CN109886937A - Defects of insulator detection method based on super-pixel segmentation image recognition - Google Patents
Defects of insulator detection method based on super-pixel segmentation image recognition Download PDFInfo
- Publication number
- CN109886937A CN109886937A CN201910086179.XA CN201910086179A CN109886937A CN 109886937 A CN109886937 A CN 109886937A CN 201910086179 A CN201910086179 A CN 201910086179A CN 109886937 A CN109886937 A CN 109886937A
- Authority
- CN
- China
- Prior art keywords
- pixel
- super
- image
- insulator
- defects
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Abstract
The present invention discloses a kind of method for diagnosing its defect based on insulation subgraph, comprising the following steps: insulation subgraph is carried out Gassian low-pass filter noise suppression preprocessing, removes the noise pixel in image;The super-pixel cluster that Color Channel weighting is carried out to the image of denoising, divides the image into certain super-pixel;By treated, image is divided into training set and test set and adds label, selects algorithm of support vector machine to image classification, obtains prediction result.Using the method for the present invention, the recognition speed of defects of insulator can be sufficiently improved under the premise of not reducing discrimination, and to improving the efficiency of electric inspection process, reducing cost with certain values, also there is certain meaning to guarantee electric power netting safe running.
Description
Technical field
The present invention relates to image identification technical fields, and in particular to the defects of insulator detection side based on super-pixel segmentation
Method.
Background technique
Increasingly developed with electric system scale, the requirement of safe operation and power supply reliability to power circuit is more next
It is higher.Power circuit plays an important role in power grid, and safe and stable operation plays certainly the integrality for guaranteeing electric network composition
Qualitative effect.Insulator is one of the emphasis of electric inspection process work as the important composition in overhead transmission line, therefore high
Effect, accurate defects of insulator detection method demand are further urgent.Traditional defects of insulator detection is usually by maintenance person edge
Line is maked an inspection tour by tower, although this method is simple, low efficiency, period length, large labor intensity, costly and risk are big.In recent years
Come, be gradually gained popularity by the method for inspecting of vision facilities, patrol officer shoots photo to route by tools such as unmanned planes
It is detected on ground, but the defect diagonsis of most of insulators is that micro-judgment is relied on by patrol officer, can still be slipped
Or judge by accident and expend a large amount of manpower and material resources, and efficiency is not improved, there are detection time length, real-time difference etc. are many
Problem.
As power grid develops towards intelligent direction, electric inspection process also develops towards automation direction, gradually knows image
Other Technology application is in defects of insulator detection.The defects detection of electric transmission line isolator mainly divides the image into abnormal and just
Normal two class, in numerous sorting algorithms, the prior art has had certain classifying quality, but itself or use high-resolution insulation
Sub- original image is identified, brings very test to the calculating speed of processing equipment and real-time is bad;Or to insulation
Sub- original image is compressed, and is identified using low-resolution image, although this ensure that speed but greatly reduces knowledge
Not rate, the above problem all hinder the development of image recognition technology and the application in related fields to a certain extent.
Summary of the invention
In order to solve some shortcomings in conventional electric power inspection and existing defects of insulator detection technique, the present invention proposes one
Defects of insulator detection method of the kind based on super-pixel segmentation image recognition can not reduce discrimination using the method for the present invention
Under the premise of sufficiently improve the recognition speed of defects of insulator, to improve the efficiency of electric inspection process, reduce cost, and then to guarantor
Barrier electric power netting safe running has a very important significance.
To realize that above scheme, technical solution of the present invention are specific as follows:
A kind of defects of insulator detection method based on super-pixel segmentation image recognition, comprising the following steps:
Step 1: the original image of the n identical insulators of resolution ratio of acquisition carries out noise suppression preprocessing;
Step 2: according to the complexity of n insulation set of sub-images, specifying its number K that classifies, obtain K initial super pictures
Element;
Step 3: K initial super-pixel central points are distributed on the regular grids for being spaced A pixel, distribution
Principle is to generate the super-pixel of equal sizes;
Step 4: each initial super-pixel center position of adjustment, the gradient where making central point be moved to it in neighborhood
On the smallest pixel;
Step 5: being directed to each super-pixel central point ζi, all pixels point in the range of 2A × 2A is traversed, is gone forward side by side
The following operation of row: if the distance d of a super-pixel point to super-pixel center ζ be less than this super-pixel point belonged to originally to it in it is first
The distance d ' of beginning super-pixel center ζ ', then this point belongs to super-pixel corresponding to the ζ of the super-pixel center;Super picture is obtained after traversal
Plain center ζiCluster areas SiAnd SiThe pixel quantity N for being includedi;
Step 6: being directed to cluster areas SiIn each super-pixel region, recalculate to obtain the super picture in each super-pixel region
Plain center ζi;
Step 7: iterative step 5~6, until the super-pixel center ζ in each super-pixel regioniIt restrains or reaches default and change
Algebra;The super-pixel center in each super-pixel region is redefined, and K updated super-pixel regions are obtained with this;
Step 8: updated super-pixel is extracted into feature on channel in the channel r of rgb color space, the channel g, b respectively;
And be normalized to feature is extracted, thus obtain the feature vector, X of insulation subgraphi;
Step 9: giving its corresponding feature vector, X according to the defect situation of every insulator super-pixeliLabel y is addedi∈{-
1,1 }, wherein yi=-1 indicates that the insulator form of the i-th width super-pixel image is normal, yi=1 the i-th width super-pixel image of expression
Insulator have defect, thus establish sample data set
Step 10: the sampling of data set D stratified random is divided into the identical exclusive subsets of k size;
Step 11: by the subset D of data set1As test set, complementary subset { D2,D3,…,DnIt is used as training set;
Step 12: by the support vector machine classifier that training set input is selected, and selecting kernel function, and according to selected
Support vector machine classifier and kernel function determine the parameter for needing to adjust, and obtain support vector cassification after training using training set
Device model;
Step 13: test set being inputted into trained support vector machine classifier model, classification results are obtained, according to support
The defect situation of the output forecast image insulator of vector machine classifier model;It assesses the error rate of supporting vector machine model, look into
Quasi- rate, recall ratio and Receiver operating curve ROC;
Step 14: subset D will be successively removed in data set D1A subset in addition is as a test set, remaining subset
As training set, k-1 progress step 12~13 are repeated, assessment models simultaneously carry out parameter tuning;
Step 15: by data set D layering, simultaneously stochastical sampling is divided into the identical exclusive subsets of k size at random again, weight
Multiple p-1 progress step 11~14, assessment models simultaneously carry out parameter tuning;
Step 16: repeating step 10~15 and carry out p k folding cross validation parameter optimization, obtain trained insulator and lack
Fall into identification model;
Step 17: insulator super-pixel image being inputted into trained defects of insulator identification model and carries out defects of insulator
Detection.
As a further improvement of the present invention, noise suppression preprocessing specific steps in step 1 are as follows:
Discrete Fourier transform is carried out to original image, two dimensional discrete Fourier transform formula is as follows:
Wherein F (u, v) is Fourier transformation result, that is, image frequency domain function, and u and v are frequency component, and f (x, y) is original image,
M and N is the width and height of image, and j is imaginary unit;Calculate dimensional Gaussian low pass filter function G (u, v) and image
The product of frequency-domain function F (u, v), i.e., filtered image frequency domain function H (u, v);The transmitting letter of dimensional Gaussian low-pass filter
Number is as follows:
δ is standard deviation in formula, chooses δ according to the characteristic of image;
The inverse discrete Fourier transform for calculating filtered image frequency domain function H (u, v), obtains image in two-dimensional surface
Pixel distribution function h (x, y);The mould of capture element distribution function h (x, y) eliminates in original image and exists as output image
Noise pixel.
As a further improvement of the present invention, in step 3, setting grid spacings areWherein N is total picture of image
Plain number.
As a further improvement of the present invention, in step 5, the calculating of the distance d of a super-pixel point to super-pixel center ζ
Method is as follows:
It takes coordinate of the pixel in rgb color space and the normalized on progress [0,1] section is named as
(r, g, b), the pixel also carry out the life of the normalized on [0,1] section in the position coordinate value in the two-dimensional space of image
Entitled (x, y);The calculation formula of distance l of the super-pixel point to super-pixel center in color space is as follows:
α, β, γ are the weight in the channel R, the channel G and channel B, (r in formula1, g1, b1)、(r2, g2, b2) it is that super-pixel point arrives
Super-pixel center carries out the coordinate after the normalized on [0,1] section in rgb color space;
Super-pixel point is as follows to distance s calculation formula of the super-pixel center in two-dimensional space:
(x in formula1,y1)、(x2,y2) it is that super-pixel point to super-pixel center carries out [0,1] area in the two-dimensional space of image
Between on normalized after position coordinates;Summation is weighted to the distance in color space and two-dimensional space, is surpassed
Actual range d of the pixel to super-pixel center:
L is the distance in color space in formula, and s is the distance in two-dimensional space, in color space and two-dimensional space
In the weight of distance be respectively μ, ρ.
As a further improvement of the present invention, in step 6, its super-pixel center ζ is recalculated according to the following formulai?
Position in rgb color space and two-dimensional space:
The position coordinates average value for the pixel for taking the region all.
As a further improvement of the present invention, in step 8, super-pixel is respectively in the channel r of rgb color space, the channel g, b
Feature is extracted on channel refers to first moment about the origin E ζ in each channel of super-pixel image, second geometric moment E ζ2With three rank moment of the orign E
ζ3, calculation formula is as follows:
Wherein ζi=(ri,gi,bi) be super-pixel image each super-pixel block central pixel point RGB channel on
Numerical value, the coordinate (x of the two-dimensional surface in super-pixel blocki,yi) it is as follows:
Wherein NjFor the pixel number for belonging to the super-pixel, SiFor cluster areas, the i.e. region of the super-pixel,For the sum of the coordinate value of all pixels two-dimensional surface of the super-pixel block;
To first moment about the origin E ζ, second geometric moment E ζ2With three rank moment of the orign E ζ3Calculated result proceed to (0,1) section
Thus normalized obtains feature vector:
As a further improvement of the present invention, in step 12, kernel function is selected from linear, multinomial, radial base RBF or S type
Growth curve sigmoid.
