CN101216436A - Fabric flaw automatic detection method based on Support Vector data description theory - Google Patents

Abstract

The invention belongs to the field of automatic detection and control of textile quality and particularly relates to a method for automatically detecting defects of textiles based on support vector data description theory. The automatic textile defect detection based on computer vision is a more difficult one-class classification task in real world. In the invention, support vector data description (SVDD) of the advanced one-class classification method is applied in the textile defect detection field for the first time to obviate the problems present in textile defect detection of conventional two-class classification support vector machine, which is difficult to collect representative defect samples completely and at larger number and further fails to effectively train the detector. Additionally, the invention provides a robust new method for solving the optimization problem of parameters, in particular to scale parameter of gauss kernel function, related in SVDD training. The automatic textile defect detection system based on SVDD can prospectively and conveniently control false alarm rate (false detection rate) in practice and can obtain lower miss ratio at lower false alarm rate.

Description

A kind of fabric flaw automatic detection method based on Support Vector data description theory

Technical field

The invention belongs to the automatic Detection ﹠ Controling of quality of textile products field, particularly relate to a kind of fabric flaw automatic detection method based on Support Vector data description theory.

Background technology

Automatically detecting based on the fabric defects of computer vision is research focus and the difficult point of using modern intellectual technology monitoring product quality over nearly one, 20 year.The automatic detection of float belongs to the pattern classification category in essence, specifically is exactly by the discriminating of various modern information technologies realizations to normal region in the textile image and defect areas.At present, most textile mills still adopt artificial perching mode, and there are many shortcomings in this mode, and are low as detection speed, can't implement real-time detection, and the labor cost that relates to, labour intensity, loss etc. are also all higher.Float automatic checkout system based on computer vision then can overcome above-mentioned defective, satisfies the needs of society to high efficiency production.

Float automatically detects the task of why being thought a difficulty by numerous association area researchers, one of them main cause be exactly it be a kind of typical single class classification problem.Single class classification (One-class Classification) also makes novel detection (Novelty Detection) or wild point detect (Outlier Detection), and it belongs to the unsupervised learning category in essence.Consider negative sample have distribute sparse, procurement cost is high or representative sample is difficult to comprehensive acquisition, in single class classification problem, the respective classified device all causes as training sample negative sample distributed with positive sample usually and knows nothing, decision function relies on the folk prescription support, more responsive to the noise in the training, and can only obtain in advance the erroneous judgement of positive sample is the error rate of positive sample for the error rate of negative sample can not obtain negative sample judged by accident in the training stage.The float sample promptly is the negative sample in the float detection problem.The float kind is extremely numerous and diverse, the mode of appearance variation is various, lacks the unified regularity of distribution, and its representative sample is difficult to comprehensive collection, can only rely on normal sample to the detection model training.

People such as researcher Kumar and Karras once attempted using two traditional class category support vector machines, and (SupportVector Machine SVM) detects research to a few float.In the training of such SVM, not only need a large amount of normal textile image samples, also need a large amount of involved float image patterns.Two class sorters are used for single class classification problem, and obviously, this is irrational: the training float sample that can not guarantee sufficient amount; Though can not guarantee to not training the kind fault and through training but show as effective detection of the similar fault of other form.And in recent years under the enlightenment of support vector machine principle, Support Vector data description (Support Vector DataDescription by Tax and Duin proposition, SVDD) and the one-class support vector machines that proposes by people such as Sch  lkopf (One-Class SupportVector Machine, OCSVM) then satisfied use in to the demand of one-class classifier.

Summary of the invention

Technical matters to be solved by this invention is that the single class sorting technique with this advanced person of Support Vector data description (SVDD) is applied to the automatic detection range of fabric defects rightly, detects the needs that this belongs to single class classification mode category task to adapt to fabric defects; And, provide a kind of new robust method with regard to the optimal selection problem of the scale parameter of parameter involved in the SVDD model training especially gaussian kernel function; Realize accurately controlling the target of the by mistake alert rate of expection and actual mistake police's rate and loss simultaneously lower target of maintenance in the actual detected practice of fault.

