
Shape model generates data set, and data set includes the sample of several scenes classification；Come using deep neural network training dataset
To training pattern；Training pattern is tested to obtain the test result of each scene type；By accuracy pair minimum in test result
The information for the scene type answered feeds back to computer graphics model；Computer graphics model modifies ginseng corresponding with the scene type
Number generates the new samples about the scene type；And new samples are added in data set.Note that in the training process, making
It is trained with the sample under scenes all in training set to obtain training pattern, rather than is instructed with the sample under each scene
Practice to obtain the training pattern under each scene.
Wherein, it includes: according to predetermined ratio by data that training pattern is obtained using deep neural network training dataset
The sample of concentration is assigned randomly in training set, verifying collection and test set；Come using the deep neural network training training set
Obtain multiple initial models；And multiple initial models are verified using verifying collection, obtain the training with optimized parameter
Model.
In addition, repeating training, the training pattern of deep neural network after new samples are added in the training set
Test, the feedback of test result, new samples generation processing, until the accuracy of the test result of each scene type is equal
Until predetermined threshold.
Wherein, the training for repeating deep neural network includes any one of following two mode: utilizing depth nerve net
Network re -training is added with the training set of new samples to obtain new training pattern；In deep neural network, using being added with
The training set of new samples is finely adjusted the model parameter for the training pattern that previous training obtains to obtain new training pattern.
Preferably, the initial proportion of each scene type sample in training set is identical.
Preferably, it further includes authentic specimen that training set and verifying, which are concentrated,.
It preferably, further include authentic specimen in training set, verifying collection and test set.
It preferably, include the demarcating file of training objective in training set, verifying collection and test set.
Preferably, test training pattern includes: to test to instruct using test set come the test result for obtaining each scene type
Practice model, respectively obtains the test result of each scene type.
Preferably, new samples are added to includes: that new samples are added in training set in data set.
Preferably, ratio of the new samples in training set is 20%.
Preferably, several scenes classification includes fine day, cloudy day, snowy day, rainy day, daytime, night, strong light etc..
According to the method for the present invention, it by generating sample using CG technology, effectively reduces and takes pictures sampling to actual environment
Workload and working strength, and reduce manpower, improve work efficiency.By using deep neural network, Neng Gouti
High training effect.By extraly increasing the verification processing of verifying collection in training, over-fitting can be prevented, so that training
Model out has versatility, the image data being applicable to outside training sample.In addition, by by the lower information of accuracy
CG is fed back to, CG is made to generate the new samples of corresponding scene type, re -training is added with the training set of new samples or to previous
The model parameter of trained deep neural network is finely adjusted, to guarantee accuracy of the sample under more scene conditions.As a result,
Classifier training method through the invention, can be improved the accuracy of classifier training and reduce classifier training when
Between.
Detailed description of the invention
It in order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, for those of ordinary skill in the art, without creative efforts, can also obtain according to these attached drawings
Obtain other attached drawings.
Fig. 1 is the flow chart of classifier training method according to the present invention；And
Fig. 2 be deep neural network according to the present invention a kind of exemplary group at schematic diagram.
Specific embodiment
The description of this illustrated embodiment should be combined with corresponding attached drawing, and attached drawing should be used as the one of complete specification
Part.In the accompanying drawings, the shape of embodiment or thickness can expand, and to simplify or facilitate mark.Furthermore it is respectively tied in attached drawing
The part of structure will be to describe to be illustrated respectively, it is notable that the member for being not shown in the figure or not being illustrated by text
Part is the form known to a person of ordinary skill in the art in technical field.
The description of embodiments herein, any reference in relation to direction and orientation, is merely for convenience of describing, and cannot manage
Solution is any restrictions to the scope of the present invention.It can be related to the combination of feature below for the explanation of preferred embodiment,
These features may be individually present or combine presence, and the present invention is not defined in preferred embodiment particularly.The present invention
Range be defined by the claims.
The digital simulation that CG graphics can use magnanimity generates the driving scene with high realism, while generating target
Accurate target position and size in the picture, mark letter big with data volume relative to the implementation manually demarcated
Cease accurate, scene advantage abundant.The CG high realism true value mark database generated is used for target object in supervised learning
Detection will be of great significance.Meanwhile there is the support of Massive Sample that can assess the various performances of sorting algorithm.
