CN110120036A - A kind of multiple dimensioned tire X-ray defect detection method - Google Patents
A kind of multiple dimensioned tire X-ray defect detection method Download PDFInfo
- Publication number
- CN110120036A CN110120036A CN201910310408.1A CN201910310408A CN110120036A CN 110120036 A CN110120036 A CN 110120036A CN 201910310408 A CN201910310408 A CN 201910310408A CN 110120036 A CN110120036 A CN 110120036A
- Authority
- CN
- China
- Prior art keywords
- training
- model
- tire
- fpn
- image
- 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
- 230000007547 defect Effects 0.000 title claims abstract description 54
- 238000001514 detection method Methods 0.000 title claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 49
- 238000000034 method Methods 0.000 claims abstract description 20
- 238000012360 testing method Methods 0.000 claims abstract description 20
- 230000004927 fusion Effects 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 230000008569 process Effects 0.000 claims description 12
- 238000013507 mapping Methods 0.000 claims description 10
- 238000011176 pooling Methods 0.000 claims description 9
- 239000000284 extract Substances 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 3
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 3
- 238000005520 cutting process Methods 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims description 3
- 230000008676 import Effects 0.000 claims description 3
- 238000007689 inspection Methods 0.000 claims description 3
- 230000000007 visual effect Effects 0.000 claims description 3
- 241000208340 Araliaceae Species 0.000 claims description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 2
- 235000008434 ginseng Nutrition 0.000 claims description 2
- 238000010606 normalization Methods 0.000 claims description 2
- 238000013527 convolutional neural network Methods 0.000 description 14
- 230000006870 function Effects 0.000 description 14
- 201000010099 disease Diseases 0.000 description 4
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000000605 extraction Methods 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 229910000831 Steel Inorganic materials 0.000 description 3
- 230000007423 decrease Effects 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000003062 neural network model Methods 0.000 description 3
- 239000010959 steel Substances 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 241000220317 Rosa Species 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 230000003466 anti-cipated effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000003490 calendering Methods 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000023077 detection of light stimulus Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000007717 exclusion Effects 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000000465 moulding Methods 0.000 description 1
- 238000004073 vulcanization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10116—X-ray 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2210/00—Indexing scheme for image generation or computer graphics
- G06T2210/22—Cropping
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Quality & Reliability (AREA)
- Analysing Materials By The Use Of Radiation (AREA)
Abstract
The invention belongs to image recognitions and detection technique field, specifically disclose a kind of multiple dimensioned tire X-ray defect detection method.The step of detection method, is included data mark, image preprocessing, image cropping, builds FPN (feature pyramid network) fusion Faster R-CNN network model, initialization model, image data set is divided into training set, verifying collection and test set, being tested based on the test set, and test result is obtained.It is as follows that it is compared with the traditional method advantage: the present invention uses multiple dimensioned tire X-ray defect detection method, by extracting the feature of image multilayer, carries out multiple dimensioned fusion, the accuracy of detection tire X-ray defect can be improved；Target identification is carried out using the network model of FPN fusion Faster R-CNN, the accuracy of multiple scale detecting is greatly improved, to accurately be supervised to tire quality, there is Practical significance very much.
Description
Technical field
The invention belongs to computer vision and industrial detection technical field, specifically a kind of multiple dimensioned tire X-ray disease
Defect detection method.
Background technique
With the fast development of economy and society, tire industry plays more and more important role in life, from
Start within 2004, the tire yield in China ranks first in the world always.The wide variety of tire, is generally divided into oblique and meridian
Tire, radial are divided into half steel and all-steel radial tyre again, and all-steel radial tyre has complicated internal structure, right
Production process requires extremely harsh.It is unreasonable once to occur misoperation, molding, calendering and vulcanization size in production process, again
Ageing equipment in other words can all influence the remaining mass of tire.And the quality of tire the safety of automobile is played it is vital
Effect, so the control to tire quality is very crucial.
The important monitoring link of detection tire quality is exactly to shoot x-ray image to tire, is then differentiated according to X light image
Whether present tire has certain defect, be most initially by artificial cognition, proposed now using deep neural network model into
Row automatic discrimination can detecte position and the class of defect by the tire X-ray defect detection method based on neural network model
Type.
But not all deep neural network model all has good effect to the detection of tire X-ray defect,
Faster R-CNN (the real-time target detection based on region candidate network) network can position simultaneously defect and sort out disease
Defect type, but Faster R-CNN is to carry out feature extraction on obtained the last layer characteristic pattern, to carry out target knowledge
It is other.It is done so that there are the drawbacks of be, some information of wisp are had ignored in top-level feature, therefore only according to top layer
Feature carries out target identification, cannot completely reflect the information of Small object object.If can be in conjunction with the feature of multi-layer, so that it may
To greatly improve the accuracy of target detection.So needing a kind of fusion multilayer feature, multiple dimensioned tire X-ray defect detection
Method.