The technical solution adopted for the present invention to solve the technical problems totally is made up of the two large divisions: obtain to be used for the data set of SVDD model training, i.e. training set; The parameter that relates in the SVDD model especially scale parameter of gaussian kernel function is carried out preferably, and on preferred parameter basis, finish training SVDD.The decision function that obtains thus promptly can be used for the detection to unknown sample, judges whether it is the flaw sample.Two parameters that relate in the SVDD model effectively control detection error rate are especially expected alert rate of mistake and the alert rate of actual mistake.

1, the acquisition of training set

Gather the flawless textile image of 256 gray levels as much as possible, they all are divided into the subgraph of 32 * 32 pixel sizes zero lap, again each subgraph is implemented histogram equalization and handle, the subgraph after each equalization is regarded (normally) sample as.

From each normal sample, extract the some features that can distinguish normal texture and flaw texture and form the detection proper vector.The present invention has extracted following four fractal characteristic composition characteristic vectors according to meter box method:

Feature 1: image is the FRACTAL DIMENSION under 3～7 the situation in box size sequence;

Feature 2: image is the FRACTAL DIMENSION under 3～16 the situation in box size sequence;

Feature 3: the gradation of image value is along the FRACTAL DIMENSION of warp, the combination of latitude both direction projection sequence, and used grid size sequence is 3～16;

Feature 4: through window size be 10 * 10, standard deviation is that 0.2 LOG is the FRACTAL DIMENSION of the filtered image of Gauss-Laplace operator, observation box size sequence is 3～7;

The proper vector that obtains from each normal sample is implemented the softmax normalized, and the set of these proper vectors after normalized promptly constitutes the training sample set of SVDD model.

2, the SVDD Model parameter preferably with the training of SVDD model

2.1SVDD the meaning of ultimate principle and related parameter

The ultimate principle of SVDD is exactly at first the training sample data to be mapped to higher-dimension (inner product) feature space from the non-linear implicit expression of the input space, search a minimum suprasphere that comprises most mapping (enum) datas then therein, if sample data to be analyzed is positioned at suprasphere after same mapping, think that then this sample data is normal, otherwise be judged as unusual.

The search of optimum suprasphere relates to some significant variables, as the radius of a ball and the centre of sphere, is positioned at the outer robustness that strengthens judgement system of spheroid thereby need introduce some slack variables in addition to allow the small part training data.The formulation of SVDD is as follows: consider training dataset X={x ₁..., x _l∈ ^N, wherein l ∈ is the training sample sum.Feature Mapping Φ: X → F arrives higher-dimension inner product space F with input space data Nonlinear Mapping, inner product between the Image Data in it can calculate by a certain simple kernel function κ: κ (x, y)=＜Φ (x) Φ (y)＞, for example can use gaussian kernel function κ (x, y)=exp (γ || x-y|| ²), γ＞0.The use of kernel function makes mapping Φ can implicit expression exist and the user need not know its concrete form.The inner product space also is feature space.Note suprasphere radius is r ∈, centre of sphere c ∈ F, slack variable ξ={ ξ ₁..., ξ _l∈ ^l, the problem of seeking optimum suprasphere so in feature space promptly is converted into, for ν ∈ (0,1),

${\min }_{r,c,\mathrm{\&xi;}}{\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="S2008100322512E00011.png"\; state="\{\{state\}\}">r</figure-callout>}}^2+\frac{1}{\mathrm{vl}}{\left|\right|\mathrm{\&xi;}\left|\right|}_1$

subject to||Φ(x _i)-c|| ²≤r ²+ξ _i

ξ _i≥0，i＝1，…，l

Its dual problem is

min _α∑ _ijα _iα _jκ(x _i，x _j)-∑ _iα _iκ(x _i，x _i)

subject to∑ _iα _i＝1， $0\&le;{\mathrm{\&alpha;}}_i\&le;\frac{1}{\mathrm{vl}},$ i＝1，…，l(1)

Wherein, α={ α ₁... α _lIt is the Lagrange multiplier of introducing.

For the optimum solution α of dual problem,, have: if Φ (x according to KKT (Karush-Kuhn-Tucker) complementarity condition _i) be positioned at suprasphere, then its corresponding α _i=0; If Φ (x _i) be positioned at suprasphere surface, then its corresponding α _iSatisfy $0<{\mathrm{\&alpha;}}_i<\frac{1}{\mathrm{vl}};$ If Φ (x _i) be positioned at outside the suprasphere its corresponding α then _iSatisfy ${\mathrm{\&alpha;}}_i=\frac{1}{\mathrm{vl}}.$ In fact, in optimum solution α, most α _iAll be 0, and all that is corresponding to α _i＞0 training data x _iPromptly be known as support vector (support vectors).The indexed set of remembering these support vector correspondences is SV, then centre of sphere optimum solution $c=\underset{i\&Element;\mathrm{SV}}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_i\mathrm{\&Phi;}\left(x_i\right),$ And by kernel function, radius of a ball optimum solution r can by calculate corresponding to $0<{\mathrm{\&alpha;}}_i<\frac{1}{\mathrm{vl}}$ Arbitrary Φ (x _i) or whole Φ (x _i) average obtain to the distance of centre of sphere c.

To arbitrary sample data x to be discriminated, its decision function is

f(x)＝sgn(r ²-||Φ(x)-c|| ²)

＝sgn(r ²-∑ _i，j∈SVα _iα _jκ(x _i，x _j)+2∑ _i∈SVα _iκ(x _i，x)-κ(x，x)) (2)

If f (x)=1 judges that then x is a normal data; If f (x)=-1 judges that then x is an abnormal data.

The observation pairs optimization problem is not difficult to find, is being found the solution α={ α ₁... α _lBefore, need the occurrence of given parameter ν ∈ (0,1) and particular type and the parameter value thereof of kernel function κ.Gaussian kernel function κ (x, y)=exp (γ || x-y|| ²), γ＞0 function admirable, highly versatile, thereby widely used in practice.The present invention also selects for use gaussian kernel function to carry out relevant research, and it only relates to unique scale parameter γ.Simplified summary once, that is exactly that finding the solution of dual problem needs ν and γ is given in advance or preferably come out.

At first introduce two evaluation of algorithm indexs:

Wherein by mistake alert rate comprises two kinds, i.e. the alert rate of actual mistake of alert rate of the expectation of training stage mistake and test phase.

Parameter ν is discussed below.ν is that exceptional sample is that those upper bounds that are positioned at the outer training sample ratio of normal estimation region also are the lower bounds of support vector ratio simultaneously.That is to say that parameter ν is controlling the ratio of support vector in the training result.And according to the demonstration of relevant document, the expectation value of support vector ratio is the upper bound of the alert rate of mistake, promptly

$\mathrm{FAR}\&le;E\left[\frac{\#\mathrm{SV}}{l}\right],$

Wherein FAR is the alert rate of mistake, and #SV is the quantity of support vector, and l is a total sample number.In actual the use, can directly use of the roughly estimation of support vector ratio as the alert rate of mistake.This shows, preestablishes the ν value and has also just roughly set the alert rate of mistake, and conversely, if proposed in advance the requirement of police's rate by mistake, so corresponding ν value also can be determined.

Next the scale parameter γ of gaussian kernel function is discussed.In fact, ν effectively is controlled at the concrete value that depends on γ to a great extent to the alert rate of mistake, and it can not independently bring into play due effect.The detecting device that design performance is superior, it is vital choosing appropriate γ value, it has determined the flexibility ratio of gaussian kernel function.If the γ value is too small, so detecting device to training data phenomenon appears owing to learn in study, this will cause decision boundary too loose and simple, the separating capacity of detecting device is very poor, loss can be very high during actual test; Otherwise, if the γ value is excessive, the study phenomenon then appearred in the detecting device training process, and this will cause decision boundary too complicated, and the generalization ability of detecting device is very poor, and the alert rate of actual test mistiming can be very high.The ν value according to the predefined situation of the alert rate of desired mistake under, our main task is exactly to finish preferred to γ.

2.2 the deficiency of existing gaussian kernel function scale parameter γ method for optimizing

Aspect the gaussian kernel function scale parameter that utilizes the cross validation method that SVDD is related to preferred, people such as Banerjee are this deductions of the upper bound of the alert rate of mistake according to the expectation value of support vector ratio, and consider the fact of in the training process loss being known nothing, utilize the search of cross validation method to make the expectation value minimum of support vector ratio promptly miss the γ value of alert rate upper bound minimum, then with the parameter value of this value as actual test model.Of the present invention studies show that only relies on the mean value of the support vector ratio in the cross validation method to come preferred γ can not guarantee to estimate the less and comparatively approaching ν value that sets of the alert rate of expectation mistake of by mistake alert rate in actual applications.To this, we are the example explanation that experimentizes with a data set, and this data set is made up of training set that comprises 4608 normal samples and the test set that comprises 2797 normal samples and 275 flaw samples respectively.To predefined ν value, adopt 8 retransposing proof methods that a series of γ values are tested.Note mrsv is the mean value of the support vector ratio that obtains of training under the above-mentioned proof method system, and mfar is the alert rate mean value of corresponding cross validation mistake, then with above-mentioned data set in ν=0.02, γ=1 * 10 ^-6～1 * 10 ³The time training be example, obtain result as shown in table 1.

By table 1 as can be seen: thus minimum support vector ratio average may be corresponding the very big ν value cross validation mistake police rate average far away that departs from, for example number the corresponding mrsv minimum of 4 γ values between 1～4, but their mfar is but very undesirable, thereby such parameter is not the best or preferable selection that is used for actual test model, to this data set under other ν value situation with the analysis under different ν value situations also has similar conclusion to other data set, no longer itemize here.This shows that the scale parameter method for optimizing of the gaussian kernel function that people such as Banerjee propose is also unreliable, promptly only relies on this index of support vector ratio average can not guarantee that selected γ value is optimum.

Cross validation result under table 1 ν=0.02 situation

The γ numbering	1	2	3	4	5	6	7	8	9	10
											The γ value	1.E-06	5.E-06	1.E-05	5.E-05	1.E-04	5.E-04	1.E-03	5.E-03	1.E-02	5.E-02
mrsv	0.020089	0.020089	0.020089	0.020089	0.020089	0.020089	0.020101	0.020208	0.020252	0.020303
											mfar	0.458810	0.520280	0.565420	0.519810	0.023365	0.020486	0.020841	0.020255	0.020298	0.020247
The γ numbering	11	12	13	14	15	16	17	18	19
											The γ value	1.E-01	5.E-01	1.E+00	5.E+00	1.E+01	5.E+01	1.E+02	5.E+02	1.E+03
mrsv	0.020325	0.020361	0.020373	0.021079	0.021990	0.035378	0.077550	0.522980	0.791470
											mfar	0.020312	0.020305	0.020312	0.021050	0.021846	0.035323	0.077257	0.528860	0.805060

2.3 the sane γ method for optimizing that the present invention proposes

From table 1, we it can also be seen that:

(1) in γ numbering 5～15 (corresponding γ value 1 * 10 ^-4～1 * 10 ¹) sizable span in, mrsv and mfar are very approaching, and mrsv and mfar in the corresponding zone of other γ value, they are more approaching with default ν=0.02.We claim that mrsv and this common long stabilized zone of mfar are the smooth transition band.We courageously infer thus, and alert rate of the mistake during actual test the and loss also all will keep relative stability in this smooth transition band.The final testing result of back will confirm this supposition.

(2) bigger in view of smooth transition band span, more careful denser in other words γ value is divided and the cross validation experiment there is no need.

(3) the minimum alert rate average of cross validation mistake not necessarily corresponding minimum support vector ratio average, as number shown in 10 the relevant data.

In theory, those pairing mrsv and mfar are that comparatively ideal parameter value is selected with the more approaching γ value of default ν value.Two indexs are of equal value to a certain extent, but can reflect the alert rate situation of actual mistake separately from different perspectives, and two indexes can obtain in the cross validation experiment simultaneously.The average of remembering both is msf, and then msf also can be used as the estimation of estimating the alert rate of mistake, and the pairing γ value of choosing among the msf of reckling is more reliable.This γ generally is arranged in the smooth transition band, is numbered No. 8 as this γ value in this example.

Therefore, the sane method for optimizing of γ that the present invention proposes promptly is to adopt msf (being the average of mrsv and mfar) index rather than mrsv to come preferred γ, specifically is exactly the optimized parameter of choosing when making the γ value of msf minimum as the SVDD model training.

In above-mentioned method for optimizing, the inventor finds and has proposed smooth transition band notion first; Point out that in the smooth transition band alert rate of mistake and loss fluctuation are little, keep relative stability; The bigger notion of smooth transition mark degree span is proposed first; Unite first and use cross validation support vector ratio average and two indexs of the alert rate average of cross validation mistake to carry out the optimization of gaussian kernel function scale parameter.

2.4 the training of SVDD model and the acquisition of decision function

On (all forming) whole training set, find the solution the corresponding optimization problem described in 2.1 joints according to ν value of preferably coming out and γ value, thereby obtain the formula that embodies of the decision function described in 2.1 joints by normal sample.Can realize differentiation according to this decision function to unknown sample ownership.

Alternate embodiments

SVDD is based on suprasphere thought, and OCSVM then is based on lineoid thought, promptly distinguishes normal and flaw by the optimum lineoid that training data is separated with largest interval and initial point at energy of high-dimensional feature space searching.People such as people such as Sch  lkopf and Tsang are verified, and for used most of kernel functions in the reality, SVDD and OCSVM are of equal value in itself.Therefore the SVDD in the alternative technique scheme of available OCSVM reaches same fabric defects detection effect.

Invention advantage and beneficial effect

1, the present invention is applied to the automatic detection range of fabric defects with advanced person's one-class classifier SVDD first, has adapted to fabric defects and has detected the needs that this belongs to single class classification mode category task.SVDD is used for fabric defects as detecting device detects, can avoid adopting that two traditional class category support vector machines are met with is difficult to all sidedly and larger amt ground representative float sample of collection and then the predicament that can not effectively train detecting device when fabric defects detects.

2, at what relate in the SVDD training the especially scale parameter optimal selection problem of gaussian kernel function of related parameter arranged, the present inventor has provided a kind of steadily and surely and easily new method, SVDD is trained and uses it for actual detected making on the parameter basis of preferably coming out in this way, can be expectedly and the alert rate (false drop rate) of control mistake easily, and can under the lower situation of the alert rate of mistake, obtain lower loss simultaneously.

Description of drawings

Fig. 1 is the relevant training and testing index of data set 2 changing trend diagram with γ when v=0.02.

Fig. 2 is the relevant training and testing index of data set 2 changing trend diagram with γ when v=0.05.

Fig. 3 is the relevant training and testing index of data set 5 changing trend diagram with γ when v=0.02.

Fig. 4 is the relevant training and testing index of data set 5 changing trend diagram with γ when v=0.05.

Fig. 5 is some routine fabric defects actual detected effect synoptic diagram.

Embodiment

Concrete implementation step

A kind of fabric flaw automatic detection method based on Support Vector data description theory comprises following concrete implementation step:

1, gather the flawless textile image of 256 gray levels as much as possible, and they all are divided into the subgraph of 32 * 32 pixel sizes, each subgraph is a training sample.

2, above-mentioned each training sample being carried out histogram equalization handles.

3, from each sample, extract the proper vector that following four fractal characteristics are formed training and detected usefulness according to meter box method:

Feature 1: image is the FRACTAL DIMENSION under 3～7 the situation in box size sequence.

Feature 2: image is the FRACTAL DIMENSION under 3～16 the situation in box size sequence.

Feature 3: the gradation of image value is along the FRACTAL DIMENSION through the combination of broadwise projection sequence, and used grid size sequence is 3～16.

Feature 4: through the FRACTAL DIMENSION that window size is 10 * 10, standard deviation is 0.2 the filtered image of LOG operator, observation box size sequence is 3～7.

