CN106097360A - A kind of strip steel surface defect identification method and device - Google Patents
A kind of strip steel surface defect identification method and device Download PDFInfo
- Publication number
- CN106097360A CN106097360A CN201610439036.9A CN201610439036A CN106097360A CN 106097360 A CN106097360 A CN 106097360A CN 201610439036 A CN201610439036 A CN 201610439036A CN 106097360 A CN106097360 A CN 106097360A
- Authority
- CN
- China
- Prior art keywords
- feature
- glcm
- image
- hog
- defect
- 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
- 229910000831 Steel Inorganic materials 0.000 title claims abstract description 186
- 239000010959 steel Substances 0.000 title claims abstract description 186
- 230000007547 defect Effects 0.000 title claims abstract description 160
- 238000000034 method Methods 0.000 title claims abstract description 66
- 238000000605 extraction Methods 0.000 claims abstract description 19
- 239000000284 extract Substances 0.000 claims abstract description 12
- 238000004422 calculation algorithm Methods 0.000 claims description 62
- 230000004927 fusion Effects 0.000 claims description 55
- 238000007637 random forest analysis Methods 0.000 claims description 43
- 238000003066 decision tree Methods 0.000 claims description 42
- 238000001914 filtration Methods 0.000 claims description 21
- 238000010606 normalization Methods 0.000 claims description 13
- 230000002068 genetic effect Effects 0.000 claims description 8
- 230000006872 improvement Effects 0.000 claims description 5
- 238000012549 training Methods 0.000 claims description 4
- 238000007635 classification algorithm Methods 0.000 claims description 2
- 230000008569 process Effects 0.000 description 7
- 230000008901 benefit Effects 0.000 description 5
- 238000005286 illumination Methods 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 4
- 238000001514 detection method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- 238000013139 quantization Methods 0.000 description 2
- 230000031068 symbiosis, encompassing mutualism through parasitism Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000007850 degeneration Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 239000012467 final product Substances 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000007620 mathematical function Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4007—Scaling of whole images or parts thereof, e.g. expanding or contracting based on interpolation, e.g. bilinear interpolation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Quality & Reliability (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a kind of strip steel surface defect identification method and device, wherein, the method includes: set up defect recognition model;Obtain belt steel surface picture;Belt steel surface picture is carried out pretreatment;Extract histograms of oriented gradients HOG feature and the gray level co-occurrence matrixes GLCM feature of the image of pretreated belt steel surface picture;By defect recognition model, HOG feature and the GLCM feature of extraction are identified, determine the defect information of belt steel surface.After realizing by the defect recognition model set up, the various features extracted being identified, the defect information of image can be fast and accurately identified, substantially increase the discrimination of defect, the accuracy of the defect identified is the highest, and can effectively overcome geometry, optics invariance and rotatory that characteristics of image describes the problem brought, the discrimination for area-type defect is the highest.
Description
Technical field
The present invention relates to technical field of image detection, particularly relate to a kind of strip steel surface defect identification method and device.
Background technology
At present, the defect that one of band hardness of steel and the topmost factor of quality are belt steel surfaces is affected.Therefore, to strip steel table
The defect in face carries out detecting becomes a very important link in steel manufacture process.
Currently, when the defect of belt steel surface is detected, first extract the image of belt steel surface, then by carrying
The image taken carries out process and determines defect information, wherein, in processing the image process extracted, first passes through the filter of neighborhood averaging value
The image extracted is filtered by ripple;Subsequently extract characteristics of image by gray level co-occurrence matrixes GLCM;Finally use based on 1-
Defect is identified by category support vector machines SVM and RBF RBF, determines RBF by the method for double web search
Parameter, the final position positioning defect.
The method of the defect of above-mentioned detection belt steel surface, for single feature extraction can not overcome illumination invariant,
The problems such as rotatory are the lowest to area-type defect recognition rate, and due to the defect kind up to 39 kinds of strip steel, area-type defect reaches
About 15 kinds, so, the lowest to steel strip surface defect discrimination, it is difficult to the demand meeting enterprise to detection.
Summary of the invention
In view of the above problems, it is proposed that the present invention in case provide one overcome the problems referred to above or at least in part solve on
State a kind of strip steel surface defect identification method and the device of problem.
According to one aspect of the present invention, it is provided that a kind of strip steel surface defect identification method, described method includes:
Set up defect recognition model;
Obtain belt steel surface picture;
Belt steel surface picture is carried out pretreatment;
Extract histograms of oriented gradients HOG feature and the gray scale symbiosis square of the image of pretreated belt steel surface picture
Battle array GLCM feature;
By described defect recognition model, described HOG feature and the described GLCM feature of extraction are identified, determine
The defect information of belt steel surface.
Alternatively, described belt steel surface picture is carried out pretreatment, including:
The image denoising of described belt steel surface picture is filtered;
Use piecewise linear transform, the image of the belt steel surface picture after noise-removed filtering is carried out grey level enhancement;
Use bilinear interpolation algorithm that the image of the belt steel surface picture after grey level enhancement is zoomed in and out and normalization.
Alternatively, described set up defect recognition model, including:
Obtain strip steel samples pictures;
The histograms of oriented gradients HOG feature and the gray level co-occurrence matrixes GLCM that extract the image of strip steel samples pictures are special
Levy;
The HOG feature of described image and the GLCM feature of image are merged by the method using stepping statistics, obtain
Many group weight factors and the recognition accuracy that often group weight factor is corresponding;
Described many group weight factors and described recognition accuracy are carried out Gauss curve fitting, obtains feature fitting function;
By genetic algorithm, described feature fitting function is solved, obtain fusion factor;
According to described fusion factor, random forests algorithm is improved, obtain Feature Fusion based on random forest classification
Algorithm;
By described Feature Fusion sorting algorithm based on random forest to extract image described HOG characteristics of image and
GLCM feature is trained, and determines the defect information of described strip steel sample;
Defect information according to described strip steel sample sets up defect recognition model.
Alternatively, described according to described fusion factor, random forests algorithm is improved, obtain based on random forest
Feature Fusion sorting algorithm includes:
According to decision tree corresponding in described fusion factor reconstruct random forests algorithm, obtain reconstructing decision tree, described heavy
Structure decision tree includes HOG decision tree and GLCM decision tree;
According to described reconstruct decision tree, described HOG feature and described GLCM feature are chosen, obtain HOG feature certainly
Plan tree group and GLCM feature decision-making array;
According to described HOG decision-making array, the HOG feature of the image of described strip steel samples pictures is identified, obtains HOG
Biometric data, and according to described GLCM decision-making array, the GLCM feature of the image of described strip steel samples pictures is known
Not, GLCM biometric data is obtained;
Described HOG biometric data and described GLCM biometric data are added up, obtains statistical classification result;
According to the principle of random forests algorithm ballot, by feature the highest for number of votes obtained in described statistical classification result, determine
Defect recognition result for described strip steel sample;
The algorithm performing above procedure is defined as Feature Fusion sorting algorithm based on random forest.
Alternatively, the HOG feature of described image and the GLCM feature of image are entered by the method for described utilization stepping statistics
Row merges, including:
The weight of HOG feature, the weight of GLCM feature and weight constraints condition are set;
Step-length is added up according to described weight constraints condition;
Described step-length according to statistics determines the weighted value span of HOG feature, and determines the weight of GLCM feature
Value span;
Weighted value span according to described HOG feature and the weighted value span of GLCM feature, be calculated
Many group weight factors and the recognition accuracy that often group weight factor is corresponding.