Compared with prior art, the invention has the following advantages that
Insulation subgraph is carried out Gassian low-pass filter noise suppression preprocessing first by the present invention, removes the noise picture in image
Element;The super-pixel cluster that Color Channel weighting is carried out to the image of denoising, divides the image into certain super-pixel;After handling
Image be divided into training set and test set and add label, select support vector machines algorithm to image classification.It is exhausted to realize
The Accurate Diagnosis of edge defect situation.Conventional method is needed when carrying out defect recognition to insulation subgraph to all pixels point
It carries out feature extraction or establishes node, the method for the present invention will have class according to the texture and color of image using the thought divided and ruled
Pixel cluster like feature is super-pixel region, and then carries out feature extraction to each regional center point, is avoided to a large amount of
Similar pixel point computes repeatedly, and reduces calculation amount, and then improve arithmetic speed.The method of the present invention is clustered in super-pixel
Weight is introduced in the process, by the weight of RGB color in regulating calculation, is conducive to the survey light for excluding the equipment of acquisition image
The interference introduced with white balance deviation;By the weight of color space coordinate in regulating calculation and 2-d spatial coordinate, facilitate
It is overlapped super-pixel region and object actual profile and texture accurately to promote Clustering Effect.It, can be not using the method for the present invention
Reduce discrimination under the premise of sufficiently improve the recognition speed of defects of insulator, thus improve the efficiency of electric inspection process, reduce at
This, and then guarantee electric power netting safe running is had a very important significance.
Detailed description of the invention
Fig. 1 is the insulation subgraph after Gassian low-pass filter denoising;
Fig. 2 is K=36, iteration 10 times insulator super-pixel images.
Specific embodiment
Being detected as example below with reference to defects of insulator, the present invention will be described in more detail.
The present invention proposes a kind of defects of insulator detection method based on super-pixel segmentation image recognition, specific as follows:
Step 1: the image of the n identical insulators of resolution ratio of acquisition carries out Gassian low-pass filter noise suppression preprocessing.To original
Image carries out discrete Fourier transform, and two dimensional discrete Fourier transform formula is as follows:
Wherein F (u, v) is Fourier transformation result, that is, image frequency domain function, and u and v are frequency component, and f (x, y) is original image,
M and N is the width and height of image, and j is imaginary unit.Calculate filter transfer function G (u, v) and image frequency domain function F (u,
V) product, i.e., filtered image frequency domain function H (u, v).The transmitting letter of the dimensional Gaussian low-pass filter of this step application
Number is as follows:
δ is standard deviation in formula, chooses suitable δ according to the characteristic of image.
The inverse discrete Fourier transform for calculating filtered image frequency domain function H (u, v), obtains image in two-dimensional surface
Pixel distribution function h (x, y).The mould of capture element distribution function h (x, y) eliminates in original image and exists as output image
Noise pixel;
Step 2: according to the complexity of this group of n insulation set of sub-images, specify its number K that classifies, obtain K it is initial super
Pixel;
3:K super-pixel initial cluster center point of step is distributed on the regular grids for being spaced A pixel, distribution
Principle is to generate the super-pixel of roughly equal size, therefore grid spacings are arranged and areWherein N is total pixel of image
Number is its resolution ratio product;
Step 4: the position of adjustment super-pixel initial cluster center point, in 3 × 3 neighborhoods of central point, Mobility Center point
Into this 9 points on the smallest point of gradient, in order to avoid super-pixel location of the core is not filtered out in the profile and border of article and dry
On net noise pixel;
Step 5: to each super-pixel center ζi, the initial range of affiliated super-pixel is A × A, i.e. elementary cell
Region.Point in the range of its 2A × 2A is traversed, and is proceeded as follows: as the distance d of fruit dot to super-pixel center ζ are less than
This point belonged to originally to it in super-pixel center ζ ' distance d ', then illustrate this put belong to corresponding to the ζ of super-pixel center
Super-pixel.Coordinate of the capture vegetarian refreshments in rgb color space and the normalized on progress [0,1] section be named as (r, g,
B), this be named as in the normalized that the position coordinate value in the two-dimensional space of image also carries out on [0,1] section (x,
y).The calculation formula of distance l of the two o'clock in color space is as follows:
α, β, γ are the weight in the channel R, the channel G and channel B, (r in formula1,g1,b1)、(r2,g2,b2) it is that two pixels exist
Coordinate after carrying out the normalized on [0,1] section in rgb color space.α, β, γ need to be according to the practical feelings of picture
Condition is selected, to exclude the interference for surveying light and the introducing of white balance deviation of the equipment of acquisition image.Two o'clock is in two-dimensional space
Distance s calculation formula it is as follows:
(x in formula1,y1)、(x2,y2) it is that two pixels carry out the normalization on [0,1] section in the two-dimensional space of image
Treated position coordinates.Summation is weighted to the distance in color space and two-dimensional space, obtain pixel it is practical away from
From d:
L is distance of the two o'clock in color space in formula, and s is distance of the two o'clock in two-dimensional space, in color space and
The weight of distance in two-dimensional space is respectively μ, ρ, and suitable weight is selected to help to make super-pixel boundary according to the actual situation
It is accurately overlapped with object actual profile.Super-pixel center ζ is updated after traversaliCluster areas SiAnd SiThe pixel number for being included
Measure Ni;
Step 6: being directed to each super-pixel region, recalculate its super-pixel center ζ according to the following formulaiIn rgb color
Position in space and two-dimensional space:
The position coordinates average value for the pixel for taking the region all;
Step 7: iterative step 5~6, until super-pixel central point ζiConverge on fixed value.Or according to required identification
The computing capability of precision and equipment selects suitable the number of iterations t;
Step 8: super-pixel image obtained in step 2~7 is led in the channel r of rgb color space, the channel g, b respectively
Feature is extracted on road, this step counts first moment about the origin E ζ, the second geometric moment E ζ for being characterized in each channel of super-pixel image2
With three rank moment of the orign E ζ3, calculation formula is as follows:
Wherein ζi=(ri,gi,bi), it is on the RGB channel of the central pixel point of each super-pixel block of super-pixel image
Numerical value, the coordinate (x of the two-dimensional surface in super-pixel blocki,yi) it is as follows:
Wherein NjFor the number of pixels of the super-pixel, SiFor the region of the super-pixel,For the super-pixel block
All pixels two-dimensional surface the sum of coordinate value.
