CN106355579A - Defect detecting method of cigarette carton surface wrinkles - Google Patents
Defect detecting method of cigarette carton surface wrinkles Download PDFInfo
- Publication number
- CN106355579A CN106355579A CN201610826839.XA CN201610826839A CN106355579A CN 106355579 A CN106355579 A CN 106355579A CN 201610826839 A CN201610826839 A CN 201610826839A CN 106355579 A CN106355579 A CN 106355579A
- Authority
- CN
- China
- Prior art keywords
- layer
- image
- output
- vector
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
- G06T7/0008—Industrial image inspection checking presence/absence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a defect detecting method of cigarette carton surface wrinkles. The method includes: collecting the image information of a collecting detecting area, and converting pixels into 256-order color depth through image graying; highlighting image surface defect features on the basis of Laplace operator slight sharpening; performing convolutional neural network training to obtain images, performing surface defect image detection, that is to say, judging whether an image is a surface defect image or not through a trained convolutional neural network. By the defect detecting method, surface wrinkle defects can be identified effectively, and a fast and accurate identification result can be achieved during image brightness change.
Description
Technical field
The invention belongs to surface defects detection technology, specifically a kind of based on image procossing and convolutional neural networks
Plume detection method of surface flaw.
Background technology
In the surface defects detection system of numerous view-based access control model, vision detection technology based on Digital Image Processing with
Its technology maturation, low cost, recognition effect are outstanding to be favored by people, the need of the technical products of corresponding surface defects detection
Ask and increasingly increase, application is also more and more extensive.
Surface testing is the whether defective process of analysis of the image from picture to be detected.Surface detection technique of the present invention
Relate generally to the steps such as image acquisition, image processing, the structure of convolutional neural networks and training, the classification of image deflects.
Wherein the digital processing of image and the structure of convolutional neural networks are two key issues in Surface testing field.
At present the method for Surface testing mainly has two kinds: characteristic processing method based on image processing algorithm and be based on convolution
Neutral net and the method for image processing algorithm.Based on the characteristic processing method of image processing algorithm, it is by the ash to image
Degreeization is processed, and calculates the average gray numerical value of image and the variance of gray processing numerical value respectively, by with pre-set mark
Quasi- value is compared, and judges to obtain whether size of the difference exceedes threshold value to determine whether surface defect with standard value.The method is led to
Often need the information such as the bright dark, position of plume image acquisition of pre-set camera exposure time, ambient light, by extraneous bar
Part impact is big, and the robustness of system is poor.Based on the method for convolutional neural networks and image processing algorithm, convolutional network is for knowing
Other two-dimensional shapes and a multilayer perceptron of particular design, this network structure to translation, proportional zoom, inclination or altogether he
The deformation of form has height invariance.And avoid complicated feature extraction data process of reconstruction in tional identification algorithm.
The recognition speed of defect characteristic also can improve a lot, but such method need substantial amounts of sample come to train reach higher
Discrimination, training sample is very few may to lead to discrimination to decline.
Content of the invention
It is an object of the invention to provide a kind of defect inspection method of plume surface folding, thus fast and accurately realizing
Surface defects detection.
The technical solution realizing the object of the invention is: a kind of defect inspection method of plume surface folding, step is such as
Under:
The first step, is weighted after average gray to the detection zone of input picture, using the calculation based on Laplace operator
Method carries out image sharpening process, prominent characteristics of image, and then normalized image size is 64*64.
Second step, builds 7 layers of convolutional neural networks, and that is, ground floor input layer input dimension is 64*64, and second arrives layer 7
It is respectively the convolutional layer that convolution kernel is 5*5, the pond layer of 2*2, convolution kernel is the convolutional layer of 5*5, the pond layer of 2*2, convolution kernel
Convolutional layer for 5*5, full articulamentum.Wherein full articulamentum output dimension is defect kind number size.
3rd step, training convolutional neural networks parameter.The image that the first step is processed is inputted with the convolution god that second step builds
Through network, oneself is built multiple images and is repeatedly trained with tag library and test, and after wherein often training 10 groups of data, tests one
Group training data is as parameter correction.
4th step, imaging surface defects detection, the image input second step inputting to be detected is built and passes through the 3rd step
Train the convolutional neural networks obtaining, you can obtain the result of surface defects detection.
The present invention compared with prior art, its remarkable advantage: (1) not only effectively inhibits ambient brightness to image procossing
The impact producing, reduces the dependency that camera is determined with time of exposure simultaneously, and it is scarce quickly to identify that surface has
Sunken image.(2) first image is carried out with gray proces and a certain degree of sharpening operation, the defect characteristic of strengthening image.(2)
Next build the model of the convolutional neural networks that can stably identify defect image, by realizing the design of multilayer perceptron,
Can have effective evident characteristics of height with the deformation to the translation of image, proportional zoom, inclination or his form common.(3) will
By this system, images to be recognized can quickly judge whether image is defective, it is to avoid complicated in tional identification algorithm
Feature extraction data process of reconstruction.
Brief description
Accompanying drawing is the convolutional neural networks structure chart of the present invention.
Specific embodiment
In conjunction with accompanying drawing, the present invention is a kind of plume surface defects detection side based on image procossing and convolutional neural networks
Pixel is converted into the color depth of 256 ranks by method, the first image information in acquisition testing region by image gray processing;Base
Slightly sharpen in Laplace operator, prominent imaging surface defect characteristic;The image being obtained by convolutional neural networks training, so
Carry out the detection of surface defect picture afterwards, the convolutional neural networks that trained will be passed through, using upper by surface defect picture
State several steps and judge whether a width picture is surface defect picture.Specifically comprise the following steps that
The first step, is weighted after average gray to the detection zone of input picture, using the calculation based on Laplace operator
Method carries out image sharpening process, prominent characteristics of image, and then normalized image size is 64*64.
Second step, builds 7 layers of convolutional neural networks, and that is, ground floor input layer input dimension is 64*64, and second arrives layer 7
It is respectively the convolutional layer that convolution kernel is 5*5, the pond layer of 2*2, convolution kernel is the convolutional layer of 5*5, the pond layer of 2*2, convolution kernel
Convolutional layer for 5*5, full articulamentum.Wherein full articulamentum output dimension is defect kind number size.
(1) build input layer, input layer does not have input vector, only output vector, and this vectorial size is exactly picture
Size, i.e. the matrix of a 64*64, wherein activation primitive be one-dimensional linear function.
(2) build c1 convolutional layer, the input of this convolutional layer derives from the output layer of last layer, altogether comprises 6 5*5's
Convolution kernel, one preserves the vector that last layer exports 64*64, a vector preserving this layer of output 60*60*6, and one represents complete
The connection table of annexation.Excitation function adopts tanh, and each convolution kernel extracts a kind of characteristics of image, altogether can have 6
Feature map.
(3) build s2 pond layer, this layer is that last layer map is carried out with the convolution kernel of 2*2 with sampling processing, a preservation
The vector of the 60*60*6 of last layer output, a vector preserving this layer of output 30*30*6, excitation function adopts tanh, each
Individual convolution kernel takes the maximum in 2*2 region.
(4) build c3 convolutional layer, the input of this convolutional layer derives from pond layer, altogether comprises the convolution of 16*6 5*5
Core, one preserve last layer output 30*30*6 vector, one preserve this layer output 16*26*26 vector, this layer with
Last layer is connected by connection matrix, is not using full connected mode, excitation function adopts tanh, altogether can obtain 16
Feature map.
(5) build s4 pond layer, this layer is similar with s2 layer, sampling processing, a preservation are carried out using the convolution kernel of 2*2
The vector of the 16*26*26 of last layer output, a vector preserving this layer of output 40*13*13, excitation function adopts tanh, often
One convolution kernel takes the maximum in 2*2 region.
(6) build c5 convolutional layer, this convolutional layer is similar with c3 layer, and input derives from pond layer, altogether comprises 40 5*5
Convolution kernel, one preserve last layer output 16*13*13 vector, one preserve this layer output 40*9*9 vector, this
Layer is equally connected using connection matrix with last layer, and excitation function adopts tanh, altogether can obtain 40 features map.
(7) build the full articulamentum of f6, this layer is connected entirely with last layer, is output as defect kind number to be detected, excitation
Function is tanh, and loss function calculates the error of back transfer with mean square deviation formula.
3rd step, training convolutional neural networks parameter.The image that the first step is processed is inputted with the convolution god that second step builds
Through network, oneself is built multiple images and is repeatedly trained with tag library and test, and after wherein often training 10 groups of data, tests one
Group training data is as parameter correction.
4th step, imaging surface defects detection, the image input second step inputting to be detected is built and passes through the 3rd step
Train the convolutional neural networks obtaining, you can obtain the result of surface defects detection.
Claims (2)
1. a kind of defect inspection method of plume surface folding is it is characterised in that step is as follows:
The first step, is weighted average gray to the detection zone of input picture, then adopts based on Laplace operator
Algorithm carries out image sharpening process, prominent characteristics of image, and then normalized image size is 64*64;
Second step, builds 7 layers of convolutional neural networks, and that is, ground floor input layer input dimension is 64*64, and second arrives layer 7 respectively
For the convolutional layer for 5*5 for the convolution kernel, the pond layer of 2*2, convolution kernel be 5*5 convolutional layer, the pond layer of 2*2, convolution kernel be 5*5
Convolutional layer, full articulamentum;Wherein full articulamentum output dimension is defect kind number size;
3rd step, training convolutional neural networks parameter;The image that the first step is processed is input to the convolutional Neural of second step structure
Network, itself is built multiple images and is repeatedly trained with tag library and test, and after wherein often training 10 groups of data, tests one group
Training data is as parameter correction;
4th step, imaging surface defects detection;Build and pass through the 3rd step training to inputting image input second step to be detected
The convolutional neural networks obtaining, you can obtain the result of surface defects detection.
2. the defect inspection method of plume surface folding according to claim 1 is it is characterised in that described second step is concrete
Process is as follows:
2.1, build input layer, input layer does not have input vector, only output vector, and this vectorial size is exactly the big of picture
Little, i.e. the matrix of a 64*64;
2.2, build c1 convolutional layer, the input of this convolutional layer derives from the output of last layer, altogether comprises the convolution kernel of 6 5*5,
One preserves the vector that last layer exports 64*64, a vector preserving this layer of output 60*60*6;Excitation function adopts tanh,
Each convolution kernel extracts a kind of characteristics of image, a total of 6 features map;
2.3, build s2 pond layer, this layer is that last layer map is carried out with the convolution kernel of 2*2 with sampling processing, one preserves upper one
The vector of the 60*60*6 of layer output, a vector preserving this layer of output 30*30*6, excitation function adopts tanh, each volume
Long-pending core takes the maximum in 2*2 region;
2.4, build c3 convolutional layer, the input of this convolutional layer derives from pond layer, altogether comprise the convolution kernel of 16*6 5*5, one
The vector of the individual 30*30*6 preserving last layer output, vector, this layer and the last layer of a 16*26*26 preserving this layer of output
Connected by connection matrix, excitation function adopts tanh, 16 features map are always obtained;
2.5, build s4 pond layer, this layer is similar with s2 layer, sampling processing is carried out using the convolution kernel of 2*2, one preserves upper one
The vector of the 16*26*26 of layer output, a vector preserving this layer of output 40*13*13, excitation function adopts tanh, each
Convolution kernel takes the maximum in 2*2 region;
2.6, build c5 convolutional layer, this convolutional layer is similar with c3 layer, input derives from pond layer, altogether comprises the volume of 40 5*5
Long-pending core, one preserve last layer output 16*13*13 vector, one preserve this layer output 40*9*9 vector, this layer with
Last layer equally adopts connection matrix to connect, and excitation function adopts tanh, 40 features map are always obtained;
2.7, build the full articulamentum of f6, this layer is connected entirely with last layer, is output as defect kind number to be detected, excitation function
For tanh, loss function calculates the error of back transfer with mean square deviation formula.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610826839.XA CN106355579A (en) | 2016-09-14 | 2016-09-14 | Defect detecting method of cigarette carton surface wrinkles |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610826839.XA CN106355579A (en) | 2016-09-14 | 2016-09-14 | Defect detecting method of cigarette carton surface wrinkles |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106355579A true CN106355579A (en) | 2017-01-25 |
Family
ID=57858021
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610826839.XA Pending CN106355579A (en) | 2016-09-14 | 2016-09-14 | Defect detecting method of cigarette carton surface wrinkles |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106355579A (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107328787A (en) * | 2017-07-05 | 2017-11-07 | 北京科技大学 | A kind of metal plate and belt surface defects detection system based on depth convolutional neural networks |
CN107561738A (en) * | 2017-08-30 | 2018-01-09 | 湖南理工学院 | TFT LCD surface defect quick determination methods based on FCN |
CN108154508A (en) * | 2018-01-09 | 2018-06-12 | 北京百度网讯科技有限公司 | Method, apparatus, storage medium and the terminal device of product defects detection positioning |
CN108427969A (en) * | 2018-03-27 | 2018-08-21 | 陕西科技大学 | A kind of paper sheet defect sorting technique of Multiscale Morphological combination convolutional neural networks |
CN109291657A (en) * | 2018-09-11 | 2019-02-01 | 东华大学 | Laser Jet system is identified based on convolutional neural networks space structure part industry Internet of Things |
CN109583295A (en) * | 2018-10-19 | 2019-04-05 | 河南辉煌科技股份有限公司 | A kind of notch of switch machine automatic testing method based on convolutional neural networks |
CN111145163A (en) * | 2019-12-30 | 2020-05-12 | 深圳市中钞科信金融科技有限公司 | Paper wrinkle defect detection method and device |
US10769774B2 (en) | 2018-01-09 | 2020-09-08 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and device for detecting a defect in a steel plate, as well as apparatus and server therefor |
CN112419291A (en) * | 2020-11-30 | 2021-02-26 | 佛山职业技术学院 | Training method of bottle blank defect detection model, storage medium and terminal equipment |
CN113362305A (en) * | 2021-06-03 | 2021-09-07 | 河南中烟工业有限责任公司 | Smoke box strip missing mixed brand detection system and method based on artificial intelligence |
CN113610795A (en) * | 2021-08-02 | 2021-11-05 | 沈阳航空航天大学 | Combustible cartridge case surface defect detection method and system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105118044A (en) * | 2015-06-16 | 2015-12-02 | 华南理工大学 | Method for automatically detecting defects of wheel-shaped cast product |
WO2016132146A1 (en) * | 2015-02-19 | 2016-08-25 | Magic Pony Technology Limited | Visual processing using sub-pixel convolutions |
-
2016
- 2016-09-14 CN CN201610826839.XA patent/CN106355579A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2016132146A1 (en) * | 2015-02-19 | 2016-08-25 | Magic Pony Technology Limited | Visual processing using sub-pixel convolutions |
CN105118044A (en) * | 2015-06-16 | 2015-12-02 | 华南理工大学 | Method for automatically detecting defects of wheel-shaped cast product |
Non-Patent Citations (3)
Title |
---|
A KRIZHEVSKY 等: "ImageNet Classification with Deep Convolutional Neural Networks", 《INTERNATIONAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS》 * |
徐姗姗 等: "基于卷积神经网络的木材缺陷识别", 《山东大学学报(工学版)》 * |
百度文库: "卷积神经网络(CNN)", 《百度文库》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107328787A (en) * | 2017-07-05 | 2017-11-07 | 北京科技大学 | A kind of metal plate and belt surface defects detection system based on depth convolutional neural networks |
CN107561738B (en) * | 2017-08-30 | 2020-06-12 | 湖南理工学院 | Fast TFT-LCD surface defect detection method based on FCN |
CN107561738A (en) * | 2017-08-30 | 2018-01-09 | 湖南理工学院 | TFT LCD surface defect quick determination methods based on FCN |
CN108154508A (en) * | 2018-01-09 | 2018-06-12 | 北京百度网讯科技有限公司 | Method, apparatus, storage medium and the terminal device of product defects detection positioning |
US10769774B2 (en) | 2018-01-09 | 2020-09-08 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and device for detecting a defect in a steel plate, as well as apparatus and server therefor |
CN108427969A (en) * | 2018-03-27 | 2018-08-21 | 陕西科技大学 | A kind of paper sheet defect sorting technique of Multiscale Morphological combination convolutional neural networks |
CN109291657B (en) * | 2018-09-11 | 2020-10-30 | 东华大学 | Convolutional neural network-based aerospace structure industrial Internet of things identification laser coding system |
CN109291657A (en) * | 2018-09-11 | 2019-02-01 | 东华大学 | Laser Jet system is identified based on convolutional neural networks space structure part industry Internet of Things |
CN109583295A (en) * | 2018-10-19 | 2019-04-05 | 河南辉煌科技股份有限公司 | A kind of notch of switch machine automatic testing method based on convolutional neural networks |
CN109583295B (en) * | 2018-10-19 | 2022-12-06 | 河南辉煌科技股份有限公司 | Automatic detection method for switch machine notch based on convolutional neural network |
CN111145163A (en) * | 2019-12-30 | 2020-05-12 | 深圳市中钞科信金融科技有限公司 | Paper wrinkle defect detection method and device |
CN112419291A (en) * | 2020-11-30 | 2021-02-26 | 佛山职业技术学院 | Training method of bottle blank defect detection model, storage medium and terminal equipment |
CN113362305A (en) * | 2021-06-03 | 2021-09-07 | 河南中烟工业有限责任公司 | Smoke box strip missing mixed brand detection system and method based on artificial intelligence |
CN113610795A (en) * | 2021-08-02 | 2021-11-05 | 沈阳航空航天大学 | Combustible cartridge case surface defect detection method and system |
CN113610795B (en) * | 2021-08-02 | 2023-09-29 | 沈阳航空航天大学 | Method and system for detecting surface defects of combustible cartridge |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106355579A (en) | Defect detecting method of cigarette carton surface wrinkles | |
CN111325713B (en) | Neural network-based wood defect detection method, system and storage medium | |
CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
CN108074231B (en) | Magnetic sheet surface defect detection method based on convolutional neural network | |
CN108918536B (en) | Tire mold surface character defect detection method, device, equipment and storage medium | |
CN111402226A (en) | Surface defect detection method based on cascade convolution neural network | |
JP5546317B2 (en) | Visual inspection device, visual inspection discriminator generation device, visual inspection discriminator generation method, and visual inspection discriminator generation computer program | |
CN111027547A (en) | Automatic detection method for multi-scale polymorphic target in two-dimensional image | |
CN110473173A (en) | A kind of defect inspection method based on deep learning semantic segmentation | |
CN107316295A (en) | A kind of fabric defects detection method based on deep neural network | |
CN111257341B (en) | Underwater building crack detection method based on multi-scale features and stacked full convolution network | |
CN105740910A (en) | Vehicle object detection method and device | |
CN112233067A (en) | Hot rolled steel coil end face quality detection method and system | |
WO2022236876A1 (en) | Cellophane defect recognition method, system and apparatus, and storage medium | |
CN112348787A (en) | Training method of object defect detection model, object defect detection method and device | |
CN111127417B (en) | Printing defect detection method based on SIFT feature matching and SSD algorithm improvement | |
CN114897816A (en) | Mask R-CNN mineral particle identification and particle size detection method based on improved Mask | |
CN105223208B (en) | A kind of circuit board detecting template and preparation method thereof, circuit board detecting method | |
CN112150460A (en) | Detection method, detection system, device, and medium | |
CN112381175A (en) | Circuit board identification and analysis method based on image processing | |
CN111080574A (en) | Fabric defect detection method based on information entropy and visual attention mechanism | |
CN114005081A (en) | Intelligent detection device and method for foreign matters in tobacco shreds | |
CN110554052A (en) | artificial board surface defect detection method and system | |
CN112200790A (en) | Cloth defect detection method, device and medium | |
CN112884741B (en) | Printing apparent defect detection method based on image similarity comparison |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170125 |