CN106204543A - Fabric defect detection method based on single category support vector machines - Google Patents

Abstract

The invention discloses a kind of fabric defect detection method based on single category support vector machines.Obtain fault-free fabric image, use the parameter of RDPSO algorithm optimization Gabor filter, the single Optimal Gabor Filters of structure adaptation fault-free fabric image texture characteristic；Use the parameter of RDPSO algorithm optimization list classification SVM；Textile image to be detected is carried out Gabor convolutional filtering；Based on extracting one group of textural characteristics on GLCM after the filtering image；Single classification SVM is used to carry out fault differentiation.The present invention uses single Optimal Gabor Filters, can be effectively improved detection speed, it is ensured that system real time requirement；Use single classification SVM as fault method of discrimination, be avoided that traditional statistical pattern recognition method local extremum, cross the problems such as study and deficient study, can effectively promote the generalization ability of system, it is ensured that the Detection accuracy requirement of system.

Description

Fabric defect detection method based on single category support vector machines

Technical field:

The invention belongs to mode identification technology, particularly to a kind of fabric defects based on single category support vector machines Detection method.

Background technology:

Textile industry is faced with acid test at present, improves product quality and reduces production cost, becoming enterprise The key factor of existence.Fabric defects drastically influence product quality, thus causes fabric price reduction, reduces enterprise's income. The method relying on machine vision automatically to detect fault, can become raising product quality and reduce the effective of production cost Approach.Compared to artificial range estimation, automatic testing method based on machine vision has the advantage that stability is high, the most examined personnel The impact of the factor such as emotion, health and environmental disturbances.It addition, after using the system of automatic defect inspection, work as system identification To there being fault to occur, machine down can be controlled, it is to avoid the expansion of fault with output signal, thus improves product quality, and one The number of machines that individual workman is responsible for can also increase, it is possible to saves employment cost, reduces production cost, and then makes the product of enterprise more There is competitiveness.

Owing to carrying out the advantage of fabric defects detection based on machine vision, many scholars have been attracted to be engaged in grinding of this respect Study carefully, and have been proposed for certain methods.But at present, major part textile enterprise still uses artificial method to carry out fault inspection Survey, one of them reason be exactly propose at present to go out its defect detection capabilities of system the most preferable, also need the most perfect.

The information being disclosed in this background section is merely intended to increase the understanding of the general background to the present invention, and should not When being considered to recognize or imply in any form this information structure prior art well known to persons skilled in the art.

Summary of the invention:

It is an object of the invention to provide a kind of fabric defect detection method based on single category support vector machines, thus gram Take above-mentioned defect of the prior art.

For achieving the above object, the invention provides a kind of fabric defects detection side based on single category support vector machines Method, the steps include:

(1) obtain fault-free fabric image, use random drift particle group optimizing (RDPSO) algorithm optimization Gabor filtering The parameter of device, the single Optimal Gabor Filters of structure adaptation fault-free fabric image texture characteristic；

(2) parameter of RDPSO algorithm optimization list category support vector machines (SVM) is used；

(3) textile image to be detected is carried out Gabor convolutional filtering；

(4) based on extracting one group of textural characteristics on gray level co-occurrence matrixes (GLCM) image after the filtering；

(5) single classification SVM is used to carry out fault differentiation.

Preferably, in technique scheme, step (1) comprises the following steps:

(1.1) structure two-dimensional space territory Gabor filter function, obtains the filtering of frequency domain Gabor through two-dimensional Fourier transform Function, the parameter needing the Gabor filter optimized is (σ_x, σ_y, λ, θ)；

(1.2) fault-free fabric image is carried out Gabor convolution transform, constructs fitness function based on Fisher criterion, RDPSO algorithm optimization is used to obtain the parameter (σ of Optimal Gabor Filters_x, σ_y, λ, θ), structure adaptation fault-free fabric image The single Optimal Gabor Filters of textural characteristics.

