CN109509187A - A kind of efficient check algorithm for the nibs in big resolution ratio cloth image - Google Patents
A kind of efficient check algorithm for the nibs in big resolution ratio cloth image Download PDFInfo
- Publication number
- CN109509187A CN109509187A CN201811336000.3A CN201811336000A CN109509187A CN 109509187 A CN109509187 A CN 109509187A CN 201811336000 A CN201811336000 A CN 201811336000A CN 109509187 A CN109509187 A CN 109509187A
- Authority
- CN
- China
- Prior art keywords
- image
- characteristic
- flaw
- model
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- 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
- G06T2207/30124—Fabrics; Textile; Paper
Abstract
The present invention relates to a kind of efficient check algorithms for the nibs in big resolution ratio cloth image, comprising: (1) by camera collection image, is then labeled using labelImg tool image；(2) image after processing is split into training set and test set, training set is used to train testing model, and test set is used to assess testing model performance；(3) training set image and corresponding classification information location information etc. are input in improved se-resnext101 testing model, training testing model；(4) using the image in the testing model processing test set after training, the approximate location and corresponding classification of flaw are obtained.Method of the invention can realize that the processing of Analysis On Multi-scale Features figure is adapted to a variety of different size of flaws in this way, increases substantially detection accuracy and speed to handle multiscale image block to the input picture of single resolution ratio；The algorithm realizes the case where there are a variety of flaws in obtaining flaw approximate location and processing image on image classification frame simultaneously.
Description
Technical field
The present invention relates to image classification fields, and in particular to a kind of nibs in big resolution ratio cloth image
Efficient check algorithm.
Background technique
Previous image classification mainly by the method for traditional machine learning, is generally divided into two parts: being based on feature
Extracting method and method based on template matching, and mainly have Statistics-Based Method based on feature extracting method, based on spectrum
Method, based on texture model method, the method based on study and structure-based method.These methods require artificially to select
Feature, while generalization is not strong.
As the development of deep learning, especially convolutional neural networks are in image classification, image detection and image segmentation etc.
The application of aspect, acquired effect are that passing traditional algorithm is incomparable.But the sorting algorithm based on deep learning needs
In the data set to be trained target area account for image area specific gravity it is larger, can just have preferable effect, if image is high score
Resolution image but the target but very little classified, it is possible to which class object is only accounted for less than total image area 1%, then directly use
Traditional sorting algorithm based on deep learning carries out image classification, and accuracy rate is very low.Meanwhile if in image there is
The target of multiple classifications, we can not also obtain the classification information of these targets.
Summary of the invention
In order to overcome the shortcomings of the prior art, the invention proposes a kind of for small in big resolution ratio cloth image
The efficient check algorithm of flaw, this method can realize the processing of Analysis On Multi-scale Features figure to the input picture of single resolution ratio, thus
Multiscale image block is handled, a variety of different size of flaws can be effectively treated in this way, increase substantially detection accuracy and speed；
Simultaneously this method can be obtained on image classification frame flaw approximate region and handle picture in there are the feelings of a variety of flaws
Condition.
In order to achieve the goal above, specific step is as follows for method proposed by the present invention:
(1) Image Acquisition shoots cloth image using the camera that resolution ratio is 2560*1920, obtains related
Data set and image is renamed, such as 1.jpg, 2.jpg, 3.jpg ..., M.jpg etc. then scales the images to 1024*
768 sizes, and the image of shooting is labeled using labelImg tool, obtain the label in image about flaw, flaw
Label contain the coordinate (x1, y1) in the flaw upper left corner in the picture, the classification of coordinate (x2, the y2) and flaw in the lower right corner
DefectN, wherein N indicates number, such as 1,2,3 ... etc., particularly, if not having flaw in the image of shooting, we will not be used
LabelImg is handled, its classification information norm is only recorded；
(2) image divides, and divides an image into training set and test set two parts, and identical figure is not present in two parts
Picture, training set are used to train testing model, and test set is used to assess the performance of testing model；
(3) image preprocessing, including spin upside down at random, left and right overturning and accidental light irradiation change etc. at random, wherein at random
Spin upside down, at random left and right overturning and accidental light irradiation change just for training set, particularly, when spun upside down at random and with
When machine or so is overturn, the coordinate information of flaw is also required to make corresponding variation；
(4) training testing model, by the training set after image preprocessing image and label information be input to inspection
It tests in model and is trained, testing model is improved on the basis of se-resnext101, and network is directed to
The input picture of single resolution ratio obtains Analysis On Multi-scale Features figure on model, obtains each feature by the propagated forward of testing model
The class probability value of each characteristic point on figure calculates Classification Loss by Focal Loss function, using under the gradient with momentum
Algorithm backpropagation training pattern drops；
(5) image in test set is input in trained testing model and extracts feature and obtain by cloth vision inspections
Take the class probability value of each characteristic point on Analysis On Multi-scale Features figure；If had in two or more characteristic patterns in three characteristic patterns
All characteristic points are all judged to norm, then it is assumed that the image category is norm, other situations then think that there are flaws in image；For
It is determined as the image there are flaw, we correspond to some image block in original image using each characteristic point, pass through characteristic point
Predict that the pixel value of class switch correspondence image block obtains relevant thermodynamic chart, the corresponding thermodynamic chart of superposition various features figure obtains
Final thermodynamic chart, obtains the approximate location of flaw by final thermodynamic chart, takes mathematical expectation of probability to obtain the image block near flaw
To the classification of the flaw, particularly, which can handle the case where there are more flaws in image, and obtain the class of each flaw
Other and approximate location.
Training includes being trained step, transfer learning based on improved se-resnext101 model in the step (4)
Step, two-stage learning rate set-up procedure, convolutional network extract characterization step, adaptive adjustment feature weight step, more rulers
Image block processing step is spent, Focal Loss step is calculated and utilizes the gradient descent algorithm backpropagation training pattern with momentum
Step.
As depicted in figs. 1 and 2, the step (4) specifically:
(4.1) by the last global pool layer of original se-resnext101 model instead of 3 by parallel characteristic block
The characteristic block pond little module that global pool layer and characteristic block maximum pond layer are constituted, each little module are at parallel pass
It is that the size of pond layer is identical in each little module but the size of the pond layer between disparate modules is different, it in addition will be last
Full articulamentum 1 1*1 size, step-length is replaced by 1 convolution operation, utilizes the improved se-resnext101 model
As testing model；
(4.2) weight that training obtains on ImageNets image set using se-resnext101 model changes to initialize
Into the se-resnext101 model of version, i.e. testing model, we are only remained except all biasing weights, last global pool
Weight outside change layer, last full articulamentum and softmax layers；
(4.3) training pattern when using two-stage learning rate adjustment network learning rate, i.e., the initial stage with
Last three layers of model of some learning rate training including the module of characteristic block pond and keep the weight of other layers of model
It is constant, to last the three of model after several iteration cycles (all images in each iteration cycle traversal training set) of training
Layer uses a biggish learning rate, other layers use a lesser learning rate, and reduce and learn according to certain rule
Practise rate；
(4.4) training image is inputted in improved se-resnext101 model, extracts feature using convolution operation, increases
The receptive field of big characteristic pattern, while making net using the extruding and actuation sub-module for including in original se-resnext101 model
Network can adaptively adjust feature weight, and prominent validity feature inhibits invalid feature, so that feature space and two, feature channel
Dimension is improved；
(4.5) 3 parallel characteristic block pond beggar's modules are utilized to the characteristic pattern of last convolutional layer output, as shown in figure 3,
The size of pond layer is identical in each little module but the size of the pond layer between disparate modules is different, to obtain 3 kinds not
With the characteristic pattern of size；Utilizing size to the characteristic pattern of acquisition is 1*1, and then the convolution operation that step-length is 1 utilizes
Softmax calculates the corresponding class probability value of each characteristic point on 3 kinds of different size of characteristic patterns；
(4.6) due to the presence of receptive field, the characteristic point on characteristic pattern corresponds to the image block of original image, we exist according to flaw
Position in image can know the true classification of each image block, so that the true classification of character pair point is obtained, according to pre-
The class probability value and true classification information of survey calculate Classification Loss using Focal Loss function, finally utilize band momentum
Gradient descent algorithm backpropagation update detection model parameter.