In the present invention, CG technology is combined with ADAS recognizer, being not only embodied in CG can provide for sorting algorithm
The data of magnanimity are also embodied in sorting algorithm while CG can be instructed to generate what type of sample, to mention in maximum efficiency
The performance of high algorithm.That is, one closed loop thinking generated based on CG data with classification algorithm training of building, CG can be used
Good classifier of data " culture " of magnanimity, and classifier " can instruct " generation of CG sample database simultaneously, below into
Row is described in detail.
Fig. 1 is the flow chart of classifier training method according to the present invention.
Referring to Fig. 1, classifier training method according to the present invention includes: to generate data set by computer graphics model,
Data set includes the sample of several scenes classification；Training pattern is obtained using deep neural network training dataset；Test instruction
Practice model to obtain the test result of each scene type；By the information of the corresponding scene type of accuracy minimum in test result
Feed back to computer graphics model；Computer graphics model modifies parameter corresponding with the scene type, generates about the scene
The new samples of classification；And new samples are added in data set.Note that in the training process, using fields all in training set
Sample under scape is trained to obtain training pattern, rather than is trained with the sample under each scene to obtain each scene
Under training pattern.
In classifier training method of the invention, data set is generated by CG technology, includes several scenes in data set
Type, and saved by scene type, scene type is for example including fine day, cloudy day, snowy day, rainy day, daytime, night, strong light etc..
It should be understood that data set used in the present invention can also classify otherwise.
In classifier training method of the invention, deep neural network training sample is used.Fig. 2 is according to the present invention
A kind of exemplary group of deep neural network at schematic diagram.Referring to Fig. 2, the deep neural network that the present invention uses is by convolution
Layer, pond layer, full articulamentum and softmax layers of composition.Wherein, image characteristics extraction depends on convolution algorithm, is transported by convolution
Calculation is filtered to obtain significant Expressive Features.
Specifically, input layer main function be by picture input be converted into output data required for model (data) and
Label (label), input layer also convert data format required for platform, such as the lmdb of caffe etc. for picture simultaneously.
Convolutional layer main function is feature extraction, and the convolution operation of deep learning is similar to filter, by important information
It extracts, main definitions convolution kernel and convolution operation.
Pond layer main function is characteristic pattern compression, and carrying out compressed characteristic pattern can be reduced calculation amount, reduce network
Computation complexity, and characteristic pattern main information can be continued to extract.Pond layer operation includes average pond and maximum
The modes such as pond.
Full articulamentum main function is to connect all features, and output valve is passed to softmax layers.Full connection
Layer parameter is more, and full articulamentum is affected to parameter optimization.
Softmax layers of main function are classifiers, are last output.
It should be noted that Fig. 2 is only a kind of exemplary composed structure of deep neural network workable for the present invention, this hair
The bright deep neural network that can also be applied to various composed structures, and can also be with using the deep neural network of other structures
It realizes function of the invention and realizes identical effect.
Preferably, it includes: that will count according to predetermined ratio that training pattern is obtained using deep neural network training dataset
It is assigned randomly in training set, verifying collection and test set according to the sample of concentration；Come using deep neural network training training set
To multiple initial models；And multiple initial models are verified using verifying collection, obtain the training mould with optimized parameter
Type.
In addition, repeating the training of deep neural network, the survey of training pattern after new samples are added in training set
Examination, the generation of the feedback of test result, new samples processing, until the accuracy of the test result of each scene type is all larger than
Until predetermined threshold.The predetermined threshold can according to circumstances use different values.Specifically, the sample in data set CG generated
This is assigned randomly in training set, verifying collection and test set according to predetermined ratio, such as all data sets can be randomly divided into
50% training set, 25% verifying collection and 25% test set, or it is divided into 60% training set, 20% verifying collection and 20% test set,
Certainly it will be understood by those skilled in the art that other ratios can also be used.
In an advantageous embodiment, when sample in the first sub-distribution data set, make each scene class in training set
Very originally initial proportion is identical, that is, the quantity of the sample of each scene type is identical, to guarantee identification of the model to each scene
Initialization weight having the same.
In addition, concentrating in training set and verifying includes authentic specimen, that is, training set and verifying collection are in addition to including what CG was generated
It further include the various information of road surface under a small amount of various weather and natural environment collected by modes such as shootings except sample
A variety of vehicle samples, ratio of the authentic specimen in data set is about 5% to 10%.
It include authentic specimen in training set, verifying collection and test set in another preferred embodiment.
In addition, including that (such as various static traffic participate in object and movement to training objective in training set, verifying collection and test set
Traffic participates in object) demarcating file, demarcating file indicates the size and location of training objective in the picture.
After establishing data set according to each scene type with same ratio, the starting model instruction of deep neural network is carried out
Practice, obtain multiple initial models, then multiple initial models are verified using verifying collection, obtain the instruction with optimized parameter
Practice model.Note that in the training process, using multiple Artificial Neural Network Structures to the sample under scenes all in training set into
Row training is to obtain multiple initial models.Wherein, training pattern can be known to the skilled in the art various training patterns.
For example, the present invention can use multi-target detection model, it can also be using models such as semantic segmentations.
In classifier training method of the invention, other than including training set and test set, also it is additionally provided with and tests
Card collection, verifying collection can carry out cross validation to the model that training set training obtains to select optimal models, and effect is to prevent
Over-fitting because for training set train come model, for training set data, the parameter of model can reach best
Effect, but when carrying out test set test, modelling effect may just have biggish change, and adding verifying collection can be well
Solution such issues that, by verify collection obtain the model with optimized parameter, then the model is tested with test set again.
Deep neural network training is repetitive exercise, and when just having started to train, the effect of model has very big error rate.With
The increase of neural metwork training number, model constantly update inherent parameters weight, the reversed biography of deep learning by backpropagation
Broadcast and error be calculated by the output of true value and last softmax in true value file, by error it is reversed pass network back.
Then, network punishes weighting parameter according to error, updates weight to the direction for reducing error, to guarantee network convergence
And obtain further feature.By the verifying of verifying collection, the model for training deep neural network has versatility, can be applied to
Image data outside sample.
As described above, the present invention carries out feature extraction and classification to the sample that CG is generated using the method for deep neural network
Judgement, after obtaining preliminary depth sorting model, does not ensure that this model can meet verification and measurement ratio under any environment
It is required that.Therefore, the present invention is also based on test result " guidance " CG and generates new samples carrying out supplementary training sample, to promote depth point
The accuracy of class model.It after obtaining a depth sorting model, is analyzed by the accuracy rate to test set, changes CG sample
The relevant parameter that library generates achievees the purpose that " instruct " generation of CG sample, and the scene ratio of sample database is updated, to update
Training sample database used in next iteration.
In the present invention, training pattern is tested using test set, respectively obtains the test result of each scene type, tested
As a result it can be indicated with modes such as accuracy, false detection rate, recall rates.That is, by evaluating model, available model
Accuracy rate, false detection rate, recall rate and other evaluation results.For example, multi-target detection model can to for each classification into
The evaluation of row accuracy rate and recall rate, while also carrying out the evaluation of consensus forecast rate.When scene type accuracy rate or call together
When the rate of returning is relatively bad, CG can be fed back so that CG generates the new samples of the scene type and new samples are added to training set
In, ratio of the scene type in training set is increased, makes the fine tuning training of depth network parameter to the effect of the scene type
Rise.
Specifically, after Sample Refreshment, it can use the training set that deep neural network re -training is added with new samples
To obtain new training pattern, that is, the process of training is identical with first time, and using deep neural network, training is new again completely
Training set.Then, test set tests the training pattern newly obtained, instructs CG to generate new sample, Zhi Houzai according to accuracy
New training set is trained using deep neural network, repeatedly iteration, until the test result of each scene type is all satisfied
Until it is required that.
However, if completely re -training deep neural network will devote a tremendous amount of time, therefore the invention also provides
Trimming scheme.That is, being carried out using a certain proportion of old sample and newly-increased sample to the old model parameter that last iteration obtains micro-
It adjusts, to guarantee accuracy of the sample under more scenes.Wherein, ratio of the new samples in training set is 20%, certainly
It can be other ratios.Specifically, it is finely adjusted, is learned with the model that the training set added with new samples obtains last training
Practise new samples.The method of fine tuning is to extract model, and the value of reserving model weight and biasing is joined using the weight and biasing as starting
Number, is then again trained new training set.Since model has had preferable testing result, so new samples start just
Just there is good testing result in training stage.And due to increasing the general sample proportion of performance in new training set, thus
In the case where guaranteeing original good scene type of performance, and the study to general scene type is showed is strengthened, thus
To new model.Then, above-mentioned various processing are repeated, the evaluation result that can be got well until guaranteeing the data of each scene type.
