CN106529579A - Improved construction method of robust AdaBoost classifier based on Ransac algorithm - Google Patents
Improved construction method of robust AdaBoost classifier based on Ransac algorithm Download PDFInfo
- Publication number
- CN106529579A CN106529579A CN201610917782.4A CN201610917782A CN106529579A CN 106529579 A CN106529579 A CN 106529579A CN 201610917782 A CN201610917782 A CN 201610917782A CN 106529579 A CN106529579 A CN 106529579A
- Authority
- CN
- China
- Prior art keywords
- model
- sample
- adaboost
- classification
- sorter model
- 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
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (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)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses an improved construction method of robust AdaBoost classifier based on Ransac algorithm, comprising: initializing parameter configuration, randomly selecting the quantity samples, adding the samples meeting the model sample, updating a classifier model, iterating to construct the next classifier model, calculating the classification correct rate, and finally selecting a classification model. According to the invention, the Ransac algorithm is introduced into the AdaBoost classifier model construction process and through the iterative modeling process, correct samples are searched. At the same time, a best classification model is selected from the classifier model completely based on AdaBoost algorithm. Through the above strategy, the invention realizes the construction of robust AdaBoost classifier based on Ransac algorithm, and the design of the obtained classifier model is completely free from the influence of an external point. Finally, the AdaBoost classifier model based on this method is used to verify the handwriting samples with a certain amount of external points. The experimental results show that the method proposed in this invention is better than the other two construction methods of AdaBoost classifier and achieves higher classification accuracy.
Description
Technical field
The present invention relates to mode identification method, the robust based on Ransac algorithms of more particularly to a kind of improvement
AdaBoost grader construction methods.
Background technology
It is that sample to be detected chooses optimal classification vacation from optional the classification that sorting algorithm is namely based on sorter model
If it belongs to machine learning category in artificial intelligence, the very big concern of field correlative study person is attracted.People put into
Substantial amounts of time and efforts research such as C4.5, SVMs, bayesian algorithm, AdaBoost algorithms and K- arest neighbors point
The sorting algorithms such as class algorithm, and they are applied to into the different necks such as face recognition, person's handwriting checking, data analysis and medical application
Domain.
Mono- words of AdaBoost are derived from the abbreviation of Adaptive Boosting (self adaptation enhancing), are by Yoav
The machine learning Meta algorithm that Freund and Robert Schapire are proposed.The guideline of its design is to guarantee currently to train sample
Originally there is highest nicety of grading.By by different Weak Classifiers, (so-called Weak Classifier refers to that nicety of grading is better than slightly here
Random guess) reasonably combine, strong classifier is formed, it is although the nicety of grading of each Weak Classifier is not high, final
Strong classifier obtains tremendous increase on classification performance.AdaBoost algorithms in the sense that be it is adaptive, by adjust
By the wrong sample weights divided of Weak Classifier before whole, attention degree of the follow-up Weak Classifier to wrong point of sample improved, is realized finally
The design of sorter model.This is based on, the appropriate design of one group of Weak Classifier can be combined into strong classifier, obtain one
Gratifying nicety of grading on the whole.But, as anything all has dual character, although AdaBoost algorithm advantages
It is numerous, but which is sensitive to exterior point, is easier in some cases to suffer from this and cause the degradation of grader overall performance, mistake
Effect.This is because constantly the sample to correctly classifying is weighted, the especially continuous weighting to exterior point so that exterior point
Weights rapid growth.Excessive exterior point weights cause classifier algorithm constantly to deviate to exterior point, and then away from most of normal sample
This, inevitably causes the degradation of designed sorter model.
The sustainable growth for finding a kind of effective means to limit exterior point weights is AdaBoost algorithms design sorter model
Middle problem demanding prompt solution.In recent years there has been proposed many different methods are used for suppressing the ceaselessly expansion of sample weights,
Great majority are confined to sample weights and arrange this direction, and algorithm or excessively simple is not added with distinguishing to exterior point and normal sample；Or calculate
Method is complicated, increases operation difficulty.And because the improper of threshold value setting has the misjudged disconnected situation of point and exterior point, this
The correct classification of algorithm is hindered, inevitably causes the degradation of grader.For this reason, it is necessary to find a kind of effective means from
The adverse effect of exterior point is removed in the structure of classification model.In this patent, mutually tied with AdaBoost algorithms using Ransac algorithms
Close, the degradation problem occurred in solving the sorter model training that exterior point causes.
The content of the invention
For the defect and the deficiency of threshold setting method of traditional AdaBoost algorithms, propose a kind of improvement based on
The robust AdaBoost grader construction methods of Ransac algorithms.Built in Weak Classifier different from other AdaBoost algorithms
The simple means for using sample weighting and control of right in journey, this algorithm is by being incorporated into AdaBoost classification by Ransac algorithms
In device model construction process, and correct sample is searched by the process of iterative model building, remove potential exterior point, effectively overcome existing
There is sample weighting method in AdaBoost algorithms.
Meanwhile, strong this advantage of exterior point ability is removed by means of Ransac algorithms, be all based on AdaBoost algorithm structures
Optimal classification model is chosen in the sorter model built, and need not only consider how the problem for sample weighting, and effectively
Eliminate and degraded by the sorter model that exterior point causes.By above-mentioned strategy, finally realized based on Ransac algorithms herein
Robust AdaBoost graders build, and the design of the sorter model of acquisition is not affected by exterior point completely.
A kind of robust AdaBoost grader construction methods based on Ransac algorithms of improvement, comprise the steps：
(1) initial parameter is arranged
This algorithm whole parameter is initialized, including：The maximum number for arranging disaggregated model is Nmax；Arrange and currently just building point
Class device model is ith, arranges i=1 when most starting；It is T that setting builds the maximum iteration time of each sorter modelmax；Arrange
The iterations for building each sorter model is j, arranges j=1 during beginning；The sample of each initial construction grader is set
Number is M.
(2) a certain amount of sample is randomly selected
Random M sample of selection from sample set.
(3) built based on AdaBoost graders initial model
The sample selected based on these, using AdaBoost Algorithm for Training strong classifiers, i.e. so that sorter model is most accorded with
This M sample before being fated.
(4) sorter model updates
In the remaining sample of sorter model Ci classification that AdaBoost algorithms build；Correctly divided by the model with all
The sample of class builds sorter model Ci based on AdaBoost algorithms again；Judge to meet this new again with this new grader
Whole samples of sorter model, update j=j+1；Repeat the above steps 4 are to step 6 until meeting the sample number of grader Ci
No longer change or iterations have been over Tmax；.
(5) build next sorter model
I=i+1 is updated, judges the sorter model set up whether more than Nmax, such as not less than returning to step (2).
(6) classification accuracy rate is calculated
Calculate the corresponding sample classification precision of all classification device model.
(7) final classification model is selected
Compare the corresponding correct nicety of grading of all classification device model, the sorter model quilt with highest classification accuracy rate
It is defined as the sorter model of final selection.
Compared with prior art, the present invention has following obvious advantage and beneficial effect：
(1) the robust AdaBoost grader construction methods based on Ransac algorithms of a kind of improvement of the invention, can be effective
Exterior point interference is removed, the setting of sample weighting threshold value need not be not only carried out, and be avoided the improper of sample weighting threshold value setting
There is the misjudged disconnected situation of point and exterior point.
(2) even if the present invention contains the situation of a large amount of exterior points in the sample using Ransac algorithms, still can be with robust
Estimate sorter model parameter and the advantage that do not affected by exterior point, realize the accurate structure of disaggregated model.For checking one's duty
Class algorithm, with the inventive method in person's handwriting checking test, identifies the handwriting and writes people's identity and judged, even if this paper algorithms exist
When occurring in that 15% exterior point in sample, nicety of grading still can reach 95.67%.
Description of the drawings
Fig. 1 is a kind of robust AdaBoost graders structure side based on Ransac algorithms of improvement proposed by the invention
Method functional block diagram；
Fig. 2 is the flow chart of method involved in the present invention；
Specific embodiment
The present invention will be further described with reference to the accompanying drawings and detailed description.
As shown in Figure 1, 2, the present invention discloses a kind of robust AdaBoost grader structures based on Ransac algorithms of improvement
Construction method, comprises the following steps：
Step 1, initial parameter are arranged
This algorithm whole parameter is initialized, including：The maximum number for arranging disaggregated model is 100；Arrange and currently just building point
Class device model is ith, arranges i=1 when most starting；It is 100 that setting builds the maximum iteration time of each sorter model；Arrange
The iterations for building each sorter model is j, arranges j=1 during beginning；The sample of each initial construction grader is set
Number is 50.In this experiment, according to experiment demand every time from the person's handwriting sample of random 200 different peoples of selection of HIT-MW Sample Storehouses
This, and according to a certain percentage which part writing sample classification is mislabeled, a certain proportion of exterior point of artificial generation.
Step 2, a certain amount of sample are randomly selected
50 initial samples of random selection are concentrated from writing sample.Based on above-mentioned sample, using current writing sample
Person's handwriting microstructure features are extracted as the sample characteristics of Weak Classifier in AdaBoost algorithms, and Weak Classifier is built based on SVM.
This paper AdaBoost algorithms and the anti-exterior point interference performance of other AdaBoost algorithms are verified, is repeated this experiment repeatedly, is asked
Take average nicety of grading of every kind of algorithm under the effect of different proportion exterior point.
Step 3, is built based on AdaBoost graders initial model
The writing sample selected based on these at random and its corresponding person's handwriting microstructure features, are built using SVM algorithm weak
Grader, at the same using AdaBoost Algorithm for Training obtain strong classifier model so that the sorter model best suit it is current this
The judgement of 50 writing samples.
Step 4, sorter model update
The remaining sample of sorter model Ci classification built based on AdaBoost algorithms；With all correct by the model
The sample of classification builds sorter model Ci based on AdaBoost algorithms again；Judge again to meet with this new grader and be somebody's turn to do
Whole samples of new sorter model, update j=j+1；Repeat the above steps 3 are to step 4 until meeting the sample of grader Ci
Number no longer change or iterations have been over 100.
Step 5, builds next sorter model
I=i+1 is updated, whether judges the sorter model number of foundation more than 100, such as not less than returning to step 2.
Step 6, classification accuracy rate are calculated
Calculating is all based on the sorter model of this patent algorithm structure to whole writing samples in the writing sample storehouse
Sample classification precision, nicety of grading Accuracy=(correct classification person's handwriting sample number)/(whole writing sample numbers).
Step 7, final classification model are selected
Compare the corresponding correct nicety of grading of all classification device model, with highest nicety of grading AmaxSorter model
It is confirmed as the sorter model of final selection.
In order to verify effectiveness of the invention, this algorithm is applied to into person's handwriting confirmatory experiment, and is entered with other two methods
Row compares, and the accuracy of three kinds of sorting techniques is as shown in table 1.
1 the inventive method of table and the comparative result of two kinds of recognition methods
1 data display of table set forth herein improvement AdaBoost sorting techniques and other two kinds of AdaBoost methods point
Class precision, three kinds of methods are respectively：Context of methods, the common AdaBoost algorithms for not having control of right and fixed weight threshold
The AdaBoost methods of control.As it can be seen from table 1 when no exterior point in sample or less exterior point, the classification of three kinds of methods
Precision very close to.But proportion increases in the sample with exterior point, two methods niceties of grading receives very big negative in addition
Affect, nicety of grading monotonic decreasing.And compared with both approaches, this paper algorithms are by the good removal exterior point of Ransac algorithms
Jamming performance, after exterior point ratio is continuously increased, nicety of grading is still stable, with the obvious advantage to other two methods.In contrast to another
Outer two methods, even if this paper algorithms occur in that 15% exterior point in the sample, nicety of grading still can reach 95.67%.
Claims (1)
1. robust AdaBoost grader construction methods based on Ransac algorithms of a kind of improvement, it is characterised in that include as
Lower step：
Step (1) initial parameter is arranged
This algorithm whole parameter is initialized, including：The maximum number for arranging disaggregated model is Nmax；Setting is currently just building grader
Model is ith, arranges i=1 when most starting；It is T that setting builds the maximum iteration time of each sorter modelmax；Arrange and build
The iterations of each sorter model is j, arranges j=1 during beginning；The number of samples for arranging each initial construction grader is
M；
Step (2) amount sample is randomly selected
Random M sample of selection from sample set；
Step (3) is built based on AdaBoost graders initial model
The sample selected based on these, using AdaBoost Algorithm for Training strong classifiers, i.e. so that sorter model is best suited working as
Front this M sample；
Step (4) sorter model updates
In the remaining sample of sorter model Ci classification that AdaBoost algorithms build；With all by the model correctly classification
Sample builds sorter model Ci based on AdaBoost algorithms again；Judge again to meet the new classification with this new grader
Whole samples of device model, update j=j+1；Repeat the above steps 4 to step 6 until meet grader Ci sample number no longer
Change or iterations have been over Tmax；
Step (5) builds next sorter model
I=i+1 is updated, judges the sorter model set up whether more than Nmax, such as not less than returning to step (2)；
Step (6) classification accuracy rate is calculated
Calculate the corresponding sample classification precision of all classification device model；
Step (7) final classification model is selected
Compare the corresponding correct nicety of grading of all classification device model, the sorter model with highest classification accuracy rate is determined
For the final sorter model chosen.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610917782.4A CN106529579A (en) | 2016-10-20 | 2016-10-20 | Improved construction method of robust AdaBoost classifier based on Ransac algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610917782.4A CN106529579A (en) | 2016-10-20 | 2016-10-20 | Improved construction method of robust AdaBoost classifier based on Ransac algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106529579A true CN106529579A (en) | 2017-03-22 |
Family
ID=58332882
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610917782.4A Pending CN106529579A (en) | 2016-10-20 | 2016-10-20 | Improved construction method of robust AdaBoost classifier based on Ransac algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106529579A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111814880A (en) * | 2020-07-10 | 2020-10-23 | 北京航空航天大学 | Fatigue prediction method based on Bayesian optimization XGboost algorithm |
-
2016
- 2016-10-20 CN CN201610917782.4A patent/CN106529579A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111814880A (en) * | 2020-07-10 | 2020-10-23 | 北京航空航天大学 | Fatigue prediction method based on Bayesian optimization XGboost algorithm |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109948658B (en) | Feature diagram attention mechanism-oriented anti-attack defense method and application | |
| CN107506799B (en) | Deep neural network-based mining and expanding method and device for categories of development | |
| CN104778448B (en) | A kind of face identification method based on structure adaptive convolutional neural networks | |
| CN111507469B (en) | Method and device for optimizing super parameters of automatic labeling device | |
| CN107578061A (en) | Based on the imbalanced data classification issue method for minimizing loss study | |
| CN105205448A (en) | Character recognition model training method based on deep learning and recognition method thereof | |
| CN102567742A (en) | Automatic classification method of support vector machine based on selection of self-adapting kernel function | |
| CN107943856A (en) | A kind of file classification method and system based on expansion marker samples | |
| EP1906369A1 (en) | Real-time method of determining eye closure state using off-line adaboost-over-genetic programming | |
| CN110929848A (en) | Training and tracking method based on multi-challenge perception learning model | |
| CN111047054A (en) | Two-stage countermeasure knowledge migration-based countermeasure sample defense method | |
| CN101710382A (en) | Gabor human face recognizing method based on simplified intelligent single-particle optimizing algorithm | |
| Zhang et al. | Evolving neural network classifiers and feature subset using artificial fish swarm | |
| CN112597993A (en) | Confrontation defense model training method based on patch detection | |
| Yang et al. | Chaotic maps in binary particle swarm optimization for feature selection | |
| De Almeida et al. | Handling concept drifts using dynamic selection of classifiers | |
| CN117940936A (en) | Method and apparatus for evaluating robustness against | |
| CN109842614B (en) | Network intrusion detection method based on data mining | |
| CN114137967B (en) | Driving behavior decision method based on multi-network joint learning | |
| CN107273922A (en) | A kind of screening sample and weighing computation method learnt towards multi-source instance migration | |
| CN104537385A (en) | Method for evaluating DAGSVM classification accuracy | |
| CN107832722B (en) | Face detection classifier construction method based on AdaBoost | |
| JP2020204909A (en) | Machine learning device | |
| CN106529579A (en) | Improved construction method of robust AdaBoost classifier based on Ransac algorithm | |
| EP1480167A1 (en) | Pattern feature selection method, classification method, judgment method, program, and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | 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: 20170322 |