CN107561738A - TFT LCD surface defect quick determination methods based on FCN - Google Patents
TFT LCD surface defect quick determination methods based on FCN Download PDFInfo
- Publication number
- CN107561738A CN107561738A CN201710764659.8A CN201710764659A CN107561738A CN 107561738 A CN107561738 A CN 107561738A CN 201710764659 A CN201710764659 A CN 201710764659A CN 107561738 A CN107561738 A CN 107561738A
- Authority
- CN
- China
- Prior art keywords
- image
- sample
- fcn
- convolution
- pixel
- 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
Abstract
Edge blurry, contrast in TFT LCD surface defects detections be low, the interference for repeating the noises such as grain background be present in image in order to overcome, and the present invention proposes a kind of quick determination method end to end based on full convolutional neural networks.This method first passes through multiple sample trainings and obtains the detection model of FCN_21 × 21, image to be detected will be pre-processed afterwards, obtain multiple subgraphs, multiple process cores of computer carry out parallel processing to multiple subgraphs, the processing is to carry out defects detection to subgraph using the detection model of FCN_21 × 21 after training, finally multiple parallel processing results are synthesized, obtain image to be detected the defects of testing result.The method of the present invention has good Detection results for high-definition picture, while also has prominent performance advantage in processing speed.
Description
Technical field
The invention belongs to TFT-LCD surface defects detection technical fields, and in particular to a kind of TFT-LCD based on FCN is lacked
Fall into quick determination method.
Background technology
Thin Film Transistor-LCD (Thin Film Transistor Liquid Crystal Display,
TFT-LCD) as a kind of material low in energy consumption, in light weight, brightness is high, classification rate is high, monitor, notes are had been widely used for
The fields such as sheet, mobile phone.Because its manufacturing process is complicated, TFT-LCD can be produced respectively by noise jamming unavoidably in the fabrication process
Kind defect.These Defect Edges are fuzzy, contrast is low, the spies such as the brightness irregularities for repeating grain background and entirety in image be present
Point, it drastically influence TFT-LCD display quality.TFT-LCD surface defect automatic detections based on machine vision are to ensure that certainly
One very important link of dynamic metaplasia production product quality, current detection method of surface flaw lead in low-resolution image
A certain type flaw often can be effectively detected, but Detection results reduce when using high-definition picture, meanwhile, processing speed
Slowly, it is unable to reach requirement of real time.
The content of the invention：
The defects of for prior art, the present invention propose one kind and are based on full convolutional neural networks (Fully Convolutional
Networks, FCN) quick TFT-LCD surface defects detections algorithm, it is intended to solve TFT-LCD surface defects detection algorithms
The problem of Detection results are low in high-definition picture, detection speed is slow.
To reach above-mentioned purpose, technical scheme provided by the invention is：
Thin Film Transistor-LCD (TFT-LCD) surface defect of one kind based on full convolutional neural networks (FCN) is quickly examined
Survey method, it is characterised in that the detection method comprises the following steps：
1) detection model of FCN_21 × 21 is obtained by multiple sample trainings；
2) image to be detected is pre-processed, obtains multiple subgraphs；
3) each process cores of computer are handled each subgraph respectively, and the processing is using the FCN_ after training
21 × 21 detection models carry out defects detection to subgraph, and it is specifically included：
（3a）Pixel Information obtains the stage, using subgraph as input picture, to ensure output characteristic figure size and input subgraph
As in the same size, first to input picture carry out mend 0 processing, then using first layer convolution by mend 0 processing after image each
Pixel and ambient area information carry out convolution algorithm and generate low contrast characteristic's figure, and the first layer convolution convolution kernel size is
21×21；
（3b）In the Nonlinear Mapping stage, the low contrast features figure one-to-one mapping of generation is contrasted to height using second layer convolution
Spend on characteristic pattern, the second layer convolution convolution kernel size is 1 × 1；
（3c）In the pixel reconstruction stage, high contrast characteristics figure is carried out to mend 0 processing, afterwards using third layer convolution to mending 0 processing
High contrast characteristics figure afterwards carries out convolution algorithm generation high-contrast image and exported, and it is right from soft image to height to complete
Than the end-to-end mapping of degree image, the convolution kernel size of the third layer convolution is 21 × 21；
Wherein, three-layer coil product mapping relations are respectively：
In formula, * is convolution operation,For input image pixels,ForLayer output image pixel,ForThe convolution of layer
Core,ForThe biasing of layer,；
（3d）By step（3c）The value of all pixels of output image is normalized, make its span [0,1] it
Between, and by the value of each pixel after normalization compared with threshold value H set in advance, if greater than equal to H, be labeled as
1, if less than H, labeled as 0, according to testing result the defects of marking acquisition subgraph；
4）Testing result the defects of multiple subgraphs is synthesized, obtain image to be detected the defects of testing result.
Further, the step 1）Specifically include following sub-step：
Step 1a）：Parameter using Xavier algorithms to the detection model of FCN_21 × 21
Initialized, obtain the detection model of initial FCN_21 × 21；
Step 1b）：Different types of screen detection image sample, the screen detection image sample are gathered from actual production workshop
Resolution ratio be W × H, the qualified samples of M are selected from the screen detection image sample of collection, each sample is cut into N
Individual size is w × h image block, wherein,, stride is cutting step-length；
Step 1c）：Data amplification, the sample set P after being expanded are carried out to the image block after cutting using data amplification method；
Each sample in sample set after amplification is labeled, defect pixel is labeled as 1, and normal pixel is labeled as 0, obtains each
The label of sample；
Step 1d）：A sample p in sample set P is selected, sample p is carried out using the detection model of initial FCN_21 × 21
Defects detection, by testing result compared with sample p label, the detection model of FCN_21 × 21 is adjusted according to comparative result
Parameter；
Step 1e）：Repeat step 1d）, until sample all in sample set P all completions are handled, the FCN_21 after being trained
× 21 detection models.
Further, step 2）Described in image to be detected is pre-processed, obtain multiple subgraphs and specifically include：
Step 2a)：Image to be detected of acquisition is converted into gray level image；
Step 2b)：Gray level image is evenly dividing and is equal to the process cores of computer for multiple subgraphs, the number of the subgraph
Number.
Further, it is specially each figure to being obtained after cutting that the image block after described pair of cutting, which carries out data amplification,
As block carries out 90 ° of rotations, 180 ° of rotations, 270 ° of rotations and the conversion of mirror image, left and right mirror image, diagonal mirror picture up and down.
Further, step（3a）And step（3c）In to image carry out mend 0 processing be specially：By the figure that size is U × V
As surrounding carries out 0 pixel filling, image size after making filling complete for (
, whereinFor convolution kernel size.
Further, the span of the threshold value set in advance is [0,1], preferably 0.5.
The beneficial effects of the invention are as follows：
1）The problem of overcoming Detection results difference under high definition case, simultaneously because the learning ability that deep learning is powerful, it
Polytype noise can be detected.
2）For high-definition picture, processing speed can be accelerated by way of parallel processing, realize processing in real time.
Brief description of the drawings
Fig. 1 is the flow chart of the TFT-LCD surface defect quick determination methods based on FCN of the present invention.
Detect structure in FCN_21 × 21 that Fig. 2 is the present invention.
Fig. 3 is the result of FCN_21 × 21 contrast of the present invention.
Embodiment
To further appreciate that present disclosure, the present invention is described in detail in conjunction with the accompanying drawings and embodiments.
A kind of schematic flow sheet such as Fig. 1 institutes of TFT-LCD surface defect quick determination methods based on FCN of the present invention
Show, it includes four steps：
Step 101：The detection model of FCN_21 × 21 is obtained by multiple sample trainings.
The specific processing of the step is as follows：
Step a：Obtain the detection model of initial FCN_21 × 21.Ginseng using Xavier algorithms to the detection model of FCN_21 × 21
NumberInitialized, obtain the detection model of initial FCN_21 × 21.
Step b：Construct training sample set.Multiple different types of high-resolution screen detections are gathered from actual production workshop
Image, the images of M opening and closing lattice is selected therefrom as sample, wherein the defects of each sample includes equally distributed different type,
It is the single channel image that resolution ratio is 6500 × 4500 by sample configuration.M=10 are selected in the present invention, due to only having used 10 height
Image in different resolution, data volume is few, directly trains FCN networks can not obtain accurately network model with it, therefore we are to sample
Collection is expanded.It is w × h that the sample that each resolution ratio is W × H first is cut into N number of size according to cutting step-length stride
Image block, wherein,。
As stride=100, then the image that 1 resolution ratio is 6500 × 4500 is cleavable into 2604 400 × 400
Image.Again to every 90 ° of image block progress, 180 °, 270 ° of rotations and upper and lower, left and right, diagonal mirror as seven kinds of conversion after stripping and slicing.
Finally, the sample of 10 6500 × 4500 expands the sample for 208320 400 × 400, forms training sample set P.
Step c：Each sample in sample set P is labeled, defect pixel is labeled as 1, and normal pixel is labeled as 0,
Obtain the label of each sample.
Step d：A sample p in sample set P is selected, sample p is entered using the detection model of initial FCN_21 × 21
Row defects detection.Detection mode is as shown in Fig. 2 it is specifically included：First using first layer convolution by sample p pixel and
Ambient area information carries out convolution algorithm generation low contrast features figure, and first layer convolution convolution kernel size is 21 × 21, to protect
The characteristic pattern size of card output is 400 × 400 to be handled, it is necessary to carry out benefit 0 to sample p before first layer convolution algorithm is carried out,
It is that the image surrounding that size is U × V is carried out into 0 pixel filling to mend 0 processing mode, and the image size after making filling complete is, whereinFor convolution kernel size, due to
Sample p size is 400 × 400 in the present invention, and first layer convolution convolution kernel size is 21 × 21, and the image size after filling is
420×420；Secondly, the low contrast features figure one-to-one mapping that first layer convolution algorithm generates is arrived using second layer convolution
On high contrast characteristics figure, second layer convolution is one-to-one Nonlinear Mapping, and convolution kernel size is 1 × 1；Then third layer is used
Convolution carries out convolution algorithm generation high-contrast image to high contrast characteristics figure, completes and is contrasted from soft image to height
The end-to-end mapping of image is spent, exports high-contrast image afterwards, the convolution kernel size of third layer convolution is 21 × 21, to ensure
The high-contrast image size of output is 400 × 400, is also required to before third time convolution algorithm is carried out to high contrast characteristics
Figure carries out mending 0 processing, and it is that the high contrast characteristics figure surrounding that size is u × v is carried out into 0 pixel filling to mend 0 processing mode, makes to fill out
Image size after charging is, wherein
For convolution kernel size；Finally, the value of all pixels of the high-contrast image of output is normalized, makes its span
Between [0,1], and by the value of each pixel after normalization compared with threshold value H set in advance, if greater than etc.
In H, the pixel is labeled as 1, if less than H, the pixel is labeled as 0, afterwards according to the defects of mark acquisition sample p
Testing result, H span is [0,1] here, it is generally the case that H=0.5.
In above-mentioned steps d, three-layer coil product mapping relations such as formula（1）-（3）It is shown：
（1）
（2）
（3）
In formula, * is convolution operation, and Y is input image pixels, F(Y) it is theLayer output image pixel,ForThe convolution of layer
Core,ForThe biasing of layer,。
By testing result the defects of sample p compared with sample p label, FCN_21 × 21 are adjusted according to comparative result
The parameter of detection model, the parameter of detection model is adjusted using Adam optimized algorithms in the application.
Step e：Step d is repeated, until sample all in sample set P all completions are handled, after being trained
The detection model of FCN_21 × 21.
Step 102：Image to be detected is pre-processed, obtains multiple subgraphs.
Pretreatment refer to image to be detected of acquisition first is converted into gray level image, then by gray level image be evenly dividing for
Multiple subgraphs, the number of division determines according to the process cores number of computer, for example, being located using four core processors
During reason, image to be detected can be divided into 4 size identical subgraphs.
In order to facilitate subgraph is synthesized in subsequent step, carrying out drawing position of each subgraph of time-sharing recording in original image
Put.
Step 103：Multiple process cores of computer carry out parallel processing to multiple subgraphs, and the processing is using training
The detection model of FCN_21 afterwards × 21 carries out defects detection to subgraph, and it specifically includes following several stages：
Pixel Information obtain the stage, input subgraph, it is carried out mend 0 processing, with ensure the size of the stage output image with
It is consistent to input subgraph.It is described mend 0 processing mode be：The subgraph surrounding that size is U × V is subjected to 0 pixel filling, makes to fill out
Image size after charging for (, wherein
For convolution kernel size.
The each pixel and ambient area information of the image after mending 0 processing are carried out at convolution using first layer convolution
Reason generation low contrast features figure, first layer convolution convolution kernel size is 21 × 21, and convolution mapping relations are shown below：
（4）
Formula（4）In, * is convolution operation,For input image pixels,For the 1st layer of output image pixel,For the 1st layer
Convolution kernel,For the 1st layer of biasing.
In the Nonlinear Mapping stage, Pixel Information is obtained to the low contrast features figure one of stage generation using second layer convolution
It is mapped to one on high contrast characteristics figure, the stage convolution is one-to-one Nonlinear Mapping, output characteristic figure and input size
Unanimously, without mending 0 filling, second layer convolution convolution kernel size is 1 × 1, and convolution mapping relations are shown below：
（5）
Formula（5）In, * is convolution operation,For 1 layer of output image pixel,For the 2nd layer of output image pixel,For
2 layers of convolution kernel,For the 2nd layer of biasing.
Pixel reconstruction stage, the high contrast characteristics figure obtained to the Nonlinear Mapping stage carry out mending 0 processing, and the stage mends
The mode of 0 processing is identical with the processing mode of benefit 0 in pixel acquisition stage, the height after then benefit 0 is handled using third layer convolution
Contrast metric figure carries out convolution algorithm generation high-contrast image, and the convolution kernel size of third layer convolution is 21 × 21, convolution
Mapping relations are shown below：
（6）
Formula（6）In, * is convolution operation,For 2 layers of output image pixel,For the 3rd layer of output image pixel,For
3 layers of convolution kernel,For the 3rd layer of biasing.
Subgraph obtains stage, Nonlinear Mapping stage, pixel reconstruction stage by Pixel Information, completes from low contrast
Spend end-to-end mapping of the image to high-contrast image.Fig. 3 is the example that image is transformed into high-contrast from low contrast, is schemed
In, difference is smaller between the pixel of original image, and contrast is low, defect point（Pixel of the pixel value apparently higher than surrounding pixel point
The point of value）Difference unobvious between normal point, difference increases between the image pixel obtained after treatment, and contrast is high,
Defect point pixel value approaches to 255, and normal point pixel value approaches to 0, this allow for the defects of follow-up mark it is more accurate.
Testing result marking phase, all pixel values of pixel reconstruction stage output image are normalized, made every
The span of individual pixel is carried out between [0,1], and by the value of each pixel after normalization and threshold value H set in advance
Compare, if greater than equal to H, labeled as 1, if less than H, labeled as 0, detected and tied according to the defects of mark acquisition subgraph
Fruit；Here H span is [0,1], it is generally the case that H=0.5.
Step 104：Multiple parallel processing results are synthesized, obtain image to be detected the defects of testing result.
In step 104, according to position of each subgraph in original image, by testing result figure the defects of each subgraph according to
Position of the corresponding subgraph in original image is synthesized, the defects of obtaining original image testing result figure, wherein, labeled as 0
Pixel is normal point, is defect point labeled as 1 pixel, in this manner it is possible to the defects of quickly obtaining image testing result.
Embodiment of above is merely to illustrate technical scheme, rather than its limitations；Although with reference to foregoing implementation
The present invention is described in detail mode, it will be understood by those within the art that：It still can be to foregoing reality
Apply the technical scheme described in mode to modify, or equivalent substitution is carried out to which part technical characteristic；And these are changed
Or replace, the essence of appropriate technical solution is departed from the spirit and scope of various embodiments of the present invention technical scheme.
Claims (6)
1. Thin Film Transistor-LCD (TFT-LCD) surface defect of one kind based on full convolutional neural networks (FCN) is quick
Detection method, it is characterised in that the detection method comprises the following steps：
1) detection model of FCN_21 × 21 is obtained by multiple sample trainings；
2) image to be detected is pre-processed, obtains multiple subgraphs；
3) each process cores of computer are handled each subgraph respectively, and the processing is using the FCN_ after training
21 × 21 detection models carry out defects detection to subgraph, and it is specifically included：
（3a）Pixel Information obtains the stage, using subgraph as input picture, to ensure output characteristic figure size and input subgraph
As in the same size, first to input picture carry out mend 0 processing, then using first layer convolution by mend 0 processing after image each
Pixel and ambient area information carry out convolution algorithm and generate low contrast characteristic's figure, and the first layer convolution convolution kernel size is
21×21；
（3b）In the Nonlinear Mapping stage, the low contrast features figure one-to-one mapping of generation is contrasted to height using second layer convolution
Spend on characteristic pattern, the second layer convolution convolution kernel size is 1 × 1；
（3c）In the pixel reconstruction stage, high contrast characteristics figure is carried out to mend 0 processing, afterwards using third layer convolution to mending 0 processing
High contrast characteristics figure afterwards carries out convolution algorithm generation high-contrast image and exported, and it is right from soft image to height to complete
Than the end-to-end mapping of degree image, the convolution kernel size of the third layer convolution is 21 × 21；
Wherein, three-layer coil product mapping relations are respectively：
In formula, * is convolution operation,For input image pixels,ForLayer output image pixel,ForThe convolution of layer
Core,ForThe biasing of layer,；
（3d）By step（3c）The value of all pixels of output image is normalized, make its span [0,1] it
Between, and by the value of each pixel after normalization compared with threshold value H set in advance, if greater than equal to H, be labeled as
1, if less than H, labeled as 0, according to testing result the defects of marking acquisition subgraph；
4）Testing result the defects of multiple subgraphs is synthesized, obtain image to be detected the defects of testing result.
2. the Thin Film Transistor-LCD surface defect according to claim 1 based on full convolutional neural networks is quick
Detection method, it is characterised in that：The step 1）Specifically include following sub-step：
1a）Parameter using Xavier algorithms to the detection model of FCN_21 × 21Carry out
Initialization, obtain the detection model of initial FCN_21 × 21；
1b）Different types of screen detection image sample, point of the screen detection image sample are gathered from actual production workshop
Resolution is W × H, and M qualified samples are selected from the screen detection image sample of collection, each sample cutting that will be singled out
The image block for being w × h for N number of size, wherein,, stride is cutting step-length；
1c）Data amplification, the sample set P after being expanded are carried out to each image block after cutting using data amplification method；
Each sample in sample set after amplification is labeled, defect pixel is labeled as 1, and normal pixel is labeled as 0, obtains each
The label of sample；
1d）A sample p in sample set P is selected, defect inspection is carried out to sample p using the detection model of initial FCN_21 × 21
Survey, by testing result with sample p label compared with, according to the parameter of the comparative result adjustment detection model of FCN_21 × 21；
1e) repeat step 1d), until sample all in sample set P all completions are handled, the FCN_21 after being trained × 21 are examined
Survey model.
3. the Thin Film Transistor-LCD surface defect according to claim 1 based on full convolutional neural networks is quick
Detection method, it is characterised in that：It is described that image to be detected is pre-processed, obtain multiple subgraphs and specifically include：
2a）Image to be detected of acquisition is converted into gray level image；
2b）Gray level image is evenly dividing and is equal to the process cores number of computer for multiple subgraphs, the number of the subgraph.
4. the Thin Film Transistor-LCD surface defect quick detection according to claim 2 based on full convolutional neural networks
Method, it is characterised in that：The step（3a）And step（3c）In to image carry out mend 0 processing be specially：By the figure that size is U × V
As surrounding carries out 0 pixel filling, image size after making filling complete for (
, whereinFor convolution kernel size.
5. the Thin Film Transistor-LCD surface defect according to claim 2 based on full convolutional neural networks is quick
Detection method, it is characterised in that：It is specially to being obtained after cutting that each image block after described pair of cutting, which carries out data amplification,
Each image block carries out 90 ° of rotations, 180 ° of rotations, 270 ° of rotations and the conversion of mirror image, left and right mirror image, diagonal mirror picture up and down.
6. the Thin Film Transistor-LCD surface based on full convolutional neural networks according to claim any one of 1-5
Defect quick determination method, it is characterised in that：The span of the threshold value set in advance is [0,1], preferably 0.5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710764659.8A CN107561738B (en) | 2017-08-30 | 2017-08-30 | Fast TFT-LCD surface defect detection method based on FCN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710764659.8A CN107561738B (en) | 2017-08-30 | 2017-08-30 | Fast TFT-LCD surface defect detection method based on FCN |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107561738A true CN107561738A (en) | 2018-01-09 |
| CN107561738B CN107561738B (en) | 2020-06-12 |
Family
ID=60978269
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710764659.8A Active CN107561738B (en) | 2017-08-30 | 2017-08-30 | Fast TFT-LCD surface defect detection method based on FCN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107561738B (en) |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108171707A (en) * | 2018-01-23 | 2018-06-15 | 武汉精测电子集团股份有限公司 | A kind of Mura defects level evaluation method and device based on deep learning |
| CN108596871A (en) * | 2018-03-08 | 2018-09-28 | 中北大学 | A kind of BGA air blister defect image detection methods based on deep learning |
| CN108830851A (en) * | 2018-06-29 | 2018-11-16 | 凌云光技术集团有限责任公司 | A kind of LCD mat patch defect inspection method |
| CN108846841A (en) * | 2018-07-02 | 2018-11-20 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN108921840A (en) * | 2018-07-02 | 2018-11-30 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| CN108961238A (en) * | 2018-07-02 | 2018-12-07 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN109003271A (en) * | 2018-07-25 | 2018-12-14 | 江苏拙术智能制造有限公司 | A kind of Wiring harness connector winding displacement quality determining method based on deep learning YOLO algorithm |
| CN109064446A (en) * | 2018-07-02 | 2018-12-21 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN109087282A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| CN109085174A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| CN109086828A (en) * | 2018-08-10 | 2018-12-25 | 北京百度网讯科技有限公司 | Method and apparatus for detecting battery pole piece |
| CN109087281A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| CN109084955A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN109583496A (en) * | 2018-11-28 | 2019-04-05 | 武汉精立电子技术有限公司 | A kind of network model and method for the classification of display panel large area defect |
| CN109683360A (en) * | 2019-02-14 | 2019-04-26 | 宁波轻蜓视觉科技有限公司 | Liquid crystal display panel defect inspection method and device |
| CN109949305A (en) * | 2019-03-29 | 2019-06-28 | 北京百度网讯科技有限公司 | Method for detecting surface defects of products, device and computer equipment |
| CN110619343A (en) * | 2018-06-20 | 2019-12-27 | 亚摩信息技术股份有限公司 | Automatic defect classification method based on machine learning |
| CN110781888A (en) * | 2019-10-25 | 2020-02-11 | 北京字节跳动网络技术有限公司 | Method and device for regressing screen in video picture, readable medium and electronic equipment |
| CN110969175A (en) * | 2018-09-29 | 2020-04-07 | 长鑫存储技术有限公司 | Wafer processing method and device, storage medium and electronic equipment |
| CN111077166A (en) * | 2018-10-19 | 2020-04-28 | 北京金山云网络技术有限公司 | Defect detection method and device of liquid crystal display and terminal equipment |
| CN111275660A (en) * | 2018-12-05 | 2020-06-12 | 合肥欣奕华智能机器有限公司 | Defect detection method and device for flat panel display |
| CN111738338A (en) * | 2020-06-23 | 2020-10-02 | 征图新视(江苏)科技股份有限公司 | Defect detection method applied to motor coil based on cascaded expansion FCN network |
| CN113268914A (en) * | 2020-02-14 | 2021-08-17 | 聚积科技股份有限公司 | Method for establishing LED screen adjustment standard judgment model |
| CN114565567A (en) * | 2022-02-15 | 2022-05-31 | 清华大学 | Method and device for detecting defects of complex texture lace cloth |
| CN115239644A (en) * | 2022-07-05 | 2022-10-25 | 港珠澳大桥管理局 | Concrete defect identification method and device, computer equipment and storage medium |
| CN116993718A (en) * | 2023-09-25 | 2023-11-03 | 深圳市东陆科技有限公司 | TFT array substrate defect detection method based on machine vision |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106355579A (en) * | 2016-09-14 | 2017-01-25 | 南京理工大学 | Defect detecting method of cigarette carton surface wrinkles |
| US20170061258A1 (en) * | 2015-08-25 | 2017-03-02 | Nokia Technologies Oy | Method, apparatus, and computer program product for precluding image capture of an image presented on a display |
| US20170177997A1 (en) * | 2015-12-22 | 2017-06-22 | Applied Materials Israel Ltd. | Method of deep learining-based examination of a semiconductor specimen and system thereof |
| CN106952250A (en) * | 2017-02-28 | 2017-07-14 | 北京科技大学 | A kind of metal plate and belt detection method of surface flaw and device based on Faster R CNN networks |
| CN106971185A (en) * | 2017-03-27 | 2017-07-21 | 深圳市丰巨泰科电子有限公司 | A kind of license plate locating method and device based on full convolutional network |
| CN107016681A (en) * | 2017-03-29 | 2017-08-04 | 浙江师范大学 | Brain MRI lesion segmentation approach based on full convolutional network |
-
2017
- 2017-08-30 CN CN201710764659.8A patent/CN107561738B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170061258A1 (en) * | 2015-08-25 | 2017-03-02 | Nokia Technologies Oy | Method, apparatus, and computer program product for precluding image capture of an image presented on a display |
| US20170177997A1 (en) * | 2015-12-22 | 2017-06-22 | Applied Materials Israel Ltd. | Method of deep learining-based examination of a semiconductor specimen and system thereof |
| CN106355579A (en) * | 2016-09-14 | 2017-01-25 | 南京理工大学 | Defect detecting method of cigarette carton surface wrinkles |
| CN106952250A (en) * | 2017-02-28 | 2017-07-14 | 北京科技大学 | A kind of metal plate and belt detection method of surface flaw and device based on Faster R CNN networks |
| CN106971185A (en) * | 2017-03-27 | 2017-07-21 | 深圳市丰巨泰科电子有限公司 | A kind of license plate locating method and device based on full convolutional network |
| CN107016681A (en) * | 2017-03-29 | 2017-08-04 | 浙江师范大学 | Brain MRI lesion segmentation approach based on full convolutional network |
Non-Patent Citations (1)
| Title |
|---|
| 欧先锋等: "基于CNN的车牌数字字符识别算法", 《成都工业学院学报》 * |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108171707A (en) * | 2018-01-23 | 2018-06-15 | 武汉精测电子集团股份有限公司 | A kind of Mura defects level evaluation method and device based on deep learning |
| CN108596871A (en) * | 2018-03-08 | 2018-09-28 | 中北大学 | A kind of BGA air blister defect image detection methods based on deep learning |
| CN110619343B (en) * | 2018-06-20 | 2023-06-06 | 亚摩信息技术(上海)有限公司 | Automatic defect classification method based on machine learning |
| CN110619343A (en) * | 2018-06-20 | 2019-12-27 | 亚摩信息技术股份有限公司 | Automatic defect classification method based on machine learning |
| CN108830851A (en) * | 2018-06-29 | 2018-11-16 | 凌云光技术集团有限责任公司 | A kind of LCD mat patch defect inspection method |
| CN108830851B (en) * | 2018-06-29 | 2020-08-28 | 凌云光技术集团有限责任公司 | LCD rough spot defect detection method |
| US11380232B2 (en) | 2018-07-02 | 2022-07-05 | Beijing Baidu Netcom Science Technology Co., Ltd. | Display screen quality detection method, apparatus, electronic device and storage medium |
| CN108961238A (en) * | 2018-07-02 | 2018-12-07 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN109087282A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| CN109085174A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| JP2020530125A (en) * | 2018-07-02 | 2020-10-15 | 北京百度網訊科技有限公司 | Display screen peripheral circuit detection method, display screen peripheral circuit detection device, electronic devices and storage media |
| CN109087281A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| CN109084955A (en) * | 2018-07-02 | 2018-12-25 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| KR102320371B1 (en) | 2018-07-02 | 2021-11-01 | 베이징 바이두 넷컴 사이언스 테크놀로지 컴퍼니 리미티드 | Display screen peripheral circuit detection method, apparatus, electronic device and storage medium |
| US11488294B2 (en) | 2018-07-02 | 2022-11-01 | Beijing Baidu Netcom Science Technology Co., Ltd. | Method for detecting display screen quality, apparatus, electronic device and storage medium |
| CN109064446A (en) * | 2018-07-02 | 2018-12-21 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN108921840A (en) * | 2018-07-02 | 2018-11-30 | 北京百度网讯科技有限公司 | Display screen peripheral circuit detection method, device, electronic equipment and storage medium |
| KR20200004822A (en) * | 2018-07-02 | 2020-01-14 | 베이징 바이두 넷컴 사이언스 테크놀로지 컴퍼니 리미티드 | Display screen peripheral circuit detection method, device, electronic device and storage medium |
| KR20200004825A (en) * | 2018-07-02 | 2020-01-14 | 베이징 바이두 넷컴 사이언스 테크놀로지 컴퍼니 리미티드 | Display device quality checking methods, devices, electronic devices and storage media |
| KR102321765B1 (en) * | 2018-07-02 | 2021-11-03 | 베이징 바이두 넷컴 사이언스 테크놀로지 컴퍼니 리미티드 | Display device quality inspection method, device, electronic device and storage medium |
| CN108846841A (en) * | 2018-07-02 | 2018-11-20 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN109003271A (en) * | 2018-07-25 | 2018-12-14 | 江苏拙术智能制造有限公司 | A kind of Wiring harness connector winding displacement quality determining method based on deep learning YOLO algorithm |
| CN109086828A (en) * | 2018-08-10 | 2018-12-25 | 北京百度网讯科技有限公司 | Method and apparatus for detecting battery pole piece |
| CN110969175A (en) * | 2018-09-29 | 2020-04-07 | 长鑫存储技术有限公司 | Wafer processing method and device, storage medium and electronic equipment |
| CN110969175B (en) * | 2018-09-29 | 2022-04-12 | 长鑫存储技术有限公司 | Wafer processing method and device, storage medium and electronic equipment |
| CN111077166A (en) * | 2018-10-19 | 2020-04-28 | 北京金山云网络技术有限公司 | Defect detection method and device of liquid crystal display and terminal equipment |
| CN109583496A (en) * | 2018-11-28 | 2019-04-05 | 武汉精立电子技术有限公司 | A kind of network model and method for the classification of display panel large area defect |
| CN111275660A (en) * | 2018-12-05 | 2020-06-12 | 合肥欣奕华智能机器有限公司 | Defect detection method and device for flat panel display |
| CN109683360B (en) * | 2019-02-14 | 2022-02-22 | 宁波轻蜓视觉科技有限公司 | Liquid crystal panel defect detection method and device |
| CN109683360A (en) * | 2019-02-14 | 2019-04-26 | 宁波轻蜓视觉科技有限公司 | Liquid crystal display panel defect inspection method and device |
| CN109949305A (en) * | 2019-03-29 | 2019-06-28 | 北京百度网讯科技有限公司 | Method for detecting surface defects of products, device and computer equipment |
| CN109949305B (en) * | 2019-03-29 | 2021-09-28 | 北京百度网讯科技有限公司 | Product surface defect detection method and device and computer equipment |
| CN110781888A (en) * | 2019-10-25 | 2020-02-11 | 北京字节跳动网络技术有限公司 | Method and device for regressing screen in video picture, readable medium and electronic equipment |
| CN113268914A (en) * | 2020-02-14 | 2021-08-17 | 聚积科技股份有限公司 | Method for establishing LED screen adjustment standard judgment model |
| CN111738338A (en) * | 2020-06-23 | 2020-10-02 | 征图新视(江苏)科技股份有限公司 | Defect detection method applied to motor coil based on cascaded expansion FCN network |
| CN114565567A (en) * | 2022-02-15 | 2022-05-31 | 清华大学 | Method and device for detecting defects of complex texture lace cloth |
| CN114565567B (en) * | 2022-02-15 | 2024-04-09 | 清华大学 | Defect detection method and device for complex texture lace cloth |
| CN115239644A (en) * | 2022-07-05 | 2022-10-25 | 港珠澳大桥管理局 | Concrete defect identification method and device, computer equipment and storage medium |
| CN115239644B (en) * | 2022-07-05 | 2024-04-05 | 港珠澳大桥管理局 | Concrete defect identification method, device, computer equipment and storage medium |
| CN116993718A (en) * | 2023-09-25 | 2023-11-03 | 深圳市东陆科技有限公司 | TFT array substrate defect detection method based on machine vision |
| CN116993718B (en) * | 2023-09-25 | 2023-12-22 | 深圳市东陆科技有限公司 | TFT array substrate defect detection method based on machine vision |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107561738B (en) | 2020-06-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107561738A (en) | TFT LCD surface defect quick determination methods based on FCN | |
| US10504209B2 (en) | Image dehazing method | |
| CN107123111B (en) | Deep residual error network construction method for mobile phone screen defect detection | |
| CN109683360B (en) | Liquid crystal panel defect detection method and device | |
| US9336578B2 (en) | Interactive tone mapping for high dynamic range video | |
| CN104992183B (en) | The automatic testing method of well-marked target in natural scene | |
| US20210390282A1 (en) | Training data increment method, electronic apparatus and computer-readable medium | |
| CN106651792A (en) | Method and device for removing strip noise of satellite image | |
| CN115829995A (en) | Cloth flaw detection method and system based on pixel-level multi-scale feature fusion | |
| CN115880271A (en) | Identification and detection method for seed crystal single crystal line edge angle in crystal growth process | |
| CN116205876A (en) | Unsupervised notebook appearance defect detection method based on multi-scale standardized flow | |
| CN107169498A (en) | It is a kind of to merge local and global sparse image significance detection method | |
| CN117011300B (en) | Micro defect detection method combining instance segmentation and secondary classification | |
| CN112017263B (en) | Intelligent test paper generation method and system based on deep learning | |
| CN105913427B (en) | Machine learning-based noise image saliency detecting method | |
| CN113763384B (en) | Defect detection method and defect detection device in industrial quality inspection | |
| CN117011222A (en) | Cable buffer layer defect detection method, device, storage medium and equipment | |
| CN106682676A (en) | Image test method and image test device | |
| CN115661022A (en) | Fabric flaw detection model establishing method and fabric flaw detection method | |
| CN115631186A (en) | Industrial element surface defect detection method based on double-branch neural network | |
| CN114627081A (en) | Wheat imperfect grain identification method based on Mask R-CNN | |
| CN105023016B (en) | Target apperception method based on compressed sensing classification | |
| CN112019774B (en) | High-quality display method of infrared high-bit-width digital image | |
| CN108537762A (en) | Secondary jpeg compressed image evidence collecting method based on the multiple dimensioned network of depth | |
| CN110930314B (en) | Image stripe noise suppression method and device, electronic equipment and storage medium |
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 |