Although the performance of entire training system is it should be noted that the mode efficiency of above two repetition training is different
It is identical.
Wherein, CG generates the process of sample by targeted species and quantity, environment type and quantity, illumination, meteorology, object
The influence of surface area object and camera parameter, such as the parameter that control CG sample generates are represented by Pcg={ VT1,…,n, VN,
ET1,…,m, EN, SA, SI, SC, SVL, OVL, EL, RI, SI, FI, WP, SP, DP, CP, CD, CS, IQ }, wherein VT1,…,nIndicate vehicle
Type, such as car, SUV, MPV, sport car, microcaloire, pick up, bus, mini-bus, trailer, tank car, lorry；VN indicates vehicle mould
Type quantity represents the degree of crowding for generating vehicle in picture, can be divided into unimpeded and crowded two kinds of situations；ET1,…,mIndicate environment kind
Class, such as urban road, expressway, tunnel, bridge road, hill path；EN indicates environmental model quantity, indicates the complexity of environment
Degree, can be divided into two kinds of situations of single environment and complex environment；Illumination effect mostlys come from the influence of sunlight in daylight environment,
SA indicates the natural environment sun angular of simulation, and being continuously can control value；SI indicates the natural environment sunlight intensity of simulation, is
It continuously can control value；SC indicates the natural environment sun light color of simulation, and being continuously can control value；Illumination effect is main in night environment
From the influence from vehicle, Ta Che and street lamp, wherein SVL indicates that being continuously can control value from the bright dark of car light；OVL indicates him
Car light angular and bright dark, for continuously can control value；The angle and intensity of EL expression street lamp light are continuously can control value；Meteorologic parameter
Middle RI indicates rainfall intensity, can be divided into weak, medium, strong three grades；SI indicates snowfall intensity, can be divided into weak, medium, strong three etc.
Grade；FI indicates drop mist intensity, can be divided into weak, medium, strong three grades；Due to being influenced by meteorology, front vehicles surface can have been accumulated
Different substances, WP indicate water accumulating volume, can be divided into weak, medium, strong three grades；SP indicates snowpack, can be divided into weak, medium, strong three
A grade；DP indicates dust accumulation amount, can be divided into weak, medium, strong three grades；In the influence of camera parameter, CP indicates the direction of camera,
For continuous controlled variable；CD indicates camera distortion and Fov, is commonly divided into pin-hole model and two kinds of flake model；CS indicates phase
The parameters such as machine ISO are generally corresponding with camera model；IQ indicates picture quality, can be divided into three kinds basic, normal, high.It is above-mentioned by changing
The value or type of parameter can simulate the samples pictures under various true environments, be supplied to classifier training disaggregated model.
It is not inventive point of the invention it should be noted that specifically how to generate sample about CG, therefore omits here detailed
Description.
As described above, by generating sample using CG technology, being effectively reduced in classifier training method of the invention
It takes pictures the workload and working strength of sampling to actual environment, and reduces manpower, improve work efficiency.By using depth
Neural network is spent, can be improved training effect.By extraly increasing the verification processing of verifying collection in training, can prevent
Fitting phenomenon, so that the model trained is applicable to the image data outside training sample.In addition, by the way that accuracy is lower
Information feed back to CG, so that CG is generated the new samples of corresponding scene type, training set of the re -training added with new samples or
The model parameter of previously trained deep neural network is finely adjusted, to guarantee that sample is accurate under more scene conditions
Property.Classifier training method through the invention as a result, can be improved the accuracy of classifier training and reduce classifier instruction
The experienced time.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (12)
1. a kind of classifier training method, which is characterized in that the classifier training method the following steps are included:
Data set is generated by computer graphics model, the data set includes the sample of several scenes classification；
Training pattern is obtained using the deep neural network training data set；
The training pattern is tested to obtain the test result of each scene type；
The information of the corresponding scene type of accuracy minimum in test result is fed back into the computer graphics model；
The computer graphics model modifies parameter corresponding with the scene type, generates the new samples about the scene type；
And
The new samples are added in the data set.