Summary of the invention
In view of above-mentioned, the present invention provides a kind of multiple dimensioned tire X-ray defect detection methods, can greatly improve defect
The accuracy of detection.
A kind of multiple dimensioned tire X-ray defect detection method is provided, is comprised the following steps that
S1, data mark: the tire X-ray check picture being collected into is labeled with LabelImg tool, marks out disease
Defect position, defect type, the defect position mark when marked with box, the defect type can there are many, mark file
Type be xml document；
S2, image preprocessing: the tire X-ray check picture is sharpened processing and obtains pretreated big picture；
S3, image cropping: by the pretreated big picture that size is 20000 × 1900 be divided into 11 1900 ×
1900 small figure, corresponding coordinate position is converted, and rewrites the xml document for recording the defect type and coordinate；
S4, FPN (feature pyramid network) fusion Faster R-CNN network model is built:；
S5, initialization model: setting parameter, the parameter include size unified after inputting picture, model in Search
The periodicity of the size of Selective (selective search) stage box and number, model training；
S6, image data set is divided into training set, verifying collection and test set: division principle is that the training set accounts for 70%,
The verifying collection and the test set respectively account for 15%；
S7, repeat the above steps S5, S6, can train to obtain multiple models, carries out test analysis to the multiple model,
It is tested based on the test set, obtains test result.
Further, Faster R-CNN network master mould mainly includes four parts in step S4:
1) Conv layers (convolutional layer) extracts characteristic pattern, uses the conv+relu+pooling (convolution on one group of basis
+ amendment linear unit+pond) layer extraction input picture feature maps (characteristic pattern), which is used in
Subsequent RPN layers and full articulamentum；
2) RPN (Region Proposal Networks, regional choice network), for generating region proposals
(candidate region) firstly generates a pile Anchor box (anchor box), (is normalized after cutting filtering is carried out to it by softmax
Exponential function) judge that anchors (anchor) belongs to prospect (foreground) or background (background)；Meanwhile Ling Yifen
Branch bounding box regression (frame recurrence) corrects anchor box, forms more accurate proposal (candidate
Frame)；
3) Roi Pooling (candidate pool area) layer, using RPN generate the proposals (candidate frame) and
The feature map that VGG16 (visual graphics generator 16) the last layer obtains, obtains the proposal of fixed size
Feature map (candidate frame characteristic pattern), target identification and positioning can be carried out using full attended operation below by entering；
4) Classifier (classifier) forms described Roi Pooling layers the feature map of fixed size
Full attended operation is carried out, the classification of specific category is carried out using Softmax, meanwhile, it is complete using L1 Loss (L1 loss function)
The exact position that operation obtains object is returned at bounding box regression (frame recurrence).
Further, FPN (feature pyramid network) concrete operation method in step S4 are as follows:
1) bottom-up path is made of multiple convolution modules, and each convolution module includes multiple convolutional layers, the bottom of from
In upward process, Spatial Dimension halves by module, and the output of each convolution module will make in top-down path
With；
2) top-down path, FPN use the Convolution Filter of a 1x1 by the channel depth of uppermost convolution module
256 dimensions are down to, image M5 is obtained, then obtain image P5 using the convolution of a 3x3, described image P5 is for target prediction
First Feature Mapping；Down along top-down path, FPN up-samples layer application arest neighbors before, meanwhile, FPN
1x1 convolution is applied to the individual features mapping in bottom-up access, then application point element is added, and obtains mesh using 3x3 convolution
Mark the Feature Mapping of detection.
Further, include: the training for carrying out model based on the training set in step S6, parameter is carried out based on verifying collection
Adjustment, the training carries out successive ignition, and the adjustment of parameter configuration is carried out using K cross validation method, and the training arrives
Certain period needs to check whether the current parameter configuration is correct, the specific steps are as follows:
1) in the training process of the model, the damage of the model above in the training set and the verifying collection is obtained
Lose the functional value of function；
2) after the training to certain period, the training is temporarily ceased, and current model is preserved, side
Just continue the training after；
3) the loss function value of the training set and the verifying collection is drawn, horizontal axis is periodicity, and the longitudinal axis is the loss
Functional value, whether the loss function value observed on the training set and the verifying collection is a correct downward trend；
If 4) be in step 3) in correct downward trend, need not adjustment parameter, at this moment import and saved in step 2)
The model continue training until reach model convergence；Otherwise it enters step 5)；
5) if the functional value of loss function is not in correct downward trend, reason is just found, and adjusts the parameter,
The parameter returns in step S5 after determining.
The beneficial effects of the present invention are:
(1) present invention uses multiple dimensioned tire X-ray defect detection method, by extracting the feature of image multilayer, carries out
The accuracy of detection tire X-ray defect can be improved in multiple dimensioned fusion；
(2) using the network model of FPN fusion Faster R-CNN, Faster R-CNN provides Target detection and identification
Frame, but it is to carry out feature extraction on obtained the last layer characteristic pattern, to carry out target identification.But in this way
Do there are the drawbacks of be, some information of wisp are had ignored in top-level feature, thus only according to top-level feature carry out target
Identification, cannot completely reflect the information of Small object object.If can be in conjunction with the feature of multi-layer, so that it may greatly improve more
The accuracy of size measurement, FPN realize this point, so the network model of FPN fusion Faster R-CNN is highly suitable for X
The detection of light defect；
(3) for tire as the important spare part on automobile, the quality of production is related to the life security of people, tire X-ray disease
Defect differentiates last one of the critical point detected as tire quality, and the domestic quality inspection personnel that generallys use at present carries out quality surveillance, this
Kind mode has low efficiency, and at high cost, reliability is low and has the shortcomings that damage to quality inspection personnel eyes.The present invention passes through more rulers
The tire X-ray defect detection method of degree detects automatically differentiates tire defect, to accurately be supervised to tire quality, it is non-often with
It is of practical meaning.
Detailed description of the invention
The drawings described herein are used to provide a further understanding of the present application, constitutes part of this application, this Shen
Illustrative embodiments and their description please are not constituted an undue limitation on the present application for explaining the application.In the accompanying drawings:
Fig. 1 is the multiple dimensioned tire X-ray defect testing process schematic diagram of one kind of the invention.
Fig. 2 is FPN fusion Faster R-CNN model structure of the invention.
Fig. 3 is FPN data flow architecture figure of the invention.
Fig. 4 is FPN concrete operations flow diagram of the invention.
Specific embodiment
Presently filed embodiment is described in detail below in conjunction with accompanying drawings and embodiments, how the application is applied whereby
Technological means solves technical problem and reaches the realization process of technical effect to fully understand and implement.
This example is intended to implement to obtain by this method a multiple dimensioned tire X-ray defect detection system.The realization process
Including collect tire X-ray check picture, picture is labeled, pre-process picture, FPN merge Faster R-CNN model take
Build with training, model loading, the specific implementation process is as follows:
(1) data mark, and mark x-ray image with LabelImg tool, mark out position, the type of defect, mark file
Type is xml document, and when marking defect, position is marked with box, and the labeling form of type is A, B, C, D......, K, specifically
Type is shown in Table 1:
1 defect type declaration of table
(2) image preprocessing.Original image is sharpened processing, original image has unsharp place, uses tool
Processing is sharpened to these parts, makes that picture lines is clearer, defect part is more prominent；
(3) image cropping.The original image that size is 20000 × 1900 is divided into 11 1900 × 1900 small figures, it will
Corresponding coordinate position is converted, and the xml document of record defect type and coordinate is rewritten；
(4) FPN fusion Faster R-CNN network model is built, Faster R-CNN network master mould mainly includes four
Point:
1) Conv layers (convolutional layer) extracts characteristic pattern: as a kind of CNN network objectives detection method, Faster R-
CNN uses conv+relu+pooling (convolution+amendment linear unit+pond) layer on one group of basis to extract input picture first
Feature maps (characteristic pattern), which is used in subsequent RPN layers and full articulamentum；
2) RPN (Region Proposal Networks, regional choice network): RPN network is mainly used for generating
Region proposals (candidate region) firstly generates a pile Anchor box (anchor box), leads to after cutting filtering is carried out to it
It crosses softmax (normalization exponential function) and judges that anchors (anchor) belongs to prospect (foreground) or background
It (background), be object or is not object, so this is one two classification；Meanwhile another branch bounding box
Regression (frame recurrence) corrects anchor box, forms more accurate proposal (candidate frame) (note: here more smart
Really for the regression of box again of full articulamentum below)；
3) Roi Pooling (candidate pool area) layer: this layer using RPN generate proposals (candidate frame) and
The feature map that VGG16 (visual graphics generator 16) the last layer obtains, obtains the proposal of fixed size
Feature map (candidate frame characteristic pattern), target identification and positioning can be carried out using full attended operation below by entering；
4) Classifier (classifier): the feature map that Pooling layers of Roi form fixed size can be carried out
Full attended operation carries out the classification of specific category using Softmax, meanwhile, it is completed using L1 Loss (L1 loss function)
Bounding box regression (frame recurrence) returns the exact position that operation obtains object.
It is based on ResNet that above-mentioned FPN, which is merged the Conv layers partial replacement in Faster R-CNN network model,
The FPN network of (depth residual error network), replaced model structure are as shown in Fig. 2.
FPN is made of bottom-up and top-down two paths.Bottom-up path is common extraction feature
Convolutional network (uses ResNet network) herein.Bottom-up, spatial resolution is successively decreased, and detects more Multistory-tall building, network layer
Semantic values are increase accordingly.Top-down path, based on the semantic more rich higher layer of layer building resolution ratio.Although rebuilding
Layer is semantic abundant enough, but passes through these down-samplings and upper sampling process, and the position of target is no longer accurate.Therefore FPN is in weight
Lateral connection is increased between build-up layers and corresponding Feature Mapping, to help detector that position is better anticipated.These lateral connections
Play the role of jump connection (skip connection) (way of similar residual error network) simultaneously.FPN data flow architecture
As shown in Fig. 3.
FPN network concrete operations are as shown in Fig. 4, comprising: 1) bottom-up path is made of, often many convolution modules
A module includes many convolutional layers.In bottom-up process, Spatial Dimension halves (step-length is double) by module.Each convolution mould
The output of block will use in top-down path.2) top-down path, in attached drawing 4, FPN uses the convolution of a 1x1
The channel depth of C5 (uppermost convolution module) is down to 256 dimensions by filter, obtains M5.Then the convolution of a 3x3 is applied
P5 is obtained, P5 is exactly first Feature Mapping for being used for target prediction.Down along top-down path, FPN is to before
Layer is using arest neighbors up-sampling (x2).Meanwhile FPN applies 1x1 convolution to the individual features mapping in bottom-up access.Then
It is added using a point element.It is last equally to obtain the Feature Mapping of target detection using 3x3 convolution.This filter alleviates up-sampling
Aliasing effect.This process stops after P2, because the Spatial Dimension of C1 is too high.If do not stopped, follow an example to do to obtain P1
If, slow-motion journey can be dragged significantly.
(5) initialization model sets parameter, including size (in the present embodiment, picture unified after input picture
Long side can be arbitrary dimension, but short is 600pix on one side), model is at Search Selective (selective search)
The periodicity of the size of stage box and number, model training；
(6) image data set is divided into training set, verifying collection and test set.Division principle is training set 70%, verifying collection
15% is respectively accounted for test set.The distribution that the distribution of defect keeps every kind of defect original in training set, test set and verifying are concentrated
Meet the distribution of original defect type and grade.The training that model is carried out based on training set is collected based on verifying and carries out parameter
Adjustment, training carry out successive ignition, and the adjustment of model parameter is carried out using K cross validation method.Certain period is arrived in training,
Need to check whether current parameter configuration is correct, the specific steps are as follows:
1) in the training process of model, the functional value of the loss function of model above in training set and verifying collection is obtained.
2) after training to certain period, training is temporarily ceased, and current model is preserved, it is convenient with subsequent
Continuous training.
3) functional value of training set and verifying collection loss function is drawn, horizontal axis is periodicity, and the longitudinal axis is the letter of loss function
Numerical value；Loss function value may not decline, it is possible to decline it is slow, it is possible to begin to decline and rose instead later, need
Whether the loss function value observed on training set and verifying collection is a correct downward trend.
If 4) be in step 3) in correct downward trend, need not adjustment parameter, at this moment import and saved in step 2)
Model continue training until reach model convergence；Otherwise it enters step 5).
5) if the functional value of loss function is not in the downward trend of health, possible reason is just found, and adjust ginseng
Number, parameter return to above-mentioned steps (5) after determining.
(7) repeat the above steps (5), (6), can train to obtain multiple models, carries out test analysis, base to multiple models
It is tested in test set, test result cannot be used as with MAP (Mean Average Precision, bat) and commented
Valence index, because we do not need, and defect position is especially accurate, and emphasis is not in this practical problem of tire X-ray check
Defect can be missed.
The center of gravity of evaluation result is put both ways, and first aspect is the accuracy of normal picture and the judgement of defect picture,
It is defect picture which can be sorted out by, which first having to,；Second Problem is the tire X-ray in the case where picture is defect picture
Picture belongs to any specific defect type.Therefore we use measurement index be mainly normal picture recall ratio with look into
The precision ratio and recall ratio of the specific defect type and grade discrimination of quasi- rate and defect picture.
The recall rate, precision rate for comparing all models is selected in recall ratio and precision ratio all higher (recall ratio and Cha Zhun
Rate is both greater than model 85%) as last model, and such model is the relatively good model of generalization ability；If without this
The model of sample then illustrates that model is unavailable, needs to return to step (5) initialization model, and re -training model.
It should also be noted that, the terms "include", "comprise" or its any other variant are intended to nonexcludability
Include, so that commodity or system including a series of elements not only include those elements, but also including not clear
The other element listed, or further include for this commodity or the intrinsic element of system.In the feelings not limited more
Under condition, the element that is limited by sentence "including a ...", it is not excluded that in the commodity or system for including the element also
There are other identical elements.
Several preferred embodiments of the invention have shown and described in above description, but as previously described, it should be understood that the present invention
Be not limited to forms disclosed herein, should not be regarded as an exclusion of other examples, and can be used for various other combinations,
Modification and environment, and the above teachings or related fields of technology or knowledge can be passed through within that scope of the inventive concept describe herein
It is modified.And changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, then it all should be in this hair
In the protection scope of bright appended claims.
Claims (4)
1. a kind of multiple dimensioned tire X-ray defect detection method, comprises the following steps that
S1, data mark: the tire X-ray check picture being collected into is labeled with LabelImg tool, marks out defect position
Set, defect type, the defect position mark when marked with box, the defect type can there are many, mark the class of file
Type is xml document；
S2, image preprocessing: the tire X-ray check picture is sharpened processing and obtains pretreated big picture；
S3, image cropping: the pretreated big picture that size is 20000 × 1900 is divided into 11 1900 × 1900
Small figure converts corresponding coordinate position, rewrites the xml document for recording the defect type and coordinate；
S4, FPN (feature pyramid network) fusion Faster R-CNN network model is built:；
S5, initialization model: setting parameter, the parameter include size unified after inputting picture, model in Search
The periodicity of the size of Selective (selective search) stage box and number, model training；
S6, image data set is divided into training set, verifying collection and test set: division principle is that the training set accounts for 70%, described
Verifying collection and the test set respectively account for 15%；
S7, repeat the above steps S5, S6, can train to obtain multiple models, carries out test analysis to the multiple model, is based on
The test set is tested, and test result is obtained.
2. multiple dimensioned tire X-ray defect detection method according to claim 1, which is characterized in that in step S4
Faster R-CNN network master mould mainly includes four parts:
1) Conv layers (convolutional layer) extracts characteristic pattern, uses the conv+relu+pooling (convolution+amendment on one group of basis
Linear unit+pond) layer extract input picture feature maps (characteristic pattern), which is used in subsequent
RPN layers and full articulamentum；
2) RPN (Region Proposal Networks, regional choice network) (is waited for generating region proposals
Favored area), a pile Anchor box (anchor box) is firstly generated, passes through softmax (normalization index after cutting filtering is carried out to it
Function) judge that anchors (anchor) belongs to prospect (foreground) or background (background)；Meanwhile another branch
Bounding box regression (frame recurrence) corrects anchor box, forms more accurate proposal (candidate frame)；
3) Roi Pooling (candidate pool area) layer, the proposals (candidate frame) and VGG16 generated using RPN
The feature map that (visual graphics generator 16) the last layer obtains obtains the proposal feature of fixed size
Map (candidate frame characteristic pattern), target identification and positioning can be carried out using full attended operation below by entering；
4) Classifier (classifier) carries out the feature map that described Roi Pooling layers form fixed size
Full attended operation carries out the classification of specific category using Softmax, meanwhile, it is completed using L1 Loss (L1 loss function)
Bounding box regression (frame recurrence) returns the exact position that operation obtains object.
3. multiple dimensioned tire X-ray defect detection method according to claim 1 or 2, which is characterized in that in step S4
FPN (feature pyramid network) concrete operation method are as follows:
1) bottom-up path is made of multiple convolution modules, and each convolution module includes multiple convolutional layers, bottom-up
During, Spatial Dimension halves by module, and the output of each convolution module will use in top-down path；
2) the channel depth of uppermost convolution module is down to by top-down path, FPN using the Convolution Filter of a 1x1
256 dimensions, obtain image M5, then obtain image P5 using the convolution of a 3x3, and described image P5 is used for the first of target prediction
A Feature Mapping；Down along top-down path, FPN up-samples layer application arest neighbors before, meanwhile, FPN is to certainly
1x1 convolution is applied in individual features mapping in the upward access in bottom, and then application point element is added, and obtains target inspection using 3x3 convolution
The Feature Mapping of survey.
4. multiple dimensioned tire X-ray defect detection method according to claim 1, which is characterized in that include: in step S6
The training that model is carried out based on the training set collects the adjustment for carrying out parameter based on verifying, and the training carries out successive ignition, and
The adjustment of parameter configuration is carried out using K cross validation method, the training needs to check the current ginseng to certain period
Whether number configuration is correct, the specific steps are as follows:
1) in the training process of the model, the loss letter of the model above in the training set and the verifying collection is obtained
Several functional values；
2) after the training to certain period, temporarily cease the training, and current model is preserved, it is convenient with
After continue the training；
3) the loss function value of the training set and the verifying collection is drawn, horizontal axis is periodicity, and the longitudinal axis is the loss function
Value, whether the loss function value observed on the training set and the verifying collection is a correct downward trend；
If 4) be in step 3) in correct downward trend, need not adjustment parameter, at this moment import the institute saved in step 2)
It states model and continues training until reaching model convergence；Otherwise it enters step 5)；
5) if the functional value of loss function is not in correct downward trend, reason is just found, and adjusts the parameter, it is described
Parameter returns in step S5 after determining.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910310408.1A CN110120036A (en) | 2019-04-17 | 2019-04-17 | A kind of multiple dimensioned tire X-ray defect detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910310408.1A CN110120036A (en) | 2019-04-17 | 2019-04-17 | A kind of multiple dimensioned tire X-ray defect detection method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110120036A true CN110120036A (en) | 2019-08-13 |
Family
ID=67521121
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910310408.1A Pending CN110120036A (en) | 2019-04-17 | 2019-04-17 | A kind of multiple dimensioned tire X-ray defect detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110120036A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110660049A (en) * | 2019-09-16 | 2020-01-07 | 青岛科技大学 | Tire defect detection method based on deep learning |
| CN110852558A (en) * | 2019-09-24 | 2020-02-28 | 连云港耀科铝业有限公司 | Computer-aided quality control method for hub production data |
| CN111179239A (en) * | 2019-12-24 | 2020-05-19 | 浙江大学 | Tire X-ray flaw detection method for performing re-ranking by using background features |
| CN111460926A (en) * | 2020-03-16 | 2020-07-28 | 华中科技大学 | Video pedestrian detection method fusing multi-target tracking clues |
| CN111476165A (en) * | 2020-04-07 | 2020-07-31 | 同方赛威讯信息技术有限公司 | Method for detecting fingerprint characteristics of title seal in electronic document based on deep learning |
| CN111582073A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation violation identification method based on ResNet101 characteristic pyramid |
| CN112132804A (en) * | 2020-09-22 | 2020-12-25 | 苏州巨能图像检测技术有限公司 | Anti-lifting detection method for hub of hub truck |
| CN112150434A (en) * | 2020-09-22 | 2020-12-29 | 霍尔果斯奇妙软件科技有限公司 | Tire defect detection method, device, equipment and storage medium |
| CN112418170A (en) * | 2020-12-11 | 2021-02-26 | 法赫光学科技(成都)有限公司 | Oral examination and identification method based on 3D scanning |
| CN112613375A (en) * | 2020-12-16 | 2021-04-06 | 中国人寿财产保险股份有限公司 | Tire damage detection and identification method and device |
| CN112668531A (en) * | 2021-01-05 | 2021-04-16 | 重庆大学 | Motion posture correction method based on motion recognition |
| CN112686152A (en) * | 2020-12-30 | 2021-04-20 | 广西慧云信息技术有限公司 | Crop pest and disease identification method with multi-size input and multi-size targets |
| CN112819748A (en) * | 2020-12-16 | 2021-05-18 | 机科发展科技股份有限公司 | Training method and device for strip steel surface defect recognition model |
| CN113486782A (en) * | 2021-07-02 | 2021-10-08 | 苏州雷泰医疗科技有限公司 | Method, system and storage medium for privacy protection monitoring of radiotherapy patient |
| CN114078106A (en) * | 2020-08-06 | 2022-02-22 | 沈阳中科数控技术股份有限公司 | Defect detection method based on improved Faster R-CNN |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108288075A (en) * | 2018-02-02 | 2018-07-17 | 沈阳工业大学 | A kind of lightweight small target detecting method improving SSD |
| CN108564563A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of tire X-ray defect detection method based on Faster R-CNN |
| CN108734705A (en) * | 2018-05-17 | 2018-11-02 | 杭州电子科技大学 | Digital galactophore fault image calcification clusters automatic testing method based on deep learning |
| CN108764462A (en) * | 2018-05-29 | 2018-11-06 | 成都视观天下科技有限公司 | A kind of convolutional neural networks optimization method of knowledge based distillation |
| CN109376681A (en) * | 2018-11-06 | 2019-02-22 | 广东工业大学 | A kind of more people's Attitude estimation method and system |
-
2019
- 2019-04-17 CN CN201910310408.1A patent/CN110120036A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108288075A (en) * | 2018-02-02 | 2018-07-17 | 沈阳工业大学 | A kind of lightweight small target detecting method improving SSD |
| CN108564563A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of tire X-ray defect detection method based on Faster R-CNN |
| CN108734705A (en) * | 2018-05-17 | 2018-11-02 | 杭州电子科技大学 | Digital galactophore fault image calcification clusters automatic testing method based on deep learning |
| CN108764462A (en) * | 2018-05-29 | 2018-11-06 | 成都视观天下科技有限公司 | A kind of convolutional neural networks optimization method of knowledge based distillation |
| CN109376681A (en) * | 2018-11-06 | 2019-02-22 | 广东工业大学 | A kind of more people's Attitude estimation method and system |
Non-Patent Citations (11)
| Title |
|---|
| JIANGMIAO PANG等: "Libra R-CNN: Towards Balanced Learning for Object Detection", 《ARXIV:1904.02701V1 [CS.CV]》 * |
| LIXIN WANG等: "Vehicle Detection Based on Drone Images with the Improved Faster R-CNN", 《RESEARCHGATE》 * |
| SHAOQING REN等: "Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks", 《ARXIV:1506.01497V3 [CS.CV]》 * |
| TSUNG-YI LIN等: "Feature Pyramid Networks for Object Detection", 《ARXIV:1612.03144V2 [CS.CV]》 * |
| 周孝佳等: "基于分块的有遮挡人脸识别算法", 《计算机应用与软件》 * |
| 宋兰霞: "《新一代多媒体系统中的关键技术》", 31 December 2018 * |
| 张杰等: "复杂环境的车牌定位方法设计与实现", 《通信技术》 * |
| 王飞: "基于深度学习的行人检测算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 许婷婷等: "基于联合分类器的非自然图像检索", 《计算机应用与软件》 * |
| 贾永红等: "《数字图像处理技巧》", 31 January 2017 * |
| 郭文强等: "《数字图像处理》", 31 May 2009 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110660049A (en) * | 2019-09-16 | 2020-01-07 | 青岛科技大学 | Tire defect detection method based on deep learning |
| CN110852558A (en) * | 2019-09-24 | 2020-02-28 | 连云港耀科铝业有限公司 | Computer-aided quality control method for hub production data |
| CN111179239B (en) * | 2019-12-24 | 2022-04-26 | 浙江大学 | Tire X-ray flaw detection method for performing re-ranking by using background features |
| CN111179239A (en) * | 2019-12-24 | 2020-05-19 | 浙江大学 | Tire X-ray flaw detection method for performing re-ranking by using background features |
| CN111460926A (en) * | 2020-03-16 | 2020-07-28 | 华中科技大学 | Video pedestrian detection method fusing multi-target tracking clues |
| CN111460926B (en) * | 2020-03-16 | 2022-10-14 | 华中科技大学 | Video pedestrian detection method fusing multi-target tracking clues |
| CN111476165A (en) * | 2020-04-07 | 2020-07-31 | 同方赛威讯信息技术有限公司 | Method for detecting fingerprint characteristics of title seal in electronic document based on deep learning |
| CN111582073A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation violation identification method based on ResNet101 characteristic pyramid |
| CN114078106B (en) * | 2020-08-06 | 2024-08-02 | 沈阳中科数控技术股份有限公司 | Defect detection method based on improved Faster R-CNN |
| CN114078106A (en) * | 2020-08-06 | 2022-02-22 | 沈阳中科数控技术股份有限公司 | Defect detection method based on improved Faster R-CNN |
| CN112132804B (en) * | 2020-09-22 | 2023-10-31 | 苏州巨能图像检测技术有限公司 | Anti-lifting detection method for hub of hub card |
| CN112150434A (en) * | 2020-09-22 | 2020-12-29 | 霍尔果斯奇妙软件科技有限公司 | Tire defect detection method, device, equipment and storage medium |
| CN112132804A (en) * | 2020-09-22 | 2020-12-25 | 苏州巨能图像检测技术有限公司 | Anti-lifting detection method for hub of hub truck |
| CN112418170A (en) * | 2020-12-11 | 2021-02-26 | 法赫光学科技(成都)有限公司 | Oral examination and identification method based on 3D scanning |
| CN112418170B (en) * | 2020-12-11 | 2024-03-01 | 法赫光学科技(成都)有限公司 | 3D scanning-based oral examination and identification method |
| CN112613375A (en) * | 2020-12-16 | 2021-04-06 | 中国人寿财产保险股份有限公司 | Tire damage detection and identification method and device |
| CN112819748A (en) * | 2020-12-16 | 2021-05-18 | 机科发展科技股份有限公司 | Training method and device for strip steel surface defect recognition model |
| CN112613375B (en) * | 2020-12-16 | 2024-05-14 | 中国人寿财产保险股份有限公司 | Tire damage detection and identification method and equipment |
| CN112819748B (en) * | 2020-12-16 | 2023-09-19 | 机科发展科技股份有限公司 | Training method and device for strip steel surface defect recognition model |
| CN112686152A (en) * | 2020-12-30 | 2021-04-20 | 广西慧云信息技术有限公司 | Crop pest and disease identification method with multi-size input and multi-size targets |
| CN112686152B (en) * | 2020-12-30 | 2023-06-09 | 广西慧云信息技术有限公司 | Crop pest identification method with multi-size input and multi-size targets |
| CN112668531A (en) * | 2021-01-05 | 2021-04-16 | 重庆大学 | Motion posture correction method based on motion recognition |
| CN113486782A (en) * | 2021-07-02 | 2021-10-08 | 苏州雷泰医疗科技有限公司 | Method, system and storage medium for privacy protection monitoring of radiotherapy patient |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110120036A (en) | A kind of multiple dimensioned tire X-ray defect detection method | |
| CN102854191B (en) | Real-time visual detection and identification method for high speed rail surface defect | |
| CN110378869B (en) | Steel rail fastener abnormity detection method with automatic sample marking function | |
| CN109344905A (en) | A kind of transmission facility automatic fault recognition methods based on integrated study | |
| CN108711148B (en) | Tire defect intelligent detection method based on deep learning | |
| CN113763312B (en) | Detection of defects in semiconductor samples using weak labels | |
| CN109064461A (en) | A kind of detection method of surface flaw of steel rail based on deep learning network | |
| CN111489339B (en) | Method for detecting backup and main defects of high-speed railway locator bolts | |
| CN109086756A (en) | A kind of text detection analysis method, device and equipment based on deep neural network | |
| CN102332089B (en) | Railway wagon brake shoe key going-out fault recognition method based on artificial neural network | |
| CN105574550A (en) | Vehicle identification method and device | |
| CN106408015A (en) | Road fork identification and depth estimation method based on convolutional neural network | |
| CN110136101A (en) | A kind of tire X-ray defect detection method compared based on twin distance | |
| CN103324937A (en) | Method and device for labeling targets | |
| CN114399672A (en) | Railway wagon brake shoe fault detection method based on deep learning | |
| CN108564563A (en) | A kind of tire X-ray defect detection method based on Faster R-CNN | |
| CN105260734A (en) | Commercial oil surface laser code recognition method with self modeling function | |
| CN111862029A (en) | Fault detection method for bolt part of vertical shock absorber of railway motor train unit | |
| CN109858414A (en) | A kind of invoice piecemeal detection method | |
| CN108596203A (en) | Optimization method of the pond layer in parallel to pantograph carbon slide surface abrasion detection model | |
| CN113506269B (en) | Turnout and non-turnout rail fastener positioning method based on deep learning | |
| CN110246134A (en) | A kind of rail defects and failures sorter | |
| CN110032935A (en) | A kind of traffic signals label detection recognition methods based on deep learning cascade network | |
| CN111353377A (en) | Elevator passenger number detection method based on deep learning | |
| CN110544253A (en) | fabric flaw detection method based on image pyramid and column template |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20211203 Address after: 310000 room 604-605, building 16, No. 57, Science Park, Baiyang street, Hangzhou Economic and Technological Development Zone, Zhejiang Province Applicant after: HANGZHOU YINGGE INFORMATION TECHNOLOGY CO.,LTD. Address before: 310026 station 9, room 1213-1218, building 17, No. 57, kekekeyuan Road, Baiyang street, Hangzhou Economic and Technological Development Zone, Zhejiang Province Applicant before: Hangzhou Data Point Gold Technology Co.,Ltd. |
|
| TA01 | Transfer of patent application right | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190813 |
|
| RJ01 | Rejection of invention patent application after publication |