CN106127756A - A kind of insulator recognition detection method based on multicharacteristic information integration technology - Google Patents

Abstract

The invention discloses a kind of insulator recognition detection method based on multicharacteristic information integration technology, specifically implement according to following steps: step 1, the picture signal of insulator on the transmission line of electricity of the focusing video camera collection that utilization is arranged on the band The Cloud Terrace on on-the-spot steel tower or iron；Step 2, collects image and carries out pretreatment step 1；Step 3, extracts the texture feature vector of insulator from pretreated image；Step 4, extracts insulator color feature vector to be identified, and the HSV color space of the insulator image collected in step 1 is carried out unequal interval quantization；Step 5, uses and extracts insulator shape facility based on Hu square algorithm in region description method；The 3 kinds of features extracting insulator in step 3, step 4, step 5 are merged by step 6, and the method for the present invention can detect the running status of insulator simply, reliably, fast and automatically, thus prevents the electric power system fault caused because of Insulator detection.

Description

A kind of insulator recognition detection method based on multicharacteristic information integration technology

Technical field

The invention belongs to power system on-line monitoring field, relate to a kind of insulator based on multicharacteristic information integration technology Recognition detection method.

Background technology

Along with the proposition of global energy the Internet theory conception, with extra-high voltage grid as bulk transmission grid, with powerful intelligence Electrical network is for relying on, and monitoring, security maintenance and the operational management in real time of strengthening transmission line of electricity are the most important things.And at high voltage transmission line Lu Zhong, defective insulation is main cause accident occur, because Insulator detection causes accident to become electric power system fault rate at present First.Insulator runs the most out of doors, and in air, the factor such as various dunghill interference, easily causes interior insulator and split The faults such as the reduction of stricture of vagina, surface fracture, dielectric strength and pollution flashover.When insulator chain exists low value or zero resistance insulator, powerful Lightning current and power frequency continued flow flow through from the porcelain body of zero resistance insulator head, easily cause zero resistance insulator to cross thermal spalling.If The running status grasping insulator early, it will the fault of a lot of power system is reduced or avoided.Traditional inspection insulator The method of running status is periodically to have a power failure or charged manual detection, and these operations not only need to step on bar and detect piecewise, and high-altitude Operation height big, dangerous affected by environment, work efficiency are relatively low.Therefore, need not step on bar the most simply, reliably, the most automatic The running status of monitoring insulator is to reduce human input, eliminates safe hidden trouble, it is ensured that safe operation of power system is to be explored An important technology difficult problem.

Summary of the invention

It is an object of the invention to provide a kind of insulator recognition detection method based on multicharacteristic information integration technology, it is possible to Detect the running status of insulator simply, reliably, fast and automatically, thus the power system that causes because of Insulator detection of prevention therefore Barrier.

The technical solution adopted in the present invention is, a kind of insulator recognition detection side based on multicharacteristic information integration technology Method, specifically implements according to following steps:

Step 1, utilizes on the transmission line of electricity of focusing video camera collection of the band The Cloud Terrace being arranged on on-the-spot steel tower or iron The picture signal of insulator；

Step 2, collects image and carries out pretreatment step 1, and concrete step is: utilize optimal entropic threshold method (OET) Carry out image segmentation, by analyzing the entropy of image grey level histogram, find optimal threshold；

Step 3, carried out extracting pretreated image the texture feature vector of insulator from step 2；

Step 4, extracts insulator color feature vector to be identified, the HSV color to the insulator image collected in step 1 Color space carries out unequal interval quantization；

Step 5, extracts insulator shape eigenvectors to be identified, and the shape facility of insulator is unrelated with external environment, The external the most stable information of object, be also image the most intuitively, the most directly visualize, use based on Hu in region description method Square algorithm extracts insulator shape facility；

The 3 kinds of features extracting insulator in step 3, step 4, step 5 are merged, three kinds of single features by step 6 During fusion, each characteristic quantity was normalized before similarity measurement, adjust target image to be identified and scheme in data base As the weight between three features.Determine its weight with analytic hierarchy process (AHP), calculate Consistency Ratio CR, verify judgment matrix three Whether the weights of Feature Fusion meet the requirements.

The feature of the present invention also resides in,

Step 2 is specifically implemented according to following steps:

Step 2.1, the tonal range of the image collected in step 1 is designated as 0,1,2 ... L-1}；

Step 2.2, if the region that the pixel that gray level is less than t is constituted is target area A, then the table of target area A entropy Reaching formula is:

$H_A\left(t\right)=-\underset{i}{\Σ}\frac{p_i}{p_t}lg\frac{p_i}{p_t}---(2-1)$

P in formula_iRepresent that gray level is the probability that i occurs, i=0,1,2 ... t；P_tRepresent that gray level is the general of t appearance Rate；

Step 2.3, if the region that the pixel that gray level is higher than t is constituted is background area B, then the table of background area B entropy Reaching formula is:

$H_B\left(t\right)=-\underset{i}{\Σ}\frac{p_i}{1-p_t}lg\frac{p_i}{1-p_t}---(2-2)$

I=t+1, t+2, t+3 ... .L-1；

P in formula_iRepresent that gray level is the probability that i occurs, i=0,1,2 ... t；P_tRepresent that gray level is the general of t appearance Rate；

Step 2.4, calculates max-thresholds,

Define according to entropy function,

$C\left(t\right)=H_A\left(t\right)+H_B\left(t\right)=\lg \frac{p_t}{1-p_t}+\frac{H_t}{p_t}+\frac{H_L-H_t}{1-P_t}---(2-3)$

In formulaI=0 therein, 1,2 ... t；I=0,1 therein, 2 ... L-1；

Then, when entropy function C (t) is maximum,

Calculate t=argmax{C (t) } (2-4)

Mean that the max-thresholds of gray scale t.

Step 3 is specifically implemented according to following steps:

Step 3.1, utilizes 3 color spaces to former RGB color image of the statistical method in texture characteristic extracting method Calculate average, variance, the degree of bias, kurtosis, energy, entropy totally 6 the color space features parameter as textural characteristics respectively,

Specific formula for calculation is as follows:

(1) average

(2) variance

(3) degree of bias

(4) kurtosis

(5) energy

(6) entropy

Wherein g (i) is the gray value of i-gray level, p_gI () is the probability of i-th gray value；

Former RGB color image is just divided into 18 characteristic vectors altogether by these 6 features.The single features that image is different is entered Carry out feature internal normalization during row feature extraction, use Gaussian normalization method these 18 characteristic vectors to be carried out inside feature Normalize to [-1,1], reduce other element value and the distribution of the element value after normalization is produced impact.Returning of characteristic vector One changes F_iComputing formula as follows:

$F_i=\frac{f_i-{\mathrm{\μ}}_i}{{\mathrm{\σ}}_i};---(3-7)$

${\mathrm{\μ}}_i=\frac{\underset{i}{\Σ}{f}_i}{i};---(3-8)$

${\mathrm{\σ}}_i=\sqrt{\frac{{(\underset{i}{\Σ}{f}_i-{\mathrm{\μ}}_i)}^2}{i-1}};---(3-9)$

Wherein μ_iRepresent the average of target image characteristics to be identified vector, σ_iRepresent target image characteristics to be identified vector Standard deviation, f_iIt is 18 one dimensional histograms initial characteristic values and f_i∈[f₁,f₂,f₃,.......f₁₈]；I=1,2,3 ... 18；

Step 3.2, by the texture feature vector extracted in step 3.1 and the stricture of vagina in the sample database established in advance Reason characteristic vector carries out contrast coupling, calculates its similarity, method particularly includes:

Chessboard distance formula is utilized to calculate the similarity that target image to be identified mates with textural characteristics in characteristic image storehouse Distance, the normalization characteristic vector of two width images is that 1 18 dimensional vector is represented by:

F_xi[F_x1,F_x2,F_x3,......F_x18]；

F_yi[F_y1,F_y2,F_y3,......F_y18]；

Then, calculate according to chessboard distance formula:

$S_1=\underset{i}{max}|F_{xi}-F_{yi}|---(3-10)$

Wherein S₁Represent the similarity distance that target image to be identified mates, F with textural characteristics in characteristic image storehouse_xiAnd F_yi Represent the normalization characteristic vector of two width images, S₁Being worth the biggest, the similarity of two width images is the highest.