4, adopt the softmax data normalization method that each proper vector is carried out normalized.This method was made up of two steps:

$y=\frac{x_{\mathrm{test}}-\stackrel{\&OverBar;}{x}}{\mathrm{r\&sigma;}},$

$x_{\mathrm{soft}\max }=\frac{1}{1+\exp (-y)}$

Among the present invention, The average of representing certain characteristic parameter of normal cloth textured image estimates from a large amount of normal samples, and σ represents the standard deviation of this characteristic parameter of normal cloth textured image, x _TestThis characteristic ginseng value of representing cloth textured image to be measured, the coefficient of r for setting up on their own by the user, unification of the present invention is taken as 2, so x _SoftmaxBe the characteristic ginseng value after the normalization.

5, the SVDD model is carried out parametric optimization and training.This comprises following a few step:

5.1 according to requirement to the alert rate of expectation mistake, the ν value of default corresponding rationally (promptly also non-trend towards extreme 1 neither trend towards extreme 0), ν ∈ (0,1);

5.2 to than in the large scale scope but the sparse a series of γ values that distribute are implemented 8 retransposing proof methods, orient making the γ value (the concrete implication of msf is referring to relevant introduction in 2.3 joints of technical scheme) of msf minimum;

5.3 whole training set is carried out the training of SVDD model with the γ value of presetting the ν value and preferably obtain, promptly find the solution the SVDD antithesis optimization problem under corresponding ν value and the γ value situation, can be referring to the introduction of antithesis optimization problem formula (1) in the background technology.

5.4 obtain the decision function that to differentiate unknown sample, referring to the formula in the background technology (2).

6, to the unknown sample of each 32 * 32 pixel to be measured, (can referring to the concrete introductions in the 2nd, 3,4 steps) be handled in its extractions, eigenwert softmax normalized etc. of implementing histogram equalization, four fractal characteristics successively, is that the formula (2) in the background technology judges whether this sample is the flaw sample then with normalization proper vector input decision function.

Specific embodiment

The present invention provides five specific embodiments, with further elaboration the present invention.Should be understood that these embodiment only to be used to the present invention is described and be not used in and limit the scope of the invention.Should be understood that in addition those skilled in the art can make various changes or modifications the present invention after the content of having read the present invention's instruction, these equivalent form of values fall within the application's appended claims institute restricted portion equally.

Table 2 has been listed five relevant data distribution condition with data set of different texture background.

The guide look of 2 experiment sample situations

Sample dispensing	Training set	Test set
					Complete is normal sample	Normal sample	The flaw sample	Add up to
	Data set 1 plain weave	6144	2544	528					3072
Data set 2 plain weaves					4608	2797	275	3072
	Data set 3 twills	3072	1960	344					2304
Data set 4 twills					2016	2018	414	2432
	Data set 5 twills	3016	2973	867					3840

In order to verify the γ parametric optimization method that is proposed better and consider the length problem that we illustrate relevant issues with Fig. 1～Fig. 4.These figure have illustrated that in the actual test to the unknown sample of data set 2 and data set 5 indexs such as msf, the alert rate of actual mistake and loss are respectively in ν=0.02 and 0.05 o'clock variation tendency with γ.In order to simplify expression, horizontal ordinate is represented with γ numbering rather than actual value among the figure, but their the occurrence table of comparisons 1.

From this four width of cloth figure as can be seen, long smooth transition band exists really, and in the smooth transition band, msf is very high with the alert rate degree of agreement of actual mistake, both are also very near default ν value, and loss also tends towards stability in this zone as the front is expected.Outside the smooth transition band, msf and the more default ν value of the alert rate of actual mistake depart from bigger.According to the optimum principle that we propose, the optimum γ value numbering of four width of cloth figure is followed successively by 8#, 6#, 12# and 9#, and they all are positioned at the smooth transition band.What merit attention in addition is exactly, and under the alert rate of lower mistake, it is quite low that actual measurement obtains loss, and this means the recall rate of fault very high simultaneously.Relevant test to the remainder data collection also can obtain similar conclusion, lists no longer in detail here.This shows that the method for optimizing of this gaussian kernel function scale parameter that we propose is that measured result is satisfactory reliably.

