CN110020691A - LCD screen defect inspection method based on the training of convolutional neural networks confrontation type - Google Patents
LCD screen defect inspection method based on the training of convolutional neural networks confrontation type Download PDFInfo
- Publication number
- CN110020691A CN110020691A CN201910290488.9A CN201910290488A CN110020691A CN 110020691 A CN110020691 A CN 110020691A CN 201910290488 A CN201910290488 A CN 201910290488A CN 110020691 A CN110020691 A CN 110020691A
- Authority
- CN
- China
- Prior art keywords
- defect
- sample
- normal
- training
- convolutional neural
- 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
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Quality & Reliability (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
- Testing Of Optical Devices Or Fibers (AREA)
Abstract
The present invention relates to a kind of LCD screen defect inspection methods based on the training of convolutional neural networks confrontation type, belong to screen detection technique field.Method includes the following steps: defect sample is converted to normal sample by the method for frequency domain gaussian filtering or time domain Gaussian Blur, it is proposed that frequency domain gaussian filtering is used, because this fuzzy manner is smaller to normal background feature extent of the destruction；During antagonistic training, defect sample and smoothed out normal sample are input in convolutional neural networks in pairs；Difference by the way that corresponding loss function is arranged, between e-learning to defect area and the feature difference and defect area and directly related normal background of the normal background being not directly relevant to.The method that algorithm proposed by the present invention uses antagonistic training, even if under conditions of a small amount of training sample, it can also be with the difference of Fast Learning to defect and background.
Description
Technical field
The invention belongs to screen detection technique field, it is related to lacking based on the LCD screen of convolutional neural networks confrontation type training
Fall into detection method.
Background technique
1, the art state of development
The production link of LCD screen, including cutting, assembling and welding etc. are realized substantially and are automated, but are examined in quality
It surveys link and still uses artificial detection.The disadvantage of artificial detection maximum is higher cost, so many Mo Zu producers are accumulating
Find relevant AOI equipment in pole.
The reason of restricting this AOI equipment volume production is that the first-pass yield of algorithm detection is low, and algorithm debugging difficulty is big, training
Time length etc..Many vision enterprises just actively research and develop this equipment now, but how to drop under the premise of guaranteeing high-accuracy
Low equipment cost, especially maintenance cost are still a problems.
2, prior art scheme
1) it is based on Digital Image Processing scheme
The program mainly by defect enhancing or the method for Background Reconstruction, is then detected by given threshold again.Wherein,
Enhancement Method mostly uses frequency domain to enhance, and Background Rebuilding Method uses the methods of PCA, ICA more.
2) the secondary judgment method based on pattern-recognition or neural network
This method passes through Digital Image Processing first and finds possible defect area, extracts region modeling, does and accurately sentence
It is disconnected.Wherein, the algorithm of modeling is mostly EM, SVM BP network.
3) method based on deep learning,
This method is a large amount of training sample of acquisition mostly, by CNN perhaps DBN or combination RNN training.
3, the deficiency of existing scheme
1) algorithm complexity is high, and detection speed is slow.
2) more training sample is needed.
3) bad to very unconspicuous Mura defects detection effect.
4) Mura defects detection effect in edge is bad.
5) detection hyper parameter is more, and threshold parameter is difficult to unification for different types of template.
Summary of the invention
In view of this, the purpose of the present invention is to provide a kind of LCD screens based on the training of convolutional neural networks confrontation type
Defect inspection method leads to too small amount of training sample, trains one to detect multiple types defect (Mura simultaneously in a short time
Defect, point defect, line defect etc.) convolutional neural networks model.
In order to achieve the above objectives, the invention provides the following technical scheme:
Based on the LCD screen defect inspection method of convolutional neural networks confrontation type training, method includes the following steps:
Flawless sample graph A is acquired in one piece of normal mould group, then make defect in this mould group and is acquired
To containing defective sample B；Purposefully learn the region P of defect in B figureBWith in A figure with PBThe region P of identical positionA
Difference；
Defect sample is converted into normal sample by the method for frequency domain gaussian filtering or time domain Gaussian Blur, this mould
Paste mode is smaller to normal background feature extent of the destruction；
During antagonistic training, defect sample and smoothed out normal sample are input to convolutional neural networks in pairs
In；
By the way that corresponding loss function, the feature of e-learning to defect area and the normal background being not directly relevant to is arranged
Difference between difference and defect area and directly related normal background；The loss function of network is as follows:
Wherein, SamplengWith SampleokRespectively indicate the normal sample of defect sample and conversion；It indicates to lack
Fall into the confidence level loss of the corresponding characteristic pattern of sample, including defect area and normal region；Indicate defect sample
Coordinate in characteristic pattern returns loss；Indicate normal sample characteristic pattern in defect sample characteristic pattern corresponding region
Confidence level loss, it is only necessary to calculate normal region；
Further, the core network of the method is based on SSD sorter network framework, it is contemplated that the shape of defect and area etc.
Factor, the number of plies and convolution mode of core network have certain variation, and model structure is as follows:
The model structure has 36448 detection blocks, is deleted according to the actual situation.
The beneficial effects of the present invention are:
1, using convolutional neural networks, characteristics of image is had no need to change, solves and needs early period to establish background model to disappear
Except the non-uniform problem of brightness of image.Meanwhile picture edge characteristic will not be destroyed in preprocessing process, it can solve back
Edge defect is easy the problem of missing inspection in scape algorithm.
2, convolutional neural networks can detect multiple types defect simultaneously with a model, not need most for different defects
Type designs different algorithms, realizes the general of algorithm, thereby reduces the quantity of hyper parameter.
3, for the mode of Enterprise Flexibility production, enterprise needs can be with the new model of Quick thread.This does not require nothing more than new mould
The type training time is short, and training samples number will be lacked.The method that algorithm proposed by the present invention uses antagonistic training, even if
It, can also be with the difference of Fast Learning to defect and background under conditions of a small amount of training sample.
Other advantages, target and feature of the invention will be illustrated in the following description to a certain extent, and
And to a certain extent, based on will be apparent to those skilled in the art to investigating hereafter, Huo Zheke
To be instructed from the practice of the present invention.Target and other advantages of the invention can be realized by following specification
And acquisition.
Detailed description of the invention
To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention is made below in conjunction with attached drawing excellent
The detailed description of choosing, in which:
Fig. 1 is trained and overhaul flow chart；
Fig. 2 is antagonistic training process；
Fig. 3 is detection core network；
Fig. 4 is each module and alternative module in core network；Fig. 4 (a) is Conv module, and figure (b) is Atrous module,
Fig. 4 (c) is ResAtrous module；
Fig. 5 is accuracy rate and sample size relational graph.
Specific embodiment
Illustrate embodiments of the present invention below by way of specific specific example, those skilled in the art can be by this specification
Other advantages and efficacy of the present invention can be easily understood for disclosed content.The present invention can also be by addition different specific
Embodiment is embodied or practiced, and the various details in this specification can also not carried on the back based on different viewpoints and application
From carrying out various modifications or alterations under spirit of the invention.It should be noted that diagram provided in following embodiment only with
Illustration illustrates basic conception of the invention, and in the absence of conflict, feature in following embodiment and embodiment can be with
It is combined with each other.
Wherein, the drawings are for illustrative purposes only and are merely schematic diagrams, rather than pictorial diagram, should not be understood as to this
The limitation of invention；Embodiment in order to better illustrate the present invention, the certain components of attached drawing have omission, zoom in or out, not
Represent the size of actual product；It will be understood by those skilled in the art that certain known features and its explanation may be omitted and be in attached drawing
It is understood that.
The same or similar label correspond to the same or similar components in the attached drawing of the embodiment of the present invention；It is retouched in of the invention
In stating, it is to be understood that if thering is the orientation of the instructions such as term " on ", "lower", "left", "right", "front", "rear" or position to close
System is merely for convenience of description of the present invention and simplification of the description to be based on the orientation or positional relationship shown in the drawings, rather than indicates
Or imply that signified device or element must have a particular orientation, be constructed and operated in a specific orientation, therefore retouch in attached drawing
The term for stating positional relationship only for illustration, is not considered as limiting the invention, for the common skill of this field
For art personnel, the concrete meaning of above-mentioned term can be understood as the case may be.
As shown in Figure 1, the comparative test in similar biology: under an identical environment, only change wherein some
Factor, observes influence degree of the variation to result of this factor, to intuitively show the importance of this factor.In view of
This, distinguishes that defect and normal background are most simple and most efficient method is: acquiring zero defect in one piece of normal mould group
Sample graph A, then make and defect and collect containing defective sample B in this mould group.Purposefully learn in B figure
The region P of defectBWith in A figure with PBThe region P of identical positionADifference.
But the mode of real kind of this devastatingly collecting sample is infeasible.But if in turn: due to
The defects of similar Mura, point defect, line defect, block can obscure if its area is smaller until elimination.It can
Defect sample is converted into normal sample by the method for frequency domain gaussian filtering or time domain Gaussian Blur, this fuzzy manner
It is smaller to normal background feature extent of the destruction.This algorithm is recommended to use frequency domain gaussian filtering to eliminate defect (if defect area
It is larger, can be handled with the defects of different aobvious or brightness irregularities).
As shown in Fig. 2, defect sample and smoothed out normal sample are input to volume in pairs during antagonistic training
In product neural network.By being arranged corresponding loss function, network not only may learn defect area and be not directly relevant to
The feature difference of normal background also can directly learn to the difference between defect area and directly related normal background.Network
Loss function it is as follows:
Wherein, SamplengWith SampleokRespectively indicate the normal sample of defect sample and conversion.It indicates to lack
Fall into the confidence level loss of the corresponding characteristic pattern of sample, including defect area and normal region.Indicate defect sample
Coordinate in characteristic pattern returns loss.Indicate normal sample characteristic pattern in defect sample characteristic pattern corresponding region
Confidence level loss, it is only necessary to calculate normal region.
The core network of the algorithm is based on SSD sorter network framework, it is contemplated that the factors such as the shape of defect and area, trunk
The number of plies and convolution mode of network have certain variation, and model structure is as shown in Figure 3.
In Fig. 4, figure (a) and figure (b) module correspond to Fig. 3, and if necessary to deepen network, residual error module map 4 can be used
(c).The model has 36448 detection blocks, can be deleted according to the actual situation.
The training sample of the algorithm more than totally 650,3 liquid crystal module manufacturing enterprises are derived from, are 7 cun of TFT-LCD respectively
More than 300, mould group sample, 5.5 cun of OLED mould groups 150 are opened, 4.5 cun of TFT-LCD mould groups more than 200.It detects accuracy and sample
The relational graph of this quantity is as figure 5 illustrates.
It can be seen that detection can be obviously improved using the method for antagonistic training under the premise of sample size is less
Accuracy rate.
The model has following characteristics:
1. training samples number needed for is less, under 650 training samples, model omission factor < 0.3%, false detection rate <
1.2%.Meet enterprise to accuracy rate requirement.
2. model volume is in 4M or so, detection time < 1.8s/p (Intel E3-12263.31Ghz)
3. edge defect can be detected preferably and than shallower Mura class defect.
Finally, it is stated that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting, although referring to compared with
Good embodiment describes the invention in detail, those skilled in the art should understand that, it can be to skill of the invention
Art scheme is modified or replaced equivalently, and without departing from the objective and range of the technical program, should all be covered in the present invention
Scope of the claims in.
Claims (2)
1. the LCD screen defect inspection method based on the training of convolutional neural networks confrontation type, it is characterised in that: this method includes
Following steps:
Flawless sample graph A is acquired in one piece of normal mould group, then defect and collecting is made in this mould group and contains
Defective sample B；Purposefully learn the region P of defect in B figureBWith in A figure with PBThe region P of identical positionADifference
It is different；
Defect sample is converted into normal sample by the method for frequency domain gaussian filtering or time domain Gaussian Blur, this fuzzy side
Formula is smaller to normal background feature extent of the destruction；
During antagonistic training, defect sample and smoothed out normal sample are input in convolutional neural networks in pairs；
By the way that corresponding loss function is arranged, e-learning to defect area and the feature for the normal background being not directly relevant to are poor
Difference between different and defect area and directly related normal background；The loss function of network is as follows:
Wherein, SamplengWith SampleokRespectively indicate the normal sample of defect sample and conversion；Indicate defect sample
The confidence level loss of this corresponding characteristic pattern, including defect area and normal region；Indicate defect sample characteristic pattern
In coordinate return loss；Indicate the confidence level in normal sample characteristic pattern with defect sample characteristic pattern corresponding region
Loss, it is only necessary to calculate normal region.
2. the LCD screen defect inspection method according to claim 1 based on the training of convolutional neural networks confrontation type,
Be characterized in that: the core network of the method is based on SSD sorter network framework, it is contemplated that the factors such as the shape of defect and area,
The number of plies and convolution mode of core network have certain variation, and model structure is as follows:
The model structure has 36448 detection blocks, is deleted according to the actual situation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910290488.9A CN110020691B (en) | 2019-04-11 | 2019-04-11 | Liquid crystal screen defect detection method based on convolutional neural network impedance type training |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910290488.9A CN110020691B (en) | 2019-04-11 | 2019-04-11 | Liquid crystal screen defect detection method based on convolutional neural network impedance type training |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110020691A true CN110020691A (en) | 2019-07-16 |
| CN110020691B CN110020691B (en) | 2022-10-11 |
Family
ID=67191185
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910290488.9A Active CN110020691B (en) | 2019-04-11 | 2019-04-11 | Liquid crystal screen defect detection method based on convolutional neural network impedance type training |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110020691B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110717378A (en) * | 2019-08-26 | 2020-01-21 | 苏州感知线智能科技有限公司 | Method and device for detecting conductive particles based on neural network algorithm |
| CN111476759A (en) * | 2020-03-13 | 2020-07-31 | 深圳市鑫信腾机器人科技有限公司 | Screen surface detection method and device, terminal and storage medium |
| CN111524120A (en) * | 2020-04-22 | 2020-08-11 | 征图新视(江苏)科技股份有限公司 | Printed matter tiny scratch and overprint deviation detection method based on deep learning |
| CN113255590A (en) * | 2021-06-25 | 2021-08-13 | 众芯汉创（北京）科技有限公司 | Defect detection model training method, defect detection method, device and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016201947A1 (en) * | 2015-06-16 | 2016-12-22 | 华南理工大学 | Method for automated detection of defects in cast wheel products |
| CN109559298A (en) * | 2018-11-14 | 2019-04-02 | 电子科技大学中山学院 | Emulsion pump defect detection method based on deep learning |
-
2019
- 2019-04-11 CN CN201910290488.9A patent/CN110020691B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016201947A1 (en) * | 2015-06-16 | 2016-12-22 | 华南理工大学 | Method for automated detection of defects in cast wheel products |
| CN109559298A (en) * | 2018-11-14 | 2019-04-02 | 电子科技大学中山学院 | Emulsion pump defect detection method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 曾毅等: "基于BP神经网络和均值差分的TFT-LCD MURA缺陷检测方法", 《湖南理工学院学报(自然科学版)》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110717378A (en) * | 2019-08-26 | 2020-01-21 | 苏州感知线智能科技有限公司 | Method and device for detecting conductive particles based on neural network algorithm |
| CN111476759A (en) * | 2020-03-13 | 2020-07-31 | 深圳市鑫信腾机器人科技有限公司 | Screen surface detection method and device, terminal and storage medium |
| CN111476759B (en) * | 2020-03-13 | 2022-03-25 | 深圳市鑫信腾机器人科技有限公司 | Screen surface detection method and device, terminal and storage medium |
| CN111524120A (en) * | 2020-04-22 | 2020-08-11 | 征图新视(江苏)科技股份有限公司 | Printed matter tiny scratch and overprint deviation detection method based on deep learning |
| CN113255590A (en) * | 2021-06-25 | 2021-08-13 | 众芯汉创（北京）科技有限公司 | Defect detection model training method, defect detection method, device and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110020691B (en) | 2022-10-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110020691A (en) | LCD screen defect inspection method based on the training of convolutional neural networks confrontation type | |
| CN109239102A (en) | A kind of flexible circuit board open defect detection method based on CNN | |
| CN111402226A (en) | Surface defect detection method based on cascade convolution neural network | |
| CN111951249A (en) | Mobile phone light guide plate defect visual detection method based on multitask learning network | |
| CN105388162B (en) | Raw material silicon chip surface scratch detection method based on machine vision | |
| CN104766097B (en) | Surface of aluminum plate defect classification method based on BP neural network and SVMs | |
| CN106875373A (en) | Mobile phone screen MURA defect inspection methods based on convolutional neural networks pruning algorithms | |
| CN107657603A (en) | A kind of industrial appearance detecting method based on intelligent vision | |
| CN101063660B (en) | Method for detecting textile defect and device thereof | |
| CN114638797A (en) | Method and device for detecting copper surface defects based on linear array camera | |
| CN108802041B (en) | Method for rapidly changing small sample set of screen detection | |
| CN108416774A (en) | A kind of fabric types recognition methods based on fine granularity neural network | |
| CN110186375A (en) | Intelligent high-speed rail white body assemble welding feature detection device and detection method | |
| CN110647875A (en) | Method for segmenting and identifying model structure of blood cells and blood cell identification method | |
| CN113962929B (en) | Photovoltaic cell assembly defect detection method and system and photovoltaic cell assembly production line | |
| CN108764134A (en) | A kind of automatic positioning of polymorphic type instrument and recognition methods suitable for crusing robot | |
| CN110555831B (en) | Deep learning-based drainage pipeline defect segmentation method | |
| CN111476307B (en) | Lithium battery surface defect detection method based on depth field adaptation | |
| CN113222913B (en) | Circuit board defect detection positioning method, device and storage medium | |
| CN106934801A (en) | A kind of fluorescentmagnetic particle(powder) defect inspection method based on Laws texture filterings | |
| CN104252056A (en) | Detection method and device of substrate | |
| CN113327243B (en) | PAD light guide plate defect visual detection method based on Ayolov3-Tiny new framework | |
| CN116228651A (en) | Cloth defect detection method, system, equipment and medium | |
| CN112819844B (en) | Image edge detection method and device | |
| CN112964732A (en) | Spinning cake defect visual detection system and method based on deep learning |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |