CN109886936A - A kind of low contrast defect inspection method and device - Google Patents
A kind of low contrast defect inspection method and device Download PDFInfo
- Publication number
- CN109886936A CN109886936A CN201910081135.8A CN201910081135A CN109886936A CN 109886936 A CN109886936 A CN 109886936A CN 201910081135 A CN201910081135 A CN 201910081135A CN 109886936 A CN109886936 A CN 109886936A
- Authority
- CN
- China
- Prior art keywords
- image
- low contrast
- images
- source images
- inspection method
- 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
- 230000007547 defect Effects 0.000 title claims abstract description 46
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000007689 inspection Methods 0.000 title claims abstract description 25
- 238000001514 detection method Methods 0.000 claims abstract description 22
- 238000013136 deep learning model Methods 0.000 claims abstract description 10
- 230000004927 fusion Effects 0.000 claims description 20
- 238000005315 distribution function Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 8
- 238000003062 neural network model Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 5
- 238000013527 convolutional neural network Methods 0.000 claims description 3
- 239000011521 glass Substances 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 4
- 230000000007 visual effect Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 9
- 238000013135 deep learning Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 239000004744 fabric Substances 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 239000012780 transparent material Substances 0.000 description 1
- 238000005406 washing Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The invention discloses a kind of low contrast defect inspection method and devices, method is comprising steps of source images acquisition, assistant images generation and image classification, the present invention generates the assistant images of examined object image by fitting algorithm, and using assistant images and source images training deep learning model, carry out image classification, overcome technical problem low using artificial Visual detection methods detection efficiency in the prior art, it is fast to realize detection speed, the good technical effect of detection effect is particularly suitable for the small defects detection of low contrast, grey scale change.
Description
Technical field
The present invention relates to depth learning technology field, especially a kind of low contrast defect inspection method and device.
Background technique
Currently, watermark defect is that one kind must carry out detection and detection difficulty is big for mobile phone camera module group detection
Defect.It is current mainly using artificial detection method in industry for such defect, i.e., it is observed by skilled worker's eyes
To determine whether that there are watermark defects.However, there is many drawbacks, such as human eye to be easy fatigue for this eye detection method,
Thus there is unstability, not can guarantee and absolutely detect accuracy；In addition, Detection task all be repeat it is simple, uninteresting,
Mechanical movement, the spirit to people is a kind of torment, and due to the limitation in the precision of human eye, speed, some high speeds, high-precision are examined
Survey at all can not be by manually completing.Therefore, the field need it is a kind of for extremely low contrast, textural characteristics are few, background is uneven
The accurate detection method of automation of even watermark defect.
Summary of the invention
The present invention is directed to solve at least some of the technical problems in related technologies.For this purpose, of the invention
One purpose is to provide a kind of a kind of low contrast defect inspection method small suitable for low contrast, grey scale change.
For this purpose, a second object of the present invention is to provide the small a kind of low contrasts of a kind of low contrast, grey scale change to lack
Fall into detection device.
The technical scheme adopted by the invention is that:
In a first aspect, the present invention provides a kind of low contrast defect inspection method, include the following steps:
Source images acquisition: the image of examined object is acquired, described image is source images；
Assistant images generate: the intensity distribution of the source images being fitted, the fitting intensity point of object under test is obtained
Cloth function generates assistant images according to the fitting intensity distribution function；
Image classification: the source images and the assistant images are inputted into deep learning model, and carry out image classification.
Further, the deep learning model includes neural network model.
Further, described image classification comprising steps of
The source images and the assistant images are separately input to carry out convolution in the convolution module with activation function
The feature of image is extracted in operation, and the feature of the source images and the assistant images is merged and generates fusion feature image,
The fusion feature image is input in the hidden layer of the neural network model.
Further, it further comprises the steps of:
Model training: model training is carried out using the source images and the assistant images.
Further, the intensity distribution by the source images is fitted, and obtains the fitting intensity point of object under test
Cloth function includes using Gaussian function fitting of distribution method or polynomial fitting method.
Further, the polynomial fitting method includes being fitted the intensity distribution using binary quartic polynomial.
Further, the neural network model includes six hidden layers and three full articulamentums.
Further, the examined object includes glass.
Second aspect, the present invention provide a kind of low contrast defect detecting device, comprising: source images acquisition module: are used for
Acquire the image of examined object；
Assistant images generation module obtains the quasi- of object under test for the intensity distribution of the source images to be fitted
Intensity distribution function is closed, assistant images are generated according to the fitting intensity distribution function；
Fusion feature image generation module: for merging the source images and the assistant images, fusion is generated
Characteristic image；
Image classification module: for carrying out image classification for fusion feature image input convolutional neural networks training.
Further, further includes:
Model training module, for carrying out mould according to the fusion feature image of the source images and assistant images generation
Type training.
The beneficial effects of the present invention are:
The present invention generates the assistant images of examined object image by fitting algorithm, and uses assistant images and source images
Training deep learning model, carries out image classification, overcomes low using artificial Visual detection methods detection efficiency in the prior art
Under technical problem, realize detection speed it is fast, the good technical effect of detection effect, be particularly suitable for low contrast, gray scale become
Change small defects detection.
Detailed description of the invention
Fig. 1 a is that there are the photos of the mobile phone camera module of dirty test failure；
Fig. 1 b is the photo of the mobile phone camera module of test passes；
Fig. 2 is a kind of flow chart of a specific embodiment in low contrast defect inspection method in the present invention；
Fig. 3 a is the tested cell-phone camera mould in the present invention in a kind of one specific embodiment of low contrast defect inspection method
The source images of group；
Fig. 3 b is the tested cell-phone camera in the present invention in a kind of a kind of specific embodiment of low contrast defect inspection method
The intensity distribution of the source images of mould group；
Fig. 4 a is the tested cell-phone camera in the present invention in a kind of a kind of specific embodiment of low contrast defect inspection method
The assistant images of mould group；
Fig. 4 b is the tested cell-phone camera in the present invention in a kind of a kind of specific embodiment of low contrast defect inspection method
The intensity distribution of the assistant images of mould group；
Fig. 5 is that image feature value extracts schematic diagram in traditional neural network；
Fig. 6 is source images and assistant images in a kind of a kind of specific embodiment of low contrast defect inspection method in the present invention
Fusion Features schematic diagram；
Fig. 7 is a kind of knot of deep learning model in a kind of specific embodiment of low contrast defect inspection method in the present invention
Structure schematic diagram；
Fig. 8 is the accuracy comparison figure of four kinds of different depth learning network model identification low contrast defects；
Fig. 9 is a kind of a kind of structural schematic diagram of specific embodiment of low contrast defect detecting device in the present invention；
Figure 10 is a kind of structural schematic diagram of another specific embodiment of low contrast defect detecting device in the present invention.
Specific embodiment
It should be noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application can phase
Mutually combination.
In defects detection field, water spots is also known as watermark, these defects typically occur in some products after washing
Surface, such as remaining water droplet on liquid crystal display and photomoduel.As illustrated in figs. 1A and ib, Fig. 1 a is that there are dirty test failures
Mobile phone camera module photo, 1 for watermark it is dirty, Fig. 1 b be test passes mobile phone camera module photo, traditional
In machine vision method, such defect is referred to as low contrast defect, and the solution of current this defect far can not also make us full
Meaning.
As shown in Fig. 2, Fig. 2 shows a kind of processes of one specific embodiment of low contrast defect inspection method of the invention
Figure, includes the following steps:
Source images acquisition: S1 acquires the image of examined object, described image is source images；Test pair in the present embodiment
As the CMOS camera module (CMOS Camera Module) for mobile phone, mobile phone camera module to be tested is installed to beta version
On, mobile phone camera module image is obtained, as shown in Figure 3a, and is pre-processed, the image of tested mobile phone camera module is obtained
Intensity distribution, as shown in Figure 3b；
S2, assistant images generate: the intensity distribution of the source images being fitted, the fitting intensity of object under test is obtained
Distribution function generates assistant images according to the fitting intensity distribution function；
Specifically, two-dimensional Gaussian function fitting and fitting of a polynomial can be used in the method for intensity fitting.
In the present embodiment, the characteristics of distribution according to image intensity in Fig. 3 b, described image intensity distribution and Gaussian Profile phase
Seemingly, it is contemplated that use Gauss Distribution Fitting intensity distribution.Due to the shade of camera lens, intensity distribution is uneven, in addition, the optics of lens
Characteristic causes the received luminous intensity in light strength ratio center of the marginal reception of sensor image area small, therefore, center and four
Brightness between angle is inconsistent, and lens itself are convex lenses, and according to the principle of convex lens, the sensitivity at center has to be larger than periphery
Sensitivity.
Image intensity distribution is fitted using using binary quartic polynomial in the present embodiment, compared with Gaussian Profile
It is smaller that mean square error (RMSE) is compared with cubic polynomial fitting, calculation amount is smaller compared with use quintic algebra curve fitting, and
Can be to avoid overfitting the problem of.
It is distributed as f (x, y) specifically, setting image intensity, x, y are respectively pixel coordinate, and p is fitting parameter, by sampled point
Intensity bring following formula into, calculating is fitted using such as least square method,
F (x, y)=p00+p10·x+p01·y+p20·x2+p11·x·y+p02·y2+p30·x3+p21·x2·y+p12·
x·y2+p03·y3+p40·x4+p31·x3·y+p22*x2*y2+p13·x·y3+p04·y4
The distribution schematic diagram of calculated fitting intensity, according to fitting intensity, it is auxiliary to carry out image restoring generation as shown in 4b
Help the schematic diagram of image as shown in fig. 4 a.
Image classification: the source images and the assistant images are inputted deep learning model, and carry out image point by S3
Class.
In traditional deep learning method, target image is unique input of network, is then extracted and is inputted with convolution kernel
The feature of image obtains image after the convolution kernel of 5*5*3 carries out convolution as shown in figure 5, picture depth is 256*256*3
Characteristic value x, then by be fully connected output extract characteristic image, such as formula:Node i and section
Weight between point j is ω ij, and the threshold value of node j is bj, and the output valve of each node is xj, the output valve of each node is based on
The output valve of all nodes in upper layer, present node and upper layer node.The threshold of the weight of all nodes and present node on first layer
Value is also activated by activation primitive, such as formula xj=f (Sj), wherein f is output layer activation primitive, general to select sigmoid function.
For low contrast defects detection, validity feature is less, and it is very difficult for directly being extracted with convolution kernel.
In the present embodiment, the input of deep learning model includes two images, and a width is source images, and another width is by source
The assistant images that image generates, as shown in fig. 6, Fig. 6 is the Fusion Features schematic diagram of source images and assistant images in the present embodiment,
It is that the source images of 256*256*3 and assistant images merge, then carry out feature by the convolution kernel of 5*5*6 by picture depth
Extraction obtains fusion feature figure.
Specifically, as shown in fig. 7, the fusion feature figure obtained after assistant images and source images convolution, merging is input to
It is neural network model in the present embodiment in deep learning model, it all includes one after every layer that hidden layer, which includes 6 convolutional layers,
Activation primitive, 3 are fully connected layer, and an average recognition time is 0.03s.
Using tri- kinds of Alexnet, Resnet and VGG representative deep learning network model deep learning networks with
The low contrast defect inspection method used in the present embodiment compares, and is obtained using test set every 1000 times trained iteration
Heterogeneous networks accuracy of identification as shown in figure 8, abscissa is the number of iterations, ordinate is precision, and curve 1 is Alexnet net
The nicety of grading curve of network, curve 2 are the nicety of grading curve of VGG network, and curve 3 is that the nicety of grading of Resnet network is bent
Line, curve 4 are the nicety of grading curve of low contrast defect inspection method in the present embodiment, it can be seen that, are passing through one from Fig. 8
After fixed the number of iterations, the precision highest for the low contrast defect inspection method that the present invention uses can reach 95% classification essence
The classification of degree and existing deep learning, which facilitates, compares, and can identify the very low defect of contrast well, has preferably special
Levy ability to express, higher accuracy of identification.
The present embodiment generates assistant images by source images, is melted source images and assistant images by the method for convolution kernel
It closes, obtains new characteristic value, the model structure tool due to providing some reference informations in assistant images, in this example
There is better ability to express, the ability for having identification grey scale change small is particularly suitable for the detection of the transparent materials such as glass
It further include model training step in the present embodiment, the principle of model training is identical with above-mentioned image recognition processes,
This is repeated no more.
Fig. 9 is a kind of structural schematic diagram of a kind of specific embodiment of low contrast defect detecting device in the present invention, source
Image capture module: for acquiring the image of examined object；
Assistant images generation module obtains the quasi- of object under test for the intensity distribution of the source images to be fitted
Intensity distribution function is closed, assistant images are generated according to the fitting intensity distribution function；
Fusion feature image generation module: for merging the source images and the assistant images, fusion is generated
Characteristic image；
Image classification module: for carrying out image classification for fusion feature image input convolutional neural networks training.
