CN108537244A - A kind of gradual deep learning method towards real-time system - Google Patents
A kind of gradual deep learning method towards real-time system Download PDFInfo
- Publication number
- CN108537244A CN108537244A CN201711259240.3A CN201711259240A CN108537244A CN 108537244 A CN108537244 A CN 108537244A CN 201711259240 A CN201711259240 A CN 201711259240A CN 108537244 A CN108537244 A CN 108537244A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- test result
- probability
- learning method
- time system
- 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
- G06F18/24—Classification techniques
-
- 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
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The gradual deep learning method towards real-time system that the invention discloses a kind of, using the deep learning frame of statistical method handling result, in the real-time system of processing image classification problem, achieve the progress for enabling to classification accuracy to improve, deep learning method is used alone compared with previous, has achieved the effect that also obtain reliable conclusion on small data quantity, on the one hand saves the space of system storage data, on the other hand system-computed pressure is reduced, improves timeliness and reliability.
Description
Technical field
The present invention relates to image processing field, especially a kind of gradual deep learning method towards real-time system.
Background technology
In the real-time system of processing image classification, due to data generate in real time, be difficult to obtain in the short time it is a large amount of reliably
The factors such as data, the stability and decision-making capability of system receive test.In the prior art, straight often by accumulation data volume
Classifying to after meeting conventional depth learning framework requirement, either by manual type or is passing through some other tradition
Method carry out broad classification.
It is applied in tradition in the deep learning method of the real-time system of processing image classification processing, is almost all based on one
Secondary property is fed in a manner of extensive standard authentic data, to train to obtain a very reliable model, then is specifically identified
And classification.It is said from effect, such method is more preferable more reliable.But it is applied in actual real-time system field effect simultaneously
Not very good, reason has several points, including authentic data amount and not enough and data are in during generating in real time, and
System can not hold to the urgent degree that conclusion obtains uses conventional depth learning method etc. by accumulating the method for data.
Therefore, restrict conventional depth learning method is in an important factor for real-time system applied to processing image classification
The problem of data volume.
Invention content
The technical problem to be solved by the present invention is to, in view of the shortcomings of the prior art, provide it is a kind of towards real-time system gradually
Into formula deep learning method,
In order to solve the above technical problems, the technical solution adopted in the present invention is:It is a kind of towards the gradual of real-time system
Deep learning method, includes the following steps:
1) image data of acquisition is cleaned, if the amount of images after cleaning is not enough to carry out primary complete instruction
Practice, then retains the image data;Otherwise, it enters step 2);
2) image data after cleaning is trained, obtains multiple trained models, image to be sorted is inputted
It in frame, is tested in trained model, the test result after record test;
3) contrast test result is identified as the probability and ranking of every one kind, by one group worst per S group automatic rejection effects
The mode of data is updated;
2) and step 3) 4) repeat the above steps, until having handled all test results;
5) the final probability that test result belongs to a certain classification is obtained using following formula;
6) K different final probability values are obtained according to the number K of class, this K final probability values have respectively represented test
Test result is grouped into that class corresponding to most probable value, at this time test result by the possibility for as a result belonging to some class
It has been completed identification, is classified.
In step 2), the image data after cleaning is trained using Inception-V3 models
In step 5), final probability F (Pi, Ri, N, K) calculation formula be:
Wherein, N indicates test result group number;G(Ri, N, K) and in table N group test results, weigh average probability reliability
Weight;PiIt is identified as certain a kind of probability, R to be correspondingiFor corresponding probability ranking, K is the classification number belonging to possible.
Compared with prior art, the advantageous effect of present invention is that:Present invention employs statistical method handling results
Deep learning frame, processing image classification problem real-time system on, achieve enable to classification accuracy improve
Progress is used alone deep learning method compared with previous, has achieved the effect that also obtain reliable conclusion on small data quantity, one
Aspect saves the space of system storage data, on the other hand reduces system-computed pressure, improves timeliness and reliability.
Description of the drawings
Fig. 1 is model modification strategy schematic diagram of the present invention;
Fig. 2 is deep learning frame diagram of the present invention.
Specific implementation mode
The present invention realizes that process is as follows:
Step 1:Data cleansing.The index of cleaning includes, and whether image resolution ratio reaches requirement, image tag complete
Whole, picture format etc..If the data volume of the data after over cleaning is not enough to carry out primary complete training, retention data
Enter to next batch data.If enough, the data after over cleaning can be trained into frame.
Step 2:It is trained using Inception-V3 models.It is directed to image classification problem in this frame, uses depth
When the Inception-V3 networks of CNN convolutional neural networks handle the image of actual needs classification in degree study, successively
It is tested in the trained model of each acquisition, records the findings data after the test obtained every time, including (1) identification
To the probability P of each classificationi(2) corresponding probability ranking Ri(3) class belonging to Number of Models N (4) test data possibility passed through
Other number K.It is synchronous to use model modification strategy, findings data is screened and is rejected, is once picked per S groups in sequence
Except (S is the hyper parameter manually set), this strategy is repeated until all model measurements finish and have handled all obtain simultaneously
Result.
Step 3:Integrated treatment is carried out using the mode of statistical weight, to obtain final findings data.It is specific public
Formula is as follows:
Weighted averaging functions F (Pi, Ri, N, K):Function is defined as follows:
Wherein, G (Ri, N) and it is defined as follows:
K represents all data category numbers
F(Pi, Ri, N, K) and it indicates in N group test results, test data belongs to the final probability of each classification.
G(Ri, N, K) and in table N group test results, weigh the weight of average probability reliability.
Wherein, PiIt is identified as certain a kind of probability, R to be correspondingiFor corresponding probability ranking, N is that Number of Models is (contemporary
Table N groups test result), K is the classification number belonging to possible, and S is artificial setting, every how many groups of super ginsengs once rejected
Number.
Below in conjunction with the specific implementation process of the description of the drawings present invention:
Such as Fig. 1, TrainData and TrainedModel:The training data of different batches enters in the training frame of model,
It can obtain corresponding TestModeli, i ∈ (1, n), wherein TestModel represent corresponding model, and i represents corresponding batch
Secondary, n represents model quantity in total.
TestData:Test data is passed sequentially through into TestModeli, record the R=[Result obtained after test1,
Result2,Result3,....,Resulti]。
ModelUpdate:Model modification strategy.In order by ResultiIt is counted using statistical method, and by per k
The mode of a worst result of result automatic rejections completes the update of model.
More new strategy starts when model starts test.
By k-th model, obtained the K result data from 1 to K after, by comparing in this K model, will survey
Examination data are identified as the probability of every one kind and corresponding probability ranking, and worst one group of automatic rejection is put down as a result, reusing weighting
Equal function obtains the result data of first stage, and using this data as the initial results data of second stage, repeats above-mentioned mistake
Journey is until that all model test results have been processed is complete, you can obtain final identification conclusion, i.e. FinalConclusion,
As shown in Figure 2.
Claims (3)
1. a kind of gradual deep learning method towards real-time system, which is characterized in that include the following steps:
1) image data of acquisition is cleaned, if the amount of images after cleaning is not enough to carry out primary complete training,
Retain the image data;Otherwise, it enters step 2);
2) image data after cleaning is trained, obtains multiple trained models, image to be sorted is inputted into frame
In, it is tested in trained model, the test result after record test;
3) contrast test result is identified as the probability and ranking of every one kind, by per one group of worst data of S group automatic rejection effects
Mode be updated;
2) and step 3) 4) repeat the above steps, until having handled all test results;
5) the final probability that test result belongs to a certain classification is obtained using following formula;
6) K different final probability values are obtained according to the number K of class, this K final probability values have respectively represented test result
Test result is grouped into that class corresponding to most probable value by the possibility for belonging to some class, and test result has been at this time
Identification is completed, is classified.
2. the gradual deep learning method according to claim 1 towards real-time system, which is characterized in that step 2)
In, the image data after cleaning is trained using Inception-V3 models
3. the gradual deep learning method according to claim 1 towards real-time system, which is characterized in that step 5)
In, final probability F (Pi,Ri, N, K) calculation formula be:
Wherein, N indicates test result group number;G(Ri, N, K) and in table N group test results, weigh the weight of average probability reliability;
PiIt is identified as certain a kind of probability, R to be correspondingiFor corresponding probability ranking, K is the classification number belonging to possible.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711259240.3A CN108537244A (en) | 2017-12-04 | 2017-12-04 | A kind of gradual deep learning method towards real-time system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711259240.3A CN108537244A (en) | 2017-12-04 | 2017-12-04 | A kind of gradual deep learning method towards real-time system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108537244A true CN108537244A (en) | 2018-09-14 |
Family
ID=63489689
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711259240.3A Pending CN108537244A (en) | 2017-12-04 | 2017-12-04 | A kind of gradual deep learning method towards real-time system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108537244A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109886083A (en) * | 2019-01-03 | 2019-06-14 | 杭州电子科技大学 | A kind of small face detecting method of real-time scene based on deep learning |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103324937A (en) * | 2012-03-21 | 2013-09-25 | 日电(中国)有限公司 | Method and device for labeling targets |
CN104537676A (en) * | 2015-01-12 | 2015-04-22 | 南京大学 | Gradual image segmentation method based on online learning |
CN105956631A (en) * | 2016-05-19 | 2016-09-21 | 南京大学 | On-line progressive image classification method facing electronic image base |
US20190266490A1 (en) * | 2016-12-15 | 2019-08-29 | WaveOne Inc. | Enhanced coding efficiency with progressive representation |
-
2017
- 2017-12-04 CN CN201711259240.3A patent/CN108537244A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103324937A (en) * | 2012-03-21 | 2013-09-25 | 日电(中国)有限公司 | Method and device for labeling targets |
CN104537676A (en) * | 2015-01-12 | 2015-04-22 | 南京大学 | Gradual image segmentation method based on online learning |
CN105956631A (en) * | 2016-05-19 | 2016-09-21 | 南京大学 | On-line progressive image classification method facing electronic image base |
US20190266490A1 (en) * | 2016-12-15 | 2019-08-29 | WaveOne Inc. | Enhanced coding efficiency with progressive representation |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109886083A (en) * | 2019-01-03 | 2019-06-14 | 杭州电子科技大学 | A kind of small face detecting method of real-time scene based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107633255A (en) | A kind of rock lithology automatic recognition classification method under deep learning pattern | |
CN106780466A (en) | A kind of cervical cell image-recognizing method based on convolutional neural networks | |
CN106096661B (en) | The zero sample image classification method based on relative priority random forest | |
CN109800811A (en) | A kind of small sample image-recognizing method based on deep learning | |
CN111178120B (en) | Pest image detection method based on crop identification cascading technology | |
CN108805196A (en) | Auto-increment learning method for image recognition | |
CN110245550A (en) | A kind of face noise data collection CNN training method based on overall cosine distribution | |
CN110688888B (en) | Pedestrian attribute identification method and system based on deep learning | |
CN109377487B (en) | Fruit surface defect detection method based on deep learning segmentation | |
CN104268552B (en) | One kind is based on the polygonal fine classification sorting technique of part | |
CN109420622A (en) | Tobacco leaf method for sorting based on convolutional neural networks | |
CN109919252A (en) | The method for generating classifier using a small number of mark images | |
CN113128335B (en) | Method, system and application for detecting, classifying and finding micro-living ancient fossil image | |
CN109598307A (en) | Data screening method, apparatus, server and storage medium | |
CN111652257A (en) | Sample data cleaning method and system | |
CN111709477A (en) | Method and tool for garbage classification based on improved MobileNet network | |
CN109003266A (en) | A method of based on fuzzy clustering statistical picture quality subjective evaluation result | |
CN109284700B (en) | Method, storage medium, device and system for detecting multiple faces in image | |
CN111784777A (en) | SMT material quantity statistical method and statistical system based on convolutional neural network | |
CN114066848A (en) | FPCA appearance defect visual inspection system | |
CN109934297B (en) | Rice seed test method based on deep learning convolutional neural network | |
CN113869463B (en) | Long tail noise learning method based on cross enhancement matching | |
CN111414930B (en) | Deep learning model training method and device, electronic equipment and storage medium | |
CN111144462A (en) | Unknown individual identification method and device for radar signals | |
CN108537244A (en) | A kind of gradual deep learning method towards real-time system |
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: 20180914 |