According to another aspect of the present invention, it is provided that a kind of steel strip surface defect identification device, described device includes:
Set up module, be used for setting up defect recognition model;
Acquisition module, is used for obtaining belt steel surface picture;
Pretreatment module, for carrying out pretreatment to belt steel surface picture;
Extraction module, for extracting the histograms of oriented gradients HOG feature of the image of pretreated belt steel surface picture
And gray level co-occurrence matrixes GLCM feature;
Identification module, for entering described HOG feature and the described GLCM feature of extraction by described defect recognition model
Row identifies, determines the defect information of belt steel surface.
Alternatively, described pretreatment module includes:
Filter unit, for filtering the image denoising of described belt steel surface picture;
Converter unit, is used for using piecewise linear transform, and the image of the belt steel surface picture after noise-removed filtering is carried out ash
Degree strengthens;
Normalization unit, for using bilinear interpolation algorithm to carry out the image of the belt steel surface picture after grey level enhancement
Scaling and normalization.
Alternatively, described module of setting up includes:
Acquiring unit, is used for obtaining strip steel samples pictures;
Extraction unit, for extracting HOG feature and the GLCM feature of the image of strip steel samples pictures;
Integrated unit, for using method that stepping the adds up HOG feature to described image and the GLCM feature of image
Merge, obtain organizing weight factor more and often organizing the recognition accuracy that weight factor is corresponding;
Fitting unit, for described many group weight factors and described recognition accuracy are carried out Gauss curve fitting, obtains feature
Fitting function;
Computing unit, for being solved described feature fitting function by genetic algorithm, obtains fusion factor;
Improve unit, for random forests algorithm being improved according to described fusion factor, obtain based on random forest
Feature Fusion sorting algorithm;
Training unit, for by described Feature Fusion sorting algorithm based on random forest to extract image described in
HOG characteristics of image and GLCM feature are trained, and determine the defect information of described strip steel sample;
Set up unit, for setting up defect recognition model according to the defect information of described strip steel sample.
Alternatively, described improvement unit includes:
Reconstruct subelement, for according to decision tree corresponding in described fusion factor reconstruct random forests algorithm, obtaining weight
Structure decision tree, described reconstruct decision tree includes HOG decision tree and GLCM decision tree;
Choose subelement, for described HOG feature and described GLCM feature being selected according to described reconstruct decision tree
Take, obtain HOG feature decision tree group and GLCM feature decision-making array;
Identify subelement, for the HOG feature of the image of described strip steel samples pictures being entered according to described HOG decision-making array
Row identifies, obtains HOG biometric data, and according to described GLCM decision-making array to the image of described strip steel samples pictures
GLCM feature is identified, and obtains GLCM biometric data;
First statistics subelement, for uniting to described HOG biometric data and described GLCM biometric data
Meter, obtains statistical classification result;
First determines subelement, for the principle according to random forests algorithm ballot, will obtain in described statistical classification result
The feature that poll is the highest, is defined as the defect recognition result of described strip steel sample;The algorithm performing above procedure is defined as base
Feature Fusion sorting algorithm in random forest.
Alternatively, described integrated unit includes:
Subelement is set, for arranging the weight of HOG feature, the weight of GLCM feature and weight constraints condition;
Second statistics subelement, for adding up step-length according to described weight constraints condition;
Second determines subelement, for determining the weighted value span of HOG feature according to the described step-length of statistics, and
Determine the weighted value span of GLCM feature;
Computation subunit, for taking according to the weighted value span of described HOG feature and the weighted value of GLCM feature
Value scope, is calculated many group weight factors and often organizes the recognition accuracy that weight factor is corresponding.
A kind of strip steel surface defect identification method of present invention offer and device, it is achieved by the defect recognition model set up
After the various features extracted is identified, the defect information of image can be fast and accurately identified, substantially increase defect
Discrimination, the accuracy of the defect identified is the highest, and can effectively overcome geometry, optics invariance and rotatory to figure
The problem brought as feature description, the discrimination for area-type defect is the highest.
Accompanying drawing explanation
By reading the detailed description of hereafter preferred implementation, various other advantage and benefit common for this area
Technical staff will be clear from understanding.Accompanying drawing is only used for illustrating the purpose of preferred implementation, and is not considered as the present invention
Restriction.And in whole accompanying drawing, it is denoted by the same reference numerals identical parts.In the accompanying drawings:
Fig. 1 is the flow chart of a kind of strip steel surface defect identification method of one embodiment of the present invention;
Fig. 2 is the flow chart setting up defect recognition model of one embodiment of the present invention;
Fig. 3 is the flow chart merging feature of one embodiment of the present invention;
Fig. 4 is the flow chart obtaining Feature Fusion sorting algorithm based on random forest of one embodiment of the present invention;
Fig. 5 is the flow chart that steel strip surface defect carries out pretreatment of one embodiment of the present invention;
Fig. 6 is the schematic diagram of a kind of steel strip surface defect identification device of one embodiment of the present invention.
Detailed description of the invention
With embodiment, embodiments of the present invention are described in further detail below in conjunction with the accompanying drawings.Following example are used for
The present invention is described, but can not be used for limiting the scope of the present invention.
Fig. 1 is the flow chart of a kind of strip steel surface defect identification method of one embodiment of the present invention.See Fig. 1, should
Method includes following S101-S105 step.
In embodiments of the present invention, before the defect of belt steel surface is identified, need to set up defect recognition model,
It is identified, so that it is determined that go out the defect of belt steel surface by the characteristics of image of this defect recognition model belt steel surface to extracting
Information.
Step S101: set up defect recognition model.
Above-mentioned defect recognition model of setting up specifically can be set up, as shown in Figure 2 by following S1011-S1018 step.
Step S1011: obtain strip steel samples pictures.
Above-mentioned strip steel samples pictures is the belt steel surface picture of existing defects, wherein, according to the difference of defect type, permissible
Obtain every kind of belt steel surface picture corresponding to defect type, using belt steel surface picture corresponding to number of drawbacks type that obtain as
Strip steel samples pictures.
Step S1012: extract histograms of oriented gradients HOG feature and the gray scale symbiosis square of the image of strip steel samples pictures
Battle array GLCM feature.
In embodiments of the present invention, when extracting the characteristics of image of strip steel model picture, need to extract strip steel samples pictures
HOG (Histogram of Oriented Gradient, histograms of oriented gradients) feature and GLCM (Gray-Level
Co-occurrence Matrix, gray level co-occurrence matrixes) feature.
Above-mentioned HOG feature and GLCM feature can describe the feature of the image of belt steel surface in detail, all sidedly, by carrying
Take HOG feature and the GLCM feature of belt steel surface, can be efficiently against the illumination invariant of the image of belt steel surface and rotation
The problem of turning property.
Step S1013: use the method for stepping statistics that HOG feature and the GLCM feature of image are merged, obtain
Many group weight factors and the recognition accuracy that often group weight factor is corresponding.
In embodiments of the present invention, it is preferable that use the method for stepping statistics that HOG feature and GLCM feature are melted
Closing, concrete is merged, as shown in Figure 3 by following S10131-S10134 step.
Step S10131: the weight of HOG feature, the weight of GLCM feature and weight constraints condition are set;
Step S10132: add up step-length according to described weight constraints condition;
Step S10133: determine the weighted value span of HOG feature according to the described step-length of statistics, and determine GLCM
The weighted value span of feature;
Step S10134: according to weighted value span and the weighted value value model of GLCM feature of described HOG feature
Enclose, be calculated many group weight factors and often organize the recognition accuracy that weight factor is corresponding.