Table 3 for adopt detection scheme proposed by the invention to these five data sets ν=0.02,0.04 with 0.06 three kind of situation under relevant training and testing result gather.The result is satisfactory, shows as: the value of the gaussian kernel function scale parameter γ that preferably comes out according to principle that we carry all is positioned at corresponding smooth transition band; Alert rate of actual mistake and the alert rate basically identical of expectation mistake; Alert rate of mistake and loss remain on lower even extremely low level simultaneously.

Provide some fabric defects actual detected effect synoptic diagram at last, show testing result intuitively, as shown in Figure 5.Among the figure, each textile image size all is 256 * 256 pixels, all comprises 64 32 * 32 subsamples, has all passed through the histogram equalization processing.Each line display is to a testing process from the textile image of 256 * 256 pixel sizes of a certain data set; First each figure of row is corresponding normal textile image, uses as a comparison; Each figure of secondary series is the image that comprises float; The 3rd each figure of row is the actual detected design sketch, and wherein each band fork lattice represents by algorithm identified that we carry to be 32 * 32 pixel sub images of flaw.

The part testing result of five data sets of table 3 gathers

Data set numbering and training normal sample given figure/test normal sample given figure	ν (promptly estimating the alert rate of mistake)	msf	The alert rate of actual mistake	Loss	Whether the corresponding γ value that optimizes is positioned at the smooth transition band
						Data set 1 6144/2544	0.02	0.02010	0.01953	0.02181	Be
0.04	0.03994	0.03906	0.01335	Be
					0.06		0.06010	0.05794	0.00781	Be
Data set
	2 4608/2797	0.02	0.02024	0.01823	0.01237	Be
							0.04	0.04012	0.04590	0.00391	Be
		0.06	0.06014	0.07194	0.00228	Be
Data set 3 3072/1960							0.02	0.02030	0.02040	0.03472	Be
	0.04	0.04005	0.03385	0.02040	Be
						0.06	0.05997	0.06033	0.01085	Be
	Data set
4 2016/2018		0.02	0.02025	0.02467	0.02344	Be
							0.04	0.04030	0.05921	0.00617	Be
		0.06	0.06060	0.09540	0.00247	Be
	Data set 5 3016/2973						0.02	0.02053	0.01615	0.02005	Be
0.04		0.04045	0.03880	0.00781	Be
						0.06	0.06037	0.06302	0.00417	Be

Claims (10)

1. the fabric flaw automatic detection method based on Support Vector data description (brief note is SVDD) theory comprises the following steps:

(1) collection of textile image and pre-service;

(2) extraction of fractal characteristic and proper vector normalized;

(3) compromise parameter ν and gaussian kernel function scale parameter γ preferred in the SVDD model;

(4) acquisition of the training of SVDD model and corresponding decision function expression;

(5) sample to be tested is tested according to the described decision function expression formula of step (4), differentiated it and whether comprise flaw.

2. a kind of fabric flaw automatic detection method based on SVDD according to claim 1 is characterized in that: the collection of image is meant the flawless textile image of collecting more 256 gray levels in the described step (1), uses as training; Pre-service is meant the subgraph that these images all is divided into a certain size, and the subgraph size is 32 * 32 pixels among the present invention, and each subgraph is represented a training sample, and each subgraph is implemented histogram equalization handle.

3. a kind of fabric flaw automatic detection method based on SVDD according to claim 1 is characterized in that the fractal characteristic in the described step (2) is to obtain following four FRACTAL DIMENSION composition characteristic vectors according to the estimation of meter box method:

Feature 1: image is the FRACTAL DIMENSION under 3～7 the situation in box size sequence;

Feature 2: image is the FRACTAL DIMENSION under 3～16 the situation in box size sequence;

Feature 3: the gradation of image value is along the FRACTAL DIMENSION of warp, the combination of latitude both direction projection sequence, and used grid size sequence is 3～16;

Feature 4: through window size be 10 * 10, standard deviation is that 0.2 LOG is the FRACTAL DIMENSION of the filtered image of Gauss-Laplace operator, observation box size sequence is 3～7;

4. a kind of fabric defects defect automatic testing method according to claim 1 based on SVDD, it is characterized in that used proper vector is not limited to the vector of step as claimed in claim 1 (2) and the described some fractal characteristics compositions of claim 3, other proper vector that can distinguish normal texture and flaw texture also can.

5. a kind of fabric flaw automatic detection method according to claim 1 based on SVDD, what it is characterized in that proper vector normalization in the described step (2) adopts is the softmax data normalization method, is made up of two steps, is defined as follows:

$y=\frac{x_{\mathrm{test}}-\stackrel{\&OverBar;}{x}}{\mathrm{r\&sigma;}}---\left(1\right)$

$x_{\mathrm{soft}\max }=\frac{1}{1+\exp (-y)}---\left(2\right)$

Wherein,

The average of representing certain characteristic parameter of normal cloth textured image estimates from a large amount of normal samples,

σ represents the standard deviation of this characteristic parameter of normal cloth textured image, x _TestThis spy who represents cloth textured image to be measured

Levy parameter value, r is unified in the present invention to be taken as 2, so x _SoftmaxBe the characteristic ginseng value after the normalization.

6. a kind of fabric flaw automatic detection method based on SVDD according to claim 1 is characterized in that the SVDD model in described step (3) and the step (4) is meant following this optimization problem:

min _α∑ _ijα _iα _jκ(x _i，x _j)-∑ _iα _iκ(x _i，x _i)

subject toΣ _iα _i＝1 $0\&le;{\mathrm{\&alpha;}}_i\&le;\frac{1}{\mathrm{vl}},$ i=1，…，l

Wherein, α={ α ₁... α _lBe the Lagrange multiplier of introducing, X={x ₁..., x _l∈ ^NBe training dataset, l ∈ is the training sample sum, and ν is compromise parameter, and (x y) is kernel function to κ.

7. (x y) refers to gaussian kernel function in the present invention according to the kernel function κ described in the claim 6

κ (x, y)=exp (γ || x-y|| ²), wherein γ＞0 is the scale parameter of this kernel function.

8. a kind of fabric flaw automatic detection method based on SVDD according to claim 1 is characterized in that two have preferably realizing as follows of related parameter in the described step (3):

(a) according to requirement to the alert rate of expectation mistake, default ν value, ν ∈ (0,1) makes it neither trend towards extreme 0 and also non-ly trends towards extreme 1;

(b) to than in the large scale scope but the sparse a series of γ values that distribute are implemented 8 retransposing proof methods, orient making that msf is the γ value of the average minimum of mrsv and mfar.

Finish preferred to these two parameters thus.Here, mrsv is the mean value of the support vector ratio that training obtains under the proof method system in above-mentioned (b), and mfar is the alert rate mean value of corresponding cross validation mistake.

9. a kind of fabric flaw automatic detection method according to claim 1 based on SVDD, it is characterized in that in the described step (4) that the training of SVDD model is meant that the ν value and the γ value of preferably coming out according to Claim 8 find the solution the corresponding optimization problem described in the claim 6 on whole training set, thus the decision function expression formula described in the acquisition claim 1:

f(x)＝sgn(r ²-||Φ(x)-c|| ²)

＝sgn(r ²-∑ _i，j∈SVα _iα _jκ(x _i，x _j)+2∑ _i∈SVα _iκ(s _i，x)-κ(x，x))

If f (x)=1 judges that then x is normal sample; Otherwise judge that x is the flaw sample.

Here SV is meant that all that is corresponding to α _i＞0 training data x _iIndexed set.

10. a kind of fabric flaw automatic detection method according to claim 1 based on SVDD, it is characterized in that: each sample to be tested in the described step (5) all need pass through the treatment steps such as extraction, eigenwert softmax normalization of histogram equalization, four fractal characteristics before by the test of the described decision function of claim 9 successively.

Images