2. classifier training method according to claim 1, which is characterized in that use the deep neural network training number
Include: to obtain training pattern according to collection
The sample in the data set is assigned randomly in training set, verifying collection and test set according to predetermined ratio；
Multiple initial models are obtained using the deep neural network training training set；And
The multiple initial model is verified using verifying collection, obtains the training pattern with optimized parameter.
3. classifier training method according to claim 2, which is characterized in that the new samples are being added to the instruction
Practice after concentrating, repeat the training of the deep neural network, the test of the training pattern, the test result feedback,
The processing of the generation of the new samples, until the accuracy of the test result of each scene type is all larger than predetermined threshold.
4. classifier training method according to claim 3, which is characterized in that repeat the training of the deep neural network
Including any one of following two mode:
The training set of the new samples is added with using the deep neural network re -training to obtain new training mould
Type；
In the deep neural network, using the training set added with the new samples to the previous training pattern trained and obtained
Model parameter be finely adjusted to obtain new training pattern.
5. classifier training method according to claim 2, which is characterized in that each scene type sample in the training set
This initial proportion is identical.
6. classifier training method according to claim 2, which is characterized in that
The training set and verifying concentration further include authentic specimen.
7. classifier training method according to claim 2, which is characterized in that
It further include authentic specimen in the training set, verifying collection and the test set.
8. classifier training method according to claim 2, which is characterized in that the training set, verifying collection and institute
State the demarcating file in test set including training objective.
9. classifier training method according to claim 2, which is characterized in that it is each to obtain to test the training pattern
The test result of scene type includes: to test the training pattern using the test set, respectively obtains each scene type
Test result.
10. classifier training method according to claim 2, which is characterized in that the new samples are added to the number
It include: that the new samples are added in the training set according to concentration.
11. classifier training method according to claim 10, which is characterized in that the new samples are in the training set
Ratio be 20%.
12. classifier training method according to any one of claim 1 to 11, which is characterized in that the several scenes
Classification includes fine day, cloudy day, snowy day, rainy day, daytime, night, strong light etc..
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711482401.5A CN109389142A (en) | 2017-08-08 | 2017-08-08 | Classifier training method |
| CN201710670289.1A CN109389136A (en) | 2017-08-08 | 2017-08-08 | Classifier training method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710670289.1A CN109389136A (en) | 2017-08-08 | 2017-08-08 | Classifier training method |
Related Child Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711482401.5A Division CN109389142A (en) | 2017-08-08 | 2017-08-08 | Classifier training method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109389136A true CN109389136A (en) | 2019-02-26 |
Family
ID=65412973
Family Applications (2)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710670289.1A Pending CN109389136A (en) | 2017-08-08 | 2017-08-08 | Classifier training method |
| CN201711482401.5A Pending CN109389142A (en) | 2017-08-08 | 2017-08-08 | Classifier training method |
Family Applications After (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711482401.5A Pending CN109389142A (en) | 2017-08-08 | 2017-08-08 | Classifier training method |
Country Status (1)
| Country | Link |
|---|---|
| CN (2) | CN109389136A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135592A (en) * | 2019-05-16 | 2019-08-16 | 腾讯科技（深圳）有限公司 | Classifying quality determines method, apparatus, intelligent terminal and storage medium |
| CN110276443A (en) * | 2019-05-29 | 2019-09-24 | 平安科技（深圳）有限公司 | Model conversion consistency checking and analysis method, device and storage medium |
| CN110765876A (en) * | 2019-09-20 | 2020-02-07 | 深圳码隆科技有限公司 | Training method and device for commodity recognition model |
| CN110991538A (en) * | 2019-12-05 | 2020-04-10 | 腾讯科技（深圳）有限公司 | Sample classification method and device, storage medium and computer equipment |
| CN111292329A (en) * | 2020-01-15 | 2020-06-16 | 北京字节跳动网络技术有限公司 | Training method and device for video segmentation network and electronic equipment |
| CN111488939A (en) * | 2020-04-15 | 2020-08-04 | 天津美腾科技股份有限公司 | Model training method, classification method, device and equipment |
| CN112132796A (en) * | 2020-09-15 | 2020-12-25 | 佛山读图科技有限公司 | Visual detection method and system for improving detection precision by means of feedback data autonomous learning |
| CN112800233A (en) * | 2021-04-13 | 2021-05-14 | 成都数联铭品科技有限公司 | Text position detection method |
| CN114418036A (en) * | 2022-03-28 | 2022-04-29 | 浙江所托瑞安科技集团有限公司 | Method, device and storage medium for testing and training performance of neural network |
| WO2024060670A1 (en) * | 2022-09-19 | 2024-03-28 | 北京沃东天骏信息技术有限公司 | Method and apparatus for training classification model, and device and storage medium |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110110652B (en) * | 2019-05-05 | 2021-10-22 | 达闼科技（北京）有限公司 | Target detection method, electronic device and storage medium |
| CN110634036A (en) * | 2019-09-27 | 2019-12-31 | 金瓜子科技发展（北京）有限公司 | Method and device for distributing configuration objects, storage medium and electronic equipment |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2014049118A (en) * | 2012-08-31 | 2014-03-17 | Fujitsu Ltd | Convolution neural network classifier system, training method for the same, classifying method, and usage |
| CN103971342A (en) * | 2014-05-21 | 2014-08-06 | 厦门美图之家科技有限公司 | Image noisy point detection method based on convolution neural network |
| CN104504376A (en) * | 2014-12-22 | 2015-04-08 | 厦门美图之家科技有限公司 | Age classification method and system for face images |
| CN104966097A (en) * | 2015-06-12 | 2015-10-07 | 成都数联铭品科技有限公司 | Complex character recognition method based on deep learning |
| CN105051755A (en) * | 2013-01-18 | 2015-11-11 | 微软技术许可有限责任公司 | Part and state detection for gesture recognition |
| CN105320957A (en) * | 2014-07-10 | 2016-02-10 | 腾讯科技（深圳）有限公司 | Classifier training method and device |
| CN106295637A (en) * | 2016-07-29 | 2017-01-04 | 电子科技大学 | A kind of vehicle identification method based on degree of depth study with intensified learning |
| CN106934319A (en) * | 2015-12-29 | 2017-07-07 | 北京大唐高鸿软件技术有限公司 | People's car objective classification method in monitor video based on convolutional neural networks |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101976461A (en) * | 2010-10-25 | 2011-02-16 | 北京理工大学 | Novel outdoor augmented reality label-free tracking registration algorithm |
| CN103679189B (en) * | 2012-09-14 | 2017-02-01 | 华为技术有限公司 | Method and device for recognizing scene |
-
2017
- 2017-08-08 CN CN201710670289.1A patent/CN109389136A/en active Pending
- 2017-08-08 CN CN201711482401.5A patent/CN109389142A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2014049118A (en) * | 2012-08-31 | 2014-03-17 | Fujitsu Ltd | Convolution neural network classifier system, training method for the same, classifying method, and usage |
| CN105051755A (en) * | 2013-01-18 | 2015-11-11 | 微软技术许可有限责任公司 | Part and state detection for gesture recognition |
| CN103971342A (en) * | 2014-05-21 | 2014-08-06 | 厦门美图之家科技有限公司 | Image noisy point detection method based on convolution neural network |
| CN105320957A (en) * | 2014-07-10 | 2016-02-10 | 腾讯科技（深圳）有限公司 | Classifier training method and device |
| CN104504376A (en) * | 2014-12-22 | 2015-04-08 | 厦门美图之家科技有限公司 | Age classification method and system for face images |
| CN104966097A (en) * | 2015-06-12 | 2015-10-07 | 成都数联铭品科技有限公司 | Complex character recognition method based on deep learning |
| CN106934319A (en) * | 2015-12-29 | 2017-07-07 | 北京大唐高鸿软件技术有限公司 | People's car objective classification method in monitor video based on convolutional neural networks |
| CN106295637A (en) * | 2016-07-29 | 2017-01-04 | 电子科技大学 | A kind of vehicle identification method based on degree of depth study with intensified learning |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135592A (en) * | 2019-05-16 | 2019-08-16 | 腾讯科技（深圳）有限公司 | Classifying quality determines method, apparatus, intelligent terminal and storage medium |
| CN110135592B (en) * | 2019-05-16 | 2023-09-19 | 腾讯科技（深圳）有限公司 | Classification effect determining method and device, intelligent terminal and storage medium |
| CN110276443A (en) * | 2019-05-29 | 2019-09-24 | 平安科技（深圳）有限公司 | Model conversion consistency checking and analysis method, device and storage medium |