Further, as shown in Figure 10, further includes:
Model training module, for carrying out mould according to the fusion feature image of the source images and assistant images generation
Type training.
It is to be illustrated to preferable implementation of the invention, but the invention is not limited to the implementation above
Example, those skilled in the art can also make various equivalent variations on the premise of without prejudice to spirit of the invention or replace
It changes, these equivalent deformations or replacement are all included in the scope defined by the claims of the present application.
Claims (10)
1. a kind of low contrast defect inspection method, which comprises the steps of:
Source images acquisition: the image of examined object is acquired, described image is source images；
Assistant images generate: the intensity distribution of the source images being fitted, the fitting intensity distribution letter of object under test is obtained
Number generates assistant images according to the fitting intensity distribution function；
Image classification: the source images and the assistant images are inputted into deep learning model, and carry out image classification.
2. a kind of low contrast defect inspection method according to claim 1, which is characterized in that the deep learning model
Including neural network model.
3. a kind of low contrast defect inspection method according to claim 2, which is characterized in that described image, which is classified, includes
Step:
The source images and the assistant images are separately input to carry out convolution algorithm in the convolution module with activation function
The feature of the source images and the assistant images is merged and generates fusion feature image, by institute by the feature for extracting image
Fusion feature image is stated to be input in the hidden layer of the neural network model.
4. a kind of low contrast defect inspection method according to claim 1, which is characterized in that further comprise the steps of:
Model training: model training is carried out using the source images and the assistant images.
5. a kind of low contrast defect inspection method according to claim 1, which is characterized in that described by the source images
Intensity distribution be fitted, obtain object under test fitting intensity distribution function include using Gaussian function fitting of distribution method or
Polynomial fitting method.
6. a kind of low contrast defect inspection method according to claim 5, which is characterized in that the polynomial fitting method
Including being fitted the intensity distribution using binary quartic polynomial.
7. a kind of low contrast defect inspection method according to claim 2, which is characterized in that the neural network model
Including six hidden layers and three full articulamentums.
8. a kind of low contrast defect inspection method according to any one of claims 1 to 7, which is characterized in that it is described to
Detection object includes glass.
9. a kind of low contrast defect detecting device characterized by comprising
Source images acquisition module: for acquiring the image of examined object；
Assistant images generation module, for the intensity distribution of the source images to be fitted, the fitting for obtaining object under test is strong
Distribution function is spent, assistant images are generated according to the fitting intensity distribution function；
Fusion feature image generation module: for merging the source images and the assistant images, fusion feature is generated
Image；
Image classification module: for carrying out image classification for fusion feature image input convolutional neural networks training.
10. a kind of low contrast defect detecting device according to claim 9, which is characterized in that further include:
Model training module, for carrying out model instruction according to the fusion feature image of the source images and assistant images generation
Practice.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910081135.8A CN109886936B (en) | 2019-01-28 | 2019-01-28 | Low-contrast defect detection method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910081135.8A CN109886936B (en) | 2019-01-28 | 2019-01-28 | Low-contrast defect detection method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109886936A true CN109886936A (en) | 2019-06-14 |
| CN109886936B CN109886936B (en) | 2023-05-16 |
Family
ID=66927107
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910081135.8A Active CN109886936B (en) | 2019-01-28 | 2019-01-28 | Low-contrast defect detection method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109886936B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110487803A (en) * | 2019-08-20 | 2019-11-22 | Oppo（重庆）智能科技有限公司 | The defect inspection method and device of infrared light-emitting component |
| CN111445507A (en) * | 2020-04-16 | 2020-07-24 | 北京深测科技有限公司 | Data processing method for non-visual field imaging |
| CN114463327A (en) * | 2022-04-08 | 2022-05-10 | 深圳市睿阳精视科技有限公司 | Multi-shooting imaging detection equipment and method for watermark defect of electronic product lining package |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017223560A1 (en) * | 2016-06-24 | 2017-12-28 | Rensselaer Polytechnic Institute | Tomographic image reconstruction via machine learning |
| CN108009628A (en) * | 2017-10-30 | 2018-05-08 | 杭州电子科技大学 | A kind of method for detecting abnormality based on generation confrontation network |
| CN108053454A (en) * | 2017-12-04 | 2018-05-18 | 华中科技大学 | A kind of graph structure data creation method that confrontation network is generated based on depth convolution |
| US20180232601A1 (en) * | 2017-02-16 | 2018-08-16 | Mitsubishi Electric Research Laboratories, Inc. | Deep Active Learning Method for Civil Infrastructure Defect Detection |
-
2019
- 2019-01-28 CN CN201910081135.8A patent/CN109886936B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017223560A1 (en) * | 2016-06-24 | 2017-12-28 | Rensselaer Polytechnic Institute | Tomographic image reconstruction via machine learning |
| US20180232601A1 (en) * | 2017-02-16 | 2018-08-16 | Mitsubishi Electric Research Laboratories, Inc. | Deep Active Learning Method for Civil Infrastructure Defect Detection |
| CN108009628A (en) * | 2017-10-30 | 2018-05-08 | 杭州电子科技大学 | A kind of method for detecting abnormality based on generation confrontation network |
| CN108053454A (en) * | 2017-12-04 | 2018-05-18 | 华中科技大学 | A kind of graph structure data creation method that confrontation network is generated based on depth convolution |
Non-Patent Citations (1)
| Title |
|---|
| JINDONG TIAN .ETC: "Dynamic Phase Measurement Based on Two-Step Spatial Carrier-Frequency Phase-Shifting Interferometry", 《IEEE PHOTONICS JOURNAL》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110487803A (en) * | 2019-08-20 | 2019-11-22 | Oppo（重庆）智能科技有限公司 | The defect inspection method and device of infrared light-emitting component |
| CN111445507A (en) * | 2020-04-16 | 2020-07-24 | 北京深测科技有限公司 | Data processing method for non-visual field imaging |
| CN111445507B (en) * | 2020-04-16 | 2023-07-18 | 北京深测科技有限公司 | Data processing method for non-visual field imaging |
| CN114463327A (en) * | 2022-04-08 | 2022-05-10 | 深圳市睿阳精视科技有限公司 | Multi-shooting imaging detection equipment and method for watermark defect of electronic product lining package |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109886936B (en) | 2023-05-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113192040B (en) | Fabric flaw detection method based on YOLO v4 improved algorithm | |
| Ikoma et al. | Depth from defocus with learned optics for imaging and occlusion-aware depth estimation | |
| CN107123111B (en) | Deep residual error network construction method for mobile phone screen defect detection | |
| CN109886936A (en) | A kind of low contrast defect inspection method and device | |
| CN114022420B (en) | Detection method for automatically identifying defects of photovoltaic cell EL (electro-luminescence) component | |
| US11790510B2 (en) | Material testing of optical test pieces | |
| CN111932511B (en) | Electronic component quality detection method and system based on deep learning | |
| TWI480578B (en) | Method for detecting optical center of wide-angle lens, and optical center detection apparatus | |
| CN105092607A (en) | Method for evaluating surface defects of spherical optical components | |
| CN104101611A (en) | Mirror-like object surface optical imaging device and imaging method thereof | |
| CN111047655A (en) | High-definition camera cloth defect detection method based on convolutional neural network | |
| CN113092489A (en) | System and method for detecting appearance defects of battery | |
| CN109685774A (en) | Varistor open defect detection method based on depth convolutional neural networks | |
| CN116026843A (en) | Battery appearance defect detection method, device and detection system | |
| CN108548825A (en) | A kind of transparent plate defect detecting device and method based on two-dimentional illumination | |
| TWI822968B (en) | Color filter inspection device, inspection device, color filter inspection method, and inspection method | |
| CN117392042A (en) | Defect detection method, defect detection apparatus, and storage medium | |
| CN112200790A (en) | Cloth defect detection method, device and medium | |
| CN111160100A (en) | Lightweight depth model aerial photography vehicle detection method based on sample generation | |
| CN110415238A (en) | Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network | |
| Ding et al. | Automatic detection of dispersed defects in resin eyeglass based on machine vision technology | |
| CN114219758A (en) | Defect detection method, system, electronic device and computer readable storage medium | |
| Xie et al. | Detection algorithm for bearing roller end surface defects based on improved YOLOv5n and image fusion | |
| CN114612418A (en) | Method, device and system for detecting surface defects of mouse shell and electronic equipment | |
| CN111753572A (en) | Complex background low-quality two-dimensional bar code detection 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 |