Because the property of the feature vector of support vector machines needs, proceed to the normalized in (0,1) section to calculation amount,
It can thus be concluded that feature vectorThen n can be obtained
Insulate feature vector { X corresponding to subgraph1, X2..., Xn};
Step 9: giving its corresponding feature vector, X according to the defect situation of every insulator super-pixel imageiLabel y is addedi
∈ { -1,1 }, wherein yi=-1 indicates that the insulator form of the i-th width super-pixel image is normal, yi=1 the i-th width super-pixel of expression
The insulator of image has defect, thus establishes sample data set
Step 10: the sampling of data set D stratified random is divided into the identical exclusive subsets of k size, i.e. D=D1∪D2
∪…∪Dk,
Step 11: by the subset D of data set1As test set, complementary subset { D2,D3,…,DnIt is used as training set;
Step 12: by the support vector machine classifier that training set input is selected, and it is (linear, more to select suitable kernel function
Item formula, radial direction base RBF, S sigmoid growth curve sigmoid), and determined according to selected support vector machine classifier and kernel function
The parameter for needing to adjust: parameter degree, parameter coef0, parameter gamma, penalty factor cost etc., after training set training
Support vector machine classifier model can be obtained;
Step 13: test set being inputted into trained supporting vector machine model, classification results are obtained, according to support vector machines
Output forecast image insulator defect situation.Assess the error rate of supporting vector machine model, precision ratio, recall ratio and tested
Person's performance curve ROC;
Step 14: successively using remaining each subset of data set D as test set, remaining subset is as training set, weight
Multiple k-1 progress step 12~13, assessment models simultaneously carry out parameter tuning;
Step 15: the sampling of data set D stratified random being divided into the identical exclusive subsets of k size at random again, is repeated
P-1 progress step 11~14, assessment models simultaneously carry out parameter tuning;
Step 16: repeating step 10~15 (p k rolls over cross validation) and carry out parameter optimization, can be obtained trained exhausted
Edge defect recognition model;
Step 17: a resolution ratio and step 1 being taken to take the image of identical insulator.Gauss low pass is generated through step 1
It is generated insulator super-pixel image by step 2 by the insulation subgraph after filtering and noise reduction.Insulator super-pixel image is defeated
Enter trained model, model outputIndicate that the insulator form of the super-pixel image is normal,
The insulator of the super-pixel image has defect.
Embodiment
The present invention is based on the defects of insulator detection methods of super-pixel segmentation image recognition, and steps are as follows:
Step 1: it is pre- to carry out Gassian low-pass filter denoising for the image for the insulator that 1000 resolution ratio of acquisition are 400 × 200
Processing.Discrete Fourier transform is carried out to original image, two dimensional discrete Fourier transform formula is as follows:
Wherein F (u, v) is Fourier transformation result, that is, image frequency domain function, and u and v are frequency component, and f (x, y) is original image,
M and N is the width and height of image, and j is imaginary unit.Calculate filter transfer function G (u, v) and image frequency domain function F (u,
V) product, i.e., filtered image frequency domain function H (u, v), selection criteria difference δ=10, the biography of dimensional Gaussian low-pass filter
Delivery function is as follows:
The inverse discrete Fourier transform for calculating filtered image frequency domain function H (u, v), obtains image in two-dimensional surface
Pixel distribution function h (x, y).The mould of capture element distribution function h (x, y) eliminates in original image and exists as output image
Noise pixel;As shown in Figure 1.
Step 2: according to the complexity of the group 1000 set of sub-images that insulate, specifying its number of classifying is K=36;Such as Fig. 2
It is shown.