The step (5) specifically: test image is input in testing model, by propagated forward, obtains receptive field
Constantly increase, the characteristic pattern that resolution ratio constantly reduces, 3 parallel characteristic blocks are utilized to the characteristic pattern of last convolutional layer output
Pond little module is handled, and 3 kinds of different size of characteristic patterns are obtained, and is carried out size against 3 kinds of different size of characteristic patterns and is
1*1, the convolution operation that step-length is 1 finally utilize the corresponding classification of characteristic point each on each characteristic pattern of softmax operation acquisition
Probability value.If there is all characteristic points in two or more characteristic patterns to be all judged to norm in three characteristic patterns, then it is assumed that the figure
Picture classification is norm, then thinks that there are flaws in image for other situations.For there are the image of flaw, we are by characteristic pattern
In be determined as the pixel value of the image block that the characteristic point of norm maps back original image and be assigned to 0, characteristic points other in characteristic pattern are reflected
The pixel value for being emitted back towards the image block of original image is assigned to 1, obtains 3 thermodynamic charts in this way, and 3 thermodynamic chart superpositions are obtained most
Whole thermodynamic chart, obtains the approximate location of flaw from final thermodynamic chart, the classification of flaw by be determined as flaw several are special
The class probability mean value of sign point determines.
Compared with prior art, the beneficial effects of the present invention are:
The present invention is based on input picture of the improved se-resnext101 model realization to single resolution ratio to realize multiple dimensioned spy
The processing of sign figure is adapted to a variety of different size of flaws in this way, increases substantially detection to handle multiscale image block
Accuracy and speed；Frame based on disaggregated model can not only provide predicted pictures with the presence or absence of flaw, can also be directed to and exist
The image of flaw provides the approximate location and corresponding classification of flaw, and there are the images of plurality of classes flaw for processing；Utilize migration
Study and the adjustment of two-stage learning rate can make model have stronger Generalization Capability, improve the inspectability to new samples
Energy；It replaces traditional cross entropy loss function to calculate Classification Loss using Focal Loss function, it is uneven effectively to solve sample
The problem of weighing apparatus, improves the concern that testing model divides hardly possible sample, improves the performance of testing model.
Detailed description of the invention
Fig. 1 is se-resnext101 model schematic
Fig. 2 is modified version se-resnext101 model schematic
Fig. 3 is characterized block pond modular structure schematic diagram
Fig. 4 is se-resnext model and modified version se-resnext101 model two classification accuracy pair on test set
Than figure
Fig. 5 is modified version se-resnext101 model flaw category identification accuracy rate histogram in test set
Specific embodiment
Invention is further explained below.
Implementation process and embodiment of the present invention are as follows:
(1) Image Acquisition, we use resolution ratio to shoot for the camera of 2560*1920 to cloth image, altogether
5000 cloth images are obtained, and by their renamed as 1.jpg, 2.jpg ..., 5000.jpg then zooms to picture
1024*768 size, and the image of shooting is labeled using labelImg tool, obtain the label in image about flaw.
The label of flaw contains the coordinate (x1, y1) in the flaw upper left corner in the picture, the class of coordinate (x2, the y2) and flaw in the lower right corner
Other defectN, wherein { 1,2,3 ..., 9 } N ∈ indicate in data set a flaw for sharing 9 seed types, flaw be respectively greasy dirt,
Skips, lack warp, hang through, knit dilute, hair hole, wipe hole, hair spot and Zha Dong, the number of N corresponds in flaw order and defectN.
In order to facilitate processing, the information in xml document that labelImg is handled is transformed into txt file by we, only saves figure
The classification of flaw and corresponding position in piece, the corresponding txt file of every picture, and image is identical with the name of txt file.
Particularly, if not having flaw in the image of shooting, we will not be handled with labelImg, only record its classification information
Norm, and be stored in txt file；
(2) image is split, and divides an image into training set and test set two parts, and training set is used to train testing model,
Test set is for assessing detection model performance, and wherein training set is preceding 4500 images in data set, and test set is data set
In rear 500 images；
(3) image preprocessing, it is by taking the image I in training set as an example, the content in image I and corresponding txt file is defeated
Enter into image pre-processing module to carry out data enhancing and content conversion on line, wherein the content of txt file is stored in lists；
For the image in image there are flaw, list content is as follows:
For the image of flaw is not present in image, list content is as follows:
[norm]
On line data enhancing include spin upside down at random, at random left and right overturning and accidental light irradiation change etc., particularly, when into
When row is spun upside down at random and left and right is overturn at random, the coordinate information of flaw is also required to make corresponding variation；
(4) testing model is constructed, as depicted in figs. 1 and 2, by the global pool that original se-resnext101 model is last
Change the characteristic block Chi Hua little mould that layer is made of instead of 3 parallel characteristic block global pool layer and characteristic block maximum pond layer
Block, the structure of characteristic block pond little module is as shown in figure 3, each little module is at concurrency relation, pond in each little module
The size for changing layer is identical but the size of pond layer between disparate modules difference, in addition by last full articulamentum with 1 1*1
Size, step-length is replaced by 1 convolution operation, using the improved se-resnext101 model as testing model；
The parameter setting of each characteristic block pond module is as shown in table 1 in specific implementation:
The parameter setting of 1 characteristic block pond module of table
The image I of resolution ratio 1024*768 is after the propagated forward of modified version se-resnext101 model, modified version
The resolution ratio for the characteristic pattern that se-resnext101 module 5 exports is 32*24, by 3 parallel characteristic block pond resume modules
The resolution ratio of characteristic pattern afterwards is respectively 10*6,12*8 and 14*10, and intermediate-resolution is the characteristic point pair on the characteristic pattern of 10*6
The size for answering the image block in original image is 448*448, sliding step 64；Resolution ratio is the characteristic point pair on the characteristic pattern of 12*8
The size for answering the image block in original image is 320*320, sliding step 64；Resolution ratio is the characteristic point pair on the characteristic pattern of 14*6
Answering the tile size in original image is 192*192, sliding step 64.Can obtain in this way 3 kinds it is different size of
Image block, thus effectively solve the problems, such as nibs accounting is too small in original image caused by, also can handle a variety of
The flaw of scale, so that the check feature of model is more powerful.
(5) training testing model, by training set image and corresponding label information be input in testing model, pass through
Propagated forward obtains the corresponding class probability value of each characteristic point on each characteristic pattern, passes through Focal Loss loss function and calculates
Classification Loss obtains the network parameter of testing model using the gradient descent algorithm training testing model with momentum；
In specific implementation, power is set as 0.9, is input into 1 image every time, and 4500 steps are 1 iteration cycle, and 50 are arranged altogether
A iteration cycle, preceding 10 iteration cycles setting last three layers of learning rate of model is 0.001, and the learning rate of other layers is
0；Iteration cycle setting last three layers of the learning rate of model later is 0.0005, and the learning rate of other layers is 0.00005,
Every 4 iteration cycles, learning rate become the 0.94 of original learning rate simultaneously；α is set as 0.25 in Focal Loss function,
β is set as 2.After training, the parameter of testing model is saved.
(6) image in test set is input in trained testing model and extracts feature and obtain by cloth vision inspections
Take the class probability value of each characteristic point on Analysis On Multi-scale Features figure；If had in two or more characteristic patterns in three characteristic patterns
All characteristic points are all judged to norm, then it is assumed that the image category is norm, other situations then think that there are flaws in image；For
It is determined as the image there are flaw, we correspond to some image block in original image using each characteristic point, pass through characteristic point
Predict that the pixel value of class switch correspondence image block obtains relevant thermodynamic chart, the corresponding thermodynamic chart of superposition various features figure obtains
Final thermodynamic chart, obtains the approximate location of flaw by final thermodynamic chart, takes mathematical expectation of probability to obtain the image block near flaw
To the classification of the flaw, particularly, which can handle the case where there are more flaws in image, and obtain the class of each flaw
Other and approximate location.
The present embodiment is finally tested on cloth test set, and Fig. 4 illustrates se-resnext101 model and modified version
Image discriminating is normal sample and flaw sample by se-resnext101 model accuracy rate in two partition test results
Accuracy rate, it can be seen that the accuracy of the se-resnext101 model of modified version is apparently higher than se-resnext101 model, from
Test result can be seen that this method can efficiently examine the nibs in big image in different resolution.Fig. 5, which is illustrated, to be changed
Into the accuracy rate of version resnext101 model defect classification on test set, it can be seen that the method for this patent can be more to existing
The image detection of kind flaw goes out flaw classification present in image submodule.
Claims (4)
1. a kind of efficient check algorithm for the nibs in big resolution ratio cloth image, which is characterized in that including walking as follows
It is rapid:
(1) Image Acquisition shoots cloth image using the camera that resolution ratio is 2560*1920, obtains relevant number
It is renamed according to collection and to image, it is big then to scale the images to 1024*768 by such as 1.jpg, 2.jpg, 3.jpg ..., M.jpg etc.
It is small, and the image of shooting is labeled using label Img tool, obtain the label in image about flaw, the mark of flaw
Label contain the coordinate (x1, y1) in the flaw upper left corner in the picture, the classification of coordinate (x2, the y2) and flaw in the lower right corner
DefectN, wherein N indicates number, such as 1,2,3 ... etc., particularly, if not having flaw in the image of shooting, we will not be used
LabelImg is handled, its classification information norm is only recorded；
(2) image divides, and divides an image into training set and test set two parts, and identical image, instruction is not present in two parts
Practice collection to be used to train testing model, test set is used to assess the performance of testing model；
(3) image preprocessing, including spin upside down at random, left and right overturning and accidental light irradiation change etc. at random, wherein above and below random
Overturning, random left and right overturning and accidental light irradiation change just for training set, particularly, when spin upside down at random and random left
When right overturning, the coordinate information of flaw is also required to make corresponding variation；
(4) training testing model, by the training set after image preprocessing image and label information be input to inspection mould
It is trained in type, testing model is improved on the basis of se-resnext101, allows network for single point
The input picture of resolution obtains Analysis On Multi-scale Features figure on model, is obtained on each characteristic pattern by the propagated forward of testing model
The class probability value of each characteristic point is calculated Classification Loss by Focal Loss function, is declined using the gradient with momentum and calculated
Method backpropagation training pattern；
(5) image in test set is input in trained testing model and extracts feature and obtain more by cloth vision inspections
The class probability value of each characteristic point on scale feature figure；Own if had in two or more characteristic patterns in three characteristic patterns
Characteristic point is all judged to norm, then it is assumed that the image category is norm, other situations then think that there are flaws in image；For differentiating
For there are the images of flaw, we correspond to some image block in original image using each characteristic point, pass through the prediction of characteristic point
The pixel value of class switch correspondence image block obtains relevant thermodynamic chart, and the corresponding thermodynamic chart of superposition various features figure obtains final
Thermodynamic chart, the approximate location of flaw is obtained by final thermodynamic chart, takes mathematical expectation of probability to be somebody's turn to do the image block near flaw
The classification of flaw, particularly, the algorithm can handle the case where there are more flaws in image, and obtain each flaw classification and
Approximate location.
2. a kind of efficient check algorithm for the nibs in big resolution ratio cloth image according to claim 1,
It is characterized in:
Training includes that step, transfer learning step are trained based on improved se-resnext101 model in the step (4)
Suddenly, two-stage learning rate set-up procedure, convolutional network extract characterization step, adaptively adjust feature weight step, is multiple dimensioned
Characteristic pattern processing step is calculated Focal Loss step and is walked using the gradient descent algorithm backpropagation training pattern with momentum
Suddenly.
3. a kind of efficient check algorithm for the nibs in big resolution ratio cloth image according to claim 2,
It is characterized in: the step (4) specifically:
(4.1) the last global pool layer of original se-resnext101 model is global by parallel characteristic block instead of 3
The characteristic block pond little module that pond layer and characteristic block maximum pond layer are constituted, each little module are at concurrency relation, often
The size of pond layer is identical in a little module but the size of the pond layer between disparate modules is different, in addition connects last entirely
1 1*1 size of layer is connect, step-length is replaced by 1 convolution operation, using the improved se-resnext101 model as inspection
Test model；
(4.2) obtained weight is trained to initialize modified version on ImageNets image set using se-resnext101 model
Se-resnext101 model, i.e. testing model, we only remain except all biasing weights, last global pool layer,
Weight outside last full articulamentum and softmax layers；
(4.3) learning rate for adjusting network when training pattern using two-stage learning rate, i.e., in the initial stage with some
Learning rate trains last three layers of the model including the module of characteristic block pond and keeps the weight of other layers of model constant,
Last three layers of model are made after several iteration cycles (all images in each iteration cycle traversal training set) of training
With a biggish learning rate, other layers use a lesser learning rate, and reduce study speed according to certain rule
Rate；
(4.4) training image is inputted in improved se-resnext101 model, extracts feature using convolution operation, increased special
The receptive field of figure is levied, while enabling network using the extruding and actuation sub-module that include in original se-resnext101 model
Enough adaptive adjustment feature weights, prominent validity feature inhibit invalid feature, so that feature space and two, feature channel dimension
Improved；
(4.5) 3 parallel characteristic block pond beggar's modules, pond in each little module are utilized to the characteristic pattern of last convolutional layer output
The size of change layer is identical but the size of the pond layer between disparate modules is different, to obtain 3 kinds of different size of characteristic patterns；
Utilizing size to the characteristic pattern of acquisition is 1*1, and then the convolution operation that step-length is 1 calculates 3 kinds of differences using softmax
The corresponding class probability value of each characteristic point on the characteristic pattern of size；
(4.6) due to the presence of receptive field, the characteristic point on characteristic pattern corresponds to the image block of original image, we are according to flaw in image
In position can know the true classification of each image block, so that the true classification of character pair point is obtained, according to prediction
Class probability value and true classification information calculate Classification Loss using Focal Loss function, finally utilize the ladder with momentum
It spends descent algorithm backpropagation and updates detection model parameter.
4. a kind of efficient check algorithm of nibs in big image in different resolution according to claim 1, feature
It is the step (5) specifically: test image is input in testing model, by propagated forward, receptive field is obtained and constantly increases
Greatly, the characteristic pattern that resolution ratio constantly reduces utilizes 3 parallel characteristic block Chi Hua little to the characteristic pattern of last convolutional layer output
Module is handled, and 3 kinds of different size of characteristic patterns are obtained, and carrying out size to this 3 kinds different size of characteristic patterns is 1*1, step
A length of 1 convolution operation finally utilizes the corresponding class probability of characteristic point each on each characteristic pattern of softmax operation acquisition
Value.If there is all characteristic points in two or more characteristic patterns to be all judged to norm in three characteristic patterns, then it is assumed that the image class
Not Wei norm, there are flaws in image is then thought for other situations.For there are the images of flaw, we will sentence in characteristic pattern
Not Wei the pixel value of the characteristic point image block that maps back original image of norm be assigned to 0, characteristic points other in characteristic pattern are mapped back
The pixel value of the image block of original image is assigned to 1, obtains 3 thermodynamic charts in this way, and 3 thermodynamic chart superpositions are obtained finally
Thermodynamic chart, obtains the approximate location of flaw from final thermodynamic chart, and the classification of flaw is by being determined as several characteristic points of flaw
Class probability mean value determine.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811336000.3A CN109509187B (en) | 2018-11-05 | 2018-11-05 | Efficient inspection algorithm for small defects in large-resolution cloth images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811336000.3A CN109509187B (en) | 2018-11-05 | 2018-11-05 | Efficient inspection algorithm for small defects in large-resolution cloth images |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109509187A true CN109509187A (en) | 2019-03-22 |
| CN109509187B CN109509187B (en) | 2022-12-13 |
Family
ID=65748135
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811336000.3A Active CN109509187B (en) | 2018-11-05 | 2018-11-05 | Efficient inspection algorithm for small defects in large-resolution cloth images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109509187B (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110009629A (en) * | 2019-04-12 | 2019-07-12 | 北京天明创新数据科技有限公司 | A kind of pneumoconiosis screening system and its data training method |
| CN110321805A (en) * | 2019-06-12 | 2019-10-11 | 华中科技大学 | A kind of dynamic expression recognition methods based on sequential relationship reasoning |
| CN110504029A (en) * | 2019-08-29 | 2019-11-26 | 腾讯医疗健康(深圳)有限公司 | A kind of medical image processing method, medical image recognition method and device |
| CN110610482A (en) * | 2019-08-12 | 2019-12-24 | 浙江工业大学 | Workpiece flaw detection method based on resnet |
| CN111028250A (en) * | 2019-12-27 | 2020-04-17 | 创新奇智(广州)科技有限公司 | Real-time intelligent cloth inspecting method and system |
| CN111325744A (en) * | 2020-03-06 | 2020-06-23 | 成都数之联科技有限公司 | Sample processing method in panel defect detection |
| CN111340783A (en) * | 2020-02-27 | 2020-06-26 | 创新奇智(广州)科技有限公司 | Real-time cloth defect detection method |
| CN112053357A (en) * | 2020-09-27 | 2020-12-08 | 同济大学 | FPN-based steel surface flaw detection method |
| CN112348776A (en) * | 2020-10-16 | 2021-02-09 | 上海布眼人工智能科技有限公司 | Fabric flaw detection method based on EfficientDet |
| CN113689381A (en) * | 2021-07-21 | 2021-11-23 | 航天晨光股份有限公司 | Detection model and detection method for flaws on inner wall of corrugated pipe |
| CN115797349A (en) * | 2023-02-07 | 2023-03-14 | 广东奥普特科技股份有限公司 | Defect detection method, device and equipment |
| CN116304811A (en) * | 2023-02-28 | 2023-06-23 | 王宇轩 | Dynamic sample weight adjustment method and system based on focus loss function |
| CN113808035B (en) * | 2021-08-25 | 2024-04-26 | 厦门微图软件科技有限公司 | Flaw detection method based on semi-supervised learning |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20010012381A1 (en) * | 1997-09-26 | 2001-08-09 | Hamed Sari-Sarraf | Vision-based, on-loom fabric inspection system |
| CN102331425A (en) * | 2011-06-28 | 2012-01-25 | 合肥工业大学 | Textile defect detection method based on defect enhancement |
| CN103456021A (en) * | 2013-09-24 | 2013-12-18 | 苏州大学 | Piece goods blemish detecting method based on morphological analysis |
| CN107316295A (en) * | 2017-07-02 | 2017-11-03 | 苏州大学 | A kind of fabric defects detection method based on deep neural network |
| US20180268257A1 (en) * | 2017-03-20 | 2018-09-20 | Rolls-Royce Plc | Surface defect detection |
| CN108647723A (en) * | 2018-05-11 | 2018-10-12 | 湖北工业大学 | A kind of image classification method based on deep learning network |
-
2018
- 2018-11-05 CN CN201811336000.3A patent/CN109509187B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20010012381A1 (en) * | 1997-09-26 | 2001-08-09 | Hamed Sari-Sarraf | Vision-based, on-loom fabric inspection system |
| CN102331425A (en) * | 2011-06-28 | 2012-01-25 | 合肥工业大学 | Textile defect detection method based on defect enhancement |
| CN103456021A (en) * | 2013-09-24 | 2013-12-18 | 苏州大学 | Piece goods blemish detecting method based on morphological analysis |
| US20180268257A1 (en) * | 2017-03-20 | 2018-09-20 | Rolls-Royce Plc | Surface defect detection |
| CN107316295A (en) * | 2017-07-02 | 2017-11-03 | 苏州大学 | A kind of fabric defects detection method based on deep neural network |
| CN108647723A (en) * | 2018-05-11 | 2018-10-12 | 湖北工业大学 | A kind of image classification method based on deep learning network |
Non-Patent Citations (1)
| Title |
|---|
| 张龙剑 等: "基于稀疏编码的双尺度布匹瑕疵检测", 《计算机应用》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110009629A (en) * | 2019-04-12 | 2019-07-12 | 北京天明创新数据科技有限公司 | A kind of pneumoconiosis screening system and its data training method |
| CN110321805A (en) * | 2019-06-12 | 2019-10-11 | 华中科技大学 | A kind of dynamic expression recognition methods based on sequential relationship reasoning |
| CN110610482A (en) * | 2019-08-12 | 2019-12-24 | 浙江工业大学 | Workpiece flaw detection method based on resnet |
| CN110504029B (en) * | 2019-08-29 | 2022-08-19 | 腾讯医疗健康(深圳)有限公司 | Medical image processing method, medical image identification method and medical image identification device |
| CN110504029A (en) * | 2019-08-29 | 2019-11-26 | 腾讯医疗健康(深圳)有限公司 | A kind of medical image processing method, medical image recognition method and device |
| CN111028250A (en) * | 2019-12-27 | 2020-04-17 | 创新奇智(广州)科技有限公司 | Real-time intelligent cloth inspecting method and system |
| CN111340783A (en) * | 2020-02-27 | 2020-06-26 | 创新奇智(广州)科技有限公司 | Real-time cloth defect detection method |
| CN111325744A (en) * | 2020-03-06 | 2020-06-23 | 成都数之联科技有限公司 | Sample processing method in panel defect detection |
| CN112053357A (en) * | 2020-09-27 | 2020-12-08 | 同济大学 | FPN-based steel surface flaw detection method |
| CN112348776A (en) * | 2020-10-16 | 2021-02-09 | 上海布眼人工智能科技有限公司 | Fabric flaw detection method based on EfficientDet |
| CN113689381A (en) * | 2021-07-21 | 2021-11-23 | 航天晨光股份有限公司 | Detection model and detection method for flaws on inner wall of corrugated pipe |
| CN113689381B (en) * | 2021-07-21 | 2024-02-27 | 航天晨光股份有限公司 | Corrugated pipe inner wall flaw detection model and detection method |
| CN113808035B (en) * | 2021-08-25 | 2024-04-26 | 厦门微图软件科技有限公司 | Flaw detection method based on semi-supervised learning |
| CN115797349A (en) * | 2023-02-07 | 2023-03-14 | 广东奥普特科技股份有限公司 | Defect detection method, device and equipment |
| CN116304811A (en) * | 2023-02-28 | 2023-06-23 | 王宇轩 | Dynamic sample weight adjustment method and system based on focus loss function |
| CN116304811B (en) * | 2023-02-28 | 2024-01-16 | 王宇轩 | Dynamic sample weight adjustment method and system based on focus loss function |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109509187B (en) | 2022-12-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109509187A (en) | A kind of efficient check algorithm for the nibs in big resolution ratio cloth image | |
| CN105512640B (en) | A kind of people flow rate statistical method based on video sequence | |
| Wu et al. | Automatic fabric defect detection using a wide-and-light network | |
| CN106295601B (en) | A kind of improved Safe belt detection method | |
| CN110033473B (en) | Moving target tracking method based on template matching and depth classification network | |
| CN109767427A (en) | The detection method of train rail fastener defect | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN104156734B (en) | A kind of complete autonomous on-line study method based on random fern grader | |
| CN108596203A (en) | Optimization method of the pond layer in parallel to pantograph carbon slide surface abrasion detection model | |
| CN111626176B (en) | Remote sensing target rapid detection method and system based on dynamic attention mechanism | |
| CN105825511A (en) | Image background definition detection method based on deep learning | |
| CN106780546B (en) | The personal identification method of motion blur encoded point based on convolutional neural networks | |
| CN109886947A (en) | The high-tension bus-bar defect inspection method of convolutional neural networks based on region | |
| CN106650823A (en) | Probability extreme learning machine integration-based foam nickel surface defect classification method | |
| CN112465057B (en) | Target detection and identification method based on deep convolutional neural network | |
| CN105930794A (en) | Indoor scene identification method based on cloud computing | |
| CN109509170A (en) | A kind of die casting defect inspection method and device | |
| CN109508741A (en) | Method based on deep learning screening training set | |
| CN106815323A (en) | A kind of cross-domain vision search method based on conspicuousness detection | |
| CN106355195A (en) | The system and method used to measure image resolution value | |
| CN112365497A (en) | High-speed target detection method and system based on Trident Net and Cascade-RCNN structures | |
| CN108460336A (en) | A kind of pedestrian detection method based on deep learning | |
| CN111340019A (en) | Grain bin pest detection method based on Faster R-CNN | |
| CN110363218A (en) | A kind of embryo's noninvasively estimating method and device | |
| Ning et al. | Research on surface defect detection algorithm of strip steel based on improved YOLOV3 |
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 | ||
| DD01 | Delivery of document by public notice |
Addressee: Chen Chucheng Document name: Notification of Qualified Procedures |
|
| DD01 | Delivery of document by public notice |