| CN110765876A (en) * | 2019-09-20 | 2020-02-07 | 深圳码隆科技有限公司 | Training method and device for commodity recognition model |
| CN110991538B (en) * | 2019-12-05 | 2021-12-03 | 腾讯科技（深圳）有限公司 | Sample classification method and device, storage medium and computer equipment |
| CN110991538A (en) * | 2019-12-05 | 2020-04-10 | 腾讯科技（深圳）有限公司 | Sample classification method and device, storage medium and computer equipment |
| CN111292329A (en) * | 2020-01-15 | 2020-06-16 | 北京字节跳动网络技术有限公司 | Training method and device for video segmentation network and electronic equipment |
| CN111488939A (en) * | 2020-04-15 | 2020-08-04 | 天津美腾科技股份有限公司 | Model training method, classification method, device and equipment |
| CN112132796A (en) * | 2020-09-15 | 2020-12-25 | 佛山读图科技有限公司 | Visual detection method and system for improving detection precision by means of feedback data autonomous learning |
| CN112800233A (en) * | 2021-04-13 | 2021-05-14 | 成都数联铭品科技有限公司 | Text position detection method |
| CN114418036A (en) * | 2022-03-28 | 2022-04-29 | 浙江所托瑞安科技集团有限公司 | Method, device and storage medium for testing and training performance of neural network |
| CN114418036B (en) * | 2022-03-28 | 2022-06-21 | 浙江所托瑞安科技集团有限公司 | Method, device and storage medium for testing and training performance of neural network |
| WO2024060670A1 (en) * | 2022-09-19 | 2024-03-28 | 北京沃东天骏信息技术有限公司 | Method and apparatus for training classification model, and device and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109389142A (en) | 2019-02-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109389136A (en) | Classifier training method | |
| CN109977812B (en) | Vehicle-mounted video target detection method based on deep learning | |
| CN111444821B (en) | Automatic identification method for urban road signs | |
| CN103279759B (en) | A kind of vehicle front trafficability analytical procedure based on convolutional neural networks | |
| CN108985194B (en) | Intelligent vehicle travelable area identification method based on image semantic segmentation | |
| CN110111335A (en) | A kind of the urban transportation Scene Semantics dividing method and system of adaptive confrontation study | |
| CN107133974A (en) | The vehicle type classification method that Gaussian Background modeling is combined with Recognition with Recurrent Neural Network | |
| CN108764298B (en) | Electric power image environment influence identification method based on single classifier | |
| CN105844257A (en) | Early warning system based on machine vision driving-in-fog road denoter missing and early warning method | |
| CN110716529A (en) | Automatic generation method and device for automatic driving test case | |
| Zhu et al. | Extreme weather recognition using convolutional neural networks | |
| CN114677507A (en) | Street view image segmentation method and system based on bidirectional attention network | |
| CN111027430A (en) | Traffic scene complexity calculation method for intelligent evaluation of unmanned vehicles | |
| CN110956146B (en) | Road background modeling method and device, electronic equipment and storage medium | |
| CN110851948B (en) | Driving environment situation assessment method and device under unstructured road condition | |
| CN103679205A (en) | Preceding car detection method based on shadow hypothesis and layered HOG (histogram of oriented gradient) symmetric characteristic verification | |
| CN106056102A (en) | Video-image-analysis-based road vehicle type classification method | |
| CN111369053B (en) | Method for predicting tracks of pedestrians and vehicles in human-vehicle interaction scene | |
| CN212009589U (en) | Video identification driving vehicle track acquisition device based on deep learning | |
| CN116205024A (en) | Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene | |
| Cheng et al. | Modeling weather and illuminations in driving views based on big-video mining | |
| CN114169444A (en) | Driving style classification method considering risk potential field distribution under vehicle following working condition | |
| CN109948419A (en) | A kind of illegal parking automatic auditing method based on deep learning | |
| Pan et al. | Winter road surface condition recognition using a pretrained deep convolutional network | |
| CN116246169A (en) | SAH-Unet-based high-resolution remote sensing image impervious surface extraction method |
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 |
Application publication date: 20190226 |
|
| RJ01 | Rejection of invention patent application after publication |