Step 3:36A super-pixel initial cluster center point is distributed on the regular grids for being spaced 110 pixels, point
The principle of cloth is to generate the super-pixel of roughly equal size, therefore grid spacings are arranged and areWherein N is total pixel of image
Number is its resolution ratio product, i.e., 80000;
Step 4: the position of adjustment super-pixel initial cluster center point, in 3 × 3 neighborhoods of central point, Mobility Center point
Into this 9 points on the smallest point of gradient, in order to avoid super-pixel location of the core is not filtered out in the profile and border of article and dry
On net noise pixel;
Step 5: to each super-pixel center ζi, the initial range of affiliated super-pixel is 110 × 110, i.e. unit grid
The region of lattice.Traverse its 220 × 220 in the range of point, and proceed as follows: as fruit dot to super-pixel center ζ away from
From d be less than this point belong to originally to it in super-pixel center ζ ' distance d ', then illustrating that this is put belongs to super-pixel center ζ institute
Corresponding super-pixel.Coordinate of the capture vegetarian refreshments in the rgb color space and normalized on progress [0,1] section is named as
(r, g, b), this also carry out the normalized on [0,1] section in the position coordinate value in the two-dimensional space of image and are named as
(x, y).The calculation formula of distance l of the two o'clock in color space is as follows:
α, β, γ are the weight in the channel R, the channel G and channel B, (r in formula1, g1, b1)、(r2, g2, b2) it is that two pixels exist
Coordinate after carrying out the normalized on [0,1] section in rgb color space.Take α=β=γ=1.Two o'clock is in two dimension
Distance s calculation formula in space is as follows:
(x in formula1, y1)、(x2, y2) it is that two pixels carry out the normalization on [0,1] section in the two-dimensional space of image
Treated position coordinates.Summation is weighted to the distance in color space and two-dimensional space, obtain pixel it is practical away from
From d:
L is distance of the two o'clock in color space in formula, and s is distance of the two o'clock in two-dimensional space, in color space and
The weight of distance in two-dimensional space is respectively μ, ρ, takes μ=1, ρ=0.8.Super-pixel center ζ is updated after traversaliCluster area
Domain SiAnd SiThe pixel quantity N for being includedi;
Step 6: being directed to each super-pixel region, recalculate its super-pixel center ζ according to the following formulaiIn rgb color
Position in space and two-dimensional space:
The position coordinates average value for the pixel for taking the region all;
Step 7: by step 5 and step 6 iteration 10 times;
Step 8: super-pixel image obtained in step 2~7 is led in the channel r of rgb color space, the channel g, b respectively
Feature is extracted on road, this step extracts first moment about the origin E ζ, the second geometric moment E ζ for being characterized in each channel of super-pixel image2With
Three rank moment of the orign E ζ3, calculation formula is as follows:
Wherein ζi=(ri, gi, bi), it is on the RGB channel of the central pixel point of each super-pixel block of super-pixel image
Numerical value, the coordinate (x of the two-dimensional surface in super-pixel blocki, yi) it is as follows:
Wherein NjFor the number of pixels of the super-pixel, SiFor the region of the super-pixel,For the super-pixel block
All pixels two-dimensional surface the sum of coordinate value.
Because the property of the feature vector of support vector machines needs, proceed to the normalized in (0,1) section to calculation amount,
It can thus be concluded that feature vectorThen it can obtain
Feature vector { X corresponding to 1000 insulation subgraphs1, X2..., Xn};
Step 9: label y is added according to the defect situation of every insulator super-pixel imagei∈ { -1,1 }, wherein yi=-1
Indicate that the insulator form of the i-th width super-pixel image is normal, yiThe insulator of=1 the i-th width super-pixel image of expression, which has, to be lacked
It falls into, thus establishes sample data set
Step 10: the sampling of data set D stratified random is divided into the identical exclusive subsets of 10 sizes, i.e. D=D1∪D2
∪…∪D10,
Step 11: by the subset D of data set1As test set, complementary subset { D2, D3..., D10It is used as training set;
Step 12: training set inputs support vector machine classifier, selectes linear kernel function k (x, xj)=xxj, adjustment returns
One changes parameter nu, parameter p, the penalty factor cost for arriving (0,1) section, is divided using support vector machines can be obtained after training set training
Class device model;
Step 13: test set being inputted into trained supporting vector machine model, classification results are obtained, according to support vector machines
Output forecast image insulator defect situation.Assess the error rate of supporting vector machine model, precision ratio, recall ratio and tested
Person's performance curve ROC;
Step 14: successively using remaining each subset of data set D as test set, remaining subset is as training set, weight
Multiple 9 progress step 12~13, according to error rate, precision ratio, recall ratio and Receiver operating curve's ROC assessment models,
And the adjustment that step-length is 0.1 is carried out respectively to the parameter nu, parameter p and penalty factor cost that normalize to (0,1) section;
Step 15: the sampling of data set D stratified random being divided into the identical exclusive subsets of 10 sizes at random again, is repeated
9 progress step 11~14, according to error rate, precision ratio, recall ratio and Receiver operating curve's ROC assessment models, and
The adjustment that step-length is 0.1 is carried out respectively to the parameter nu, parameter p and penalty factor cost that normalize to (0,1) section;
Step 16: repeating step 10~15 (10 10 folding cross validations).Table 1 is to select in 10 10 folding cross validations
The value of parameter.Finally selected parameter nu is 0.6, parameter p is 0.1, penalty factor cost is 0.3.It has obtained trained exhausted
Edge defect recognition model;
Table 1
Step 17: taking a resolution ratio is 400 × 200 insulation subgraphs, after step 1 generates Gassian low-pass filter denoising
Insulation subgraph, as Fig. 1.Fig. 1 is generated into K=3 by step 26, iteration 10 times insulator super-pixel images, as
Fig. 2.Fig. 2 is inputted into trained model, model outputThe insulator for representing the super-pixel image has defect.
The defects of insulator detection method software layer based on super-pixel segmentation image recognition uses C++&opencv
Computer vision library, hardware layer form defects of insulator detection system using Intel core-m3 7y30.
Note: " point " and " pixel " occurred in above all refers to pixel, and " central point " refers to central pixel point.