Preferably, in technique scheme, step (2) comprises the following steps:

(2.1) textile image without fault is carried out convolution with optimum Gabor filter, obtain filtered image；

(2.2) based on texture feature extraction on GLCM after the filtering image, constitutive characteristic vector, as single classification SVM's Training sample；

(2.3) use Support Vector data description based on suprasphere thought (SVDD) as single classification SVM method, pass through RDPSO algorithm optimization determines two parameters C and the σ of SVDD.

Compared with prior art, there is advantages that

Use single Optimal Gabor Filters, detection speed can be effectively improved, it is ensured that system real time requirement；Use single Classification SVM, as fault method of discrimination, is avoided that traditional statistical pattern recognition method local extremum, excessively study and deficient study etc. Problem, can effectively promote the generalization ability of system, it is ensured that the Detection accuracy requirement of system.

Accompanying drawing illustrates:

Fig. 1 is the schematic flow sheet of present invention fabric defect detection method based on single category support vector machines；

Fig. 2 is four range direction figures that the present invention seeks gray level co-occurrence matrixes.

Detailed description of the invention:

Below the detailed description of the invention of the present invention is described in detail, it is to be understood that protection scope of the present invention is not Limited by detailed description of the invention.

Explicitly indicating that unless otherwise other, otherwise in entire disclosure and claims, term " includes " or it becomes Change and such as " comprising " or " including " etc. will be understood to comprise stated element or ingredient, and do not get rid of other yuan Part or other ingredient.

As it is shown in figure 1, fabric defect detection method based on single category support vector machines, the steps include:

(1.1) structure two-dimensional space territory Gabor filter function, obtains the filtering of frequency domain Gabor through two-dimensional Fourier transform Function, the parameter needing the Gabor filter optimized is (σ_x, σ_y, λ, θ).

Two-dimensional space territory Gabor filter function is represented by:

$G(x,y)=\exp (-\frac{1}{2}(\frac{x^{\′2}}{{\mathrm{\σ}}_x^2}+\frac{y^{\′2}}{{\mathrm{\σ}}_y^2}\left)\right)\exp \left(2\mathrm{\π}j\frac{x^{\′}}{\mathrm{\λ}}\right)---\left(1\right)$

Wherein

X '=x cos θ-y sin θ (2)

Y '=x sin θ+y cos θ (3)

σ_xAnd σ_yIt is respectively Gauss window standard deviation in time domain x-axis with y-axis.λ is wavelength, and θ is the anglec of rotation.Therefore, The parameter needing the Gabor filter optimized is s=(σ_x, σ_y, λ, θ).

When territory, application space Gabor filter function, amount of calculation is relatively big, and for ensureing requirement of real-time, spatial domain Gabor is filtered (x, y), through two-dimensional Fourier transform, frequency domain Gabor filter function to be converted into for wave function G

$\hat{G}(u,v)=DFT\left(G\right(x,y\left)\right)=A\exp (-\frac{1}{2}(\frac{{(u^{\′}-f)}^2}{{\mathrm{\σ}}_u^2}+\frac{v^{\′2}}{{\mathrm{\σ}}_v^2}\left)\right)---\left(4\right)$

Wherein,

U '=u cos θ-υ sin θ (5)

υ '=u sin θ+υ cos θ (6)

A=2 π σ_xσ_y (7)

DFT represents two-dimensional Fourier transform.WithIt is respectively Fourier transformation to exist Frequency domain u axle and the standard deviation on v axle.Mid frequency for two-dimensional Gabor function.

(1.2) fault-free fabric image is carried out Gabor convolution transform, constructs fitness function based on Fisher criterion, RDPSO algorithm optimization is used to obtain the parameter (σ of Optimal Gabor Filters_x, σ_y, λ, θ), thus structure most adapts to fault-free fabric The single Optimal Gabor Filters of image texture characteristic；

Assume there are two sample t₁(x, y) and t₂(x, y), can construct fitness function based on Fisher criterion is:

$F\left(s\right)=\frac{{({\mathrm{\μ}}_1-{\mathrm{\μ}}_2)}^2}{{\mathrm{\σ}}_1^2+{\mathrm{\σ}}_2^2}---\left(8\right)$

Wherein (μ₁, σ₁) and (μ₂, σ₂) it is respectively two sample t₁(x, y) and t₂(x, y) energy value of filtered image Average and standard deviation.

Assume the fault-free fabric image of a width N × N size be f (x, y), through the filtered image of Gabor be r (x, y). The product representation in the convolution usable frequency territory of spatial domain, then r (x, y) can be obtained by following formula:

$r(x,y)=f(x,y)*G(x,y)=IDFT\left(\hat{f}\right(u,v)\·\hat{G}(u,v\left)\right)---\left(9\right)$

Wherein " * " represents convolution, and IDFT represents Fourier inversion,It is image f (x, Fourier transformation y). Generally, filtered image r (x, y) is complex image, and its energy diagram picture can be obtained by following formula:

$E_r(x,y)={\[r_{\mathrm{Re}}^2(x,y)+r_{\mathrm{Im}}^2(x,y)\]}^{\frac{1}{2}}---\left(10\right)$

Wherein r_Re(x, y) and r_Im(x y) is respectively image r (x, real part y) and imaginary part.The mean μ of energy diagram picture₁And side DifferenceMay be defined as:

${\mathrm{\μ}}_1=\frac{1}{N\×N}\underset{x=1}{\overset{N}{\Σ}}\underset{y=1}{\overset{N}{\Σ}}{E}_r(x,y),---\left(11\right)$

${\mathrm{\σ}}_1^2=\frac{1}{N\×N-1}\underset{x=1}{\overset{N}{\Σ}}\underset{y=1}{\overset{N}{\Σ}}{\[E_r(x,y)-{\mathrm{\μ}}_1\]}^2---\left(12\right)$

Using similar method, (x, y) with filtered image r (x, error image f y) for fault-free fabric image f_c (x, y)=((x, y)-r (x, y)) can be used to calculate mean μ f₂And varianceImage f_c(x, y) through same Gabor filter G (x, y) filtered image is r_c(x, y), at frequency domain, r_c(x, y) can be obtained by following formula:

$r_c(x,y)=f_c(c,y)*G(x,y)=IDFT\left(\hat{f}\right(u,v)\·{\hat{G}}_c(u,v)\·\hat{G}(u,v\left)\right)---\left(13\right)$

WhereinMay be defined asWith calculating mean μ₁And varianceSimilar, mean μ₂With VarianceCan be obtained by formula (11) and formula (12), simply energy diagram is as E_r(x, y) by filtered image r_c(x, y) according to formula (10) It is calculated.

For obtaining the Gabor filter of optimum so that it is adaptation fault-free fabric image texture characteristic, Gabor need to be optimized Filter parameter s=(σ_x, σ_y, λ, θ), make fitness function F (s) of formula (8) maximize.

After having constructed fitness function, so that it may use RDPSO algorithm optimization to obtain the parameter (σ of Optimal Gabor Filters_x, σ_y, λ, θ).RDPSO algorithm is on the basis of particle group optimizing (PSO) algorithm, according to external electrical at a temperature of random standard finite Metallic conductor free-electron model and the optimization method that proposes.During optimizing, particle can be as one outside conductor The electronics of movement in portion's electric field, its motion can be collectively constituted by random warm-up movement and the drift motion affected by electric field.RDPSO Algorithm has been demonstrated to ensure global convergence, can find globally optimal solution.

Using RDPSO algorithm optimization Gabor filter parameter, its detailed process can be described as follows:

If iterations n=1.Initialize population, including determining maximum iteration time, the search volume (dimension of particle D), number M of particle, the position of random initializtion particle(position of particle is Gabor filtering One class value of device parameter s), i=1,2 ..., M.

It is calculated the fitness function value that each particle is corresponding by formula (8)

If current iteration frequency n=1, it is iteration, the individual desired positions of each particle for the first timeIt is this grain The initial position of sonThe individual desired positions of each particle is otherwise updated by formula (14)The overall desired positions of whole populationIts value isWherein g by Formula (15) determines.