In embodiments of the present invention, the weighted value of HOG feature can be set to w1, the weighted value of GLCM feature is set to w2,
The scope [0,10] of total weighted value is divided into N number of decile, and w1,w2∈ [0,10], and ensure w1+w2=10.The step-length then added up
Being (10-0)/N, if N=10, then step-length is 1, and get final product correspondence obtains 11 groups of weighted values, the weight that wherein HOG feature is corresponding
Value w1Value be respectively (0,1,2 ..., 10), and the weighted value w of GLCM feature2Correspond to w1Value be respectively (10,9,
8 ..., 0), and to define weight factor be θ=w1/(w2+w1) 100%, thus obtain organizing weight factor more.
In embodiments of the present invention, after two kinds of features are merged, obtain organizing θ more, much more i.e. to organize weight factor, according to
The different integration percentages of stepping statistics are often organized the recognition accuracy that weight factor is corresponding.
In embodiments of the present invention, the method describing Operator Fusion by HOG and GLCM both textural characteristics, effectively
Overcome single features and describe the defect of operator so that the feature after fusion had both possessed the ability that details describes, and possessed again whole
The ability that body describes.HOG feature and GLCM feature are merged, can very well overcome geometry, optics invariance and
Rotatory describes the problem brought to characteristics of image, such that it is able to the difference at large described between various complicated defect, with
Reaching to identify the purpose of classification, defect recognition rate is the highest, and the accuracy of the defect identified is the highest.
Step S1014: many group weight factors and recognition accuracy are carried out Gauss curve fitting, obtains feature fitting function.
In embodiments of the present invention, feature fitting function obtained above is that recognition accuracy f (θ) is about weight factor θ
Function.Wherein, this feature fitting function f (θ) is as shown in following formula (24).
In formula (24), f (θ) is recognition accuracy, and θ is weight factor, a1=0.672, b1=0.7831, c1=
0.1293, a2=0.04927, b2=0.4083, c2=0.05036, a3=0.5622, b3=0.5956, c3=0.15, a4=
0.3254, b4=0.3844, c4=0.1819, a5=0.8072, b5=0.04857, c5=0.4316, a6=0.7621, b6
=0.9892, c6=0.09563, a7=0.2256, b7=0.8753, c7=0.05123.
Step S1015: by genetic algorithm, feature fitting function is solved, obtain fusion factor.
Solving feature fitting function above by genetic algorithm, concrete solution procedure is as follows:
First, to individuality each in features described above fitting function with binary coding, and utilization is uniformly distributed at the beginning of method definition
Beginning population, arranges initial weight factor θ.
Then, using feature fitting function f (θ) as fitness function, calculate fitness, and with f (θ) as optimization aim,
Replicate outstanding individuality, eliminate inferior position individuality, and retain optimum individual, thus according to the fitness calculated to above-mentioned initial power
Repeated factor θ screens, and determines the weight factor that individual feature that fitness is the strongest is corresponding, by the strongest for this fitness
Weight factor corresponding to the feature of body is as fusion factor.In embodiments of the present invention, fitness can be with the identification calculated
Accuracy rate represents, weight factor corresponding when recognition accuracy is the highest, be fitness the strongest time corresponding weight factor.
Above procedure is carried out the iteration of preset times, in iterative process, if iterations reaches preset times, the most defeated
Go out above-mentioned fusion factor, if iterations is not up to preset times, then continue iteration.
In embodiments of the present invention, above-mentioned default iterations is 100 times, i.e. can be obtained by 100 iteration merging because of
Son.
Step S1016: improve random forests algorithm according to fusion factor, obtains the based on random forest of improvement
Feature Fusion sorting algorithm.
Above-mentioned according to fusion factor, random forests algorithm is improved, obtain Feature Fusion based on random forest classification
Algorithm, specifically can be obtained, as shown in Figure 4 by following S10161-S10166 step.
Step S10161: according to decision tree corresponding in fusion factor reconstruct random forests algorithm, obtain reconstructing decision tree,
Reconstruct decision tree includes HOG decision tree and GLCM decision tree;
In embodiments of the present invention, above-mentioned reconstruct decision tree includes two parts, and a part is HOG decision tree, another part
For GLCM decision tree, wherein, the quantity for the decision tree of HOG characteristic allocation is 100 × θ, for the decision tree of GLCM characteristic allocation
Quantity be 100 × (1-θ).
Step S10162: choose HOG feature and GLCM feature according to reconstruct decision tree, obtains HOG feature certainly
Plan tree group and GLCM feature decision-making array;
Step S10163: according to HOG decision-making array, the HOG feature of the image of strip steel samples pictures is identified, obtains
HOG biometric data, and according to GLCM decision-making array, the GLCM feature of the image of strip steel samples pictures is identified,
To GLCM biometric data;
Step S10164: add up HOG biometric data and GLCM biometric data, obtains statistical classification knot
Really;
Step S10165: according to the principle of random forests algorithm ballot, by spy the highest for number of votes obtained in statistical classification result
Levy, be defined as the defect recognition result of strip steel sample;
Step S10166: the algorithm performing above procedure is defined as Feature Fusion sorting algorithm based on random forest.
Fusion HOG feature that above-mentioned Feature Fusion sorting algorithm based on random forest can equalize and GLCM feature, and
And the integration percentage of two kinds of features can be optimized according to the result identified, to reach good Feature Fusion, fully send out
Wave HOG feature and Defect Edge is described the advantages such as careful, played GLCM feature the most simultaneously and the globality of defect is held, statistics
The advantages such as the defect texture variations of entire image.Feature Fusion sorting algorithm based on random forest can extraordinary fusion two
Plant characteristic superiority, reach well to identify classifying quality, improve the accuracy rate identified.
Step S1017: the characteristics of image extracted is trained by Feature Fusion sorting algorithm based on random forest,
Determine the defect information of strip steel samples pictures.
Step S1018: set up defect recognition model according to the defect information of strip steel samples pictures.
After setting up defect recognition model, when need treat survey belt steel surface defect detect time, need first pass through with
Lower S102 step obtains belt steel surface picture.
Step S102: obtain belt steel surface picture.
After obtaining belt steel surface picture, during due to shooting picture, there is environmental disturbances, at the figure extracting belt steel surface picture
As, before feature, needing, by following S103 step, belt steel surface picture is carried out pretreatment.
Step S103: belt steel surface picture is carried out pretreatment.
Above-mentioned belt steel surface picture is carried out pretreatment specifically can be processed by following S1031-S1033 step, as
Shown in Fig. 5.
Step S1031: the image denoising of belt steel surface picture is filtered.
In embodiments of the present invention, main by band steel surface image denoising being filtered based on grey level histogram medium filtering
Ripple, concrete noise-removed filtering process is as follows:
It is first provided with one-dimensional sequence S1,S2,S3,S4,…,Sn, and take the Square Neighborhood of the image of belt steel surface picture
A length of m=2k+1;
Then, by this above-mentioned one-dimensional sequence of medium filtering, i.e. by slip Square Neighborhood, in succession extract out from this one-dimensional sequence
Continuous print m number, then the neighborhood obtaining sliding is Si-k,Si-k+1,Si-k+2,…,Si+k, m number obtained above is arranged size
After, taking the value that size is middle is filtered output data.The mathematic(al) representation such as following formula (1) of these filtered output data
Shown in:
yi=Med (Si-k,Si-k+1,Si-k+2,…,Si+k)i∈N (1)
In formula (1), yiFor filtered output data, Si-kFor the one-dimensional sequence of image, N is neighborhood window.
When needs are to two-dimensional sequence { XijWhen carrying out medium filtering, the neighborhood of filtering is also two dimension, but the neighborhood of this two dimension
Can be various different shapes, the most square, rhombus, cross, annular, M shape etc..The medium filtering of two dimensional image
After output data mathematic(al) representation such as following formula (2) shown in:
In formula (2), YijFor filtered output data, XijFor the two-dimensional sequence of image, N is neighborhood window.
Finally, the grey level histogram of statistical picture, determine image intensity value concentrated area according to rectangular histogram, be the image back of the body
The scope of scape gray value.Wherein, the scope of image background gray value specifically can be added up by procedure below and determine:
First, each gray value correspondence frequency of occurrences of statistical picture, solve maximum frequency values fmax.Then, with
fmax/ 2 is threshold value, if the frequency of occurrences that image intensity value is correspondingThen this gray value of image will not be filtered by intermediate value
The replacement values that ripple is tried to achieve is removed, and is to maintain constant;If the frequency of occurrences that image intensity value is correspondingThen this figure
As this gray value can be replaced.Wherein, image intensity value can be by calculating with following formula (3).
In formula (3), YijFor image intensity value, XijFor image sequence, N is neighborhood window,For each gray value pair of image
Answer the frequency of occurrences, fmaxFor maximum frequency values.
Above-mentioned fine based on grey level histogram medium filtering filtering and noise reduction effect, the most at utmost remain strip steel table simultaneously
The information of the former defect image in face.
Step S1032: use piecewise linear transform, the image of the belt steel surface picture after noise-removed filtering is carried out gray scale increasing
By force.
Above-mentioned employing piecewise linear transform, when the image of the belt steel surface picture after noise-removed filtering is carried out grey level enhancement,
Specifically can be by carrying out the enhancing of image intensity value with following formula (4).
In formula (4), (x, y) is input picture to f, g (x, y) output of piecewise linear transform, x1, x2It is respectively gradation of image
The interval end points of stretching, to gradation of image at [x1,x2] interval is extended, and interval [0, x1] and interval [x2,255]
Gray scale is shunk.Wherein x1 is 15 to 30, and x2 is 35 to 45, and y1=is 190 to 210, and y2 is 210 to 230.
Step S1033: use bilinear interpolation algorithm that the image of the belt steel surface picture after grey level enhancement is zoomed in and out
And normalization.
In embodiments of the present invention, enhanced for above-mentioned image intensity value picture can be carried out bilinearity difference arithmetic contracting
Put to a size of: wide a height of 64 pixels take advantage of 64 pixels.Then the image intensity value after scaling is normalized.
After belt steel surface picture being carried out pretreatment by above S103 step, can be extracted pre-by following S104 step
The characteristics of image of the belt steel surface picture after process.
Step S104: extract HOG feature and the GLCM feature of the image of pretreated belt steel surface picture.
Above-mentioned HOG feature and GLCM feature can describe the feature of the image of belt steel surface in detail, all sidedly, by carrying
Take HOG feature and the GLCM feature of belt steel surface, can be efficiently against the illumination invariant of the image of belt steel surface and rotation
The problem of turning property.
During the HOG feature of the image of said extracted belt steel surface picture, specifically can be extracted by procedure below:
First, image level direction and the gradient of vertical direction are calculated, according to image level direction and the ladder of vertical direction
Degree, calculate pixel in image (x, gradient G y) (x, y) and each location of pixels gradient direction value α (x, y), wherein,
Can be by the gradient G with following formula (5) calculated level directionx(x, y), can be by calculating the gradient G of vertical direction with following formula (6)y
(x, y), can by calculating with following formula (7) (x, gradient G y) (x, y), can by with following formula (8) calculate gradient direction value α (x,
y);
Gx(x, y)=H (x+1, y)-H (x-1, y) (5)
In formula (5), Gx(x y), represents pixel in input picture (x, y) the horizontal direction gradient at place, vertical direction ladder
Degree, (x+1, y) (x-1 y) represents pixel (x, y) pixel value at place in input picture to H with H;
Gy(x, y)=H (x, y+1)-H (x, y-1) (6)
In formula (6), Gy(x y) represents pixel (x, y) vertical gradient at place, H (x, y+1) and H in input picture
(x, y-1) represents pixel (x, y) pixel value at place in input picture;
Then, dividing the image into 64 " cell factory ", the size of each cell factory is 8 × 8 pixels, thin at one
In born of the same parents' unit, the histogram of gradients in 9 directions of statistics, is namely divided into 9 direction blocks by the gradient direction 360 degree of cell factory.
Finally, cell factory being combined into big block, normalized gradient rectangular histogram in block, the size of block is 2 × 2 cells
Unit, chooses normalization factor L2-norm,||v||2Represent the 2 rank norms of v, represent one very with ε
Little constant, each cell factory has 9 features, and each piece has 4 cell factory, and step-length is 8 pixels, so each row and column
Having 7 blocks, one has 1764 HOG features.
During the GLCM feature of the image of said extracted belt steel surface picture, specifically can be extracted by procedure below:
In order to be more clearly understood that the GLCM feature of the image extracting belt steel surface picture that the embodiment of the present invention provides
Process, first simply introduces the gray level co-occurrence matrixes of image.Wherein, the gray level co-occurrence matrixes of image can be by following mistake
Cheng Shengcheng:
If in image, any point coordinate is (x, y) and deviation (x, y) other some pixel coordinates of certain distance
For (x+a, y+b), wherein, (x, y) corresponding gray value is i, and the gray value that (x+a, y+b) point is corresponding is j, i.e. (x, y) right
The pixel value g answeredi∈ G, the pixel value g that (x+a, y+b) point is correspondingj∈ G, wherein, G is grey level quantization layer collection.The value making a and b is solid
Fixed the most constant, and point (x, y) travels through whole image, i.e. available corresponding (i, j) value.Owing to grey level quantization layer collection G contains H ash
Degree level, so the combining form of i Yu j just has H2Kind;Add up all (g in whole imagei,gj) corresponding frequency of occurrence, and use
The form of matrix represents, finally with (i, j) total frequency of occurrence normalization obtain each (i, probability P j) occurred (i, j, d,
θ), then the gray level co-occurrence matrixes P (i, j, d, θ) of image is obtainedH×H。
The essence of gray level co-occurrence matrixes is exactly the joint histogram of a pixel pair, according to different deviation values a and b, i.e.
Available corresponding deflection is θ, and distance isGray level co-occurrence matrixes.
If understanding that gray level co-occurrence matrixes, gray level co-occurrence matrixes refer to that gray value is as i in image from the angle of mathematics
Pixel (x, y) is starting point, statistics with this some distance beGray value is the pixel (x+a, y+b) of j, institute
Have probability density P (i, j, d, θ) simultaneously occurred, wherein, the gray level co-occurrence matrixes P (i, j, d, θ) represented with mathematical function as
Shown in following formula (9):
P (i, j, d, θ)=[(x, y), (x+a, y+b) | f (x, y)=i;F (x+a, y+b)=j] } (9)
In formula (9), θ is the direction of gray level co-occurrence matrixes, i.e. the direction of two pixel lines and the angle of transverse axis, generally
Take (0 °, 35 °, 90 °, 135 °) four direction, different deviation values a and b can get different gray level co-occurrence matrixes;(a,b)
Value be typically based on the texture of image and intensity variation to carry out value, if the texture variations of image is relatively slow,
Value on its gray level co-occurrence matrixes diagonal is relatively large, preferably uses diagonal to be distributed, i.e. the value of a and b is less;If image
Quickly, the matrix value of its diagonal both sides increases texture variations, preferably uses and is uniformly distributed, i.e. the value of a and b is bigger.
In embodiments of the present invention, the gray level co-occurrence matrixes of the image of belt steel surface picture can be generated by above method,
When, after the gray level co-occurrence matrixes generating image, the gray level co-occurrence matrixes of the image by generating extracts the textural characteristics of image, its
In, the textural characteristics of the image of extraction include texture contrast, maximum of probability, texture correlation, entropy, average and, variance, variance
With, unfavourable balance square, the variance of difference and entropy, difference entropy, cluster shade, notable cluster, angle second moment.Rotation for the sample that overcomes one's shortcomings
Turning property, the texture of 4 directions (0 °, 45 °, 90 ° 135 °) the composition tranining database extracting each of the above textural characteristics respectively is special
Levy, thus obtain 56 vector characteristics, wherein, specifically can be by calculating correspondence with following formula (10)-formula (23) during extraction
Textural characteristics.Wherein, can be able to pass through to calculate texture contrast with following formula (11) by calculating angle second-order matrix with following formula (10),
Can be by calculating texture correlation with following formula (12), can be by calculating entropy with following formula (13), can be by with following formula (14) calculating side
Difference, can by with following formula (15) calculate average and, can by with following formula (16) calculate variance and, can be by with following formula (17) unfavourable balance
Square calculates, can be by calculating the variance of difference with following formula (18), can be by calculating and entropy with following formula (19), can be by with following formula (20)
Calculate difference entropy, can be by calculating notable cluster with following formula (21), can be by calculating cluster shade with following formula (22), can be by following
Formula (23) calculates maximum of probability.
In formula (14), μ is P (i, j, d, θ) average.
After extracting the HOG feature of image of belt steel surface picture and GLCM feature, can be known by following S105 step
Do not go out the defect information of belt steel surface.
Step S105: by defect recognition model, HOG feature and the GLCM feature of extraction are identified, determine strip steel
The defect information on surface.
In embodiments of the present invention, the method describing Operator Fusion by HOG and GLCM both textural characteristics, effectively
Overcome single features and describe the defect of operator so that the feature after fusion had both possessed the ability that details describes, and possessed again whole
The ability that body describes.HOG feature and GLCM feature are merged, can very well overcome geometry, optics invariance and
Rotatory describes the problem brought to characteristics of image, such that it is able to the difference at large described between various complicated defect, with
Reaching to identify the purpose of classification, defect recognition rate is the highest, and the accuracy of the defect identified is the highest.
Drawbacks described above information describe in detail belt steel surface exist defect, by drawbacks described above information can quickly,
Accurately determine out the various defects of belt steel surface, substantially increase discrimination and the accuracy rate of defect, and can effectively overcome
Geometry, optics invariance and rotatory describe the problem brought to characteristics of image, for the identification ability of area-type defect
The strongest.
In embodiments of the present invention, this strip steel surface defect identification method discrimination to the area-type defect of belt steel surface
The highest.Wherein, area-type defect is the complexity defect including multiple characteristics of image.
The strip steel surface defect identification method that the embodiment of the present invention provides, it is achieved by the defect recognition model of foundation to carrying
After the various features taken is identified, the defect information of image can be fast and accurately identified, substantially increase the knowledge of defect
Not rate, the accuracy of the defect identified is the highest, and can effectively overcome geometry, optics invariance and rotatory special to image
Levying the problem that description is brought, the discrimination for area-type defect is the highest.
Fig. 6 is the schematic diagram of a kind of steel strip surface defect identification device of one embodiment of the present invention.See Fig. 6, should
Device includes:
Set up module S1, be used for setting up defect recognition model;
Acquisition module S2, is used for obtaining belt steel surface picture;
Pretreatment module S3, for carrying out pretreatment to belt steel surface picture;
Extraction module S4, special for extracting the histograms of oriented gradients HOG of the image of pretreated belt steel surface picture
Levy and gray level co-occurrence matrixes GLCM feature;
Identification module S5, for being identified HOG feature and the GLCM feature of extraction by defect recognition model, is determined
Go out the defect information of belt steel surface.
In embodiments of the present invention, before the defect of belt steel surface is identified, need to set up module by above-mentioned
S1 sets up defect recognition model, is processed by the characteristics of image of this defect recognition model belt steel surface to extracting, thus
Identify the defect information of belt steel surface.
Above-mentioned module S1 of setting up includes acquiring unit, extraction unit, integrated unit, fitting unit, computing unit, improvement list
Unit, training unit and set up unit.
Acquiring unit, is used for obtaining strip steel samples pictures.
Above-mentioned strip steel samples pictures is the belt steel surface picture of existing defects, wherein, according to the difference of defect type, above-mentioned
Acquiring unit can obtain every kind of belt steel surface picture corresponding to defect type, by strip steel corresponding for the number of drawbacks type of acquisition
Surface picture is as strip steel samples pictures.Extraction unit, for extracting HOG feature and the GLCM of the image of strip steel samples pictures
Feature.
In embodiments of the present invention, said extracted unit, when extracting the characteristics of image of strip steel model picture, needs to extract
The HOG feature of strip steel samples pictures and GLCM feature.
Above-mentioned HOG feature and GLCM feature can describe the feature of the image of belt steel surface in detail, all sidedly, by carrying
Take HOG feature and the GLCM feature of belt steel surface, can be efficiently against the illumination invariant of the image of belt steel surface and rotation
The problem of turning property.
Integrated unit, is carried out the HOG feature of image and the GLCM feature of image for the method using stepping to add up
Merge, obtain organizing weight factor more and often organizing the recognition accuracy that weight factor is corresponding.
Above-mentioned integrated unit includes:
Subelement is set, for arranging the weight of HOG feature, the weight of GLCM feature and weight constraints condition;
Second statistics subelement, for adding up step-length according to weight constraints condition;
Second determines subelement, for determining the weighted value span of HOG feature according to the step-length of statistics, and determines
The weighted value span of GLCM feature;
Computation subunit, for the weighted value span according to HOG feature and the weighted value value model of GLCM feature
Enclose, be calculated many group weight factors and often organize the recognition accuracy that weight factor is corresponding.
In embodiments of the present invention, it is preferable that integrated unit uses the method for stepping statistics to HOG feature and GLCM spy
Levy and merge, specifically can be by the fusion side provided in a kind of strip steel surface defect identification method of offer in above-described embodiment
Method merges, and does not repeats them here.
In embodiments of the present invention, the method describing Operator Fusion by HOG and GLCM both textural characteristics, effectively
Overcome single features and describe the defect of operator so that the feature after fusion had both possessed the ability that details describes, and possessed again whole
The ability that body describes.HOG feature and GLCM feature are merged, can very well overcome geometry, optics invariance and
Rotatory describes the problem brought to characteristics of image, such that it is able to the difference at large described between various complicated defect, with
Reaching to identify the purpose of classification, defect recognition rate is the highest, and the accuracy of the defect identified is the highest.
Fitting unit, for many group weight factors and recognition accuracy are carried out Gauss curve fitting, obtains feature fitting function.
In embodiments of the present invention, the feature fitting function that above-mentioned fitting unit obtains is that recognition accuracy f (θ) is about power
The function of repeated factor θ.Wherein, a kind of strip steel surface defect identification method provided in this fitting function such as above-described embodiment carries
The fitting function of confession, does not repeats them here.
Computing unit, for being solved feature fitting function by genetic algorithm, obtains fusion factor.
Feature fitting function is solved by above-mentioned computing unit by genetic algorithm, specifically can be by providing in above-described embodiment
A kind of strip steel surface defect identification method in the process that solves of feature fitting function solve, do not repeat them here.
Improve unit, for random forests algorithm being improved according to fusion factor, obtain spy based on random forest
Levy integrated classification algorithm.
Above-mentioned improvement unit includes:
Reconstruct subelement, for according to decision tree corresponding in fusion factor reconstruct random forests algorithm, obtaining reconstruct certainly
Plan tree, reconstruct decision tree includes HOG decision tree and GLCM decision tree;
Choose subelement, for HOG feature and GLCM feature being chosen according to reconstruct decision tree, obtain HOG special
Levy decision tree group and GLCM feature decision-making array;
Identify subelement, for the HOG feature of the image of strip steel samples pictures being identified according to HOG decision-making array,
Obtain HOG biometric data, and according to GLCM decision-making array, the GLCM feature of the image of strip steel samples pictures is known
Not, GLCM biometric data is obtained;
First statistics subelement, for adding up HOG biometric data and GLCM biometric data, is united
Meter classification results;
First determines subelement, for the principle according to random forests algorithm ballot, by number of votes obtained in statistical classification result
The highest feature, is defined as the defect recognition result of strip steel sample;The algorithm performing above procedure is defined as based on the most gloomy
The Feature Fusion sorting algorithm of woods.Training unit, is used for by Feature Fusion sorting algorithm based on random forest extraction
The described HOG characteristics of image of image and GLCM feature are trained, and determine the defect information of described strip steel sample;
Set up unit, for setting up defect recognition model according to the defect information of described strip steel sample.
Set up after module S1 sets up defect recognition model when above-mentioned, when belt steel surface is detected by needs, above-mentioned
Acquisition module S2 obtains the picture of belt steel surface to be measured.Environmental disturbances is there is, in said extracted module S4 during due to shooting picture
Before extracting the characteristics of image of belt steel surface picture, need, by above-mentioned pretreatment module S3, belt steel surface picture is carried out pre-place
Reason.
Above-mentioned pretreatment module S3 includes:
Filter unit, for filtering the image denoising of belt steel surface picture;
Converter unit, is used for using piecewise linear transform, and the image of the belt steel surface picture after noise-removed filtering is carried out ash
Degree strengthens;
Normalization unit, for using bilinear interpolation algorithm to carry out the image of the belt steel surface picture after grey level enhancement
Scaling and normalization.
Above-mentioned filter unit, can be by the filtering in a kind of strip steel surface defect identification method of offer in above-described embodiment
Denoising method, carries out noise-removed filtering to the image of belt steel surface picture;Above-mentioned converter unit, can be by providing in above-described embodiment
A kind of strip steel surface defect identification method in alternative approach, the image of the belt steel surface picture after noise-removed filtering is carried out ash
Degree strengthens;Above-mentioned normalization unit, can be by the contracting in a kind of strip steel surface defect identification method of offer in above-described embodiment
Put and method for normalizing, the image of the belt steel surface picture after grey level enhancement is zoomed in and out and normalization;The most superfluous at this
State.
After the image of belt steel surface picture is processed by above-mentioned pretreatment module S3, can by above-mentioned identification module S5
To identify the defect information of belt steel surface.The each of belt steel surface can be determined quickly and accurately by drawbacks described above information
Plant defect, substantially increase discrimination and the accuracy rate of defect, and can effectively overcome geometry, optics invariance and rotatory
Characteristics of image describes the problem brought, and the identification ability for area-type defect is the strongest.
In embodiments of the present invention, this steel strip surface defect identification device discrimination to the area-type defect of belt steel surface
The highest.Wherein, area-type defect is the complexity defect including multiple characteristics of image.
The steel strip surface defect identification device that the embodiment of the present invention provides, by the defect recognition model set up, to extraction
Various features be identified, such that it is able to quickly, accurately, identify the defect information of image, this defect information can be detailed
Ground describes various defects present in image, substantially increases the discrimination of defect, and the accuracy of the defect identified is the highest, and
Can effectively overcome geometry, optics invariance and rotatory that characteristics of image describes the problem brought, area-type is lacked
The identification ability fallen into is the strongest.
In sum, a kind of strip steel surface defect identification method and device are embodiments provided, it is achieved by building
After the various features extracted is identified by vertical defect recognition model, the defect letter of image can be fast and accurately identified
Breath, substantially increases the discrimination of defect, and the accuracy of the defect identified is the highest, and can effectively overcome geometry, optics not
Degeneration and rotatory describe the problem brought to characteristics of image, and the discrimination for area-type defect is the highest.
Embodiments of the invention are given for example with for the sake of describing, and are not exhaustively or by this
Bright it is limited to disclosed form.Many modifications and variations are apparent from for the ordinary skill in the art.Choosing
Selecting and describe embodiment is in order to the principle of the present invention and actual application are more preferably described, and makes those of ordinary skill in the art
It will be appreciated that the present invention thus design is suitable to the various embodiments with various amendments of special-purpose.
Claims (10)
1. a strip steel surface defect identification method, it is characterised in that described method includes:
Set up defect recognition model;
Obtain belt steel surface picture;
Belt steel surface picture is carried out pretreatment;
Extract histograms of oriented gradients HOG feature and the gray level co-occurrence matrixes of the image of pretreated belt steel surface picture
GLCM feature;
By described defect recognition model, described HOG feature and the described GLCM feature of extraction are identified, determine strip steel
The defect information on surface.
Method the most according to claim 1, it is characterised in that described belt steel surface picture is carried out pretreatment, including:
The image denoising of described belt steel surface picture is filtered;
Use piecewise linear transform, the image of the belt steel surface picture after noise-removed filtering is carried out grey level enhancement;
Use bilinear interpolation algorithm that the image of the belt steel surface picture after grey level enhancement is zoomed in and out and normalization.
Method the most according to claim 1, it is characterised in that described set up defect recognition model, including:
Obtain strip steel samples pictures;
Extract histograms of oriented gradients HOG feature and the gray level co-occurrence matrixes GLCM feature of the image of strip steel samples pictures;
The HOG feature of described image and the GLCM feature of image are merged by the method using stepping statistics, obtain many groups
Weight factor and the recognition accuracy that often group weight factor is corresponding;
Described many group weight factors and described recognition accuracy are carried out Gauss curve fitting, obtains feature fitting function;
By genetic algorithm, described feature fitting function is solved, obtain fusion factor;
According to described fusion factor, random forests algorithm is improved, obtain Feature Fusion based on random forest classification and calculate
Method;
By described HOG feature and the GLCM feature of the described Feature Fusion sorting algorithm based on the random forest image to extracting
It is trained, determines the defect information of described strip steel sample;
Defect information according to described strip steel sample sets up defect recognition model.
Method the most according to claim 3, it is characterised in that described according to described fusion factor, random forests algorithm is entered
Row improves, and obtains Feature Fusion sorting algorithm based on random forest and includes:
According to decision tree corresponding in described fusion factor reconstruct random forests algorithm, obtaining reconstructing decision tree, described reconstruct is certainly
Plan tree includes HOG decision tree and GLCM decision tree;
According to described reconstruct decision tree, described HOG feature and described GLCM feature are chosen, obtain HOG feature decision tree
Group and GLCM feature decision-making array;
According to described HOG decision-making array, the HOG feature of the image of described strip steel samples pictures is identified, obtains HOG feature
Identify data, and according to described GLCM decision-making array, the GLCM feature of the image of described strip steel samples pictures be identified,
Obtain GLCM biometric data;
Described HOG biometric data and described GLCM biometric data are added up, obtains statistical classification result;
According to the principle of random forests algorithm ballot, by feature the highest for number of votes obtained in described statistical classification result, it is defined as institute
State the defect recognition result of strip steel sample;
The algorithm performing above procedure is defined as Feature Fusion sorting algorithm based on random forest.
Method the most according to claim 3, it is characterised in that the method for described utilization stepping statistics is to described image
The GLCM feature of HOG feature and image merges, including:
The weight of HOG feature, the weight of GLCM feature and weight constraints condition are set;
Step-length is added up according to described weight constraints condition;
Described step-length according to statistics determines the weighted value span of HOG feature, and determines that the weighted value of GLCM feature takes
Value scope;
Weighted value span according to described HOG feature and the weighted value span of GLCM feature, be calculated many groups
Weight factor and the recognition accuracy that often group weight factor is corresponding.
6. a steel strip surface defect identification device, it is characterised in that described device includes:
Set up module, be used for setting up defect recognition model;
Acquisition module, is used for obtaining belt steel surface picture;
Pretreatment module, for carrying out pretreatment to belt steel surface picture;
Extraction module, for extract pretreated belt steel surface picture image histograms of oriented gradients HOG feature and
Gray level co-occurrence matrixes GLCM feature;
Identification module, for knowing described HOG feature and the described GLCM feature of extraction by described defect recognition model
Not, the defect information of belt steel surface is determined.
Device the most according to claim 6, it is characterised in that described pretreatment module includes:
Filter unit, for filtering the image denoising of described belt steel surface picture;
Converter unit, is used for using piecewise linear transform, and the image of the belt steel surface picture after noise-removed filtering is carried out gray scale increasing
By force;
Normalization unit, for using bilinear interpolation algorithm to zoom in and out the image of the belt steel surface picture after grey level enhancement
And normalization.
Device the most according to claim 6, it is characterised in that described module of setting up includes:
Acquiring unit, is used for obtaining strip steel samples pictures;
Extraction unit, for extracting HOG feature and the GLCM feature of the image of strip steel samples pictures;
Integrated unit, is carried out the HOG feature of described image and the GLCM feature of image for the method using stepping to add up
Merge, obtain organizing weight factor more and often organizing the recognition accuracy that weight factor is corresponding;
Fitting unit, for described many group weight factors and described recognition accuracy are carried out Gauss curve fitting, obtains feature fitting
Function;
Computing unit, for being solved described feature fitting function by genetic algorithm, obtains fusion factor;
Improve unit, for random forests algorithm being improved according to described fusion factor, obtain spy based on random forest
Levy integrated classification algorithm;
Training unit, for by the described Feature Fusion sorting algorithm based on the random forest described HOG of image to extracting
Characteristics of image and GLCM feature are trained, and determine the defect information of described strip steel sample;
Set up unit, for setting up defect recognition model according to the defect information of described strip steel sample.
Device the most according to claim 8, it is characterised in that described improvement unit includes:
Reconstruct subelement, for according to decision tree corresponding in described fusion factor reconstruct random forests algorithm, obtaining reconstruct certainly
Plan tree, described reconstruct decision tree includes HOG decision tree and GLCM decision tree;
Choose subelement, for described HOG feature and described GLCM feature being chosen according to described reconstruct decision tree,
To HOG feature decision tree group and GLCM feature decision-making array;
Identify subelement, for the HOG feature of the image of described strip steel samples pictures being known according to described HOG decision-making array
Not, obtain HOG biometric data, and according to the described GLCM decision-making array GLCM to the image of described strip steel samples pictures
Feature is identified, and obtains GLCM biometric data;
First statistics subelement, for described HOG biometric data and described GLCM biometric data are added up,
To statistical classification result;
First determines subelement, for the principle according to random forests algorithm ballot, by number of votes obtained in described statistical classification result
The highest feature, is defined as the defect recognition result of described strip steel sample;By perform above procedure algorithm be defined as based on
The Feature Fusion sorting algorithm of machine forest.
Device the most according to claim 8, it is characterised in that described integrated unit includes:
Subelement is set, for arranging the weight of HOG feature, the weight of GLCM feature and weight constraints condition;
Second statistics subelement, for adding up step-length according to described weight constraints condition;
Second determines subelement, for determining the weighted value span of HOG feature according to the described step-length of statistics, and determines
The weighted value span of GLCM feature;
Computation subunit, for the weighted value span according to described HOG feature and the weighted value value model of GLCM feature
Enclose, be calculated many group weight factors and often organize the recognition accuracy that weight factor is corresponding.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610439036.9A CN106097360A (en) | 2016-06-17 | 2016-06-17 | A kind of strip steel surface defect identification method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610439036.9A CN106097360A (en) | 2016-06-17 | 2016-06-17 | A kind of strip steel surface defect identification method and device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106097360A true CN106097360A (en) | 2016-11-09 |
Family
ID=57236854
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610439036.9A Pending CN106097360A (en) | 2016-06-17 | 2016-06-17 | A kind of strip steel surface defect identification method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106097360A (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106709502A (en) * | 2016-11-18 | 2017-05-24 | 深圳信息职业技术学院 | Multi-feature fusion recognition method based on voting method |
CN108074025A (en) * | 2016-11-10 | 2018-05-25 | 上海宝信软件股份有限公司 | Coil of strip surface defect determination method based on surface defect distribution characteristics |
CN108447051A (en) * | 2018-03-09 | 2018-08-24 | 东北大学 | A kind of metal product surface defect rank assessment method based on computer vision |
CN109444169A (en) * | 2019-01-17 | 2019-03-08 | 南昌航空大学 | A kind of bearing defect detection method and system |
CN109858485A (en) * | 2019-01-25 | 2019-06-07 | 东华大学 | A kind of fabric defects detection method based on LBP and GLCM |
CN110046630A (en) * | 2018-01-16 | 2019-07-23 | 上海电缆研究所有限公司 | Defect mode identification method/the systems/devices and readable storage medium storing program for executing of object |
CN110211125A (en) * | 2019-07-31 | 2019-09-06 | 征图新视(江苏)科技股份有限公司 | A kind of detection method and its filming apparatus of low contrast image's difference |
CN110560376A (en) * | 2019-07-19 | 2019-12-13 | 华瑞新智科技(北京)有限公司 | Product surface defect detection method and device |
CN110705584A (en) * | 2019-08-21 | 2020-01-17 | 深圳壹账通智能科技有限公司 | Emotion recognition method, emotion recognition device, computer device and storage medium |
CN110717883A (en) * | 2019-07-19 | 2020-01-21 | 广东电网有限责任公司 | Cable part automatic identification method based on infrared image |
CN111353425A (en) * | 2020-02-28 | 2020-06-30 | 河北工业大学 | Sleeping posture monitoring method based on feature fusion and artificial neural network |
CN113207509A (en) * | 2021-06-08 | 2021-08-06 | 重庆市农业科学院 | Vegetable unmanned seedling culture logistics system |
CN115375676A (en) * | 2022-10-24 | 2022-11-22 | 微山三利特不锈钢有限公司 | Stainless steel product quality detection method based on image recognition |
CN115937199A (en) * | 2023-01-06 | 2023-04-07 | 山东济宁圣地电业集团有限公司 | Spraying quality detection method for insulating layer of power distribution cabinet |
CN116805416A (en) * | 2023-08-21 | 2023-09-26 | 中国电建集团华东勘测设计研究院有限公司 | Drainage pipeline defect identification model training method and drainage pipeline defect identification method |
CN117094604A (en) * | 2023-10-18 | 2023-11-21 | 南通双耀冲压有限公司 | Surface quality management method and system for motor stator and rotor stamping die |
CN117173179A (en) * | 2023-11-02 | 2023-12-05 | 南通市通州兴辰机械有限公司 | Method and system for rapidly detecting production quality of sound-proof cloth |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103745234A (en) * | 2014-01-23 | 2014-04-23 | 东北大学 | Band steel surface defect feature extraction and classification method |
US20150178560A1 (en) * | 2013-01-11 | 2015-06-25 | Grg Banking Equipment Co., Ltd. | Recognition method and recognition device for sheet-type medium |
CN104794491A (en) * | 2015-04-28 | 2015-07-22 | 重庆大学 | Fuzzy clustering steel plate surface defect detection method based on pre classification |
CN104866862A (en) * | 2015-04-27 | 2015-08-26 | 中南大学 | Strip steel surface area type defect identification and classification method |
EP2990998A1 (en) * | 2014-08-29 | 2016-03-02 | Institute of Electronics, Chinese Academy of Sciences | Method and device for assessing damage in disaster area |
-
2016
- 2016-06-17 CN CN201610439036.9A patent/CN106097360A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150178560A1 (en) * | 2013-01-11 | 2015-06-25 | Grg Banking Equipment Co., Ltd. | Recognition method and recognition device for sheet-type medium |
CN103745234A (en) * | 2014-01-23 | 2014-04-23 | 东北大学 | Band steel surface defect feature extraction and classification method |
EP2990998A1 (en) * | 2014-08-29 | 2016-03-02 | Institute of Electronics, Chinese Academy of Sciences | Method and device for assessing damage in disaster area |
CN104866862A (en) * | 2015-04-27 | 2015-08-26 | 中南大学 | Strip steel surface area type defect identification and classification method |
CN104794491A (en) * | 2015-04-28 | 2015-07-22 | 重庆大学 | Fuzzy clustering steel plate surface defect detection method based on pre classification |
Non-Patent Citations (1)
Title |
---|
马强等: ""基于多特征融合的织物瑕疵检测研究"", 《微型机与应用》 * |
Cited By (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108074025A (en) * | 2016-11-10 | 2018-05-25 | 上海宝信软件股份有限公司 | Coil of strip surface defect determination method based on surface defect distribution characteristics |
CN108074025B (en) * | 2016-11-10 | 2021-12-31 | 上海宝信软件股份有限公司 | Steel coil surface defect judgment method based on surface defect distribution characteristics |
CN106709502A (en) * | 2016-11-18 | 2017-05-24 | 深圳信息职业技术学院 | Multi-feature fusion recognition method based on voting method |
CN110046630A (en) * | 2018-01-16 | 2019-07-23 | 上海电缆研究所有限公司 | Defect mode identification method/the systems/devices and readable storage medium storing program for executing of object |
CN108447051A (en) * | 2018-03-09 | 2018-08-24 | 东北大学 | A kind of metal product surface defect rank assessment method based on computer vision |
CN108447051B (en) * | 2018-03-09 | 2019-12-24 | 东北大学 | Metal product surface defect grade evaluation method based on computer vision |
CN109444169A (en) * | 2019-01-17 | 2019-03-08 | 南昌航空大学 | A kind of bearing defect detection method and system |
CN109444169B (en) * | 2019-01-17 | 2021-06-15 | 南昌航空大学 | Bearing defect detection method and system |
CN109858485A (en) * | 2019-01-25 | 2019-06-07 | 东华大学 | A kind of fabric defects detection method based on LBP and GLCM |
CN110717883A (en) * | 2019-07-19 | 2020-01-21 | 广东电网有限责任公司 | Cable part automatic identification method based on infrared image |
CN110560376A (en) * | 2019-07-19 | 2019-12-13 | 华瑞新智科技(北京)有限公司 | Product surface defect detection method and device |
CN110211125A (en) * | 2019-07-31 | 2019-09-06 | 征图新视(江苏)科技股份有限公司 | A kind of detection method and its filming apparatus of low contrast image's difference |
CN110705584A (en) * | 2019-08-21 | 2020-01-17 | 深圳壹账通智能科技有限公司 | Emotion recognition method, emotion recognition device, computer device and storage medium |
CN111353425A (en) * | 2020-02-28 | 2020-06-30 | 河北工业大学 | Sleeping posture monitoring method based on feature fusion and artificial neural network |
CN113207509A (en) * | 2021-06-08 | 2021-08-06 | 重庆市农业科学院 | Vegetable unmanned seedling culture logistics system |
CN115375676A (en) * | 2022-10-24 | 2022-11-22 | 微山三利特不锈钢有限公司 | Stainless steel product quality detection method based on image recognition |
CN115937199A (en) * | 2023-01-06 | 2023-04-07 | 山东济宁圣地电业集团有限公司 | Spraying quality detection method for insulating layer of power distribution cabinet |
CN116805416A (en) * | 2023-08-21 | 2023-09-26 | 中国电建集团华东勘测设计研究院有限公司 | Drainage pipeline defect identification model training method and drainage pipeline defect identification method |
CN117094604A (en) * | 2023-10-18 | 2023-11-21 | 南通双耀冲压有限公司 | Surface quality management method and system for motor stator and rotor stamping die |
CN117094604B (en) * | 2023-10-18 | 2023-12-26 | 南通双耀冲压有限公司 | Surface quality management method and system for motor stator and rotor stamping die |
CN117173179A (en) * | 2023-11-02 | 2023-12-05 | 南通市通州兴辰机械有限公司 | Method and system for rapidly detecting production quality of sound-proof cloth |
CN117173179B (en) * | 2023-11-02 | 2024-03-05 | 南通市通州兴辰机械有限公司 | Method and system for rapidly detecting production quality of sound-proof cloth |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106097360A (en) | A kind of strip steel surface defect identification method and device | |
CN110211045B (en) | Super-resolution face image reconstruction method based on SRGAN network | |
CN104866862B (en) | A kind of method of belt steel surface area-type defect recognition classification | |
CN101206715B (en) | Face recognition apparatus, face recognition method, Gabor filter application apparatus, and computer program | |
CN112633382B (en) | Method and system for classifying few sample images based on mutual neighbor | |
CN111860171A (en) | Method and system for detecting irregular-shaped target in large-scale remote sensing image | |
CN101430759A (en) | Optimized recognition pretreatment method for human face | |
CN102779157B (en) | Method and device for searching images | |
CN104077742B (en) | Human face sketch synthetic method and system based on Gabor characteristic | |
CN109949200B (en) | Filter subset selection and CNN-based steganalysis framework construction method | |
CN106650667A (en) | Pedestrian detection method and system based on support vector machine | |
CN103235947A (en) | Handwriting digital recognition method and device | |
CN102982343A (en) | Handwritten number recognition and incremental type obscure support vector machine method | |
CN109740539A (en) | 3D object identification method based on transfinite learning machine and fusion convolutional network | |
CN115457568B (en) | Historical document image noise reduction method and system based on generation countermeasure network | |
CN106846354A (en) | A kind of Book Inventory method on frame converted based on image segmentation and random hough | |
CN103839066A (en) | Feature extraction method from biological vision | |
CN113344110A (en) | Fuzzy image classification method based on super-resolution reconstruction | |
CN107133579A (en) | Based on CSGF (2D)2The face identification method of PCANet convolutional networks | |
CN111126169A (en) | Face recognition method and system based on orthogonalization graph regular nonnegative matrix decomposition | |
CN110674822A (en) | Handwritten digit recognition method based on feature dimension reduction | |
CN116503398B (en) | Insulator pollution flashover detection method and device, electronic equipment and storage medium | |
CN117237954A (en) | Text intelligent scoring method and system based on ordering learning | |
CN111127407B (en) | Fourier transform-based style migration forged image detection device and method | |
CN108960285B (en) | Classification model generation method, tongue image classification method and tongue image classification device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into 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: 20161109 |