More than, only presently preferred embodiments of the present invention is not limited only to practical range of the invention, all according to the invention patent
The equivalence changes and modification that the content of range is done all should be technology scope of the invention.
Claims (7)
1. a kind of defects of insulator detection method based on super-pixel segmentation image recognition, which comprises the following steps:
Step 1: the original image of the n identical insulators of resolution ratio of acquisition carries out noise suppression preprocessing;
Step 2: according to the complexity of n insulation set of sub-images, specifying its number K that classifies, obtain K initial super-pixel;
Step 3: K initial super-pixel central points being distributed on the regular grids for being spaced A pixel, the principle of distribution
It is the super-pixel for generating equal sizes;
Step 4: each initial super-pixel center position of adjustment, the gradient where making central point be moved to it in neighborhood are minimum
Pixel on;
Step 5: being directed to each super-pixel central point ζi, all pixels point in the range of 2A × 2A is traversed, and carry out as follows
Operation: if a super-pixel point to super-pixel center ζ distance d less than this super-pixel point belong to originally to it in initial super picture
The distance d ' of plain center ζ ', then this point belongs to super-pixel corresponding to the ζ of the super-pixel center;Super-pixel center is obtained after traversal
ζiCluster areas SiAnd SiThe pixel quantity N for being includedi;
Step 6: being directed to cluster areas SiIn each super-pixel region, recalculate to obtain in the super-pixel in each super-pixel region
Heart ζi;
Step 7: iterative step 5~6, until the super-pixel center ζ in each super-pixel regioniRestrain or reach default number of iterations;
The super-pixel center in each super-pixel region is redefined, and K updated super-pixel regions are obtained with this;
Step 8: updated super-pixel is extracted into feature on channel in the channel r of rgb color space, the channel g, b respectively;And it is right
It extracts feature to be normalized, thus obtains the feature vector, X of insulation subgraphi;
Step 9: giving its corresponding feature vector, X according to the defect situation of every insulator super-pixeliLabel y is addedi∈{-1,
1 }, wherein yi=-1 indicates that the insulator form of the i-th width super-pixel image is normal, yi=1 indicates the i-th width super-pixel image
Insulator has defect, thus establishes sample data set
Step 10: the sampling of data set D stratified random is divided into the identical exclusive subsets of k size;
Step 11: by the subset D of data set1As test set, complementary subset { D2,D3,···,DnIt is used as training set;
Step 12: by the support vector machine classifier that training set input is selected, and selecting kernel function, and according to selected support
Vector machine classifier and kernel function determine the parameter for needing to adjust, and obtain support vector machine classifier mould after training using training set
Type;
Step 13: test set being inputted into trained support vector machine classifier model, classification results are obtained, according to supporting vector
The defect situation of the output forecast image insulator of machine sorter model;
Step 14: subset D will be successively removed in data set D1A subset in addition is as a test set, remaining subset conduct
Training set, repeats k-1 progress step 12~13, and assessment models simultaneously carry out parameter tuning;
Step 15: by data set D layering, simultaneously stochastical sampling is divided into the identical exclusive subsets of k size at random again, repeats p-1
Secondary progress step 11~14, assessment models simultaneously carry out parameter tuning;
Step 16: repeating step 10~15 and carry out p k folding cross validation parameter optimization, obtain trained defects of insulator and know
Other model;
Step 17: insulator super-pixel image being inputted into trained defects of insulator identification model and carries out defects of insulator inspection
It surveys.
2. the defects of insulator detection method according to claim 1 based on super-pixel segmentation image recognition, feature exist
In noise suppression preprocessing specific steps in step 1 are as follows:
Discrete Fourier transform is carried out to original image, two dimensional discrete Fourier transform formula is as follows:
Wherein F (u, v) is Fourier transformation result, that is, image frequency domain function, and u and v are frequency component, and f (x, y) is original image, M and N
For the width and height of image, j is imaginary unit;Calculate dimensional Gaussian low pass filter function G (u, v) and image frequency domain
The product of function F (u, v), i.e., filtered image frequency domain function H (u, v);The transmission function of dimensional Gaussian low-pass filter is such as
Under:
δ is standard deviation in formula, chooses δ according to the characteristic of image;
The inverse discrete Fourier transform for calculating filtered image frequency domain function H (u, v) obtains image in the pixel of two-dimensional surface
Distribution function h (x, y);The mould of capture element distribution function h (x, y) is eliminated present in original image and is made an uproar as output image
Acoustic image element.
3. the defects of insulator detection method according to claim 1 based on super-pixel segmentation image recognition, feature exist
In in step 3, setting grid spacings areWherein N is total number of pixels of image.
4. the defects of insulator detection method according to claim 1 based on super-pixel segmentation image recognition, feature exist
In in step 5, the calculation method of the distance d of a super-pixel point to super-pixel center ζ is as follows:
Take coordinate of the pixel in rgb color space and the normalized on progress [0,1] section be named as (r, g,
B), which also carries out the normalized on [0,1] section in the position coordinate value in the two-dimensional space of image and is named as
(x,y);The calculation formula of distance l of the super-pixel point to super-pixel center in color space is as follows:
α, β, γ are the weight in the channel R, the channel G and channel B, (r in formula1, g1, b1)、(r2, g2, b2) it is super-pixel point to super picture
Plain center carries out the coordinate after the normalized on [0,1] section in rgb color space;
Super-pixel point is as follows to distance s calculation formula of the super-pixel center in two-dimensional space:
(x in formula1, y1)、(x2, y2) it is that super-pixel point carries out on [0,1] section in the two-dimensional space of image to super-pixel center
Normalized after position coordinates;Summation is weighted to the distance in color space and two-dimensional space, obtains super-pixel
Point arrives the actual range d at super-pixel center:
L is the distance in color space in formula, and s is the distance in two-dimensional space, in color space and two-dimensional space
The weight of distance is respectively μ, ρ.