$P_i^n=\left\{\begin{array}{ccc}{x}_i^n,& if& F\left(x_i^n\right)>F\left(P_i^{n-1}\right)\\ P_i^{n-1},& if& F\left(x_i^n\right)\≤F\left(P_i^{n-1}\right)\end{array}\right.---\left(14\right)$

$g=\arg \underset{1\≤i\≤M}{\max }F\left(P_i^n\right)---\left(15\right)$

If end condition meets (reaching maximum iteration time), training terminates, the overall desired positions of whole population GⁿIt is one group of Gabor filter parameter s of optimum；Otherwise, formula (16) and (17) speed and the position of each particle are entered Row updates, and forwards formula (8) to and is calculated the fitness function value that each particle is corresponding

$v_{i,j}^{n+1}=\mathrm{\α}|C_j^n-x_{i,j}^n|{\mathrm{\φ}}_{i,j}^{n+1}+\mathrm{\β}(p_{i,j}^n-x_{i,j}^n),---\left(16\right)$

$x_{i,j}^{t+1}=x_{i,j}^t+v_{i,j}^{t+1}---\left(17\right)$

Wherein α and β is respectively hot coefficient and coefficient of deviation.It is the random of normal distribution on (0, a 1) interval Number.CⁿFor average desired positions, can be calculated as follows:

$C^n=(C_1^n,...,C_D^n)=\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}{P}_i^n---\left(18\right)$

For the attractor in particle convergence process, when n iteration, its coordinate is:

$p_{i,j}^n=r_{i,j}^n{P}_{i,j}^n+(1-r_{i,j}^n)G_j^n---\left(19\right)$

WhereinIt it is equally distributed random number on (0, a 1) interval.

After being obtained Optimal Gabor Filters parameter by RDPSO algorithm optimization, so that it may structure Optimal Gabor Filters G^*(x, y)。

(2.1) textile image without fault is carried out convolution with optimum Gabor filter, obtain filtered image；

Without the textile image f of fault, (x, y) with optimum Gabor filter G^*(x, y) carries out convolution, after can being filtered Image r (x, y):

WhereinWithIt is that (x, y) with Optimal Gabor Filters G for image f respectively^*(x, Fourier y) becomes Change.

(2.2) based on texture feature extraction on GLCM after the filtering image, constitutive characteristic vector, as single classification SVM's Training sample；

The gray level co-occurrence matrixes (GLCM) of piece image can reflect the texture information of diagram picture.Assume that a width size is (x, y), its tonal gradation is G (tonal gradation of general gray level image is 256) to the image I of M × N, then the GLCM of diagram picture Matrix for G × G size.For same width gray level image, in different distances and direction, different GLCM can be produced, instead Answer different textural characteristics.At range direction vectorOn, (i, j) element of position, its value of this matrixCan be calculated by following formula:

$P_{\stackrel{\→}{d}}(i,j)=\underset{x=1}{\overset{M}{\Σ}}\underset{y=1}{\overset{N}{\Σ}}\mathrm{\δ}\{I(x,y)=iandI(x+d_x,y+d_y)=j\}---\left(21\right)$

Wherein δ True}=1, δ False}=0, i.e.Represent that the pixel that gray level is i is positioned at another gray scale Level is the pixel of jThe number of times occurred on direction.

When calculating the GLCM of piece image, range direction vectorIt it is an important parameter.For same piece image, because of ForDifference, can produce multiple GLCM.The general method used is first by differentObtain multiple GLCM, then to these GLCM is averaged and obtains a GLCM unrelated with direction.

Shown in Fig. 2 four range direction figure, vectorIt is respectively (-1 ,-1), (-1,0), (-1,1) and (0,1).In reality During border uses, 8 vectors can be usedIt is respectively (-1 ,-1), (-1,0), (-1,1), (0,1), (1,1), (1,0), (1 ,-1) (0 ,-1), obtains 8 GLCM, is finally averaged these 8 GLCM and obtains a GLCM.

