CN107154041A - A kind of learning method classified for defects of display panel - Google Patents
A kind of learning method classified for defects of display panel Download PDFInfo
- Publication number
- CN107154041A CN107154041A CN201710343201.5A CN201710343201A CN107154041A CN 107154041 A CN107154041 A CN 107154041A CN 201710343201 A CN201710343201 A CN 201710343201A CN 107154041 A CN107154041 A CN 107154041A
- Authority
- CN
- China
- Prior art keywords
- mtd
- defect
- msub
- characteristic attribute
- mtr
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
- G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
Abstract
The invention discloses a kind of learning method classified for defects of display panel, including:Introduce defect characteristic attribute description parameter and defect characteristic property gradient parameter, wherein, the defect characteristic attribute description parameter includes the characteristic attribute description vectors of number of drawbacks type, and the defect characteristic property gradient parameter includes the characteristic attribute gradient scope α of number of drawbacks type and the one-to-one gradient factor factors of characteristic attribute gradient scope α with the number of drawbacks type;And by adjusting defect characteristic attribute description parameter or defect characteristic property gradient parameter so that the defect characteristic attribute description parameter and defect characteristic property gradient parameter of introducing are capable of the display defect type of Auto-matching display panel.It is of the invention that defect characteristic attribute description parameter, and constantly convergence optimization defect characteristic property gradient parameter, the defect recognition detection efficiency and accuracy of energy significant increase display panel are introduced by the study constantly improve to known defect type.
Description
Technical field
The present invention relates to display panel detection technique field, and in particular to a kind of classify for defects of display panel
Learning method.
Background technology
Flat-panel screens has the advantages that high-resolution, high gray scale and without geometry deformation, simultaneously because its small volume, again
Amount is light and low in energy consumption, thus is widely used in people's consumption electronic product used in everyday, such as TV, computer, hand
Machine, flat board etc..Display panel is the main body part of flat-panel screens tool, and its manufacturing process is complicated, and with display surface
The size of plate is done bigger and bigger, and the uniformity of its gray scale is also increasingly difficult to control, therefore occurs unavoidably in the fabrication process various
Display defect, such as bright spot/dim spot/foreign matter be bright/BL foreign matters (backlight foreign matter)/white point/bright concealed wire/Mura display defects.
At present, it is general in display panel producing line to be recognized using naked eyes by the way of display defect quantity and type to display panel
Grade judgement is carried out, detection efficiency is low, false drop rate is high.
The content of the invention
For above-mentioned the deficiencies in the prior art, the present invention discloses a kind of learning method classified for defects of display panel,
By introducing defect characteristic attribute description parameter and defect characteristic property gradient parameter, and pass through the study to known defect type
Constantly improve introduces defect characteristic attribute description parameter, and constantly convergence optimization defect characteristic property gradient parameter so that introduce
Defect characteristic attribute description parameter and defect characteristic property gradient parameter be capable of the display defect class of Auto-matching display panel
Type, the defect recognition detection efficiency and accuracy of energy significant increase display panel.
To achieve the above object, the present invention provides a kind of learning method classified for defects of display panel, including following
Step:
1) defect characteristic attribute vector d is extracted from the image of a known defect type;
2) a defect characteristic property gradient coefficient sets G is provided, defect characteristic property gradient coefficient sets G has included this
Know the characteristic attribute gradient scope α of defect typeijAnd with this feature property gradient scope αijThe one-to-one gradient factor factor
βij;Defect characteristic attribute vector d is updated to this feature property gradient scope α respectivelyijIn, obtain the gradient factor factor to
Measure β;
3) a defect characteristic attribute description set M is provided, defect characteristic attribute description set M includes number of drawbacks type
Multiple characteristic attribute description vectors m;Gradient factor factor vector β is entered with the plurality of characteristic attribute description vectors m respectively
Row convolution, generates one group of convolution value;
4) if the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is the known defect type, tie
Harness defects type identification flow;
If the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is not the known defect type, repair
Change the characteristic attribute gradient scope α of the known defect typeijAnd with this feature property gradient scope αijOne-to-one gradient system
Number factor-betaij, repeat step 1 to 4, until maximum convolution is worth the defect type of corresponding characteristic attribute description vectors for this
Know defect type, obtain the characteristic attribute gradient scope of the new known defect type and with this feature property gradient scope one by one
The corresponding gradient factor factor.
Further, above-mentioned technical proposal carries out convolutional calculation using below equation:
Wherein, miIt is characteristic attribute description vectors m (m1,m2,…mn) (0 < mi< 1) in element, for description defect characteristic
The weighted factor of information, characteristic attribute description vectors m is converted into matrixParticipate in convolutional calculation;βi
It is gradient factor factor vector β (β1,β2,…βn) (0 < βi< 1) in element, the gradient factor factor vector β conversion in column to
AmountParticipate in convolutional calculation.
Further, in above-mentioned technical proposal defect characteristic include picture title, and/or area, and/or length, and/or
Wide, and/or length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate.
Further, in above-mentioned technical proposal this feature attribute description vector m be to area, and/or length, and/or width,
And/or the normalized set of numerical value of length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate.
Further, the characteristic attribute gradient model of the known defect type is changed described in above-mentioned technical proposal step 4
Enclose the characteristic attribute gradient scope for specially reducing the known defect type.
In addition, the present invention still further provides a kind of learning method classified for defects of display panel, comprise the following steps:
1) defect characteristic attribute vector d is extracted from the image of a known defect type;
2) a defect characteristic property gradient coefficient sets G is provided, defect characteristic property gradient coefficient sets G has included this
Know the characteristic attribute gradient scope α of defect typeijAnd with this feature property gradient scope αijThe one-to-one gradient factor factor
βij;Defect characteristic attribute vector d is updated to this feature property gradient scope α respectivelyijIn, obtain the gradient factor factor to
Measure β;
3) a defect characteristic attribute description set M is provided, defect characteristic attribute description set M includes number of drawbacks type
Multiple characteristic attribute description vectors m, defect characteristic attribute description set M also including the known defect type feature belong to
Property description vectors m1;By gradient factor factor vector β respectively with the plurality of characteristic attribute description vectors m, the known defect class
The characteristic attribute description vectors m1 of type carries out convolution, generates one group of convolution value;
4) if the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is the known defect type, tie
Harness defects type identification flow;
If the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is not the known defect type, repair
Change this feature attribute description vector m1, repeat step 1 to 4, until maximum convolution is worth corresponding characteristic attribute description vectors
Defect type is the known defect type, obtains the characteristic attribute description vectors m2 of the new known defect type.
Further, above-mentioned technical proposal carries out convolutional calculation using below equation:
Wherein, miIt is characteristic attribute description vectors m (m1,m2,…mn) (0 < mi< 1) in element, for description defect characteristic
The weighted factor of information, characteristic attribute description vectors m is converted into matrixParticipate in convolutional calculation;βi
It is gradient factor factor vector β (β1,β2,…βn) (0 < βi< 1) in element, the gradient factor factor vector β conversion in column to
AmountParticipate in convolutional calculation.
Further, in above-mentioned technical proposal defect characteristic include picture title, and/or area, and/or length, and/or
Wide, and/or length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate.
Further, in above-mentioned technical proposal this feature attribute description vector m be to area, and/or length, and/or width,
And/or the normalized set of numerical value of length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate.
Further, it is specially increase that this feature attribute description vector m1 is changed described in above-mentioned technical proposal step 4
The known defect type characteristic attribute description, or redescribe the known defect type defect characteristic information weighting because
Son.
The present invention is by introducing defect characteristic attribute description parameter and defect characteristic property gradient coefficient, and by known
The study constantly improve of defect type introduces defect characteristic attribute description parameter, and constantly convergence optimization defect characteristic property gradient
Coefficient so that the defect characteristic attribute description parameter and defect characteristic property gradient parameter of introducing being capable of Auto-matching display panels
Display defect type, can significant increase display panel defect recognition detection efficiency and accuracy.
Brief description of the drawings
Fig. 1 uses the structural representation of the defects of display panel sorter embodiment of the present invention;
The flow chart for the learning method that Fig. 2 present invention classifies for defects of display panel.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in each embodiment of invention described below
Not constituting conflict each other can just be mutually combined.
Embodiment:The defect type of display panel can be learnt, classify and grade judge defects of display panel
Grade decision maker
As shown in figure 1, a kind of defects of display panel grade decision maker that the embodiment of the present invention is provided mainly includes industry
Camera, image algorithm processing module, defect grader study module, defect sort module and defect rank determination module.Wherein,
Industrial camera is mainly used in gathering the display picture of display panel;Image algorithm processing module is mainly used in carrying from display picture
Take out defect characteristic information;Defect grader study module is used to provide defect type configuration file and defect characteristic property gradient
File;Defect sort module is used for according to defect type configuration file and defect characteristic property gradient file to defect characteristic information
Carry out the identification of defect type;Defect rank determination module is used to judge display according to the defects count and defect type that identify
The grade of panel.
In above-mentioned technical proposal, the defect characteristic information that image algorithm processing module is extracted refers to include multigroup defect spy
Levy attribute vector d defect characteristic attribute set D;Wherein, defect characteristic refers to defect point in display picture or defect area
The information such as area, length and width, length-width ratio, center gray scale, contrast, coordinate, and display picture title;It is every in display picture
One defect point or the actual defect characteristic information of defect area are all represented with one group of defect characteristic attribute vector d.
In above-mentioned technical proposal, defective spy is included in the defect type configuration file that defect grader study module is provided
Levy include in attribute description set M, defect characteristic attribute description set M various defect types multiple characteristic attributes describe to
M (a kind of one characteristic attribute description vectors of defect type correspondence) is measured, characteristic attribute description vectors m is one defect type of description
The set of the weighted factor of defect characteristic information.Wrapped in the defect characteristic property gradient file that defect grader study module is provided
Include in defective characteristic attribute gradient factor set G, defect characteristic property gradient coefficient sets G and include various defect types
Characteristic attribute gradient scope αijAnd with characteristic attribute gradient scope αijOne-to-one gradient factor factor-betaij, as shown in table 1.
Table 1
In above-mentioned technical proposal, defective grader kernel function 1 is built in defect grader study module:
Or defect grader kernel function 2:
Wherein, miIt is characteristic attribute description vectors m (m1,m2,…mn) (0 < mi< 1) in element, it is special to describe each defect
The weighted factor levied, characteristic attribute description vectors m is converted into matrixParticipate in convolutional calculation, diIt is
Defect characteristic attribute vector d (d1,d2,…dn) in element, be the actual value of each defect characteristic, defect characteristic attribute vector d turns
Change column vector intoParticipate in convolutional calculation, βiFor the gradient factor factor of the defect characteristic of introducing, gradient factor factor vector β
(β1,β2,…βn) (0 < βi< 1) it is converted into column vectorParticipate in convolutional calculation.Defect grader study module is used to combine
Defect grader kernel function is bright to known bright spot/dim spot/foreign matter/BL foreign matters (backlight foreign matter)/white point/bright concealed wire/Mura etc.
Display defect is learnt, generation defect type configuration file and defect characteristic property gradient file.
In above-described embodiment, the mistake that defects of display panel grade decision maker is learnt to the defect type of display panel
Journey includes defect type configuration file acquisition process and defect characteristic property gradient file acquisition process.
As shown in Fig. 2 in above-described embodiment, the acquisition process of defect type configuration file comprises the following steps, to learn
Exemplified by BL foreign matter defects:
S11) image algorithm processing module loads the image that a width includes a BL foreign matter defects, and is extracted from the image
Go out the defect characteristic information of BL foreign matter defects, obtain one group of defect characteristic attribute vector d, as shown in table 2.
Table 2
S12 the weighted factor of the defect characteristic information of BL foreign matter defects, such as table) are configured in defect grader study module
Shown in 3, the characteristic attribute description vectors m1 of BL foreign matter defects is obtained, and generates defect type configuration file.It should be noted that
Defect grader study module has included various other common deficiency classes by the study of early stage in defect type configuration file
The characteristic attribute description vectors of type, as shown in table 3.
Table 3
Characteristic attribute | Area | Length-width ratio | White pictures | Black pictures | L63 pictures | Contrast |
Point | 0.45 | 0.45 | 0.1 | (0.1) | 0.0 | 0.0 |
Line | 0.0 | 0.9 | 0.1 | (0.1) | 0.0 | 0.0 |
BL foreign matters | 0.8 | 0.0 | 0.2 | 0.0 | 0.0 | 0.0 |
S13) defect sort module reads defect type configuration file from defect grader study module, and it is scarce to parse this
Type configuration file generated defect characteristic attribute description set M is fallen into, as shown in table 3;Meanwhile, defect sort module is classified from defect
Defect characteristic property gradient file is read in device study module, and parses the defect characteristic property gradient file generated defect characteristic
Property gradient coefficient sets G, as shown in table 4.
Table 4
S14) feature corresponding with BL foreign matters defect in table 4 of the defect characteristic attribute vector of BL foreign matters defect in table 2 is belonged to
Property gradient scope αijIt is compared, obtains the gradient factor factor vector of BL foreign matter defects
S15) using defect grader kernel function 2, by gradient factor factor vectorRespectively with 3 spies in table 3
Levy attribute description vector and carry out convolution, generate one group of convolution value.
S16) if it is BL foreign matter defects that maximum convolution, which is worth corresponding defect type, Classifcation of flaws flow is terminated;
If it is not BL foreign matter defects that maximum convolution, which is worth corresponding defect type, repaiied in defect grader study module
Changing the weighted factor of the defect characteristic information of BL foreign matter defects (specially increases the characteristic attribute description of BL foreign matter defects, such as increases
Plus the description of the defect characteristic information such as BL foreign matters defect center gray scale, coordinate;Or redescribe the defect characteristic of BL foreign matter defects
The weighted factor of information), repeat step S11 to S16, until it is BL foreign matter defects that maximum convolution, which is worth corresponding defect type,
Obtain the weighted factor of the defect characteristic information of amended BL foreign matters defect.
As shown in Fig. 2 in above-described embodiment, the acquisition process of defect characteristic property gradient file comprises the following steps, with
Exemplified by study BL foreign matter defects:
S21) image algorithm processing module loads the image that a width includes a BL foreign matter defects, and is extracted from the image
Go out the defect characteristic information of BL foreign matter defects, obtain one group of defect characteristic attribute vector d, as shown in table 5.
Table 5
S22 the characteristic attribute gradient scope α of BL foreign matter defects) is configured in defect grader study moduleijAnd it is right therewith
The gradient factor factor-beta answeredij, as shown in table 6, and generate defect characteristic property gradient file.
Table 6
S23) defect sort module reads defect type configuration file from defect grader study module, and it is scarce to parse this
Type configuration file generated defect characteristic attribute description set M is fallen into, as shown in table 7;Meanwhile, defect sort module is classified from defect
Defect characteristic property gradient file is read in device study module, and parses the defect characteristic property gradient file generated defect characteristic
Property gradient coefficient sets G, as shown in table 6.
Table 7
Characteristic attribute | Area | Length-width ratio | White pictures | Black pictures | L63 pictures | Contrast |
Point | 0.45 | 0.45 | 0.1 | (0.1) | 0.0 | 0.0 |
Line | 0.0 | 0.9 | 0.1 | (0.1) | 0.0 | 0.0 |
BL foreign matters | 0.8 | 0.0 | 0.2 | 0.0 | 0.0 | 0.0 |
S24) feature corresponding with BL foreign matters defect in table 6 of the defect characteristic attribute vector of BL foreign matters defect in table 5 is belonged to
Property gradient scope αijIt is compared, obtains the gradient factor factor vector of BL foreign matter defects
S25) using defect grader kernel function 2, by gradient factor factor vectorRespectively with 3 spies in table 7
Levy attribute description vector and carry out convolution, generate one group of convolution value.,
S26) if it is BL foreign matter defects that maximum convolution, which is worth corresponding defect type, Classifcation of flaws flow is terminated;
If it is not BL foreign matter defects that maximum convolution, which is worth corresponding defect type, repaiied in defect grader study module
Change the characteristic attribute gradient scope α of BL foreign matter defectsijAnd corresponding gradient factor factor-betaij, or reduce BL foreign matters and lack
Sunken characteristic attribute gradient scope (provides characteristic attribute gradient scope α after reducing the scopeijAnd the corresponding gradient factor factor
βij), repeat step S21 to S26, until it is BL foreign matter defects that maximum convolution, which is worth corresponding defect type, is obtained amended
The characteristic attribute gradient scope α of BL foreign matter defectsijAnd corresponding gradient factor factor-betaij。
Defects of display panel grade decision maker is classified and waited to the defect type of display panel in above-described embodiment
Level judges that embodiment 1 comprises the following steps:
S31) image signal generator sends solid-color image (it should be noted that pure in the present embodiment to display panel
Color image includes but is not limited to the solid-color images such as R, G, B, white, black), then industrial camera is adopted to the solid-color image
Collection, and be sent in image algorithm processing module.
S32) image algorithm processing module extracts defect characteristic information from the solid-color image, obtains including multigroup defect
The sunken characteristic attribute set D of type, includes multigroup sunken characteristic attribute vector, such as the institute of table 8 wherein falling into characteristic attribute set D
Show.
Table 8
S33) defect sort module reads defect type configuration file and defect characteristic from defect grader study module
Property gradient file, and defect type configuration file is parsed into defect characteristic attribute description set M, as shown in table 9;By defect
Characteristic attribute gradient document analysis is into defect characteristic property gradient coefficient sets G, as shown in table 10.
Table 9
Table 10
S34) defect sort module by each group of defect characteristic attribute vector shown in table 8 respectively with the midpoint of table 10, line, BL
The corresponding characteristic attribute gradient scope α of foreign matter defectijIt is compared so that the vectorial point of acquisition respectively of each group of sunken characteristic attribute,
The gradient factor factor vector of line, 3 kinds of defect types of BL foreign matters.For example, the defect characteristic attribute vector d1 of defect 1 respectively with table
10 midpoints, line, the corresponding characteristic attribute gradient scope α of BL foreign matter defectsijIt is compared, so that defect characteristic attribute vector
D1 obtains the gradient factor factor vector β of point defect type respectively1, line defect type gradient factor factor vector β2, BL foreign matters
The gradient factor factor vector β of defect type3。
S35) defect sort module falls into 3 groups of vectorial gradients of characteristic attribute using defect grader kernel function 2 by each group
Coefficient factor vector β1、β2、β3Convolution is carried out with the characteristic attribute description vectors of the corresponding defect type of table 9 respectively, so that, it is each
Group defect characteristic attribute vector generates 3 convolution values respectively.Wherein, in 3 convolution values of each group of defect characteristic attribute vector,
The maximum corresponding defect type of convolution value is the defect type of the defect characteristic attribute vector.
S36) defect rank determination module read defect rank decision rule file, according to the defects count identified and lack
Fall into the grade of type decision display panel.
Defects of display panel grade decision maker is classified and waited to the defect type of display panel in above-described embodiment
Level judges that embodiment 2 comprises the following steps:
S41) image signal generator sends solid-color image to display panel, and then industrial camera is carried out to the solid-color image
Collection, and be sent in image algorithm processing module.
S42) image algorithm processing module extracts defect characteristic information from the solid-color image, obtains including multigroup defect
The sunken characteristic attribute set D of type, includes multigroup sunken characteristic attribute vector, such as the institute of table 8 wherein falling into characteristic attribute set D
Show.
S43) defect sort module reads defect type configuration file from defect grader study module, and by defect class
Type configuration file is parsed into defect characteristic attribute description set M, as shown in table 9.
S44) defect sort module by each group of defect characteristic attribute vector shown in table 8 respectively with spy all in table 9
Levy attribute description vector and carry out convolution, so that, each group of defect characteristic attribute vector generates 3 convolution values respectively.Wherein, it is each
In 3 convolution values of group defect characteristic attribute vector, the maximum corresponding defect type of convolution value is the defect characteristic attribute
The defect type of vector.
S46) defect rank determination module read defect rank decision rule file, according to the defects count identified and lack
Fall into the grade of type decision display panel.
As it will be easily appreciated by one skilled in the art that the content that this specification is not described in detail belongs to this area professional technique
Prior art known to personnel, these are only presently preferred embodiments of the present invention, be not intended to limit the invention, all in this hair
Any modifications, equivalent substitutions and improvements made within bright spirit and principle etc., should be included in protection scope of the present invention
Within.
Claims (10)
1. a kind of learning method classified for defects of display panel, it is characterised in that comprise the following steps:
1) defect characteristic attribute vector d is extracted from the image of a known defect type;
2) a defect characteristic property gradient coefficient sets G is provided, it is known scarce that defect characteristic property gradient coefficient sets G includes this
Fall into the characteristic attribute gradient scope α of typeijAnd with this feature property gradient scope αijOne-to-one gradient factor factor-betaij;
Defect characteristic attribute vector d is updated to this feature property gradient scope α respectivelyijIn, obtain gradient factor factor vector β;
3) a defect characteristic attribute description set M is provided, defect characteristic attribute description set M includes many of number of drawbacks type
Individual characteristic attribute description vectors m;By gradient factor factor vector β respectively with the plurality of m volumes of characteristic attribute description vectors
Product, generates one group of convolution value;
4) if the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is the known defect type, terminate to lack
Fall into type identification flow;
If the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is not the known defect type, modification should
The characteristic attribute gradient scope α of known defect typeijAnd with this feature property gradient scope αijOne-to-one gradient factor because
Sub- βij, repeat step 1 to 4, until the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is that this is known scarce
Type is fallen into, the characteristic attribute gradient scope of the new known defect type is obtained and is corresponded with this feature property gradient scope
The gradient factor factor.
2. learning method according to claim 1, it is characterised in that convolutional calculation is carried out using below equation:
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>&beta;</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>&beta;</mi>
<mn>2</mn>
</msub>
<mo>,</mo>
<mo>...</mo>
<msub>
<mi>&beta;</mi>
<mi>n</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mi>i</mi>
<mi>n</mi>
</munderover>
<msub>
<mi>m</mi>
<mi>i</mi>
</msub>
<msub>
<mi>&beta;</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msub>
<mi>m</mi>
<mn>1</mn>
</msub>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<msub>
<mi>m</mi>
<mn>2</mn>
</msub>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mo>...</mo>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<msub>
<mi>m</mi>
<mi>n</mi>
</msub>
</mtd>
</mtr>
</mtable>
</mfenced>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msub>
<mi>&beta;</mi>
<mn>1</mn>
</msub>
</mtd>
</mtr>
<mtr>
<mtd>
<msub>
<mi>&beta;</mi>
<mn>2</mn>
</msub>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>...</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msub>
<mi>&beta;</mi>
<mi>n</mi>
</msub>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
Wherein, miIt is characteristic attribute description vectors m (m1,m2,…mn) (0 < mi< 1) in element, for description defect characteristic information
Weighted factor, characteristic attribute description vectors m is converted into matrixParticipate in convolutional calculation;βiIt is ladder
Spend coefficient factor vector β (β1,β2,…βn) (0 < βi< 1) in element, gradient factor factor vector β is converted into column vectorParticipate in convolutional calculation.
3. learning method according to claim 1 or 2, it is characterised in that defect characteristic includes picture title, and/or face
Product, and/or long, and/or wide, and/or length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate.
4. learning method according to claim 3, it is characterised in that this feature attribute description vector m be to area and/
Or long, and/or wide, and/or length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate the normalized collection of numerical value
Close.
5. learning method according to claim 1, it is characterised in that the known defect type is changed described in step 4
Characteristic attribute gradient scope is specially to reduce the characteristic attribute gradient scope of the known defect type.
6. a kind of learning method classified for defects of display panel, it is characterised in that comprise the following steps:
1) defect characteristic attribute vector d is extracted from the image of a known defect type;
2) a defect characteristic property gradient coefficient sets G is provided, it is known scarce that defect characteristic property gradient coefficient sets G includes this
Fall into the characteristic attribute gradient scope α of typeijAnd with this feature property gradient scope αijOne-to-one gradient factor factor-betaij;
Defect characteristic attribute vector d is updated to this feature property gradient scope α respectivelyijIn, obtain gradient factor factor vector β;
3) a defect characteristic attribute description set M is provided, defect characteristic attribute description set M includes many of number of drawbacks type
Individual characteristic attribute description vectors m, defect characteristic attribute description set M also characteristic attribute including the known defect type are retouched
State vectorial m1;By gradient factor factor vector β respectively with the plurality of characteristic attribute description vectors m, the known defect type
Characteristic attribute description vectors m1 carries out convolution, generates one group of convolution value;
4) if the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is the known defect type, terminate to lack
Fall into type identification flow;
If the defect type that maximum convolution is worth corresponding characteristic attribute description vectors is not the known defect type, modification should
Characteristic attribute description vectors m1, repeat step 1 to 4, until maximum convolution is worth the defect of corresponding characteristic attribute description vectors
Type is the known defect type, obtains the characteristic attribute description vectors m2 of the new known defect type.
7. learning method according to claim 6, it is characterised in that convolutional calculation is carried out using below equation:
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>&beta;</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>&beta;</mi>
<mn>2</mn>
</msub>
<mo>,</mo>
<mo>...</mo>
<msub>
<mi>&beta;</mi>
<mi>n</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mi>i</mi>
<mi>n</mi>
</munderover>
<msub>
<mi>m</mi>
<mi>i</mi>
</msub>
<msub>
<mi>&beta;</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msub>
<mi>m</mi>
<mn>1</mn>
</msub>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<msub>
<mi>m</mi>
<mn>2</mn>
</msub>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mo>...</mo>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<msub>
<mi>m</mi>
<mi>n</mi>
</msub>
</mtd>
</mtr>
</mtable>
</mfenced>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msub>
<mi>&beta;</mi>
<mn>1</mn>
</msub>
</mtd>
</mtr>
<mtr>
<mtd>
<msub>
<mi>&beta;</mi>
<mn>2</mn>
</msub>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>...</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msub>
<mi>&beta;</mi>
<mi>n</mi>
</msub>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
Wherein, miIt is characteristic attribute description vectors m (m1,m2,…mn) (0 < mi< 1) in element, for description defect characteristic information
Weighted factor, characteristic attribute description vectors m is converted into matrixParticipate in convolutional calculation;βiIt is ladder
Spend coefficient factor vector β (β1,β2,…βn) (0 < βi< 1) in element, gradient factor factor vector β is converted into column vectorParticipate in convolutional calculation.
8. the learning method according to claim 6 or 7, it is characterised in that defect characteristic includes picture title, and/or face
Product, and/or long, and/or wide, and/or length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate.
9. learning method according to claim 8, it is characterised in that this feature attribute description vector m be to area and/
Or long, and/or wide, and/or length-width ratio, and/or center gray scale, and/or contrast, and/or coordinate the normalized collection of numerical value
Close.
10. learning method according to claim 7, it is characterised in that changed described in step 4 this feature attribute description to
Amount m1 is specially to increase the characteristic attribute description of the known defect type, or redescribes the defect characteristic of the known defect type
The weighted factor of information.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710343201.5A CN107154041B (en) | 2017-05-16 | 2017-05-16 | A kind of learning method for defects of display panel classification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710343201.5A CN107154041B (en) | 2017-05-16 | 2017-05-16 | A kind of learning method for defects of display panel classification |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107154041A true CN107154041A (en) | 2017-09-12 |
CN107154041B CN107154041B (en) | 2019-08-09 |
Family
ID=59793020
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710343201.5A Active CN107154041B (en) | 2017-05-16 | 2017-05-16 | A kind of learning method for defects of display panel classification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107154041B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108109136A (en) * | 2017-12-12 | 2018-06-01 | 武汉精测电子集团股份有限公司 | Surface dirt fast filtering method and device in a kind of panel detection |
CN111311572A (en) * | 2020-02-13 | 2020-06-19 | 桂林理工大学 | Pipeline detection method and device, storage medium and robot |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101918818A (en) * | 2007-11-12 | 2010-12-15 | 麦克罗尼克激光系统公司 | Methods and apparatuses for detecting pattern errors |
US20120113293A1 (en) * | 1997-07-15 | 2012-05-10 | Silverbrook Research Pty Ltd | Image capture device with four processing units |
CN103392125A (en) * | 2011-02-24 | 2013-11-13 | 3M创新有限公司 | System for detection of non-uniformities in web-based materials |
CN104392432A (en) * | 2014-11-03 | 2015-03-04 | 深圳市华星光电技术有限公司 | Histogram of oriented gradient-based display panel defect detection method |
-
2017
- 2017-05-16 CN CN201710343201.5A patent/CN107154041B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120113293A1 (en) * | 1997-07-15 | 2012-05-10 | Silverbrook Research Pty Ltd | Image capture device with four processing units |
CN101918818A (en) * | 2007-11-12 | 2010-12-15 | 麦克罗尼克激光系统公司 | Methods and apparatuses for detecting pattern errors |
CN103392125A (en) * | 2011-02-24 | 2013-11-13 | 3M创新有限公司 | System for detection of non-uniformities in web-based materials |
CN104392432A (en) * | 2014-11-03 | 2015-03-04 | 深圳市华星光电技术有限公司 | Histogram of oriented gradient-based display panel defect detection method |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108109136A (en) * | 2017-12-12 | 2018-06-01 | 武汉精测电子集团股份有限公司 | Surface dirt fast filtering method and device in a kind of panel detection |
CN111311572A (en) * | 2020-02-13 | 2020-06-19 | 桂林理工大学 | Pipeline detection method and device, storage medium and robot |
Also Published As
Publication number | Publication date |
---|---|
CN107154041B (en) | 2019-08-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107290345A (en) | A kind of defects of display panel sorting technique and device based on AOI | |
CN106875373B (en) | Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm | |
CN103544506B (en) | A kind of image classification method and device based on convolutional neural networks | |
CN105825511B (en) | A kind of picture background clarity detection method based on deep learning | |
CN112884712B (en) | Method and related device for classifying defects of display panel | |
CN109829900A (en) | A kind of steel coil end-face defect inspection method based on deep learning | |
Pan et al. | Automatic recognition of woven fabric pattern based on image processing and BP neural network | |
CN108229580A (en) | Sugared net ranking of features device in a kind of eyeground figure based on attention mechanism and Fusion Features | |
CN106651057A (en) | Mobile terminal user age prediction method based on installation package sequence table | |
CN114092389A (en) | Glass panel surface defect detection method based on small sample learning | |
CN107154041A (en) | A kind of learning method classified for defects of display panel | |
CN106650823A (en) | Probability extreme learning machine integration-based foam nickel surface defect classification method | |
CN108805343A (en) | A kind of Scientech Service Development horizontal forecast method based on multiple linear regression | |
CN108491856A (en) | A kind of image scene classification method based on Analysis On Multi-scale Features convolutional neural networks | |
CN109145964A (en) | A kind of method and system for realizing color of image cluster | |
CN108427970A (en) | Picture mask method and device | |
CN110633739B (en) | Polarizer defect image real-time classification method based on parallel module deep learning | |
CN110110845A (en) | A kind of learning method based on parallel multi-level width neural network | |
CN103903009A (en) | Industrial product detection method based on machine vision | |
CN107330454A (en) | The non-linear visualization of magnanimity higher-dimension sequence data sort feature and quantitative analysis method | |
CN110400307A (en) | A kind of screen picture method for evaluating quality based on area differentiation | |
CN110188662A (en) | A kind of AI intelligent identification Method of water meter number | |
CN106250871A (en) | City management case classification method and device | |
CN206897873U (en) | A kind of image procossing and detecting system based on detection product performance | |
CN108961270A (en) | A kind of Bridge Crack Image Segmentation Model based on semantic segmentation |
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 | ||
CB02 | Change of applicant information | ||
CB02 | Change of applicant information |
Address after: 430070 Hubei City, Hongshan Province, South Lake Road, No. 53, Hongshan Venture Center, building on the 4 floor, No. Applicant after: Wuhan fine test electronics group Limited by Share Ltd Address before: 430070 Hubei City, Hongshan Province, South Lake Road, No. 53, Hongshan Venture Center, building on the 4 floor, No. Applicant before: Wuhan Jingce Electronic Technology Co., Ltd. |
|
GR01 | Patent grant | ||
GR01 | Patent grant |