Step 4 is specifically implemented according to following steps:

Step 4.1, is divided into 8 parts by tone H space, and saturation S and brightness V space are divided into 3 parts, the tone H of image Scope is that [0,360 °], saturation S and brightness V are in the range of [0,1]；

Step 4.2, carries out color quantization according to the different range of H, S, V, specific as follows shown:

$S=\left\{\begin{array}{cc}0,& 0\≤S\≤0.2\\ 1,& 0.2\≤S\≤0.7\\ 2,& 0.7\≤S\≤1.0\end{array}\right.---(4-2)$

$r=\left\{\begin{array}{cc}0,& 0\≤V\≤0.2\\ 1,& 0.2\≤V\≤0.7\\ 2,& 0.7\≤V\≤1.0\end{array}\right.---(4-3)$

3 color components after quantization are synthesized one-dimensional characteristic vectorial:

I=HQ_sQ_v+SQ_v+V (4-4)

I=HQ_sQ_vQ in+SQ formula_s=3 is the quantization progression of component S；Q_v=3 is component V

Quantization progression time 3-4 be represented by:

I=9H+3S+V (4-5)

Quantify the subspace that the HSV space after terminating resolves into, it is thus achieved that 72 one dimensional histograms；

Step 4.3, the characteristic vector Gaussian normalization of the color histogram after quantifying in step 4.2, to [-1,1], subtracts Few other element value produces impact, the normalized F of characteristic vector to the distribution of the element value after normalization_lComputing formula is such as Under:

$F_l=\frac{f_l-{\mathrm{\μ}}_l}{{\mathrm{\σ}}_l};$

${\mathrm{\μ}}_l=\frac{\underset{l}{\Σ}{f}_l}{l};$

${\mathrm{\σ}}_l=\sqrt{\frac{{(\underset{l}{\Σ}{f}_l-{\mathrm{\μ}}_l)}^2}{l-1}}---(4-6)$

Wherein μ_lRepresent the average of target image characteristics to be identified vector, σ_lRepresent target image characteristics to be identified vector Standard deviation, f_lIt is 72 one dimensional histograms initial characteristic values and f_l∈[f₁,f₂,f₃,.......f₇₂]；L=1,2,3 ... 72,

Step 4.4, utilizes chessboard distance formula to calculate target image to be identified and mates with color characteristic in characteristic image storehouse Similarity distance.The normalization characteristic of two width images is represented by for 2 72 dimensional vectors:

F_pl[F_p1,F_p2,F_p3,......F_p72]；F_ql[F_q1,F_q2,F_q3,......F_q72]；

Wherein chessboard distance computing formula:

$S_2=\underset{l}{max}|F_{pl}-F_{ql}|---(4-7)$

Wherein S₂Represent the similarity distance that target image to be identified mates, F with textural characteristics in characteristic image storehouse_plAnd F_ql Represent the normalization characteristic vector of two width images, S₂Value maximum, the similarity of two width images is the highest.

Step 5 is specifically implemented according to following steps:

Step 5.1, definition collect image f (x, p+q rank square y) is:

$m_{pq}=\underset{x}{\Σ}{x}^p{y}^{q}f(x,y)---(5-1)$

Then, the central moment of p+q rank square is:

${\mathrm{\μ}}_{pq}=\underset{x}{\Σ}\underset{y}{\Σ}{(x-x_0)}^p{(y-y_0)}^q---(5-2)$

In formulaRepresent the center of gravity of image-region；

Step 5.2, in order to the central moment in (5-2) formula can be returned by the character obtaining image itself unrelated with scaling One changes, and the central moment after normalization is expressed as:

${\mathrm{\η}}_{pq}=\frac{{\mathrm{\μ}}_{pq}}{{{\mathrm{\μ}}^r}_{00}}---(5-3)$

In formulaP+q=2,3,4 ... normalized centre-to-centre spacing is to the translation of object, scaling

All keep constant with rotating；

Step 5.3, (x, y) second order and third central moment, obtain to utilize formula (5-2) to calculate image f

${\mathrm{\μ}}_{02}=\underset{x}{\Σ}\underset{y}{\Σ}{(x-x_0)}^0{(y-y_0)}^2$

${\mathrm{\μ}}_{20}=\underset{x}{\Σ}\underset{y}{\Σ}{(x-x_0)}^2{(y-y_0)}^0---(5-4)$

${\mathrm{\μ}}_{03}=\underset{x}{\Σ}\underset{y}{\Σ}{(x-x_0)}^0{(y-y_0)}^3$

${\mathrm{\μ}}_{30}=\underset{x}{\Σ}\underset{y}{\Σ}{(x-x_0)}^3{(y-y_0)}^0---(5-5)$

Wherein second-order moment around mean μ₀₂And μ₂₀Represent respectively around the inertia by the vertically and horizontally axis of gray scale barycenter Square, third central moment μ₀₃And μ₃₀The amplitude mensurable institute analyzed area asymmetric degree to vertically and horizontally axis；

Construct 7 not bending moments, translation, scaling and invariable rotary can be kept under conditions of consecutive image.Described 7 are not Bending moment is defined respectively as:

Central moment expression formula after second order normalized:

${\mathrm{\η}}_{02}=\frac{{\mathrm{\μ}}_{02}}{{{\mathrm{\μ}}^2}_{00}};{\mathrm{\η}}_{20}=\frac{{\mathrm{\μ}}_{20}}{{{\mathrm{\μ}}^2}_{00}};{\mathrm{\η}}_{11}=\frac{{\mathrm{\μ}}_{11}}{{{\mathrm{\μ}}^2}_{00}};---(5-6)$

Central moment expression formula after three rank normalization:

${\mathrm{\η}}_{12}=\frac{{\mathrm{\μ}}_{12}}{{{\mathrm{\μ}}^{\frac{5}{2}}}_{00}};{\mathrm{\η}}_{21}=\frac{{\mathrm{\μ}}_{21}}{{{\mathrm{\μ}}^{\frac{5}{2}}}_{00}};{\mathrm{\η}}_{03}=\frac{{\mathrm{\μ}}_{03}}{{{\mathrm{\μ}}^{\frac{5}{2}}}_{00}};{\mathrm{\η}}_{30}=\frac{{\mathrm{\μ}}_{30}}{{{\mathrm{\μ}}^{\frac{5}{2}}}_{00}}---(5-7)$

φ₁=η₂₀+η₀₂ (5-8)

φ₂=(η₂₀+η₀₂)²+4η² ₁₁ (5-9)

φ₃=(η₃₀-3η₁₂)²+(3η₂₁-η₀₃)² (5-10)

φ₄=(η₃₀+η₁₂)²+(η₂₁+η₀₃)² (5-11)

φ₅=(η₃₀-3η₁₂)(η₃₀+η₁₂)[(η₂₀+η₁₂)²-3(η₂₁+η₀₃)²]+(3η₂₁+η₀₃)(η₂₁+η₀₃)[3(η₃₀+η₁₂ )²-(η₂₁+η₀₃)] (5-12)

φ₆=(η₂₀-η₀₂)[(η₃₀+η₁₂)²-(η₂₁+η₀₃)²]+4η₁₁(η₃₀+η₁₂)(η₂₁+η₀₃) (5-13)

φ₇=(3 η₂₁-η₀₃)(η₃₀+η₁₂)[(η₃₀+η₁₂)²-3(η₂₁+η₀₃)²]+(3η₂₁-η₀₃)(η₂₁+η₀₃)[3(η₃₀+η₁₂ )²-(η₂₁+η₀₃)²] (5-14)

φ therein₁～φ₇Being the computing formula of 7 not bending moments, each η value represents the center after second order, three rank normalization Square；

Step 5.4, extracts 7 shape spies of insulator to be identified by the computing formula of Hu square algorithm (5-8)～(5-14) Levy vector, then to each characteristic vector Gaussian normalization to [-1,1].By target image to be identified and shape in characteristic image storehouse Feature normalization is that 27 dimensional vectors are represented by:

F_si[F_s1,F_s2,F_s3,......F_s7]；F_ti[F_t1,F_t2,F_t3,......F_t7]；

Equally, chessboard distance formula is utilized to calculate target image to be identified and the mating of shape facility in characteristic image storehouse Similarity distance S₃:

$S_3=\underset{i}{max}|F_{si}-F_{ti}|---(5-15)$

S₃Value maximum, the similarity of two width images is the highest；

Step 5.5, carries out Feature Fusion by the color of insulator, shape and the textural characteristics that extract.If three kinds of features The similarity (distance) of Vector Fusion is D, the similarity of three kinds of single features vectors (feature of color, shape and texture) (away from From) S₁,S₂,S₃Corresponding weight is respectively w₁,w₂,w₃, then the similarity (distance) of three kinds of Feature Fusion vectors is,

D=s₁w₁+s₂w₂+s₃w₃ (5-16)

∑ w in formula_i=1；I=1,2,3.

Step 6 is specifically implemented according to following steps:

Step 6.1, if texture, color and the shape eigenvectors of a target image i to be identified are [O_i1,O_i2,O_i3], Treating N number of image in recognition target image i and data base by chessboard distance and carry out Similarity Measure, computed range value is remembered respectively For D_i1,D_i2,......D_iN；D_iNAlso it is target image to be identified and three characteristic distances of any one image in data picture library:

D_iN=[d_1,iN,d_2,iN,d_3,iN] (6-1)

Step 6.2, calculates the mean μ of 3N distance value_DAnd standard deviation sigma_D, specific formula for calculation is:

${\mathrm{\μ}}_D=\frac{1}{3N}(\underset{a=1}{\overset{N}{\Σ}}{f}_a+\underset{b=1}{\overset{N}{\Σ}}{f}_b+\underset{c=1}{\overset{N}{\Σ}}{f}_c)$

${\mathrm{\σ}}_D=\sqrt{\frac{{(\underset{a=1}{\overset{N}{\Σ}}{f}_a-{\mathrm{\μ}}_D)}^2+{(\underset{b=1}{\overset{N}{\Σ}}{f}_b-{\mathrm{\μ}}_D)}^2+{(\underset{c=1}{\overset{N}{\Σ}}{f}_c-{\mathrm{\μ}}_D)}^2}{N-1}}---(6-2)$

F therein_a、f_b、f_cRepresent texture, color and shape eigenvectors respectively；

Step 6.3, treats recognition target image and N width image (N takes any positive integer) similarity distance D in feature database_i1, D_i2,......D_iNIt is normalized, makes three kinds of different characteristic vectors roughly the same on Similarity Measure, D_iNValue normalizing After change, major part falls interval in [0,1], and its expression formula is:

${D^{*}}_{iN}=\frac{\frac{D_{iN}-{\mathrm{\μ}}_D}{3{\mathrm{\σ}}_D}+1}{2}---(6-3)$

Step 6.4, uses analytic hierarchy process (AHP) to determine the weight of three Feature Fusion.With 1-9 scaling law to Feature Fusion after Three feature description index importance of insulator image judge, the judgment matrix A utilizing policymaker to provide derives three Weight s of feature_i, then the concordance met with Consistency Ratio CR test and judge matrix, thus verify and derived three features Weighted value accuracy；

Step 6.5, calculates Consistency Ratio CR, and whether utilize Consistency Ratio CR test and judge matrix to meet consistent Property, and then determine the weighted value of Feature Fusion.

Step 6.4 is specifically implemented according to following steps:

Step 6.4.1, according to 1-9 scaling law to three feature weight s₁,s₂,s₃Importance proportion quotiety two-by-two is carried out Assignment, formation judgment matrix:

A₁=(a_pq)_n×nWherein a_pqFor pth feature relative to the proportion quotiety of q feature importance,

Step 6.4.2, takes a width insulator image, in hierarchy Model, this insulator image is only taken ground floor Secondary, and this level can be divided into three classifications, i.e. texture (C), color (S) and shape (T), then corresponding judgment matrix represents For:

$A_1={\left(a_{pq}\right)}_{3\×3}=\left[\begin{array}{ccc}\frac{H_C}{H_C}& \frac{H_C}{H_S}& \frac{H_C}{H_T}\\ \frac{H_S}{H_C}& \frac{H_S}{H_S}& \frac{H_S}{H_T}\\ \frac{H_T}{H_C}& \frac{H_T}{H_S}& \frac{H_T}{H_T}\end{array}\right]---(6-4)$

ThereinIt is respectively color and shape facility proportion quotiety, texture and color characteristic ratio Scale, texture and shape facility proportion quotiety.It is typically based on user's priori and extracts the interpretation of result of three features, shape Other two aspect ratios of aspect ratio are great,

When starting, provide one initial value of scale value, i.e. setJudge consistent Sex rate CR.

Step 6.4.3, to A₁In the element multiplication often gone open cube and obtain vectorial Y_i=(y₁,y₂,y₃), computing formula is such as Under:

$y_1=\sqrt[3]{\frac{H_C}{H_C}\×\frac{H_C}{H_S}\×\frac{H_C}{H_T}}---(6-5)$

$y_2=\sqrt[3]{\frac{H_S}{H_C}\×\frac{H_S}{H_S}\×\frac{H_S}{H_T}}---(6-6)$

$y_3=\sqrt[3]{\frac{H_T}{H_C}\×\frac{H_T}{H_S}\×\frac{H_T}{H_T}}---(6-7)$

To Y_iIt is normalized and obtains normalized weight vector s_i=(s₁,s₂,s₃)

$s_i=\frac{Y_i}{\underset{i}{\overset{3}{\Σ}}{Y}_i}---(6-8)$

Step 6.4.4, to A₁In every column element sue for peace to obtain Z_q=(Z₁,Z₂,Z₃)

$Z_q=\underset{i}{\overset{3}{\Σ}}{a}_{pq}---(6-9)$

Step 6.4.5, calculates judgment matrix A₁Maximum eigenvalue

${\mathrm{\λ}}_{max}=\underset{i}{\overset{3}{\Σ}}{s}_i{Z}_q---(6-10).$

Step 6.5 is specifically implemented according to following steps:

Step 6.5.1, calculates judgment matrix A₁Coincident indicator CI,

$CI=\frac{{\mathrm{\λ}}_{max}-n}{n-1}---(6-11)$

Wherein n is judgment matrix A₁Exponent number, the Maximum characteristic root λ of the least judgment matrix of CI_maxMore meet completely the same Property, the degree of CI the biggest explanation judgment matrix deviation crash consistency is the biggest.

Judgment matrix A is calculated according to 6-10₁Other two eigenvalue is designated as respectively: λ₂、λ₃. seek these three eigenvalue of maximum Meansigma methodsComputing formula is as follows:

${\widetilde{\mathrm{\λ}}}_{max}=\frac{{\mathrm{\λ}}_{max}+{\mathrm{\λ}}_2+{\mathrm{\λ}}_3}{3}---(6-12),$

Step 6.5.2, calculates Aver-age Random Consistency Index RI according to formula (9-10)

$RI=\frac{{\widetilde{\mathrm{\λ}}}_{max}-n}{n-1}---(6-13)$

Step 6.5.3, CI and RI that utilization calculates, calculating Consistency Ratio CR:

$CR=\frac{CI}{RI}---(6-14)$

As CR < judgment matrix A when 0.1₁There is satisfactory consistency, i.e. weight vectors s_iMeet matrix equation: A₁s_i=λ_maxs_i There is the eigenvalue λ of maximum_max, so judgment matrix A₁The weighted value of three Feature Fusion meets the requirements, the power used when now calculating The weighted value that weight values uses when being image co-registration in this method, the insulator identification image after output fusion；

When CR >=0.1, it is judged that matrix A₁Not there is satisfactory consistency, then go to step 6.4 and re-start assignment decision-making Person needs the most rightThree scales carry out assignment according to 1-9 scaling law, with front once calculate time assignment It is adjusted after contrast, constructs new judgment matrix A₁, then calculate weight vector s respectively_i, maximum feature λ of judgment matrix_maxAnd one Cause property index CI, finally calculates Consistency Ratio CR, until CR < 0.1 sets up.

The invention has the beneficial effects as follows, compared with existing system, implement the equipment used in the present invention less, structure letter Single, with low cost, this method can make full use of multiple features fusion thought in image processing techniques, by real-time interception in video flowing Insulator target image carry out three Feature Fusion match cognization with insulator feature picture library, can real time on-line monitoring insulator Running status.It is easy to Surveillance center's teleprocessing data and the ruuning situation to the insulator of electric power networks carries out overall prison Control, is advantageously implemented the security monitoring of automatization.

Accompanying drawing explanation

Fig. 1 is the flow chart of a kind of based on multicharacteristic information integration technology the insulator recognition detection method of the present invention.

Detailed description of the invention

The present invention is described in detail with detailed description of the invention below in conjunction with the accompanying drawings.

The invention provides a kind of insulator recognition detection method based on multicharacteristic information integration technology, as it is shown in figure 1, Specifically implement according to following steps:

Step 1, utilizes on the transmission line of electricity of focusing video camera collection of the band The Cloud Terrace being arranged on on-the-spot steel tower or iron The picture signal of insulator；

Step 2, collects image and carries out pretreatment step 1, and concrete step is: utilize optimal entropic threshold method (OET) Carry out image segmentation, by analyzing the entropy of image grey level histogram, find the optimal threshold, concrete step to be:

Step 2.1, the tonal range of the image collected in step 1 is designated as 0,1,2 ... L-1}.

Step 2.2, if the region that the pixel that gray level is less than t is constituted is target area A, then the table of target area A entropy Reaching formula is:

$H_A\left(t\right)=-\underset{i}{\&Sigma;}\frac{p_i}{p_t}lg\frac{p_i}{p_t}---(2-1)$

P in formula_iRepresent that gray level is the probability that i occurs, i=0,1,2 ... t；P_tRepresent that gray level is the general of t appearance Rate.

Step 2.3, if the region that the pixel that gray level is higher than t is constituted is background area B, then the table of background area B entropy Reaching formula is:

$H_B\left(t\right)=-\underset{i}{\&Sigma;}\frac{p_i}{1-p_t}lg\frac{p_i}{1-p_t}---(2-2)$

I=t+1, t+2, t+3 ... .L-1；

P in formula_iRepresent that gray level is the probability that i occurs, i=0,1,2 ... t；P_tRepresent that gray level is the general of t appearance Rate.

Step 2.4, calculates max-thresholds,

Define according to entropy function,

$C\left(t\right)=H_A\left(t\right)+H_B\left(t\right)=\lg \frac{p_t}{1-p_t}+\frac{H_t}{p_t}+\frac{H_L-H_t}{1-P_t}---(2-3)$

In formulaI=0 therein, 1,2 ... t；I=0,1 therein, 2 ... L-1；

Then, when entropy function C (t) is maximum,

Calculate t=arg max{C (t) } (2-4)

Mean that the max-thresholds of gray scale t.

Step 3, carried out extracting pretreated image the texture feature vector of insulator, concrete step from step 2 Rapid as follows:

Step 3.1, utilizes 3 color spaces to former RGB color image of the statistical method in texture characteristic extracting method Calculate average, variance, the degree of bias, kurtosis, energy, entropy totally 6 the color space features parameter as textural characteristics respectively.

Specific formula for calculation is as follows:

(1) average

(2) variance

(3) degree of bias

(4) kurtosis

(5) energy

(6) entropyIts Middle g (i) is the gray value of i-gray level, p_gI () is the probability of i-th gray value.

Former RGB color image is just divided into 18 characteristic vectors altogether by these 6 features.The single features that image is different is entered Carry out feature internal normalization during row feature extraction, use Gaussian normalization method these 18 characteristic vectors to be carried out inside feature Normalize to [-1,1], reduce other element value and the distribution of the element value after normalization is produced impact.Returning of characteristic vector One changes F_iComputing formula as follows:

$F_i=\frac{f_i-{\mathrm{\&mu;}}_i}{{\mathrm{\&sigma;}}_i};---(3-7)$

${\mathrm{\&mu;}}_i=\frac{\underset{i}{\&Sigma;}{h}_i}{i};---(3-8)$

${\mathrm{\&sigma;}}_i=\sqrt{\frac{{(\underset{i}{\&Sigma;}{f}_i-{\mathrm{\&mu;}}_i)}^2}{i-1}};---(3-9)$

Wherein μ_iRepresent the average of target image characteristics to be identified vector, σ_iRepresent target image characteristics to be identified vector Standard deviation, f_iIt is 18 one dimensional histograms initial characteristic values and f_i∈[f₁,f₂,f₃,.......f₁₈]；I=1,2,3 ... 18.

Step 3.2, by the texture feature vector extracted in step 3.1 and the stricture of vagina in the sample database established in advance Reason characteristic vector carries out contrast coupling, calculates its similarity, method particularly includes:

Chessboard distance formula is utilized to calculate the similarity that target image to be identified mates with textural characteristics in characteristic image storehouse Distance, the normalization characteristic vector of two width images is that 1 18 dimensional vector is represented by:

F_xi[F_x1,F_x2,F_x3,......F_x18]；

F_yi[F_y1,F_y2,F_y3,......F_y18]；

Then, calculate according to chessboard distance formula:

$S_1=\underset{i}{max}|F_{xi}-F_{yi}|---(3-10)$

Wherein S₁Represent the similarity distance that target image to be identified mates, F with textural characteristics in characteristic image storehouse_xiAnd F_yi Represent the normalization characteristic vector of two width images, S₁Being worth the biggest, the similarity of two width images is the highest.

Step 4, extracts insulator color feature vector to be identified, the HSV color to the insulator image collected in step 1 Color space carries out unequal interval quantization, specifically implements according to following steps:

Step 4.1, is divided into 8 parts by tone H space, and saturation S and brightness V space are divided into 3 parts, the tone H of image Scope is that [0,360 °], saturation S and brightness V are in the range of [0,1].

Step 4.2, carries out color quantization according to the different range of H, S, V, specific as follows shown:

$S=\left\{\begin{array}{cc}0,& 0\&le;S\&le;0.2\\ 1,& 0.2\&le;S\&le;0.7\\ 2,& 0.7\&le;S\&le;1.0\end{array}\right.---(4-2)$

$V=\{\begin{array}{cc}0,& 0\&le;V\&le;0.2\\ 1,& 0.2\&le;V\&le;0.7\\ 2,& 0.7\&le;V\&le;1.0\end{array}---(4-3)$

3 color components after quantization are synthesized one-dimensional characteristic vectorial:

I=HQ_sQ_v+SQ_v+V (4-4)

I=HQ_sQ_vQ in+SQ formula_s=3 is the quantization progression of component S；Q_v=3 3-4 when being the quantization progression of component V can table It is shown as:

I=9H+3S+V (4-5)

Quantify the subspace that the HSV space after terminating resolves into, it is thus achieved that 72 one dimensional histograms.

Step 4.3, the characteristic vector Gaussian normalization of the color histogram after quantifying in step 4.2, to [-1,1], subtracts Few other element value produces impact, the normalized F of characteristic vector to the distribution of the element value after normalization_lComputing formula is such as Under:

$F_l=\frac{f_l-{\mathrm{\&mu;}}_l}{{\mathrm{\&sigma;}}_l};$

${\mathrm{\&mu;}}_l=\frac{\underset{l}{\&Sigma;}{f}_l}{l};$

${\mathrm{\&sigma;}}_l=\sqrt{\frac{{(\underset{l}{\&Sigma;}{f}_l-{\mathrm{\&mu;}}_l)}^2}{l-1}}---(4-6)$

Wherein μ_lRepresent the average of target image characteristics to be identified vector, σ_lRepresent target image characteristics to be identified vector Standard deviation, f_lIt is 72 one dimensional histograms initial characteristic values and f_l∈[f₁,f₂,f₃,.......f₇₂]；L=1,2,3 ... 72.

Step 4.4, utilizes chessboard distance formula to calculate target image to be identified and mates with color characteristic in characteristic image storehouse Similarity distance.The normalization characteristic of two width images is represented by for 2 72 dimensional vectors:

F_pl[F_p1,F_p2,F_p3,......F_p72]；F_ql[F_q1,F_q2,F_q3,......F_q72]；

Wherein chessboard distance computing formula:

$S_2=\underset{l}{max}|F_{pl}-F_{ql}|---(4-7)$

Wherein S₂Represent the similarity distance that target image to be identified mates, F with textural characteristics in characteristic image storehouse_plAnd F_ql Represent the normalization characteristic vector of two width images, S₂Value maximum, the similarity of two width images is the highest.

Step 5, extracts insulator shape eigenvectors to be identified, and the shape facility of insulator is unrelated with external environment, The external the most stable information of object, be also image the most intuitively, the most directly visualize, use based on Hu in region description method Square algorithm extracts insulator shape facility, specifically implements according to following steps:

Step 5.1, definition collect image f (x, p+q rank square y) is:

$m_{pq}=\underset{x}{\&Sigma;}{x}^p{y}^{q}f(x,y)---(5-1)$

Then, the central moment of p+q rank square is:

${\mathrm{\&mu;}}_{pq}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^p{(y-y_0)}^q---(5-2)$

In formulaRepresent the center of gravity of image-region.

Step 5.2, in order to the central moment in (5-2) formula can be returned by the character obtaining image itself unrelated with scaling One changes, and the central moment after normalization is expressed as:

${\mathrm{\&eta;}}_{pq}=\frac{{\mathrm{\&mu;}}_{pq}}{{{\mathrm{\&mu;}}^r}_{00}}---(5-3)$

In formulaP+q=2,3,4 ... normalized centre-to-centre spacing to the translation of object, scale and rotate All keep constant.

Step 5.3, utilize formula (5-2) calculate image f (x, y) second order and third central moment, obtain,

${\mathrm{\&mu;}}_{02}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^0{(y-y_0)}^2$

${\mathrm{\&mu;}}_{20}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^2{(y-y_0)}^0---(5-4)$

${\mathrm{\&mu;}}_{03}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^0{(y-y_0)}^3$

${\mathrm{\&mu;}}_{30}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^3{(y-y_0)}^0---(5-5)$

Wherein second-order moment around mean μ₀₂And μ₂₀Represent respectively around the inertia by the vertically and horizontally axis of gray scale barycenter Square, third central moment μ₀₃And μ₃₀The amplitude mensurable institute analyzed area asymmetric degree to vertically and horizontally axis.

Construct 7 not bending moments, translation, scaling and invariable rotary can be kept under conditions of consecutive image.Described 7 are not Bending moment is defined respectively as:

Central moment expression formula after second order normalized:

${\mathrm{\&eta;}}_{02}=\frac{{\mathrm{\&mu;}}_{02}}{{{\mathrm{\&mu;}}^2}_{00}};{\mathrm{\&eta;}}_{20}=\frac{{\mathrm{\&mu;}}_{20}}{{{\mathrm{\&mu;}}^2}_{00}};{\mathrm{\&eta;}}_{11}=\frac{{\mathrm{\&mu;}}_{11}}{{{\mathrm{\&mu;}}^2}_{00}};---(5-6)$

Central moment expression formula after three rank normalization:

${\mathrm{\&eta;}}_{12}=\frac{{\mathrm{\&mu;}}_{12}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}};{\mathrm{\&eta;}}_{21}=\frac{{\mathrm{\&mu;}}_{21}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}};{\mathrm{\&eta;}}_{03}=\frac{{\mathrm{\&mu;}}_{03}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}};{\mathrm{\&eta;}}_{30}=\frac{{\mathrm{\&mu;}}_{30}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}}---(5-7)$

φ₁=η₂₀+η₀₂ (5-8)

φ₂=(η₂₀+η₀₂)²+4η² ₁₁ (5-9)

φ₃=(η₃₀-3η₁₂)²+(3η₂₁-η₀₃)² (5-10)

φ₄=(η₃₀+η₁₂)²+(η₂₁+η₀₃)² (5-11)

φ₅=(η₃₀-3η₁₂)(η₃₀+η₁₂)[(η₂₀+η₁₂)²-3(η₂₁+η₀₃)²]+(3η₂₁+η₀₃)(η₂₁+η₀₃)[3(η₃₀+η₁₂ )²-(η₂₁+η₀₃)] (5-12)

φ₆=(η₂₀-η₀₂)[(η₃₀+η₁₂)²-(η₂₁+η₀₃)²]+4η₁₁(η₃₀+η₁₂)(η₂₁+η₀₃) (5-13)

φ₇=(3 η₂₁-η₀₃)(η₃₀+η₁₂)[(η₃₀+η₁₂)²-3(η₂₁+η₀₃)²]+(3η₂₁-η₀₃)(η₂₁+η₀₃)[3(η₃₀+η₁₂ )²-(η₂₁+η₀₃)²] (5-14)

φ therein₁～φ₇Being the computing formula of 7 not bending moments, each η value represents the center after second order, three rank normalization Square.

Step 5.4, extracts 7 shape spies of insulator to be identified by the computing formula of Hu square algorithm (5-8)～(5-14) Levy vector, then to each characteristic vector Gaussian normalization to [-1,1].By target image to be identified and shape in characteristic image storehouse Feature normalization is that 27 dimensional vectors are represented by:

F_si[F_s1,F_s2,F_s3,......F_s7]；F_ti[F_t1,F_t2,F_t3,......F_t7]；

Equally, chessboard distance formula is utilized to calculate target image to be identified and the mating of shape facility in characteristic image storehouse Similarity distance S₃:

$S_3=\underset{i}{max}|F_{si}-F_{ti}|---(5-15)$

S₃Value maximum, the similarity of two width images is the highest.

Step 5.5, carries out Feature Fusion by the color of insulator, shape and the textural characteristics that extract.If three kinds of features The similarity (distance) of Vector Fusion is D, the similarity of three kinds of single features vectors (feature of color, shape and texture) (away from From) S₁,S₂,S₃Corresponding weight is respectively w₁,w₂,w₃, then the similarity (distance) of three kinds of Feature Fusion vectors is,

D=s₁w₁+s₂w₂+s₃w₃ (5-16)

∑ w in formula_i=1；I=1,2,3.

The 3 kinds of features extracting insulator in step 3, step 4, step 5 are merged, three kinds of single features by step 6 During fusion, each characteristic quantity was normalized before similarity measurement, adjust target image to be identified and scheme in data base As the weight between three features.Determine its weight with analytic hierarchy process (AHP), calculate Consistency Ratio CR, verify judgment matrix three Whether the weights of Feature Fusion meet the requirements, and specifically implement according to following steps:

Step 6.1, if texture, color and the shape eigenvectors of a target image i to be identified are [O_i1,O_i2,O_i3], Treating N number of image in recognition target image i and data base by chessboard distance and carry out Similarity Measure, computed range value is remembered respectively For D_i1,D_i2,......D_iN；D_iNAlso it is target image to be identified and three characteristic distances of any one image in data picture library:

D_iN=[d_1,iN,d_2,iN,d_3,iN] (6-1)

Step 6.2, calculates the mean μ of 3N distance value_DAnd standard deviation sigma_D, specific formula for calculation is:

${\mathrm{\&mu;}}_D=\frac{1}{3N}(\underset{a=1}{\overset{N}{\&Sigma;}}{f}_a+\underset{b=1}{\overset{N}{\&Sigma;}}{f}_b+\underset{c=1}{\overset{N}{\&Sigma;}}{f}_c)$

${\mathrm{\&sigma;}}_D=\sqrt{\frac{{(\underset{a=1}{\overset{N}{\&Sigma;}}{f}_a-{\mathrm{\&mu;}}_D)}^2+{(\underset{b=1}{\overset{N}{\&Sigma;}}{f}_b-{\mathrm{\&mu;}}_D)}^2+{(\underset{c=1}{\overset{N}{\&Sigma;}}{f}_c-{\mathrm{\&mu;}}_D)}^2}{N-1}}---(6-2)$

F therein_a、f_b、f_cRepresent texture, color and shape eigenvectors respectively；

Step 6.3, treats recognition target image and N width image (N takes any positive integer) similarity distance D in feature database_i1, D_i2,......D_iNIt is normalized, makes three kinds of different characteristic vectors roughly the same on Similarity Measure, D_iNValue normalizing After change, major part falls interval in [0,1], and its expression formula is:

${D^{*}}_{iN}=\frac{\frac{D_{iN}-{\mathrm{\&mu;}}_D}{3{\mathrm{\&sigma;}}_D}+1}{2}---(6-3)$

Step 6.4, uses analytic hierarchy process (AHP) to determine the weight of three Feature Fusion.With 1-9 scaling law to Feature Fusion after Three feature description index importance of insulator image judge, the judgment matrix A utilizing policymaker to provide derives three Weight s of feature_i, then the concordance met with Consistency Ratio CR test and judge matrix, thus verify and derived three features Weighted value accuracy, concrete step is:

Step 6.4.1, according to 1-9 scaling law to three feature weight s₁,s₂,s₃Importance proportion quotiety two-by-two is carried out Assignment, formation judgment matrix:

A₁=(a_pq)_n×nWherein a_pqFor pth feature relative to the proportion quotiety of q feature importance,

Wherein, 1-9 scaling law is as shown in the table:

Proportion quotiety	Definition explanation
		1	Two feature weights are the strongest
3	One feature weight is stronger than another
		5	One feature weight is stronger than another
7	One feature weight is important stronger than another
		9	One feature weight is more absolute by force than another
2,4,6,8	Trade off between two feature weight scales

Step 6.4.2, takes a width insulator image, in hierarchy Model, this insulator image is only taken ground floor Secondary, and this level can be divided into three classifications, i.e. texture (C), color (S) and shape (T), then corresponding judgment matrix represents For:

$A_1={\left(a_{pq}\right)}_{3\&times;3}=\left[\begin{array}{ccc}\frac{H_C}{H_C}& \frac{H_C}{H_S}& \frac{H_C}{H_T}\\ \frac{H_S}{H_C}& \frac{H_S}{H_S}& \frac{H_S}{H_T}\\ \frac{H_T}{H_C}& \frac{H_T}{H_S}& \frac{H_T}{H_T}\end{array}\right]---(6-4)$

ThereinIt is respectively color and shape facility proportion quotiety, texture and color characteristic ratio Scale, texture and shape facility proportion quotiety.It is typically based on user's priori and extracts the interpretation of result of three features, shape Other two aspect ratios of aspect ratio are great.

When starting, provide one initial value of scale value, i.e. setJudge consistent Sex rate CR.

Step 6.4.3, to A₁In the element multiplication often gone open cube and obtain vectorial Y_i=(y₁,y₂,y₃), computing formula is such as Under:

$y_1=\sqrt[3]{\frac{H_C}{H_C}\&times;\frac{H_C}{H_S}\&times;\frac{H_C}{H_T}}---(6-5)$

$y_2=\sqrt[3]{\frac{H_S}{H_C}\&times;\frac{H_S}{H_S}\&times;\frac{H_S}{H_T}}---(6-6)$

$y_3=\sqrt[3]{\frac{H_T}{H_C}\&times;\frac{H_T}{H_S}\&times;\frac{H_T}{H_T}}---(6-7)$

To Y_iIt is normalized and obtains normalized weight vector s_i=(s₁,s₂,s₃)

$s_i=\frac{Y_i}{\underset{i}{\overset{3}{\&Sigma;}}{Y}_i}---(6-8)$

Step 6.4.4, to A₁In every column element sue for peace to obtain Z_q=(Z₁,Z₂,Z₃)

$Z_q=\underset{i}{\overset{3}{\&Sigma;}}{a}_{pq}---(6-9)$

Step 6.4.5, calculates judgment matrix A₁Maximum eigenvalue

${\mathrm{\&lambda;}}_{max}=\underset{i}{\overset{3}{\&Sigma;}}{s}_i{Z}_q---(6-10)$

Step 6.5, calculates Consistency Ratio CR, and whether utilize Consistency Ratio CR test and judge matrix to meet consistent Property, and then determine the weighted value of Feature Fusion,

Concrete step is:

Step 6.5.1, calculates judgment matrix A₁Coincident indicator CI,

$CI=\frac{{\mathrm{\&lambda;}}_{max}-n}{n-1}---(6-11)$

Wherein n is judgment matrix A₁Exponent number, the Maximum characteristic root λ of the least judgment matrix of CI_maxMore meet completely the same Property, the degree of CI the biggest explanation judgment matrix deviation crash consistency is the biggest.

Judgment matrix A is calculated according to 6-10₁Other two eigenvalue is designated as respectively: λ₂、λ₃. seek these three eigenvalue of maximum Meansigma methodsComputing formula is as follows:

${\widetilde{\mathrm{\&lambda;}}}_{max}=\frac{{\mathrm{\&lambda;}}_{max}+{\mathrm{\&lambda;}}_2+{\mathrm{\&lambda;}}_3}{3}---(6-12)$

Step 6.5.2, calculates Aver-age Random Consistency Index RI according to formula (9-10)

$RI=\frac{{\widetilde{\mathrm{\&lambda;}}}_{max}-n}{n-1}---(6-13)$

Step 6.5.3, CI and RI that utilization calculates, calculating Consistency Ratio CR:

$CR=\frac{CI}{RI}---(6-14)$

As CR < judgment matrix A when 0.1₁There is satisfactory consistency, i.e. weight vectors s_iMeet matrix equation: A₁s_i=λ_maxs_i There is the eigenvalue λ of maximum_max, so judgment matrix A₁The weighted value of three Feature Fusion meets the requirements, the power used when now calculating The weighted value that weight values uses when being image co-registration in this method, the insulator identification image after output fusion.

When CR >=0.1, it is judged that matrix A₁Not there is satisfactory consistency, then go to step 6.4 and re-start assignment decision-making Person needs the most rightThree scales carry out assignment according to 1-9 scaling law, with front once calculate time assignment It is adjusted after contrast, constructs new judgment matrix A₁, then calculate weight vector s respectively_i, maximum feature λ of judgment matrix_maxAnd one Cause property index CI, finally calculates Consistency Ratio CR, until CR < 0.1 sets up.

A kind of based on multicharacteristic information integration technology the insulator recognition detection method of the present invention, by insulator Textural characteristics shape facility color characteristic identification, can detect insulator effectively with or without falling string, inside and outside have flawless damaged, be The open defects such as no existence self-destruction, by the identification of color characteristic, degree filthy on detection insulator.The texture of insulator Feature, color characteristic and shape facility effectively combine, and target image carries out multicharacteristic information fusion recognition detection, and passes through Go out final target insulator with three property data base similarity matching identifications of the insulator set up, and judge it online Whether running status exists filth, reveals, falls the faults such as string, crackle be damaged.With this solve during insulator identification due to The object recognition rate that acquisition of information inaccuracy, the factor such as uncertain and incomplete cause is low and wastes time and energy in fault diagnosis etc. asks Topic.

Claims (8)

1. an insulator recognition detection method based on multicharacteristic information integration technology, it is characterised in that specifically according to following Step is implemented:

Step 1, the transmission line of electricity of the focusing video camera collection that utilization is arranged on the band The Cloud Terrace on on-the-spot steel tower or iron insulate The picture signal of son；

Step 2, collects image and carries out pretreatment step 1, utilizes optimal entropic threshold method to carry out image segmentation, passes through analysis chart As the entropy of grey level histogram, find optimal threshold；

Step 3, carried out extracting pretreated image the texture feature vector of insulator from step 2；

Step 4, extracts insulator color feature vector to be identified, empty to the HSV color of the insulator image collected in step 1 Between carry out unequal interval quantization；

Step 5, extracts insulator shape eigenvectors to be identified, and the shape facility of insulator is unrelated with external environment, is thing External the most stable information, be also image the most intuitively, the most directly visualize, use and calculate based on Hu square in region description method Method extracts insulator shape facility；

The 3 kinds of features extracting insulator in step 3, step 4, step 5 are merged by step 6, and three kinds of single features merge Time, each characteristic quantity was normalized before similarity measurement, adjust target image to be identified and image three in data base Weight between individual feature；Determine its weight with analytic hierarchy process (AHP), calculate Consistency Ratio CR, verify three features of judgment matrix Whether the weights merged meet the requirements.

Insulator recognition detection method based on multicharacteristic information integration technology the most according to claim 1, its feature exists In, described step 2 is specifically implemented according to following steps:

Step 2.1, the tonal range of the image collected in step 1 is designated as 0,1,2 ... L-1}；

Step 2.2, if the region that the pixel that gray level is less than t is constituted is target area A, then the expression formula of target area A entropy For:

$H_A\left(t\right)=-\underset{i}{\&Sigma;}\frac{p_i}{p_t}\lg \frac{p_i}{p_t}---(2-1)$

P in formula_iRepresent that gray level is the probability that i occurs, i=0,1,2 ... t；P_tRepresent that gray level is the probability that t occurs；

Step 2.3, if the region that the pixel that gray level is higher than t is constituted is background area B, then the expression formula of background area B entropy For:

$H_B\left(t\right)=-\underset{i}{\&Sigma;}\frac{p_i}{1-p_t}\lg \frac{p_i}{1-p_t}---(2-2)$

I=t+1, t+2, t+3 ... .L-1；

P in formula_iRepresent that gray level is the probability that i occurs, i=0,1,2 ... t；P_tRepresent that gray level is the probability that t occurs；

Step 2.4, calculates max-thresholds,

Define according to entropy function,

$C\left(t\right)=H_A\left(t\right)+H_B\left(t\right)=\lg \frac{p_t}{1-p_t}+\frac{H_t}{p_t}+\frac{H_L-H_t}{1-P_t}---(2-3)$

In formulaI=0 therein, 1,2 ... t；I=0 therein, 1,2 ... L-1；

Then, when entropy function C (t) is maximum,

Calculate t=argmax{C (t) } (2-4)

Mean that the max-thresholds of gray scale t.

Insulator recognition detection method based on multicharacteristic information integration technology the most according to claim 1, its feature exists In, described step 3 is specifically implemented according to following steps:

Step 3.1, utilizes 3 the color spaces difference to former RGB color image of the statistical method in texture characteristic extracting method Calculate average, variance, the degree of bias, kurtosis, energy, entropy totally 6 the color space features parameter as textural characteristics,

Specific formula for calculation is as follows:

(1) average

(2) variance

(3) degree of bias

(4) kurtosis

(5) energy

(6) entropy

Wherein g (i) is the gray value of i-gray level, p_gI () is the probability of i-th gray value；

Former RGB color image is just divided into 18 characteristic vectors altogether by these 6 features；The single features that image is different is carried out spy Carry out feature internal normalization when levying extraction, use Gaussian normalization method that these 18 characteristic vectors carry out the internal normalizing of feature Change to [-1,1], reduce other element value and the distribution of the element value after normalization is produced impact；The normalization of characteristic vector F_iComputing formula as follows:

$F_i=\frac{f_i-{\mathrm{\&mu;}}_i}{{\mathrm{\&sigma;}}_i};---(3-7)$

${\mathrm{\&mu;}}_i=\frac{\underset{i}{\&Sigma;}{f}_i}{i};---(3-8)$

${\mathrm{\&sigma;}}_i=\sqrt{\frac{{(\underset{i}{\mathrm{\&Sigma;}}{f}_i-{\mathrm{\&mu;}}_i)}^2}{i-1}};---(3-9)$

Wherein μ_iRepresent the average of target image characteristics to be identified vector, σ_iRepresent the standard of target image characteristics to be identified vector Difference, f_iIt is 18 one dimensional histograms initial characteristic values and f_i∈[f₁,f₂,f₃,.......f₁₈]；I=1,2,3 ... 18；

Step 3.2, the texture feature vector extracted in step 3.1 is special with the texture in the sample database established in advance Levy vector and carry out contrast coupling, calculate its similarity, method particularly includes:

Chessboard distance formula is utilized to calculate the similarity distance that target image to be identified mates with textural characteristics in characteristic image storehouse, The normalization characteristic vector of two width images is that 1 18 dimensional vector is represented by:

F_xi[F_x1,F_x2,F_x3,......F_x18]；

F_yi[F_y1,F_y2,F_y3,......F_y18]；

Then, calculate according to chessboard distance formula:

$S_1=\underset{i}{max}|F_{xi}-F_{yi}|---(3-10)$

Wherein S₁Represent the similarity distance that target image to be identified mates, F with textural characteristics in characteristic image storehouse_xiAnd F_yiRepresent The normalization characteristic vector of two width images, S₁Being worth the biggest, the similarity of two width images is the highest.

Insulator recognition detection method based on multicharacteristic information integration technology the most according to claim 1, its feature exists In, described step 4 is specifically implemented according to following steps:

Step 4.1, is divided into 8 parts by tone H space, and saturation S and brightness V space are divided into 3 parts, the tone H scope of image For [0,360 °], saturation S and brightness V in the range of [0,1]；

Step 4.2, carries out color quantization according to the different range of H, S, V, specific as follows shown:

$S=\left\{\begin{array}{cc}0,& 0\&le;S\&le;0.2\\ 1,& 0.2\&le;S\&le;0.7\\ 2,& 0.7\&le;S\&le;1.0\end{array}\right.---(4-2)$

$V=\left\{\begin{array}{cc}0,& 0\&le;V\&le;0.2\\ 1,& 0.2\&le;V\&le;0.7\\ 2,& 0.7\&le;V\&le;1.0\end{array}\right.---(4-3)$

3 color components after quantization are synthesized one-dimensional characteristic vectorial:

I=HQ_sQ_v+SQ_v+V (4-4)

I=HQ_sQ_vQ in+SQ formula_s=3 is the quantization progression of component S；Q_vDuring the quantization progression that=3 is component V, 3-4 is represented by:

I=9H+3S+V (4-5)

Quantify the subspace that the HSV space after terminating resolves into, it is thus achieved that 72 one dimensional histograms；

Step 4.3, the characteristic vector Gaussian normalization of the color histogram after quantifying in step 4.2, to [-1,1], reduces each Other element value produces impact, the normalized F of characteristic vector to the distribution of the element value after normalization_lComputing formula is as follows:

$F_l=\frac{f_l-{\mathrm{\&mu;}}_l}{{\mathrm{\&sigma;}}_l};$

${\mathrm{\&mu;}}_l=\frac{\underset{l}{\&Sigma;}{f}_l}{l};$

${\mathrm{\&sigma;}}_l=\sqrt{\frac{{(\underset{l}{\mathrm{\&Sigma;}}{f}_l-{\mathrm{\&mu;}}_l)}^2}{l-1}}---(4-6)$

Wherein μ_lRepresent the average of target image characteristics to be identified vector, σ_lRepresent the standard of target image characteristics to be identified vector Difference, f_lIt is 72 one dimensional histograms initial characteristic values and f_l∈[f₁,f₂,f₃,.......f₇₂]；L=1,2,3 ... 72,

Step 4.4, utilizes chessboard distance formula to calculate the phase that target image to be identified mates with color characteristic in characteristic image storehouse Seemingly spend distance；The normalization characteristic of two width images is represented by for 2 72 dimensional vectors:

F_pl[F_p1,F_p2,F_p3,......F_p72]；F_ql[F_q1,F_q2,F_q3,......F_q72]；

Wherein chessboard distance computing formula:

$S_2=\underset{l}{max}|F_{pl}-F_{ql}|---(4-7)$

Wherein S₂Represent the similarity distance that target image to be identified mates, F with textural characteristics in characteristic image storehouse_plAnd F_qlRepresent The normalization characteristic vector of two width images, S₂Value maximum, the similarity of two width images is the highest.

Insulator recognition detection method based on multicharacteristic information integration technology the most according to claim 1, its feature exists In, described step 5 is specifically implemented according to following steps:

Step 5.1, definition collect image f (x, p+q rank square y) is:

$m_{pq}=\underset{x}{\&Sigma;}{x}^p{y}^{q}f(x,y)---(5-1)$

Then, the central moment of p+q rank square is:

${\mathrm{\&mu;}}_{pq}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^p{(y-y_0)}^q---(5-2)$

In formulaRepresent the center of gravity of image-region；

Step 5.2, in order to the character obtaining image itself unrelated with scaling can carry out normalizing to the central moment in formula (5-2) Changing, the central moment after normalization is expressed as:

${\mathrm{\&eta;}}_{pq}=\frac{{\mathrm{\&mu;}}_{pq}}{{{\mathrm{\&mu;}}^r}_{00}}---(5-3)$

In formulaP+q=2,3,4 ... normalized centre-to-centre spacing to the translation of object, scale and rotate and all keep Constant；

Step 5.3, (x, y) second order and third central moment, obtain to utilize formula (5-2) to calculate image f

${\mathrm{\&mu;}}_{02}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^0{(y-y_0)}^2$

${\mathrm{\&mu;}}_{20}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^2{(y-y_0)}^0---(5-4)$

${\mathrm{\&mu;}}_{03}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^0{(y-y_0)}^3$

${\mathrm{\&mu;}}_{30}=\underset{x}{\&Sigma;}\underset{y}{\&Sigma;}{(x-x_0)}^3{(y-y_0)}^0---(5-5)$

Wherein second-order moment around mean μ₀₂And μ₂₀Represent respectively around the moment of inertia by the vertically and horizontally axis of gray scale barycenter, three rank Central moment μ₀₃And μ₃₀The amplitude mensurable institute analyzed area asymmetric degree to vertically and horizontally axis；

Construct 7 not bending moments, translation, scaling and invariable rotary can be kept under conditions of consecutive image；Described 7 not bending moment It is defined respectively as:

Central moment expression formula after second order normalized:

${\mathrm{\&eta;}}_{02}=\frac{{\mathrm{\&mu;}}_{02}}{{{\mathrm{\&mu;}}^2}_{00}};{\mathrm{\&eta;}}_{20}=\frac{{\mathrm{\&mu;}}_{20}}{{{\mathrm{\&mu;}}^2}_{00}};{\mathrm{\&eta;}}_{11}=\frac{{\mathrm{\&mu;}}_{11}}{{{\mathrm{\&mu;}}^2}_{00}};---(5-6)$

Central moment expression formula after three rank normalization:

${\mathrm{\&eta;}}_{12}=\frac{{\mathrm{\&mu;}}_{12}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}};{\mathrm{\&eta;}}_{21}=\frac{{\mathrm{\&mu;}}_{21}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}};{\mathrm{\&eta;}}_{03}=\frac{{\mathrm{\&mu;}}_{03}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}};{\mathrm{\&eta;}}_{30}=\frac{{\mathrm{\&mu;}}_{30}}{{{\mathrm{\&mu;}}^{\frac{5}{2}}}_{00}}---(5-7)$

φ₁=η₂₀+η₀₂ (5-8)

φ₂=(η₂₀+η₀₂)²+4η² ₁₁ (5-9)

φ₃=(η₃₀-3η₁₂)²+(3η₂₁-η₀₃)² (5-10)

φ₄=(η₃₀+η₁₂)²+(η₂₁+η₀₃)² (5-11)

φ₅=(η₃₀-3η₁₂)(η₃₀+η₁₂)[(η₂₀+η₁₂)²-3(η₂₁+η₀₃)²]+(3η₂₁+η₀₃)(η₂₁+η₀₃)[3(η₃₀+η₁₂)²-(η₂₁ +η₀₃)] (5-12)

φ₆=(η₂₀-η₀₂)[(η₃₀+η₁₂)²-(η₂₁+η₀₃)²]+4η₁₁(η₃₀+η₁₂)(η₂₁+η₀₃) (5-13)

φ₇=(3 η₂₁-η₀₃)(η₃₀+η₁₂)[(η₃₀+η₁₂)²-3(η₂₁+η₀₃)²]+(3η₂₁-η₀₃)(η₂₁+η₀₃)[3(η₃₀+η₁₂)²-(η₂₁ +η₀₃)²] (5-14)

φ therein₁～φ₇Being the computing formula of 7 not bending moments, each η value represents the central moment after second order, three rank normalization；

Step 5.4, by the computing formula of Hu square algorithm (5-8)～(5-14) extract 7 shape facilities of insulator to be identified to Amount, then to each characteristic vector Gaussian normalization to [-1,1]；By target image to be identified and shape facility in characteristic image storehouse It is normalized to 27 dimensional vectors be represented by:

F_si[F_s1,F_s2,F_s3,......F_s7]；F_ti[F_t1,F_t2,F_t3,......F_t7]；

Equally, utilize chessboard distance formula calculate target image to be identified with in characteristic image storehouse shape facility mate similar Degree distance S₃:

$S_3=\underset{i}{max}|F_{si}-F_{ti}|---(5-15)$

S₃Value maximum, the similarity of two width images is the highest；

Step 5.5, carries out Feature Fusion by the color of insulator, shape and the textural characteristics that extract；If three feature vectors The similarity (distance) merged is D, the similarity (distance) of three kinds of single features vectors (feature of color, shape and texture) S₁,S₂,S₃Corresponding weight is respectively w₁,w₂,w₃, then the similarity (distance) of three kinds of Feature Fusion vectors is,

D=s₁w₁+s₂w₂+s₃w₃ (5-16)

∑ w in formula_i=1；I=1,2,3..

Insulator recognition detection method based on multicharacteristic information integration technology the most according to claim 1, its feature exists In, described step 6 is specifically implemented according to following steps:

Step 6.1, if texture, color and the shape eigenvectors of a target image i to be identified are [O_i1,O_i2,O_i3], pass through Chessboard distance is treated N number of image in recognition target image i and data base and is carried out Similarity Measure, and computed range value is designated as respectively D_i1,D_i2,......D_iN；D_iNAlso it is target image to be identified and three characteristic distances of any one image in data picture library:

D_iN=[d_1,iN,d_2,iN,d_3,iN] (6-1)

Step 6.2, calculates the mean μ of 3N distance value_DAnd standard deviation sigma_D, specific formula for calculation is:

${\mathrm{\&mu;}}_D=\frac{1}{3N}(\underset{a=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_a+\underset{b=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_b+\underset{c=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_c);$

${\mathrm{\&sigma;}}_D=\sqrt{\frac{{(\underset{a=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_a-{\mathrm{\&mu;}}_D)}^2+{(\underset{b=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_b-{\mathrm{\&mu;}}_D)}^2+{(\underset{c=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_c-{\mathrm{\&mu;}}_D)}^2}{N-1}}---(6-2)$

F therein_a、f_b、f_cRepresent texture, color and shape eigenvectors respectively；

Step 6.3, treats recognition target image and N width image (N takes any positive integer) similarity distance D in feature database_i1, D_i2,......D_iNIt is normalized, makes three kinds of different characteristic vectors roughly the same on Similarity Measure, D_iNValue normalizing After change, major part falls interval in [0,1], and its expression formula is:

${D^{*}}_{iN}=\frac{\frac{D_{iN}-{\mathrm{\&mu;}}_D}{3{\mathrm{\&sigma;}}_D}+1}{2}---(6-3)$

Step 6.4, uses analytic hierarchy process (AHP) to determine the weight of three Feature Fusion；Insulate to after Feature Fusion with 1-9 scaling law Three feature description index importance of subimage judge, the judgment matrix A utilizing policymaker to provide derives three features Weight s_i, then the concordance met with Consistency Ratio CR test and judge matrix, thus verify the power being derived three features Weight values accuracy；

Step 6.5, calculates Consistency Ratio CR, and utilizes whether Consistency Ratio CR test and judge matrix meets concordance, enter And determine the weighted value of Feature Fusion.

Insulator recognition detection method based on multicharacteristic information integration technology the most according to claim 6, its feature exists In, described step 6.4 is specifically implemented according to following steps:

Step 6.4.1, according to 1-9 scaling law to three feature weight s₁,s₂,s₃Importance proportion quotiety two-by-two carries out assignment, Formation judgment matrix:

A₁=(a_pq)_n×nWherein a_pqFor pth feature relative to the proportion quotiety of q feature importance,

Step 6.4.2, takes a width insulator image, in hierarchy Model, this insulator image is only taken the first level, and This level can be divided into three classifications, i.e. texture (C), color (S) and shape (T), then corresponding judgment matrix are expressed as:

$A_1={\left(a_{pq}\right)}_{3\&times;3}=\left[\begin{array}{ccc}\frac{H_C}{H_C}& \frac{H_C}{H_S}& \frac{H_C}{H_T}\\ \frac{H_S}{H_C}& \frac{H_S}{H_S}& \frac{H_S}{H_T}\\ \frac{H_T}{H_C}& \frac{H_T}{H_S}& \frac{H_T}{H_T}\end{array}\right]---(6-4)$

ThereinBe respectively color and shape facility proportion quotiety, texture and color characteristic proportion quotiety, Texture and shape facility proportion quotiety；It is typically based on user's priori and extracts the interpretation of result of three features, shape facility It is more great than other two aspect ratios,

When starting, provide one initial value of scale value, i.e. setJudge concordance ratio Rate CR；

Step 6.4.3, to A₁In the element multiplication often gone open cube and obtain vectorial Y_i=(y₁,y₂,y₃), computing formula is as follows:

$y_1=\sqrt[3]{\frac{H_C}{H_C}\&times;\frac{H_C}{H_S}\&times;\frac{H_C}{H_T}}---(6-5)$

$y_2=\sqrt[3]{\frac{H_S}{H_C}\&times;\frac{H_S}{H_S}\&times;\frac{H_S}{H_T}}---(6-6)$

$y_3=\sqrt[3]{\frac{H_T}{H_C}\&times;\frac{H_T}{H_S}\&times;\frac{H_T}{H_T}}---(6-7)$

To Y_iIt is normalized and obtains normalized weight vector s_i=(s₁,s₂,s₃)

$s_i=\frac{Y_i}{\underset{i}{\overset{3}{\&Sigma;}}{Y}_i}---(6-8)$

Step 6.4.4, to A₁In every column element sue for peace to obtain Z_q=(Z₁,Z₂,Z₃),

$Z_q=\underset{i}{\overset{3}{\&Sigma;}}{a}_{pq}---(6-9)$

Step 6.4.5, calculates judgment matrix A₁Maximum eigenvalue

${\mathrm{\&lambda;}}_{max}=\underset{i}{\overset{3}{\&Sigma;}}{s}_i{Z}_q---(6-10).$

Insulator recognition detection method based on multicharacteristic information integration technology the most according to claim 6, its feature exists In, described step 6.5 is specifically implemented according to following steps:

Step 6.5.1, calculates judgment matrix A₁Coincident indicator CI,

$CI=\frac{{\mathrm{\&lambda;}}_{max}-n}{n-1}---(6-11)$

Wherein n is judgment matrix A₁Exponent number, the Maximum characteristic root λ of the least judgment matrix of CI_maxMore meeting crash consistency, CI is more The degree of big explanation judgment matrix deviation crash consistency is the biggest；

Judgment matrix A is calculated according to 6-10₁Other two eigenvalue is designated as respectively: λ₂、λ₃. seek the flat of these three eigenvalue of maximum AverageComputing formula is as follows:

${\widetilde{\mathrm{\&lambda;}}}_{max}=\frac{{\mathrm{\&lambda;}}_{max}+{\mathrm{\&lambda;}}_2+{\mathrm{\&lambda;}}_3}{3}---(6-12),$

Step 6.5.2, calculates Aver-age Random Consistency Index RI according to formula (9-10),

$RI=\frac{{\widetilde{\mathrm{\&lambda;}}}_{\max }-n}{n-1}---(6-13)$

Step 6.5.3, CI and RI that utilization calculates, calculating Consistency Ratio CR:

$CR=\frac{CI}{RI}---(6-14)$

As CR < judgment matrix A when 0.1₁There is satisfactory consistency, i.e. weight vectors s_iMeet matrix equation: A₁s_i=λ_maxs_iHave Big eigenvalue λ_max, so judgment matrix A₁The weighted value of three Feature Fusion meets the requirements, the weighted value used when now calculating The weighted value used when being image co-registration in this method, the insulator identification image after output fusion；

When CR >=0.1, it is judged that matrix A₁Not there is satisfactory consistency, then go to step 6.4 and re-start assignment policymaker's needs The most rightThree scales carry out assignment according to 1-9 scaling law, with front once calculate time assignment contrast after It is adjusted, constructs new judgment matrix A₁, then calculate weight vector s respectively_i, maximum feature λ of judgment matrix_maxAnd concordance refers to Mark CI, finally calculates Consistency Ratio CR, until CR < 0.1 sets up.