After obtaining a GLCM unrelated with direction of piece image, based on this GLCM, so that it may obtain 5 Haralick Textural characteristics, respectively energy (Energy), entropy (Entropy), contrast (Contrast), homogeneity degree (Homogeneity) With dependency (Correlation), can be calculated as follows:

$Energy=\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}{P}_{i,j}^2---\left(22\right)$

$Entropy=\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}{P}_{i,j}(-\ln P_{i,j})---\left(23\right)$

$Contrast=\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}{P}_{i,j}{(i-j)}^2---\left(24\right)$

$Homogeneity=\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}\frac{P_{i,j}}{1+|i-j|}---\left(25\right)$

$Correlation=\frac{\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}(i-{\mathrm{\μ}}_i)(j-{\mathrm{\μ}}_j)P_{i,j}}{{\mathrm{\σ}}_i{\mathrm{\σ}}_j}---\left(26\right)$

Wherein P_{I, j}For (i, j) element value of position on the GLCM of the independent of direction finally given.Average in formula (26) μ_iAnd μ_j, standard deviation sigma_iAnd σ_jCan be defined respectively as:

${\mathrm{\μ}}_i=\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}{\mathrm{iP}}_{i,j}---\left(27\right)$

${\mathrm{\μ}}_j=\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}{\mathrm{jP}}_{i,j}---\left(28\right)$

${\mathrm{\σ}}_i=\sqrt{\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}{(i-{\mathrm{\μ}}_i)}^2{P}_{i,j}}---\left(29\right)$

${\mathrm{\σ}}_j=\sqrt{\underset{i=1}{\overset{G}{\Σ}}\underset{j=1}{\overset{G}{\Σ}}{(j-{\mathrm{\μ}}_j)}^2{P}_{i,j}}---\left(30\right)$

(2.3) use Support Vector data description based on suprasphere thought (SVDD) as single classification SVM method, pass through RDPSO algorithm optimization determines two parameters C and the σ of SVDD.

Support vector machine (SVM) is the machine learning method on the basis of a kind of Statistical Learning Theory, with Structural risk minization Turn to purpose.SVM it can be avoided that traditional statistical pattern recognition method such as neutral net, decision tree etc., local extremum, crosses study With problems such as deficient study, can effectively promote the generalization ability of system, it is ensured that the Detection accuracy requirement of system.Generally, SVM, for solving two classification problems, i.e. finds an optimal hyperlane in two class samples, can not only be by accurate for two class samples Distinguish, and this hyperplane those sample points distances nearest in two class samples are farthest.And for fabric defects detection, May be considered " abnormality detection " or " single classification problem ".In the fabric course of processing, it is normal (nothing in most cases Fault) sample, the number of fault sample is little.Further, the type of fault has a lot, such as broken yarn, broken hole, greasy dirt etc., its table Differ in shape on fabric, size, direction now, and the textural characteristics of the most dissimilar fault also can be different.Therefore, single classification side Method should be more suitable for solving fabric defects detection problem.

Support Vector data description (SVDD) is a kind of single classification SVM method proposed based on suprasphere thought.SVDD's Ultimate principle is to find a suprasphere, while minimizing its radius, allows training sample be enclosed in hypersphere as much as possible In body, i.e. minimize following object function:

$F(R,a)=R^2+C\underset{i}{\Σ}{\mathrm{\ξ}}_i---\left(31\right)$

Constraints is:

||f_i-a||²≤R²+ξ_i, ξ_i≥0 (32)

Wherein f_iFor i-th training sample, the one group of textural characteristics i.e. tried to achieve by (2.2), suprasphere radius is R, the centre of sphere For a, ξ_iFor slack variable, C is punishment parameter.Slack variable ξ_iIntroducing assume that in training sample, there may be abnormal sample This appearance, these samples can be located at outside optimum suprasphere, but introduce punishment parameter C on target function value, for all simultaneously Weighing apparatus hypersphere size and the ratio of the exceptional sample outside spheroid.

Foregoing content assumes that training sample is linear separability, when training sample Nonlinear separability, permissible Training data is projected to high-dimensional feature space by a nonlinear mapping so that these data can be linear in higher dimensional space Can divide, the method thus needing to introduce kernel function φ.Therefore, formula (32) is rewritable is:

||φ(f_i)-a||²≤R²+ξ_i, ξ_i≥0 (33)

The Lagrangian formulation that be can be defined as follows by formula (31) and (33):

$\begin{array}{c}L(R,a,{\mathrm{\α}}_i,{\mathrm{\γ}}_i,{\mathrm{\ξ}}_i)=R^2+C\underset{i}{\Σ}{\mathrm{\ξ}}_i\\ -\underset{i}{\Σ}{\mathrm{\α}}_i(R^2+{\mathrm{\ξ}}_i-|\left|\mathrm{\φ}\right(f_i)-a\left|{|}^2\right)-\underset{i}{\Σ}{\mathrm{\γ}}_i{\mathrm{\ξ}}_i\end{array}---\left(34\right)$

Wherein α_i>=0 and γ_i>=0 is respectively Lagrange multiplier.It is R that formula (34) requires at variable, α, ξ_iIn the case of minimum Change, and be α at variable_iAnd γ_iIn the case of maximize.Such problem can be converted into a convex double optimization problem, can pass through Fast algorithm based on Sequential minimal optimization algorithm solves and obtains optimum Lagrange multiplier α_iAnd meet following three condition:

||φ(f_i)-a||²＜ R²→α_i=0, γ_i=0 (35)

||φ(f_i)-a||²=R²→ 0 ＜ α_i＜ C, γ_i=0 (36)

||φ(f_i)-a||²＞ R²→α_i=C, γ_i＞ 0 (37)

It practice, most α_iBe all 0, those minorities corresponding to α_iThe training sample f of ＞ 0_iIt is referred to as supporting vector. Suprasphere centre of sphere a after training can be obtained by following formula:

$a=\underset{i}{\Σ}{\mathrm{\α}}_i{f}_i---\left(38\right)$

The radius R of suprasphere can be supported vector (i.e. 0 ＜ α by arbitrary on sphere_iThe training sample that ＜ C is corresponding) to the centre of sphere The distance of a obtains.

After determining the centre of sphere a and radius R of suprasphere, to a sample z to be tested, if normal sample, then under needing to meet Formula:

||φ(z)-a||²≤R² (39)

Before solving convex double optimization problem, need the value of given parameters C and the concrete type function of kernel function φ and Parameter value.When selecting kernel function, gaussian kernel function is because of its highly versatile, and pertains only to a scale parameter σ, applies in reality In be widely used.

The SVDD that design performance is superior, selects suitably punishment parameter C and Gauss scale parameter σ most important.Punishment ginseng Number C controls to support in training result the ratio of vector.Bigger C and less σ can cause bigger suprasphere, is enclosed in super Training sample in spheroid is more, and during actual test, loss can height；Along with reducing and the increase of σ of C, suprasphere radius becomes Little, appear in outside suprasphere from cluster centre sample point farther out, during actual test, false drop rate can height.

The parameter of SVDD can use RDPSO algorithm to be optimized to obtain.When constructing object function, classification accuracy is with super The radius of ball is two factors needing to consider.It is to say, for certain particle, if being obtained classification accuracy by this particle Height, the radius of hypersphere is little, then the value of corresponding object function is the biggest.Therefore, when training, available following formula is as object function F:

$f=w_A\×acc+\frac{w_R}{R}---\left(40\right)$

Wherein w_AAnd w_RIt is respectively corresponding classification accuracy acc and the weights of radius of hypersphere R, in use can be according to reality Situation sets.Classification accuracy acc is classify sample and the ratio of total sample accurately.

When the process of optimization, cross validation method can be used, N number of subset will be divided at random by training sample, select every time Selecting a subset as test set, remaining subset is incorporated as training set, carries out n times cross validation, finally according to stayed son The meansigma methods of collection test result carrys out the quality of evaluating value.Specifically, it is simply that according to the value (parameter C and σ) of a particle, In each cross-validation experiments, solve formula (34) and obtain optimum Lagrange multiplier α_i, then solve hypersphere centre of sphere a and radius R, tests the sample in training set further according to formula (39), available classification accuracy, is the most this time intersected The target function value of confirmatory experiment.Carrying out n times cross validation, the meansigma methods of target function value n times obtained is as this grain The target function value that son is corresponding.

The detailed process using the parameter of RDPSO algorithm optimization SVDD can be described as follows:

If iterations n=1.Initialize population, including determining maximum iteration time, the search volume (dimension of particle D), number M of particle, the position of random initializtion particle(position of particle is the parameter of SVDD C and σ), i=1,2 ..., M.

The fitness function value that each particle is corresponding is tried to achieve according to cross-validation method

If current iteration frequency n=1, it is iteration, the individual desired positions of each particle for the first timeIt is this particle Initial positionThe individual desired positions of each particle is otherwise updated by formula (14) The overall desired positions of whole populationIts value isWherein g is determined by formula (15).

If end condition meets (reaching maximum iteration time), training terminates, the overall desired positions of whole population GⁿIt is parameter C and the σ of one group of SVDD of optimum；Otherwise, formula (16) and (17) speed and the position of each particle are carried out Update, forward to try to achieve, according to cross-validation method, the fitness function value that each particle is corresponding

(3) textile image to be detected is carried out Gabor convolutional filtering, according to formula (20), by textile image to be detected With Optimal Gabor Filters G trained^*(x y) carries out convolution transform, obtains filtered image.

(4) based on extracting one group of textural characteristics on gray level co-occurrence matrixes (GLCM) image after the filtering；

Similar with (2.2), based on extracting 5 textural characteristics on GLCM after the filtering image, constitutive characteristic vector, as list The input sample of classification SVM.

(5) single classification SVM method is used to carry out fault differentiation.

For input sample, if formula (39) is set up, in the range of i.e. this sample packages is contained in SVDD suprasphere, then it is flawless point sample This (normal sample)；If the formula of being unsatisfactory for (39), then this sample is outside SVDD suprasphere scope, then be fault sample.

The aforementioned description to the specific illustrative embodiment of the present invention illustrates that and the purpose of illustration.These describe It is not wishing to limit the invention to disclosed precise forms, and it will be apparent that according to above-mentioned teaching, can much change And change.The purpose selected exemplary embodiment and describe is to explain that the certain principles of the present invention and reality thereof should With so that those skilled in the art be capable of and utilize the present invention various different exemplary and Various different selections and change.The scope of the present invention is intended to be limited by claims and equivalents thereof.

Claims (3)

1. fabric defect detection method based on single category support vector machines, it is characterised in that: the steps include:

(1) obtain fault-free fabric image, use random drift particle group optimizing (RDPSO) algorithm optimization Gabor filter Parameter, the single Optimal Gabor Filters of structure adaptation fault-free fabric image texture characteristic；

(2) parameter of RDPSO algorithm optimization list category support vector machines (SVM) is used；

(3) textile image to be detected is carried out Gabor convolutional filtering；

(4) based on extracting one group of textural characteristics on gray level co-occurrence matrixes (GLCM) image after the filtering；

(5) single classification SVM is used to carry out fault differentiation.

Method the most according to claim 1, it is characterised in that: step (1) comprises the following steps:

(1.1) structure two-dimensional space territory Gabor filter function, obtains frequency domain Gabor filter function through two-dimensional Fourier transform, The parameter needing the Gabor filter optimized is (σ_x, σ_y, λ, θ)；

(1.2) fault-free fabric image is carried out Gabor convolution transform, construct fitness function based on Fisher criterion, use RDPSO algorithm optimization obtains the parameter (σ of Optimal Gabor Filters_x, σ_y, λ, θ), structure adaptation fault-free fabric image texture The single Optimal Gabor Filters of feature.

Method the most according to claim 1, it is characterised in that: step (2) comprises the following steps:

(2.1) textile image without fault is carried out convolution with optimum Gabor filter, obtain filtered image；

(2.2) based on texture feature extraction on GLCM after the filtering image, constitutive characteristic vector, as the training of single SVM that classifies Sample；

(2.3) use Support Vector data description based on suprasphere thought (SVDD) as single classification SVM method, pass through RDPSO algorithm optimization determines two parameters C and the σ of SVDD.