CN109583496A - A kind of network model and method for the classification of display panel large area defect - Google Patents
A kind of network model and method for the classification of display panel large area defect Download PDFInfo
- Publication number
- CN109583496A CN109583496A CN201811437120.2A CN201811437120A CN109583496A CN 109583496 A CN109583496 A CN 109583496A CN 201811437120 A CN201811437120 A CN 201811437120A CN 109583496 A CN109583496 A CN 109583496A
- Authority
- CN
- China
- Prior art keywords
- layer
- classification
- large area
- network model
- defect
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Biodiversity & Conservation Biology (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Multimedia (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The invention belongs to display panel test technique automatic fields, disclose a kind of network model and method for the classification of display panel large area defect, and network model includes one layer of input layer, seven layers of convolutional layer, two layers of full articulamentum and one layer of output layer；Input layer uses big resolution ratio, and first in seven layers of convolutional layer to layer 5 uses large-sized convolution kernel；Based on the large area defect classification method that the network model provides, defect image is normalized to the sample with network model input layer resolution match by (1)；(2) sample is divided into training set, verifying collection, test set, classified to the sample in set according to defect type；(3) network model is trained with training set, and is verified to obtain disaggregated model with verifying collection；(4) classified using disaggregated model to the sample in test set；It solves the problems, such as existing method not and can be carried out the classification of large area defect image, greatly improve the accuracy rate of large area defect image Classification and Identification.
Description
Technical field
The invention belongs to display panel test technique automatic fields, and in particular to one kind is lacked for display panel large area
Fall into the network model and method of classification.
Background technique
During display panel production, AOI defects detection directly affects the final quality of panel and Yield Grade
Division.Mura defects in panel are often a kind of large area Mura defects macroscopically, and are had in irregular shape, big
Small feature uneven, contrast is low.
The method for carrying out defects of display panel classification by human eye, manual sort one by one, by human eye sensing capability and
The limitation of subjective factor is difficult that fast and accurately the large area Mura defects on display panel are accurately identified and divided
Class, and inefficiency；With the increase of human eye degree of fatigue, the mistake that will also result in minor defect is divided.
The prior art further includes for example being shown based on the convolutional neural networks in deep learning by image processing algorithm
Show the method for panel defect classification, but on the one hand, existing deep learning is primarily directed to the classification of natural image, Small object
Identification, such as the image resolution ratio of the categories of model network input layer such as AlexNet, VGG, GoogLeNet, ResNet are special
It is small, be not suitable for the Classification and Identification of big image in different resolution large area defect, and the convolution kernel of existing sorter network be all 3*3 or
The convolution kernel of 5*5, lesser convolution kernel can extract the grain details of object, therefore relatively abundanter for natural image, texture
Image classification recognition effect it is preferable, but this kind of convolution kernel be not appropriate for extracting big image in different resolution, global nature, it is macro
Large area defect in sight.
Summary of the invention
Aiming at the above defects or improvement requirements of the prior art, the present invention provides one kind lacks for display panel large area
The network model for falling into classification is used for using empty convolution building network model to global nature, big face on big image in different resolution
Product defect is accurately identified.
To achieve the above object, according to one aspect of the present invention, it provides a kind of for display panel large area defect
The network model of classification, including one layer of input layer, seven layers of convolutional layer, two layers of full articulamentum and one layer of output layer；Input layer is adopted
With big resolution ratio, first layer, the second layer in seven layers of convolutional layer use large-sized convolution kernel；Big resolution ratio refers to not less than 1024
1024 pixel of pixel *, large scale refer to not less than 9 pixel *, 9 pixel.
Preferably, the above-mentioned network model for the classification of display panel large area defect, in seven layers of convolutional layer first to
Layer 5 is using empty convolution kernel.
Preferably, the above-mentioned network model for the classification of display panel large area defect, the first convolutional layer and the second volume layer
Convolution kernel use the empty convolution kernel having a size of 15 pixel *, 15 pixel size；Third convolutional layer, Volume Four lamination, volume five
The convolution kernel of lamination uses the empty convolution kernel having a size of 7 pixel *, 7 pixel.
Preferably, the above-mentioned network model for the classification of display panel large area defect, the 6th convolutional layer, the 7th convolutional layer
Convolution kernel be 3 pixel of 5 pixel *, 5 pixel or 3 pixel *.
Preferably, the above-mentioned network model for the classification of display panel large area defect, the resolution ratio of input layer are 3701
3701 pixel of pixel *.
Preferably, the above-mentioned network model for the classification of display panel large area defect, first in two layers of full articulamentum
The node number of layer is that the interstitial content of the 1024, second layer is adjusted according to the class number of classification.
To achieve the above object, other side according to the invention provides a kind of display based on above-mentioned network model
Panel large area defect classification method, includes the following steps:
(1) defect image of collection is normalized to the image with network model input layer resolution match as sample；
(2) by sample be divided into training set, verifying collection, test set, to these three set in sample according to defect type into
Row classification, is arranged class label；
(3) network model is trained using training set, and is verified using verifying collection, obtain disaggregated model；
(4) classified using disaggregated model to the sample in test set.
Preferably, above-mentioned display panel large area defect classification method, defect type include gridiron pattern Mura defects, grayscale
Grid Mura defects, horizontal leukorrhea Mura defects and vertical white band Mura defects.
Preferably, above-mentioned display panel large area defect classification method, using ID number as classification mark corresponding with defect
Label, each ID number represent a kind of defect classification；Obtained training set includes a sets of image data and a label text
File；What is saved in sets of image data is in trained defect image sample, label file comprising each image
Specific store path and defect category IDs.
Preferably, picture is normalized to 3701 pixel *, 3701 picture by above-mentioned display panel large area defect classification method
Element.
Above-mentioned network model provided by the invention and the method classified for display panel large area defect, pass through expansion
The resolution ratio of input layer solves the problems, such as that picture of large image scale inputs, and solves mainstream sorter network model using large-sized convolution kernel
Algorithm experiences the lesser problem in the visual field, solves expanding input layer resolution ratio, using large-sized convolution kernel using empty convolution
Bring parameter increases additive model and is difficult to trained problem.In general, through the invention contemplated above technical scheme with
The prior art is compared, and can achieve the following beneficial effects:
(1) using the network model of display panel large area defect provided by the invention classification, by deep learning (Deep
Learning, DL) domain classification network model input layer input image resolution be promoted to 3701 pixel *, 3701 pixel this
The super large resolution ratio of sample ensures that large area Mura defects in the LCD image of 6480*3840 resolution ratio big in this way or not
The gross imperfection only showed on big image in different resolution can be lost because compression of images degree is too big；
(2) existing sorter network model is substantially the classification for animal or object progress in natural image；
It is not directed to the Classification and Identification model based on deep learning of this specific scene of panel detection, display surface provided by the invention
The network model of plate large area defect classification has used point in the first convolutional layer and the second convolutional layer for the first time for panel detection
Resolution is large size convolution kernel as 15 pixel *, 15 pixel；Achieve the purpose that extract global nature, defect macroscopically；It can be right
The defect image that large area, global nature, mainstream sorter network can not classify is classified；
(3) network model of display panel large area defect classification provided by the invention, introduces empty Convolution Properties
Convolution kernel, the weighting parameter in the convolution kernel of empty convolution are that interlacing is even arranged every several rows, thus reduce actual needs
Trained weighting parameter reduces the difficulty of network model training while obtaining gross imperfection；
(4) mura defects detection is carried out using the network model of display panel large area defect provided by the invention classification,
It does not need to increase any hardware there is no need to modify to current AOI structure in algorithm core software view due to improving
Cost has the characteristics that easy to accomplish, at low cost, practicability is high.
Detailed description of the invention
Fig. 1 is the first convolutional layer and the in the network model for the classification of display panel large area defect that embodiment provides
The empty convolution kernel schematic diagram that two convolutional layers use；
Fig. 2 is third and fourth in the network model for the classification of display panel large area defect for apply example offer, five convolutional layers
The empty convolution kernel schematic diagram used；
Fig. 3 is the flow diagram for the display panel large area defect classification method that embodiment provides.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
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 the various embodiments of the present invention described below
Not constituting a conflict with each other can be combined with each other.
The network model for the display panel large area defect classification that embodiment provides, including 1 layer of input layer, 7 layers of convolutional layer,
Two layers of full articulamentum and output layer；
Wherein, the convolution kernel of the first convolutional layer C1 and the second volume layer C2 use size as schematically shown in Figure 1 for 15 pixel * 15
The convolution kernel of the empty convolution kernel of pixel size, C3, C4, C5 convolutional layer uses the sky shown in Fig. 2 having a size of 7 pixel *, 7 pixel
Hole convolution kernel.
In embodiment, the method that the classification of display panel large area defect is carried out using the network model that embodiment provides,
Process specifically comprises the following steps: referring to Fig. 3
(1) by the image that all defect image normalization being collected into is with network model input layer resolution match, make
For sample；It is that 6483 pixel *, 3840 pixel is normalized to 3701 pictures by the Mura defects sample acquired in producing line in embodiment
The image of plain * 3701 pixel resolution；This resolution ratio is corresponding with the last full interstitial content of articulamentum；
(2) sample is divided into training set, verifying collection, test set；Category label is carried out to the sample in each set；One
It is to use to be randomly assigned for sample to be divided into training set, verifying collection, test set in a embodiment；
(3) classify respectively according to defect type to training set, verifying collection, test set, class label is set；
In some embodiments, defect classification includes gridiron pattern Mura defects, grayscale grid Mura defects, horizontal leukorrhea
Mura defects, vertical white band Mura defects, class label corresponding with these four types of defects are respectively 0,1,2,3, each number
As soon as being exactly an ID, each ID represents a kind of defect classification；Obtained training set includes a sets of image data and one
Label text file；What is saved in sets of image data is in trained defect image sample, label file comprising each
Open the specific store path and defect category IDs of image；
(4) the above-mentioned network model that embodiment provides is trained using training set, and is verified using verifying collection,
Obtain disaggregated model；
(5) classified using disaggregated model to the sample in test set.
The method for this display panel large area defect classification that embodiment provides uses large-sized convolution based on above-mentioned
Core, the network model network model for the classification of display panel large area defect for introducing empty convolution kernel building, due to expanding
The input image resolution of input layer, therefore can guarantee the large area Mura defects in the display panel image of big resolution ratio
The gross imperfection only showed on big image in different resolution will not be lost because compression of images degree is too big.
Due to having used large-sized convolution kernel in the first convolutional layer and the second convolutional layer, network model can be greatly increased
Receptive field range, such network model can remove the gross imperfection in observation piece image, can solve similar gridiron pattern
This kind of classification and identification with macroscopic view, global nature Mura defects.
On the other hand, in order to avoid directly causing network model parameter amount sharply to increase using large-sized common convolution kernel
Add, and then model is caused to be difficult to train, present invention introduces the buildings of empty convolution kernel for the classification of display panel large area defect
Network model can reduce the training difficulty and instruction of model while increasing network model receptive field range to the greatest extent
Practice the number of parameters of training, efficiently solving traditional images Processing Algorithm and mainstream sorter network model cannot effectively carry out greatly
The problem of Area defect image classification, while cost of labor is reduced, greatly improve large area defect image Classification and Identification
Accuracy rate.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (10)
1. a kind of network model for the classification of display panel large area defect, which is characterized in that including one layer of input layer, seven layers
Convolutional layer, two layers of full articulamentum and one layer of output layer；Input layer uses big resolution ratio, first layer in seven layers of convolutional layer, the
Two layers use large-sized convolution kernel.
2. the network model for the classification of display panel large area defect as described in claim 1, which is characterized in that seven layers of volume
In lamination first to layer 5 using empty convolution kernel.
3. the network model for the classification of display panel large area defect as claimed in claim 2, which is characterized in that the first volume
Lamination and the convolution kernel of the second volume layer are used having a size of 15 pixel *, 15 pixel cavity convolution kernel；Third convolutional layer, Volume Four product
Layer, the 5th convolutional layer convolution kernel use the empty convolution kernel having a size of 7 pixel *, 7 pixel.
4. the network model for the classification of display panel large area defect as claimed in claim 1 or 2, which is characterized in that the
Six convolutional layers, the 7th convolutional layer convolution kernel be 3 pixel of 5 pixel *, 5 pixel or 3 pixel *.
5. the network model for the classification of display panel large area defect as claimed in claim 1 or 2, which is characterized in that defeated
The resolution ratio for entering layer is 3701 pixel *, 3701 pixel.
6. the network model for the classification of display panel large area defect as claimed in claim 1 or 2, which is characterized in that two
The node number of first layer in the full articulamentum of layer is that the interstitial content of the 1024, second layer is adjusted according to the class number of classification.
7. a kind of display panel large area defect classification method based on the described in any item network models of claim 1~6,
It is characterized in that, includes the following steps:
(1) defect image of collection is normalized to the image with network model input layer resolution match as sample；
(2) sample is divided into training set, verifying collection, test set, the sample in these three set is divided according to defect type
Class label is arranged in class；
(3) network model is trained using training set, and is verified using verifying collection, obtain disaggregated model；
(4) classified using disaggregated model to the sample in test set.
8. display panel large area defect classification method as claimed in claim 7, which is characterized in that defect type includes chessboard
Lattice Mura defects, grayscale grid Mura defects, horizontal leukorrhea Mura defects and vertical white band Mura defects.
9. display panel large area defect classification method as claimed in claim 7 or 8, which is characterized in that use ID number conduct
Class label corresponding with defect, each ID number represent a kind of defect classification；Obtained training set includes an image data
Set and a label text file；What is saved in sets of image data is for trained defect image sample, label file
In include each image specific store path and defect category IDs.
10. display panel large area defect classification method as claimed in claim 7 or 8, which is characterized in that normalize picture
For 3701 pixel *, 3701 pixel.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811437120.2A CN109583496A (en) | 2018-11-28 | 2018-11-28 | A kind of network model and method for the classification of display panel large area defect |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811437120.2A CN109583496A (en) | 2018-11-28 | 2018-11-28 | A kind of network model and method for the classification of display panel large area defect |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109583496A true CN109583496A (en) | 2019-04-05 |
Family
ID=65925402
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811437120.2A Pending CN109583496A (en) | 2018-11-28 | 2018-11-28 | A kind of network model and method for the classification of display panel large area defect |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109583496A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110567967A (en) * | 2019-08-20 | 2019-12-13 | 武汉精立电子技术有限公司 | Display panel detection method, system, terminal device and computer readable medium |
| CN111753916A (en) * | 2020-06-30 | 2020-10-09 | 苏州慧维智能医疗科技有限公司 | Automatic marking method and system for small lesions in high-definition medical images |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160378912A1 (en) * | 2013-07-02 | 2016-12-29 | Epigenetx, Llc | Structure-based modeling and target-selectivity prediction |
| CN107561738A (en) * | 2017-08-30 | 2018-01-09 | 湖南理工学院 | TFT LCD surface defect quick determination methods based on FCN |
| CN107665492A (en) * | 2017-06-29 | 2018-02-06 | 南京信息工程大学 | Colon and rectum panorama numeral pathological image tissue segmentation methods based on depth network |
| CN108376392A (en) * | 2018-01-30 | 2018-08-07 | 复旦大学 | A kind of image motion ambiguity removal method based on convolutional neural networks |
| CN108427754A (en) * | 2018-03-15 | 2018-08-21 | 京东方科技集团股份有限公司 | A kind of information-pushing method, computer storage media and terminal |
| CN108492272A (en) * | 2018-03-26 | 2018-09-04 | 西安交通大学 | Cardiovascular vulnerable plaque recognition methods based on attention model and multitask neural network and system |
| CN108573683A (en) * | 2018-04-16 | 2018-09-25 | 武汉精测电子集团股份有限公司 | A kind of method and system promoting display panel quality |
| CN108596226A (en) * | 2018-04-12 | 2018-09-28 | 武汉精测电子集团股份有限公司 | A kind of defects of display panel training method and system based on deep learning |
-
2018
- 2018-11-28 CN CN201811437120.2A patent/CN109583496A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160378912A1 (en) * | 2013-07-02 | 2016-12-29 | Epigenetx, Llc | Structure-based modeling and target-selectivity prediction |
| CN107665492A (en) * | 2017-06-29 | 2018-02-06 | 南京信息工程大学 | Colon and rectum panorama numeral pathological image tissue segmentation methods based on depth network |
| CN107561738A (en) * | 2017-08-30 | 2018-01-09 | 湖南理工学院 | TFT LCD surface defect quick determination methods based on FCN |
| CN108376392A (en) * | 2018-01-30 | 2018-08-07 | 复旦大学 | A kind of image motion ambiguity removal method based on convolutional neural networks |
| CN108427754A (en) * | 2018-03-15 | 2018-08-21 | 京东方科技集团股份有限公司 | A kind of information-pushing method, computer storage media and terminal |
| CN108492272A (en) * | 2018-03-26 | 2018-09-04 | 西安交通大学 | Cardiovascular vulnerable plaque recognition methods based on attention model and multitask neural network and system |
| CN108596226A (en) * | 2018-04-12 | 2018-09-28 | 武汉精测电子集团股份有限公司 | A kind of defects of display panel training method and system based on deep learning |
| CN108573683A (en) * | 2018-04-16 | 2018-09-25 | 武汉精测电子集团股份有限公司 | A kind of method and system promoting display panel quality |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110567967A (en) * | 2019-08-20 | 2019-12-13 | 武汉精立电子技术有限公司 | Display panel detection method, system, terminal device and computer readable medium |
| CN110567967B (en) * | 2019-08-20 | 2022-06-17 | 武汉精立电子技术有限公司 | Display panel detection method, system, terminal device and computer readable medium |
| CN111753916A (en) * | 2020-06-30 | 2020-10-09 | 苏州慧维智能医疗科技有限公司 | Automatic marking method and system for small lesions in high-definition medical images |
| CN111753916B (en) * | 2020-06-30 | 2021-06-22 | 苏州慧维智能医疗科技有限公司 | Automatic marking method and system for small lesions in high-definition medical images |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107437092B (en) | The classification method of retina OCT image based on Three dimensional convolution neural network | |
| Huang et al. | Research on classification method of maize seed defect based on machine vision | |
| CN105046277B (en) | Robust mechanism study method of the feature significance in image quality evaluation | |
| CN109711474A (en) | A kind of aluminium material surface defects detection algorithm based on deep learning | |
| CN102013021B (en) | Tea tender shoot segmentation and identification method based on color and region growth | |
| CN110263705A (en) | Towards two phase of remote sensing technology field high-resolution remote sensing image change detecting method | |
| CN105740773A (en) | Deep learning and multi-scale information based behavior identification method | |
| CN107844797A (en) | A kind of method of the milking sow posture automatic identification based on depth image | |
| CN108009591A (en) | A kind of contact network key component identification method based on deep learning | |
| CN108416353A (en) | Crop field spike of rice fast partition method based on the full convolutional neural networks of depth | |
| CN107369155A (en) | A kind of cloth surface defect detection method and its system based on machine vision | |
| CN106447646A (en) | Quality blind evaluation method for unmanned aerial vehicle image | |
| CN106658169A (en) | Universal method for segmenting video news in multi-layered manner based on deep learning | |
| CN104198498B (en) | Fabric Defects Inspection detection method based on adaptive orthogonal wavelet conversion and device | |
| CN109816625A (en) | A kind of video quality score implementation method | |
| CN110516716A (en) | Non-reference picture quality appraisement method based on multiple-limb similarity network | |
| CN111652326A (en) | Improved fruit maturity identification method and identification system based on MobileNet v2 network | |
| CN109508741A (en) | Method based on deep learning screening training set | |
| CN104077612B (en) | A kind of insect image-recognizing method based on multiple features rarefaction representation technology | |
| CN104036493B (en) | No-reference image quality evaluation method based on multifractal spectrum | |
| CN110309808A (en) | A kind of adaptive smog root node detection method under a wide range of scale space | |
| CN110428412A (en) | The evaluation of picture quality and model generating method, device, equipment and storage medium | |
| CN106650823A (en) | Probability extreme learning machine integration-based foam nickel surface defect classification method | |
| CN109191460A (en) | A kind of quality evaluating method for tone mapping image | |
| CN107909588A (en) | Partition system under MRI cortex based on three-dimensional full convolutional neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190405 |