5. the defects of insulator detection method according to claim 1 based on super-pixel segmentation image recognition, feature exist
In recalculating its super-pixel center ζ according to the following formula in step 6iPosition in rgb color space and two-dimensional space:
The position coordinates average value for the pixel for taking the region all.
6. the defects of insulator detection method according to claim 5 based on super-pixel segmentation image recognition, feature exist
In in step 8, super-pixel extracts feature respectively on the channel r of rgb color space, the channel g, the channel b and refers to super-pixel image
First moment about the origin E ζ, the second geometric moment E ζ in each channel2With three rank moment of the orign E ζ3, calculation formula is as follows:
Wherein ζi=(ri, gi, bi) be super-pixel image each super-pixel block central pixel point RGB channel on numerical value,
Coordinate (the x of its two-dimensional surface in super-pixel blocki,yi) it is as follows:
Wherein NjFor the pixel number for belonging to the super-pixel, SiFor cluster areas, the i.e. region of the super-pixel,For the sum of the coordinate value of all pixels two-dimensional surface of the super-pixel block;
To first moment about the origin E ζ, second geometric moment E ζ2With three rank moment of the orign E ζ3Calculated result proceed to the normalizing in (0,1) section
Change processing, thus obtains feature vector:
7. the defects of insulator detection method according to claim 1 based on super-pixel segmentation image recognition, feature exist
In in step 12, kernel function is selected from linear, multinomial, radial base RBF or S sigmoid growth curve sigmoid.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910086179.XA CN109886937B (en) | 2019-01-29 | 2019-01-29 | Insulator defect detection method based on super-pixel segmentation image recognition |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910086179.XA CN109886937B (en) | 2019-01-29 | 2019-01-29 | Insulator defect detection method based on super-pixel segmentation image recognition |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109886937A true CN109886937A (en) | 2019-06-14 |
CN109886937B CN109886937B (en) | 2021-08-03 |
Family
ID=66927258
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910086179.XA Active CN109886937B (en) | 2019-01-29 | 2019-01-29 | Insulator defect detection method based on super-pixel segmentation image recognition |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109886937B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111563556A (en) * | 2020-05-11 | 2020-08-21 | 国网陕西省电力公司电力科学研究院 | Transformer substation cabinet equipment abnormity identification method and system based on color gradient weight |
CN112232138A (en) * | 2020-09-25 | 2021-01-15 | 天津大学 | Channel slope damage intelligent identification method based on superpixel characteristics |
CN112991302A (en) * | 2021-03-22 | 2021-06-18 | 华南理工大学 | Flexible IC substrate color-changing defect detection method and device based on super-pixels |
CN113901868A (en) * | 2021-08-25 | 2022-01-07 | 国网四川省电力公司电力科学研究院 | Substation site safety monitoring method and system |
CN114140462A (en) * | 2021-12-10 | 2022-03-04 | 江苏牛犇轴承有限公司 | Bearing wear degree evaluation method based on image processing |
CN114211168A (en) * | 2022-02-21 | 2022-03-22 | 江苏天健智能装备制造有限公司 | Method for correcting plane welding seam track based on image subtraction |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103093470A (en) * | 2013-01-23 | 2013-05-08 | 天津大学 | Rapid multi-modal image synergy segmentation method with unrelated scale feature |
CN103824079A (en) * | 2014-02-08 | 2014-05-28 | 重庆市国土资源和房屋勘测规划院 | Multi-level mode sub block division-based image classification method |
CN104392228A (en) * | 2014-12-19 | 2015-03-04 | 中国人民解放军国防科学技术大学 | Unmanned aerial vehicle image target class detection method based on conditional random field model |
CN104573719A (en) * | 2014-12-31 | 2015-04-29 | 国家电网公司 | Mountain fire detection method based on intelligent image analysis |
CN106097335A (en) * | 2016-06-08 | 2016-11-09 | 安翰光电技术(武汉)有限公司 | Digestive tract focus image identification system and recognition methods |
CN106934418A (en) * | 2017-03-09 | 2017-07-07 | 国家电网公司 | A kind of insulator infrared diagnostics method based on convolution Recursive Networks |
CN107239777A (en) * | 2017-05-13 | 2017-10-10 | 大连理工大学 | A kind of tableware detection and recognition methods based on various visual angles graph model |
US9922425B2 (en) * | 2014-12-02 | 2018-03-20 | Canon Kabushiki Kaisha | Video segmentation method |
CN108399430A (en) * | 2018-02-28 | 2018-08-14 | 电子科技大学 | A kind of SAR image Ship Target Detection method based on super-pixel and random forest |
-
2019
- 2019-01-29 CN CN201910086179.XA patent/CN109886937B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103093470A (en) * | 2013-01-23 | 2013-05-08 | 天津大学 | Rapid multi-modal image synergy segmentation method with unrelated scale feature |
CN103824079A (en) * | 2014-02-08 | 2014-05-28 | 重庆市国土资源和房屋勘测规划院 | Multi-level mode sub block division-based image classification method |
US9922425B2 (en) * | 2014-12-02 | 2018-03-20 | Canon Kabushiki Kaisha | Video segmentation method |
CN104392228A (en) * | 2014-12-19 | 2015-03-04 | 中国人民解放军国防科学技术大学 | Unmanned aerial vehicle image target class detection method based on conditional random field model |
CN104573719A (en) * | 2014-12-31 | 2015-04-29 | 国家电网公司 | Mountain fire detection method based on intelligent image analysis |
CN106097335A (en) * | 2016-06-08 | 2016-11-09 | 安翰光电技术(武汉)有限公司 | Digestive tract focus image identification system and recognition methods |
CN106934418A (en) * | 2017-03-09 | 2017-07-07 | 国家电网公司 | A kind of insulator infrared diagnostics method based on convolution Recursive Networks |
CN107239777A (en) * | 2017-05-13 | 2017-10-10 | 大连理工大学 | A kind of tableware detection and recognition methods based on various visual angles graph model |
CN108399430A (en) * | 2018-02-28 | 2018-08-14 | 电子科技大学 | A kind of SAR image Ship Target Detection method based on super-pixel and random forest |
Non-Patent Citations (1)
Title |
---|
RADHAKRISHNA ACHANTA ET AL: ""SLIC Superpixels"", 《RESEARCHGATE》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111563556A (en) * | 2020-05-11 | 2020-08-21 | 国网陕西省电力公司电力科学研究院 | Transformer substation cabinet equipment abnormity identification method and system based on color gradient weight |
CN112232138A (en) * | 2020-09-25 | 2021-01-15 | 天津大学 | Channel slope damage intelligent identification method based on superpixel characteristics |
CN112991302A (en) * | 2021-03-22 | 2021-06-18 | 华南理工大学 | Flexible IC substrate color-changing defect detection method and device based on super-pixels |
CN112991302B (en) * | 2021-03-22 | 2023-04-07 | 华南理工大学 | Flexible IC substrate color-changing defect detection method and device based on super-pixels |
CN113901868A (en) * | 2021-08-25 | 2022-01-07 | 国网四川省电力公司电力科学研究院 | Substation site safety monitoring method and system |
CN114140462A (en) * | 2021-12-10 | 2022-03-04 | 江苏牛犇轴承有限公司 | Bearing wear degree evaluation method based on image processing |
CN114140462B (en) * | 2021-12-10 | 2023-09-08 | 江苏牛犇轴承有限公司 | Bearing wear degree assessment method based on image processing |
CN114211168A (en) * | 2022-02-21 | 2022-03-22 | 江苏天健智能装备制造有限公司 | Method for correcting plane welding seam track based on image subtraction |
Also Published As
Publication number | Publication date |
---|---|
CN109886937B (en) | 2021-08-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109886937A (en) | Defects of insulator detection method based on super-pixel segmentation image recognition | |
CN106874894B (en) | Human body target detection method based on regional full convolution neural network | |
CN109118479B (en) | Capsule network-based insulator defect identification and positioning device and method | |
CN106780612B (en) | Object detecting method and device in a kind of image | |
Huijuan et al. | Fast image matching based-on improved SURF algorithm | |
CN103839065B (en) | Extraction method for dynamic crowd gathering characteristics | |
CN105678788B (en) | A kind of fabric defect detection method based on HOG and low-rank decomposition | |
CN109711288A (en) | Remote sensing ship detecting method based on feature pyramid and distance restraint FCN | |
CN108710913A (en) | A kind of switchgear presentation switch state automatic identification method based on deep learning | |
CN110930387A (en) | Fabric defect detection method based on depth separable convolutional neural network | |
CN113920107A (en) | Insulator damage detection method based on improved yolov5 algorithm | |
CN112818969A (en) | Knowledge distillation-based face pose estimation method and system | |
CN110399820B (en) | Visual recognition analysis method for roadside scene of highway | |
CN109829881A (en) | Bird's Nest detection method and system based on deep learning | |
CN109785288A (en) | Transmission facility defect inspection method and system based on deep learning | |
CN109829412A (en) | The Partial Discharge Pattern Recognition Method of fractal characteristic is decomposed based on dynamic mode | |
CN111915583A (en) | Vehicle and pedestrian detection method based on vehicle-mounted thermal infrared imager in complex scene | |
CN114565594A (en) | Image anomaly detection method based on soft mask contrast loss | |
Zhang et al. | Research on surface defect detection of rare-earth magnetic materials based on improved SSD | |
CN112270289A (en) | Intelligent monitoring method based on graph convolution attention network | |
CN114863258B (en) | Method for detecting small target based on visual angle conversion in sea-sky-line scene | |
CN109829511B (en) | Texture classification-based method for detecting cloud layer area in downward-looking infrared image | |
CN116402769A (en) | High-precision intelligent detection method for textile flaws considering size targets | |
CN116109849A (en) | SURF feature matching-based high-voltage isolating switch positioning and state identification method | |
Tan et al. | Automobile component recognition based on deep learning network with coarse-fine-grained feature fusion |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |