CN103500336B - The entropy method that optical filter defect characteristic parameter selects - Google Patents
The entropy method that optical filter defect characteristic parameter selects Download PDFInfo
- Publication number
- CN103500336B CN103500336B CN201310449476.9A CN201310449476A CN103500336B CN 103500336 B CN103500336 B CN 103500336B CN 201310449476 A CN201310449476 A CN 201310449476A CN 103500336 B CN103500336 B CN 103500336B
- Authority
- CN
- China
- Prior art keywords
- feature
- candidate
- defect
- optical filter
- characteristic
- 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.)
- Active
Links
Landscapes
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of entropy method that optical filter defect characteristic parameter selects, including: from defect optical filter image, it is partitioned into the boundary rectangle comprising defect, forms defect ROI；Arranging the element of candidate characteristic set F, arranging and having selected feature set S is empty set；Calculate defect ROI feature value, construct sample set；Calculate candidate feature fifkNormalized mutual information SU (f with class Cifk,C)；According to SU (fifk, C) and maximum selects first element s of S1；Remove feature and the normalized mutual information SU (f being selected into S in candidate characteristic set Fifk, C) and less than the candidate feature of threshold value；Calculate each candidate feature f in candidate characteristic set FifkEvaluation function J (fifk, C, S) value；According to evaluation function J (fifk, C, S) and maximum selects the next element selecting feature set S；Remove feature and the evaluation function J (f being selected into S in candidate characteristic set Fifk, C, S) and less than the candidate feature of threshold value；Judge candidate characteristic set F whether empty set；Feature has been selected in output.
Description
Technical field
The present invention relates to optical filter defective vision identification technical field, particularly relate to one and utilize Entropy principle pair
The method that the characteristic parameter of optical filter defect carries out selecting.
Background technology
Optical film filter is widely used in the fields such as optic communication, laser technology, optical imagery and detection,
Minisize pick-up head, Biomedical Instruments, advanced laser system play an important role.As in optic communication
Field, optical filter is not only in wavelength-division multiplex system the Primary Component realizing multiplexing and demultiplexing, the most conventional
In the flat gain of erbium-doped fiber amplifier, full light Add/drop Voice Channel, lambda switch.In photovoltaic,
Each mobile phone camera must be equipped with a tablet filter.Domestic, international market is huge to the demand of optical filter,
The annual requirement of domestic only mobile phone camera optical filter just has 700,000,000.
The manufacture process of optical filter is typically complex, and quick and precisely detects optical filter open defect
It it is the premise improving product quality.Currently mainly taking bore hole detection method piecewise, cost is high, detection
Speed, precision are difficult to ensure that.Vision-based detection shoots measured object image by video camera, utilizes image procossing
Noncontact, high-precision real-time measurement is realized etc. technology.Utilize the one of vision technique detection optical filter defect
Planting common method is that optical filter defect image is carried out feature selection, instructs by the characteristic vector after feature selection
Practice grader, then carry out defect recognition.When multiple candidate feature are selected, evaluate candidate feature
The working standard of significance level includes criterion distance, consistency criterion, dependency standard and information standard.
Wherein information standard utilizes the degree of uncertainty between the concept quantization characteristic of comentropy, and it need not in advance
Knowing that data are distributed, the non-linear relation between energy valid metric feature, many simulation tests have proven to information
Standard has better performance than other standard as a rule.
When using information standard evaluating characteristic to the separating capacity of classification, one method simply and intuitively is
Calculate each candidate feature f1..., fDAnd the mutual information I (f between classification C1;C),…,I(fD;C), mutual information
Showing that the most greatly candidate feature is the strongest with the dependency of classification, this candidate feature is the most important, therefore that these are mutual
The value of information arranges in descending order, and several features above select result exactly.The shortcoming of this method is not
There is the dependency or redundancy considered between feature, such as feature f1, f2All there is strong correlation with classification C,
But f1With f2Between there is strong correlation (as met f1=2f2), then f1, f2In only one should be had to enter
Choosing, but by above method f1, f2To be selected in simultaneously.For reducing the dependency between selected feature as far as possible, one
The method of kind is by mutual information I (fi;C) conditional mutual information I (f it is revised asi;C/S), conditional mutual information I (fi;C/S) table
Show candidate feature f under conditions of currently having selected feature set SiAnd the mutual information between classification C.Select feature
In collection S, selected feature is the most, conditional mutual information I (fi;C/S) amount of calculation and computation complexity are the biggest, very
Calculate to being difficult to.Another kind of method is by mutual information I (fi;C) J (f it is revised asi)=I(fi;C)-βI(fi;S), wherein
β is adjustment factor, I (fi;S) candidate feature f is representediAnd select the mutual information between feature set S.I(fi;S)
Accurately calculate the most extremely difficult, frequently with approximation method such as J (fi)=I(fi,C)-max[I(fi,s1),…,I(fi,sR)]
Or J (fi)=I(fi;C)-β(I(fi;s1)+…+I(fi;sR)), wherein s1,…,sRIt is currently selected in feature set S complete
Portion is selected in feature.When calculating mutual information, for mutual information span between solution different characteristic and class not
Same problem, can use Cambridge University Press to publish " Numerical Recipes in C " in 1988
The normalization representation of mutual information in one book, i.e. the normalized mutual information computing formula of A with B is:
Wherein mutual information I (A;And the computing formula of entropy H (A), H (B) can be found in theory of information pertinent literature B).
In sum, the method for existing employing information standard investigates the single features difference energy to classification one by one
Power, but the composite behaviour of these independent optimal characteristics may not be optimum, and easily some independent role time zones are selected in leakage
Point ability and the feature strong with separating capacity during further feature synergy.
Summary of the invention
For solving above-mentioned technical problem, it is an object of the invention to provide a kind of optical filter defect characteristic parameter choosing
The entropy method selected, distinguishes ability when the method considers some feature independent role and combines with further feature
The situation that during effect, separating capacity is strong, according to dependency, the candidate feature of candidate feature with defect classification be
The amount of new information that defect classification provides, the evaluation function of design judgment candidate feature importance, this evaluation letter
Number can being on the increase and adjust accordingly with selected feature, easily and effectively realize optical filter defect characteristic ginseng
The selection of number.
The purpose of the present invention is realized by following technical scheme:
Segmentation defect optical filter image comprises the boundary rectangle of defect, forms defect ROI；
Arranging the element of candidate characteristic set F, arranging and having selected feature set S is empty set；
Calculate the eigenvalue of described defect ROI, construct sample set by eigenvalue；
Calculate candidate feature f of all samples in sample setifkNormalized mutual information SU (f with class Cifk,
C)；
According to normalized mutual information SU (fifk, C) and maximum selects first element selecting feature set S
s1；Remove feature and the normalized mutual information SU (f being selected into S in candidate characteristic set Fifk, C) and less than threshold
The candidate feature of value；
Calculate each candidate feature f in candidate characteristic set FifkEvaluation function J (fifk, C, S) value；Root
According to evaluation function J (fifk, C, S) and maximum selects the next element selecting feature set S；Remove candidate special
Collection F has been selected into feature and the evaluation function J (f of Sifk, C, S) and less than the candidate feature of threshold value；Repeat
This step, until candidate characteristic set F is empty set；
The element having selected feature set S is optical filter defect characteristic parameter.
Compared with prior art, one or more embodiments of the invention can have the advantage that
Candidate characteristic set, except including single features, is also added into assemblage characteristic so that during independent role pair
Class discrimination ability is weak and strong with separating capacity during further feature synergy feature can correctly be selected in,
Overcome in existing method and only single features is compared one by one, the problem that overall selection result may not be optimum.
For assemblage characteristic, need only consider the combination of any two single features, it is special that this not only conforms with optical filter defect
The practical situation levied, it also avoid complicated calculating during more features combination.Judging that candidate feature is important
During degree, evaluation function J (fifk, C, S) consider simultaneously this candidate feature and classification dependency and with select
The redundancy of feature, and use the normalization representation of information, accurately reflect it and divide for defect
The amount of new information that class provides.By the present invention can the significance level of correct evaluating characteristic, improve optical filter and lack
Fall into feature selection accurateness.
Other features and advantages of the present invention will illustrate in the following description, and, partly from froming the perspective of
Bright book becomes apparent, or understands by implementing the present invention.The purpose of the present invention is excellent with other
Point can be realized by structure specifically noted in description, claims and accompanying drawing and be obtained.
Accompanying drawing explanation
Accompanying drawing is for providing a further understanding of the present invention, and constitutes a part for description, with this
Inventive embodiment is provided commonly for explaining the present invention, is not intended that limitation of the present invention.In the accompanying drawings:
Fig. 1 is the entropy method flow diagram that optical filter defect characteristic parameter selects.
Detailed description of the invention
Easy to understand, according to technical scheme, under the connotation not changing the present invention, this
The those skilled in the art in field can propose multiple frame modes and the manufacture method of the present invention.Therefore below
Detailed description of the invention and accompanying drawing are only illustrating of technical scheme, and are not to be construed as this
Invention whole or be considered as defining or limiting of technical solution of the present invention.
Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail.
Fig. 1 is the entropy method flow that optical filter defect characteristic parameter selects, and the method includes:
Step 101 is partitioned into the boundary rectangle comprising defect from defect optical filter image, forms defect
ROI；
Above-mentioned from defect optical filter image, it is partitioned into defect ROI, is that optical filter image is carried out binaryzation
Process, recycle morphologic expansion, corrosion make the often place defect in optical filter image form connection respectively
Region, and then it is partitioned into the boundary rectangle often locating defect, form defect ROI；
Step 102 arranges the element of candidate characteristic set F, and arranging and having selected feature set S is empty set；
The above-mentioned element arranging candidate characteristic set F, is 16 features specifying optical filter defect ROI,
It is barycenter vertical coordinate, barycenter abscissa, defect area, defect Ratio of long radius to short radius, eccentricity, ROI respectively
Rectangular degree, defect are constant with the first seven low order of ROI area ratio, region compactness, Euler's numbers and ROI
Square, the combination of these single features and two-by-two feature totally 136 elements constituting candidate characteristic set F.
Step 103 calculates the eigenvalue of described defect ROI, constructs sample set by eigenvalue；
Each defect ROI is first calculated 16 features specified, combined obtains whole candidate feature,
Forming a sample, the class label of sample corresponds to defect type, and the sample of all defect ROI constitutes sample
This set；
Step 104 calculates candidate feature f of all samples in sample setifkNormalization mutual trust with class C
Breath SU (fifk,C)；
Above-mentioned candidate feature fifk, the f as i=kifkRepresent single features fi, the f as i ≠ kifkRepresent fiWith
fkAssemblage characteristic, calculate each candidate feature f in candidate characteristic set F successivelyifkDefect class corresponding with sample
Normalized mutual information SU (f between other Cifk,C)。
Step 105 is according to normalized mutual information SU (fifk, C) and maximum selects and selects the first of feature set S
Individual element s1；
Maximum SU is selected from the result of calculation of above-mentioned steps 104max(fifk, C) and corresponding candidate feature
fifk, by this candidate feature fifkRemove from candidate characteristic set F, and add to and select in feature set S,
Become first element s of S1.For step 104 result of calculation less than threshold value i.e. SU (fifk,C)<SUth
Candidate feature fifk, remove from candidate characteristic set F.
Above-mentioned normalized mutual information SU (fifk, C) and being chosen as of maximum: by normalized mutual information SU (fifk,
C) value is ranked up from big to small, and wherein, the value of sequence first is normalized mutual information SU (fifk,C)
Maximum.
Step 106 calculates each candidate feature f in candidate characteristic set FifkEvaluation function, this evaluation letter
The computing formula of number is as follows:
J(fifk,C,S)=SU(fifk, C) and-max [SU (fifk,s1),…,SU(fifk,sR)]
Described evaluation function J (fifk, C, S) in s1,…,sRIt is currently to have selected the whole elements in feature set S；
Step 107 adjusts candidate characteristic set F and has selected the element of feature set S；
The element of described adjustment candidate characteristic set F, is by the maximum J of step 106 result of calculationmax(fifk,
C, S) corresponding candidate feature fifkRemove from candidate characteristic set F, by step 106 result of calculation less than threshold
Value i.e. J (fifk,C,S)<JthCandidate feature fifkRemove from candidate characteristic set F；Described adjustment is selected
The element of feature set S, is by the maximum J of step 106 result of calculationmax(fifk, C, S) and corresponding candidate
Feature fifkAdd to and select in feature set S.
Step 108 judges whether candidate characteristic set F is empty set；
If not empty set, perform step 106, step 107, otherwise perform step 109；
Step 109 exports selects feature；
The element having selected feature set S is the characteristic parameter of optical filter defect classification, and feature selection is complete,
The selected feature of output.
Optical filter defect characteristic parameter selection experiments collapses angle, pit, scuffing, sliver four at 50 containing collapsing limit
Plant in the optical filter image of defect and carry out, utilize above-mentioned steps 101～step 109 to carry out defect characteristic ginseng
The selection of number, selects result to be followed successively by defect Ratio of long radius to short radius, eccentricity, Euler's numbers by significance level arrangement,
Wherein two features of defect Ratio of long radius to short radius and eccentricity work in a joint manner.
In order to verify the feature of the described method choice effectiveness in optical filter defect is classified, by described side
Method and common feature select (being designated as CM) method to carry out contrast experiment.CM method is with described method not
It is with part: only investigate the single features discrimination to classification, to candidate feature fiEvaluation function
For J (fi)=I(fi,C)-max[I(fi,s1),…,I(fi,sR)], in formula, I represents mutual information.Described method and CM
The contrast experiment of method includes two parts: (1) carries out feature selection respectively by two kinds of methods；(2) with two
The feature selection result of the method for kind carries out optical filter defect classification experiments.Experimental result is as follows:
(1) utilize CM method that same optical filter image carries out feature selection, select result by important
Degree arrangement is followed successively by defect and ROI area ratio, Euler's numbers, eccentricity.
(2) classification of optical filter defect is carried out by the feature selection result of described method and CM method respectively real
Testing, grader uses the support vector machine in Matlab workbox, and 64 defects ROI are as training sample
This, 36 defects ROI are as test sample.All kinds of defects are institute's accounting in training sample and test sample
Example meets the probability that all kinds of defect occurs in producing reality.Classification results is described classification accuracy
Being 92%, CM classification accuracy is 87%.
The feature selection result of the described method of visible use makes the classification accuracy rate of support vector machine improve
5%。
Although the embodiment that disclosed herein is as above, but described content is only to facilitate understand this
The embodiment invented and use, is not limited to the present invention.In any the technical field of the invention
Technical staff, on the premise of without departing from the spirit and scope that disclosed herein, can implement
And make any amendment and change in details in form, but the scope of patent protection of the present invention, still must be with institute
Attached claims are defined in the range of standard.
Claims (3)
1. the entropy method that an optical filter defect characteristic parameter selects, it is characterised in that described method includes:
Segmentation defect optical filter image comprises the boundary rectangle of defect, forms defect ROI；
Arranging the element of candidate characteristic set F, arranging and having selected feature set S is empty set；
Calculate the eigenvalue of described defect ROI, construct sample set by eigenvalue；
Calculate candidate feature f of all samples in sample setifkNormalized mutual information SU (f with class Cifk,
C)；
According to normalized mutual information SU (fifk, C) and maximum selects first element selecting feature set S
s1；Remove feature and the normalized mutual information SU (f being selected into S in candidate characteristic set Fifk, C) and less than threshold
The candidate feature of value；
Calculate each candidate feature f in candidate characteristic set FifkEvaluation function J (fifk, C, S) value；Root
According to evaluation function J (fifk, C, S) and maximum selects the next element selecting feature set S；Remove candidate special
Collection F has been selected into feature and the evaluation function J (f of Sifk, C, S) and less than the candidate feature of threshold value；Repeat
This step, until candidate characteristic set F is empty set；
The element having selected feature set S is optical filter defect characteristic parameter；
J(fifk, C, S) and=SU (fifk, C) and-max [SU (fifk,s1),…,SU(fifk,sR)]
Described evaluation function J (fifk, C, S) in s1,…,sRIt is currently to have selected the whole elements in feature set S；
Described normalized mutual information SU (fifk, C) and evaluation function J (fifk, C, S) all utilize entropy to calculate, calculate
Time first by the continuous value discretization of candidate feature, then the formula pressing entropy calculates.
The entropy method that optical filter defect characteristic parameter the most according to claim 1 selects, its feature exists
In, the element in described candidate characteristic set F is referred to as candidate feature fifk, the f as i=kifkRepresent single spy
Levy fi, the f as i ≠ kifkRepresent fiWith fkAssemblage characteristic.
The entropy method that optical filter defect characteristic parameter the most according to claim 1 selects, its feature exists
In, described is that each defect ROI is calculated D the feature specified, by D by eigenvalue structure sample
The combination of two of individual eigenvalue and eigenvalue constitutes a sample of candidate feature.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310449476.9A CN103500336B (en) | 2013-09-24 | 2013-09-24 | The entropy method that optical filter defect characteristic parameter selects |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310449476.9A CN103500336B (en) | 2013-09-24 | 2013-09-24 | The entropy method that optical filter defect characteristic parameter selects |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103500336A CN103500336A (en) | 2014-01-08 |
| CN103500336B true CN103500336B (en) | 2016-08-17 |
Family
ID=49865541
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310449476.9A Active CN103500336B (en) | 2013-09-24 | 2013-09-24 | The entropy method that optical filter defect characteristic parameter selects |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103500336B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110942149B (en) * | 2019-10-31 | 2020-09-22 | 河海大学 | Feature variable selection method based on information change rate and condition mutual information |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1393690A (en) * | 2001-06-21 | 2003-01-29 | 株式会社理光 | Defect inspector and method thereof |
| CN101558292A (en) * | 2006-12-14 | 2009-10-14 | 日本电气硝子株式会社 | Glass sheet defect detection device, glass sheet manufacturing method, glass sheet, glass sheet quality judging device, and glass sheet inspection method |
| CN102803917A (en) * | 2009-06-18 | 2012-11-28 | 夏普株式会社 | Defect inspection method and defect inspection device for display panel |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPWO2010146733A1 (en) * | 2009-06-18 | 2012-11-29 | シャープ株式会社 | Display panel defect inspection method and defect inspection apparatus |
-
2013
- 2013-09-24 CN CN201310449476.9A patent/CN103500336B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1393690A (en) * | 2001-06-21 | 2003-01-29 | 株式会社理光 | Defect inspector and method thereof |
| CN101558292A (en) * | 2006-12-14 | 2009-10-14 | 日本电气硝子株式会社 | Glass sheet defect detection device, glass sheet manufacturing method, glass sheet, glass sheet quality judging device, and glass sheet inspection method |
| CN102803917A (en) * | 2009-06-18 | 2012-11-28 | 夏普株式会社 | Defect inspection method and defect inspection device for display panel |
Non-Patent Citations (1)
| Title |
|---|
| "基于联合信息熵特征选择的滤光片外观缺陷检测研究";付梦瑶;《中国优秀硕士学位论文全文数据库-信息科技辑》;20150215(第2期);I138-871 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103500336A (en) | 2014-01-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106097361A (en) | A kind of defective area detection method and device | |
| KR20200057618A (en) | Foreground-background-aware atrous multiscale network for disparity estimation | |
| CN112508044A (en) | Artificial intelligence AI model evaluation method, system and equipment | |
| CN108875602A (en) | Monitor the face identification method based on deep learning under environment | |
| CN110276264A (en) | A kind of crowd density estimation method based on foreground segmentation figure | |
| CN107027023A (en) | VoIP based on neutral net is without reference video communication quality method for objectively evaluating | |
| CN104820841B (en) | Hyperspectral classification method based on low order mutual information and spectrum context waveband selection | |
| US20230034994A1 (en) | Channel Identification Method and Apparatus, Transmission Method, Transmission Device, Base Station, and Medium | |
| CN103888541A (en) | Method and system for discovering cells fused with topology potential and spectral clustering | |
| Zeng et al. | UWB NLOS identification with feature combination selection based on genetic algorithm | |
| US20210349004A1 (en) | Methods and Systems for Characterizing Spillover Spreading in Flow Cytometer Data | |
| CN104243815A (en) | Focusing method and electronic equipment | |
| CN107679469A (en) | A kind of non-maxima suppression method based on deep learning | |
| CN105895089A (en) | Speech recognition method and device | |
| CN106888344A (en) | Camera module and its inclined acquisition methods of image planes and method of adjustment | |
| CN103500336B (en) | The entropy method that optical filter defect characteristic parameter selects | |
| CN105825288A (en) | Optimization analysis method for eliminating regression data colinearity problem of complex system | |
| CN110275909A (en) | Multivariate correlation method and system is detected based on DE-MIC algorithm | |
| CN104598921A (en) | Video preview selecting method and device | |
| CN106611339B (en) | Seed user screening method, and product user influence evaluation method and device | |
| CN116188374A (en) | Socket detection method, device, computer equipment and storage medium | |
| CN104899893A (en) | Image quality detection method based on vision attention | |
| CN105488521B (en) | A kind of dilatation screening sample method based on kernel function | |
| US20190196445A1 (en) | Method and system for sensing fine changes in processing/equipment measurement data | |
| CN108646688A (en) | A kind of process parameter optimizing analysis method based on recurrence learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |