CN103604809A - Pattern cloth flaw online visual inspection method - Google Patents

Abstract

The present invention provides a loom pattern cloth flaw online visual inspection method, which comprises: accurately calculating a texture primitive period, constructing a flaw-free primitive image offset sequence, and extracting pattern cloth characteristics so as to achieve real-time monitoring and flaw generation shutdown during a loom pattern cloth weaving process. According to the present invention, the texture primitive period extraction scheme is optimized, and the extreme value weight analysis is adopted to remove the interferential extreme value point so as to increase the period extraction accuracy; offline training on the standard flaw-free image is matched, the bi-layer classification mechanism in the online detection process is established, and detection accuracy is substantially improved while the real-time property is ensured; and requirements of high real-time property and high accuracy of the loom pattern cloth during the online detection process can be completely met.

Description

The online visible detection method of a kind of pattern Fabric Defect

Technical field

The present invention relates to utilize machine vision to loom at the weave cotton cloth online detection field of process of pattern cloth, specifically refer to a kind of image processing method that loom industry spot, that high real-time requires occurs that in complex pattern cloth is weaved cotton cloth process flaw is shut down in time that is applied to.

Background technology

Along with the raising of people to cloth quality requirements, to cloth, detect further important, Traditional Man detection technique cost is high, speed slow and the easy flase drop that causes tired out of human eye long-term work, very easily cause and occur flaw and accident that loom is still worked, when causing cloth quality to decline, also cause the factory to strict production control cost improves the quality of products in modern economy market to bring larger pressure.Loom in the process of weaving cotton cloth intelligence and find that in time the Awaiting Overhauls such as flaw auto stop are problems needing solution badly.

Currently loom is weaved cotton cloth in process to Defect Detection technical research is more to be concentrated on for the texture rules such as nonwoven fabrics or twill and the very little cloth detection field of texture, for have pattern or texture more greatly cloth such as a line, striped detect and but seldom relate to.This type of cloth detection difficulty is larger on the one hand for reason, and the less cloth of texture can be eliminated little texture impact and then be isolated defect areas by setting threshold by frequency domain methods such as Gabor or wavelet transformations, and large texture pattern cloth is difficult to eliminate texture information; Pattern cloth is various and conversion is complicated on the other hand, and design is difficult to form the unified detection scheme that can be adapted to other patterns to specific a kind of pattern detection scheme.

Summary of the invention

The object of the invention is to be difficult to for the pattern cloth compared with large texture the present situation detecting, and provides a kind of for the online vision-based detection scheme of pattern cloth (comprising the knitted fabrics such as sweater) flaw, can adapt to the requirement of industry spot high real-time, high-accuracy.

For reaching this object, technical scheme of the present invention is as follows: be divided into off-line training process and online testing process, for any a collection of texture cloth, indefectible cloth training once just can detect online to standard after camera Installation and Debugging, only to need off-line.Specifically comprise the following steps:

Off-line training process:

(1) the indefectible cloth image of Real-time Obtaining loom, adjusts the mechanical parameters such as camera lens aperture, focal length at the about 5cm of loom exit slot place, high-power strip source is installed below exit slot, pictures taken, Real-time Obtaining cloth image sequence.

(2) to the indefectible cloth image pre-service of obtaining, by gaussian filtering, remove image disruption point; By histogram equalization stretching image, strengthen contrast, highlight image texture.

(3) the cloth image texture primitive cycle accurately asks for

(3.1) build image distance stack equation, ask for respectively level and vertical direction apart from superpositing function, add up respectively and ask for minimal value.

(3.2) respectively level and vertical direction institute minimizing are done to weight analysis to eliminate the impact of the interference burr producing due to noise.First minimal value is done to ascending order by bubble sort method and arrange, the minimum value of the distance of the point after calculating every bit and arranging before it, does descending sort to minimum value by bubble sort method, and weight reduces gradually from front to back, gives up weight compared with low spot.

(3_3) calculated level and vertical direction minimal value spacing poor respectively, the difference of choosing the distance that occurrence number is maximum is the direction cycle for this reason, and note period profile is m and n.

(3.4) the correction cycle is for detection of striped cloth, if period m or n have one be 0 or the two differ greatly, the correction cycle.

(4) on the indefectible image of standard, choose of m * n size as template image M _{m * n}.This template image is offset successively, can obtains m * n skew subimage sequence, be designated as L _std.

(5) image block of choosing at random N width primitive cycle size on the indefectible image of standard is for feature extraction.Ask for respectively each image block and skew Sequence of Primitive Elements L _stdbetween energy poor, the average of the difference of calculating energy and variance, can obtain the indefectible image M of indefectible primitive and standard _{m * n}, N average energy value and N energy variance, be designated as respectively ENY and VAR.

(6) build classification policy, create double-deck classification mechanism.First ground floor fuzzy classification, determines boundary value according to the Two dimensional Distribution of average energy value and energy variance, determines and may contain primitive defective, secondly in the online testing process of second layer precise classification, completes.

Online testing process:

(1) Real-time Obtaining loom cloth image, at the about 5cm of loom exit slot place pictures taken, Real-time Obtaining loom cloth image to be detected.By gaussian filtering, remove image disruption point; By histogram equalization stretching image, strengthen contrast, highlight image texture, extract the stripe region that vertical direction height is identical with the vertical cycle size of primitive and detect.

(2), in stripe region to be detected, choose successively from left to right the big or small image block for primitive period m * n in order to detect.Ask for respectively every and be offset Sequence of Primitive Elements L with standard form _std, an energy poor, the average of the difference of calculating energy and variance.

(3) ground floor fuzzy classification, judges that according to off-line training gained boundary value this image block possibility contains flaw, if all without containing flaw, detection of end enters next frame image detection, if there is image block may contain flaw, does second layer precise classification.

(4) second layer precise classification, cumulative image block to be detected and standard form skew Sequence of Primitive Elements L _stdbetween energy profile, search minimum energy point, this point coordinate is image block to be detected and standard primitive departs from size.Analytical standard template primitive is at this migrated image and primitive to be detected, and the variance by difference between row image block to be detected and standard primitive judges whether it has higher and concentrated energy block, if there is this primitive to have flaw, if without this primitive is indefectible.

(5) if there is flaw to mark flaw location, shut down, if indefectible, enter next frame image detection.

Beneficial effect of the present invention: the present invention is by the indefectible image of off-line learning, accurately ask for the primitive cycle, ask for cycle method accurately errorless and repeated detection stable, the indefectible image shift cyclic sequence of structure standard is as testing tool, and that sets up that two layers of classified mechanism fuzzy classification and precise classification guarantee judges that whether image block is containing correctness defective.Preferred embodiment, this detection method algorithm is real-time, and accuracy is high, has overcome pattern cloth and has detected difficult problem, can be generalized to the pattern detection such as wallpaper, printing.

Accompanying drawing explanation

Fig. 1 total system design of graphics of the present invention

Fig. 2 algorithm overall flow of the present invention figure

Embodiment

For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in further detail.

The present invention is a kind of flaw detection method of pattern cloth, and testing process is divided into off-line training process and online testing process, and as shown in Figure 1, algorithm overall flow figure as shown in Figure 2 for smart camera mounting means.In off-line training process, by the accurate cycle, ask for, the structure of the indefectible primitive offset sequence of standard, extracts feature and sets up ground floor Fuzzy Classifier.In line process, analyze the difference of the indefectible primitive offset sequence of image block to be detected and standard and extract energy and Variance feature, first utilize ground floor sorter fuzzy classification, if likely there is flaw image block to build second layer precise classification device.

Further, off-line training process specific implementation step is:

Step 1, image sequence obtain

(1.1) according to the Fabric Defect real-time online accuracy of detection requirement of loom, at the about 5cm of exit slot place, the indefectible image sequence of continuous acquisition standard.

(1.2) preserve the image sequence that gathers, as indefectible standard picture, be used for follow-up off-line test.If conditions permit, regularly gathers indefectible image training, can prevent because system mechanics vibration causes the impact that the small displacement of camera causes accuracy of detection to decline.

Step 2, indefectible image pre-service

(2.1) the indefectible cloth image sequence being obtained by loom is utilized to gaussian filtering denoising, Gaussian convolution template is as follows:

$h=\frac{1}{16}*\left[\begin{array}{ccc}1& 2& 1\\ 2& 4& 2\\ 1& 2& 1\end{array}\right]$

(2.2) then image after gaussian filtering is done to histogram equalization and process, first gray level r in computed image _kthe probability occurring is approximately:

N is the summation of pixel in image, n _kthat gray level is r _knumber of pixels, L is gray level possible in image sum.Output gray level s _kfor:

$s_k=T\left(r_k\right)=\underset{1=0}{\overset{k}{\mathrm{\Σ}}}{P}_r\left(r_j\right)=\underset{j=0}{\overset{k}{\mathrm{\Σ}}}\frac{n_j}{n}$ k＝0，1，2，…，L-1

1 mistake above formula is r by gray level in input picture _keach pixel to be mapped to gray level in output image be s _krespective pixel.

Step 3, accurately extract the texture primitive cycle

(3.1) the indefectible image of note standard is [f (i, j)], i (1≤i≤p) wherein, and j (1≤j≤q) is respectively row and column coordinate, asks for the distance function of every row of the indefectible image of standard:

${\mathrm{\λ}1}_j\left(\mathrm{\δ}\right)=\underset{j=1}{\overset{q-\mathrm{\δ}}{\mathrm{\Σ}}}{[f(i,j)-f(i,j+\mathrm{\δ})]}^2$

Wherein: δ is pixel displacement, and λ 1, value (δ) can increase progressively and order computation from 1 to q-δ successively along with δ.

Each row distance superpositing function is cumulative, and the row distance superpositing function that obtains entire image is:

$\mathrm{\λ}1\left(\mathrm{\δ}\right)=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{\mathrm{\λ}1}_1=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}\underset{j=1}{\overset{q-\mathrm{\δ}}{\mathrm{\Σ}}}{[g(i,j)-g(i,j+\mathrm{\δ})]}^2$

Same method obtains column distance sum function:

$\mathrm{\λ}2\left(\mathrm{\δ}\right)=\underset{j=1}{\overset{1}{\mathrm{\Σ}}}{\mathrm{\λ}2}_1=\underset{j=1}{\overset{q}{\mathrm{\Σ}}}\underset{i=1}{\overset{p-\mathrm{\δ}}{\mathrm{\Σ}}}{[g(i,j)-g(i+\mathrm{\δ},j)]}^2$

Calculate respectively the minimum point of row distance sum function, N minimum point of note row distance superpositing function is respectively P ₁, P ₂..., P _n, same method application and column distance sum function.

(3.2) minimum point weight analysis

First use bubble sort method to the N of row distance sum function minimum point P ₁, P ₂..., P _nascending order is arranged, and after note row's example, sequence Q is Q ₁, Q ₂..., Q _n.

Then for Q ₁calculate itself and each Q ₁distance, Q wherein ₁after arranging, ascending order compares Q in sequence Q ₁first there is i.e. 1≤i≤j, note Q ₁the minimum value of required distance is D ₁, can be D apart from minimum value sequence D ₁, D ₂... D _n, this distance i.e. minimum point weight for this reason.

Finally to D ₁descending sort, after arranging, the corresponding order of extreme point is its weight descending sort table, gives up the relatively little extreme point of weight.

Same method application and column distance sum function.

(3.3) calculate the poor of row distance superpositing function retains after weight analysis minimum point spacing, the difference of choosing the distance that occurrence number is maximum is the horizontal texture primitive cycle, is designated as m; Same method application and column distance sum function, obtain the vertical texture primitive cycle, is designated as n.

(3.4) cycle correction

If m=0 or m < n/5, make m=n; If same n=0 or n < m/5, make n < m.

Step 4, the indefectible Sequence of Primitive Elements of structure standard

(4.1) on the indefectible image of standard, choose with the primitive cycle with size one as template image, remember that indefectible primitive is M _{m * n}=[f (c, d)].

(4.2) by indefectible primitive M _{m * n}skew successively, can obtain the primitive bunch of the indefectible primitive of standard after m * n translation.

For $M=\left[f(c,d)\right]=\left[\begin{array}{cccc}f\left(\mathrm{1,1}\right)& f\left(\mathrm{1,2}\right)& ...& f(1,n)\\ f\left(\mathrm{2,1}\right)& f\left(\mathrm{2,2}\right)& ...& f(2,n)\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ f(m,1)& f(m,2)& ...& f(m,n)\end{array}\right]$

Note $M_{11}=\left[\begin{array}{cccc}f\left(\mathrm{1,1}\right)& f\left(\mathrm{1,2}\right)& ...& f(1,n)\\ f\left(\mathrm{2,1}\right)& f\left(\mathrm{2,2}\right)& ...& f(2,n)\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ f(m,1)& f(m,2)& ...& f(m,n)\end{array}\right],$ $M_{12}=\left[\begin{array}{cccc}f\left(\mathrm{1,2}\right)& ...& f(1,n)& f\left(\mathrm{1,1}\right)\\ f\left(\mathrm{2,2}\right)& ...& f(2,n)& f\left(\mathrm{2,1}\right)\\ .& & .& .\\ .& & .& .\\ .& & .& .\\ f(m,2)& ...& f(m,n)& f(1,n)\end{array}\right],...,$

$M_{1n}=\left[\begin{array}{cccc}f(1,n)& f\left(\mathrm{1,1}\right)& ...& f(1,n-1)\\ f(2,n)& f\left(\mathrm{2,1}\right)& ...& f(2,n-1)\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ f(m,n)& f(m,1)& ...& f(m,n-1)\end{array}\right],...,$ $M_{22}=\left[\begin{array}{cccc}f\left(\mathrm{2,2}\right)& ...& f(2,n)& f\left(\mathrm{2,1}\right)\\ .& & .& .\\ .& & .& .\\ .& & .& .\\ f(m,2)& ...& f(m,n)& f(m,1)\\ f\left(\mathrm{1,2}\right)& ...& f(1,n)& f\left(\mathrm{1,1}\right)\end{array}\right],...,$

$M_{\mathrm{mn}}=\left[\begin{array}{cccc}f(m,n)& f(m,1)& ...& f(m,n-1)\\ f(1,n)& f\left(\mathrm{1,1}\right)& ...& f(1,n-1)\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ f(m-1,n)& f(m-\mathrm{1,1})& ...& f(m-1,n-1)\end{array}\right]$

Preserve this m * n image sequence to L _std, as feature extraction instrument.

Step 5, indefectible primitive image characteristics extraction

(5.1) on the indefectible image of standard, choose at random N (N > 100) individual with the primitive cycle with big or small image block in order to feature extraction.

(5.2) each image block that calculating is chosen at random and the energy between m * n the indefectible primitive migrated image of standard, ask for this each image block and the primitive bunch average of gained energy and the variance of energy, N image block can obtain N average energy value and N energy variance, is designated as respectively ENY _nand VAR _n

Step 6, structure fuzzy classification mechanism

Create the ground floor fuzzy classification in double-deck classification mechanism, according to the Two dimensional Distribution of average energy value and energy variance, determine boundary value, determine and may contain primitive defective.Boundary value is asked for as follows:

T11＝?(1-λ)*min(ENY _n)，T12＝?(1+λ)*max(ENY _n)

T21＝?(1-λ)*min(VAR _n)，T11＝?(1+λ)*max(VAR _n)

Wherein T11, T12, T21 and T22 represent that respectively in Two dimensional Distribution, energy lower limit, upper energy limit, variance roll off the production line and the variance upper limit, and λ is for limiting coefficient.

Further, in line process specific implementation step, be:

Step 1, detected image are obtained and pre-service

(1.1) Real-time Obtaining loom cloth image, at the about 5cm of loom exit slot place pictures taken, Real-time Obtaining loom cloth image to be detected.

(1.2) preprocessing process, utilizes gaussian filtering denoising to the indefectible cloth image sequence being obtained by loom, and Gaussian convolution template is as follows:

$h=\frac{1}{16}*\left[\begin{array}{ccc}1& 2& 1\\ 2& 4& 2\\ 1& 2& 1\end{array}\right]$

Then image after gaussian filtering is done to histogram equalization and process, increase contrast.

(1.3) extract a stripe region that wherein width is greater than the vertical cycle n of primitive as Defect Detection region.

Step 2, feature extraction

(2.1), in stripe region to be detected, choose successively from left to right with the primitive cycle with big or small image block sequence in order to detect note M ₁ ^rbe f image block, 1≤i≤NUM, NUM is separable graph picture piece number.

(2.2) ask for M ' ₁with standard form skew Sequence of Primitive Elements L _stdbetween energy poor, note standard form skew Sequence of Primitive Elements L _std(a, b) individual element is M _ab=[f _ab(c, d)], primitive to be detected is M ^r _i=[f ' ₁(c, d)] wherein m and n be primitive level and vertical cycle, f _ab(c, d) and

for element in primitive, 1≤c≤m, 1≤d≤n.The energy between two primitives is designated as:

$u_1(a,b)=\left|\right|M_{\mathrm{ab}}-{M^{\&prime;}}_1\left|\right|/N=\left(\underset{c=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{d=1}{\overset{n}{\mathrm{\&Sigma;}}}\right|f_{\mathrm{ab}}(c,d)-f_1^{\&prime;}(c,d)\left|\right)/N$ N=m * n wherein

To each primitive M ' to be detected ₁all can obtain m * n energy value.

The average energy value of this primitive to be detected and standard form skew Sequence of Primitive Elements is:

$u_1=\underset{a=1}{\overset{a=m}{\mathrm{\&Sigma;}}}\underset{b=1}{\overset{b=n}{\mathrm{\&Sigma;}}}{u}_1(a,b)/N$

The average energy value of this primitive to be detected and standard form skew Sequence of Primitive Elements is:

$v_1=\underset{a=1}{\overset{a=m}{\mathrm{\&Sigma;}}}\underset{b=1}{\overset{b=n}{\mathrm{\&Sigma;}}}{(u_1(a,b)-u_1)}^2/N$

The ground floor fuzzy classification of step 3, classification mechanism

Determining step two gained u ₁and v ₁whether all in the Two dimensional Distribution space of the training feature of extracting.

If T11≤u ₁≤ T12 and T21≤v ₁≤ T22, to any i all set up the Sequence of Primitive Elements that detects all indefectible.Otherwise the Sequence of Primitive Elements that detects may have flaw.

Note u=[u ₁, u ₂..., u _mun], v=[v ₁, v ₂..., v _mun], may contain flaw primitive index and be:

Logicl＝u≤T11||u≥T12||V≤T21||V≥T22

If sum (Logicl)=0, illustrates that detected Sequence of Primitive Elements is all indefectible, return to step 1 and obtain next frame image detection.

If sum (Logicl) ≠ 0, illustrates that detected Sequence of Primitive Elements is all indefectible, continue step 4 and do precise classification.

The second layer precise classification of step 4, classification mechanism

(4.1) in calculation procedure two in gained stripe region to be detected all with standard form skew Sequence of Primitive Elements L _stdin each element gained energy matrix u ₁(a, b) sum, obtaining big or small is the stripe region energy matrix U of m * n:

$U=\underset{1=1}{\overset{\mathrm{num}}{\mathrm{\&Sigma;}}}{u}_1(a,b)$

(4.2) ask for minimum value place coordinate (x in matrix U ₀, y ₀), (x ₀-1, y ₀-1) be primitive to be detected and standard primitive position offset, M ' with

texture is identical, f primitive M to be detected ₁ ^rwith

between difference can further react M ₁' whether for there being flaw.

I primitive M to be detected ₁ ^rwith

between difference matrix D IF ₁and DIF ₁middle all elements variance yields VAR ₁for:

${\mathrm{DIF}}_1=\mathrm{abs}(M_1^{\&prime;}-M_{(x_0-1)(v_0-1)}),$ ${\mathrm{VAR}}_1=\underset{c=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{d=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{DIF}}_1(c,d)-u_1(a,b))$

Note variance yields sequence VAR=[VAR ₁, VAR ₂..., VAR _num], ask for two-way difference and obtain VAR_differ, relatively VAR_differ and its mean value relationship obtain:

Logic2＝VAR_differ＞2*mean(VAR_differ)

Step 5, finally judgement

Note Logic=Logicl & Logic2, if sum (Logic) ≠ 0 exists flaw, marks defect areas, stopping alarm; Otherwise enter step 1 and detect next frame image.

Claims (5)

1. an online visible detection method for pattern Fabric Defect, is characterized in that, by apart from superpositing function and extreme value weight analysis, accurately asks for the pattern cloth texture primitive cycle; Extraction standard texture primitive, and create indefectible primitive migrated image bunch; Finally energy information between extraction and image block to be detected, builds according to this double-deck classification mechanism and whether contains flaw in order to decision-making, preferably resolves pattern cloth feature and is difficult to the problem of extracting, flaw is difficult to detection; Specifically comprise following step:

(1) in off-line learning process, the texture primitive cycle accurately asks for, and comprises apart from superpositing function and asking for and extreme value weight analysis;

(2) the indefectible primitive migrated image bunch L of creating a standard in off-line learning process _std, and extract energy and variance information architecture ground floor classification Fuzzy Classifier;

(3) in online testing process, extract primitive feature to be detected, with Fuzzy Classifier, classify;

(4) in online testing process, build precise classification device, determine final detection result.

2. a kind of online visible detection method of pattern Fabric Defect according to claim 1, is characterized in that: in described step (1), the texture primitive cycle accurately asks for, and comprises the following steps:

The first step, calculating row (column) are apart from superpositing function;

The indefectible image of note standard is [f (i, j)], i (1≤i≤p) wherein, and j (1≤j≤q) is respectively row and column coordinate, asks for the distance function of every row of the indefectible image of standard:

each row distance superpositing function is cumulative, and the row distance superpositing function that obtains entire image is:

$\mathrm{\&lambda;}1\left(\mathrm{\&delta;}\right)=\underset{i=1}{\overset{p}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}1}_i=\underset{i=1}{\overset{p}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{q-\mathrm{\&delta;}}{\mathrm{\&Sigma;}}}[g(i,j)-g(i,j+\mathrm{\&delta;}){]}^2$

Same method show that column distance sum function is: $\mathrm{\&lambda;}2\left(\mathrm{\&delta;}\right)=\underset{j=1}{\overset{q}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}2}_i=\underset{j=1}{\overset{1}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{p-\mathrm{\&delta;}}{\mathrm{\&Sigma;}}}{[g(i,j)-g(i+\mathrm{\&delta;},j)]}^2;$

Then calculate respectively the minimum point of row distance sum function, N minimum point of note row distance superpositing function is respectively p ₁, p ₂..., P _n;

Second step, extreme point weight analysis

First use bubble sort method to the N of row distance sum function minimum point P ₁, P ₂..., P _nascending order is arranged, and the routine postorder example of note row Q is Q ₁, Q ₂..., Q _n;

Then for Q ₁calculate itself and each Q ₁distance, Q wherein ₁after arranging, ascending order compares Q in sequence Q ₁first there is i.e. 1≤i≤j, note Q ₁the minimum value of required distance is D ₁, can be D apart from minimum value sequence D ₁, D ₂..., D _n, this distance i.e. minimum point weight for this reason;

Finally to D ₁descending sort, after arranging, the corresponding order of extreme point is its weight descending sort table, gives up the relatively little extreme point of weight;

After the 3rd step, calculating weight analysis, the minimum point spacing of row distance superpositing function reservation is poor, and the difference of choosing the distance that occurrence number is maximum is the horizontal texture primitive cycle, is designated as m; Same method application and column distance sum function, obtain the vertical texture primitive cycle, is designated as n;

The 4th step, comparison level and vertical period-luminosity relation, the correction cycle.

3. a kind of online visible detection method of pattern Fabric Defect according to claim 1, is characterized in that: in described step (2), Fuzzy Classifier builds, and comprises the following steps:

The first step, first on the indefectible image of standard, choose of primitive cycle size as template image, remember that indefectible primitive is M _{m * n}=[f (c, d)], by this indefectible primitive M _{m * n}skew successively, can obtain the primitive bunch of the indefectible primitive of standard after m * n translation, preserves this m * n image sequence to L _std, as feature extraction instrument;

Each image block that second step, calculating are chosen at random and the energy between m * n the indefectible primitive migrated image of standard, ask for this each image block and the primitive bunch average of gained energy and the variance of energy, N image block can obtain N average energy value and N energy variance, is designated as respectively ENY _nand VAR _n;

The 3rd step, according to the Two dimensional Distribution of average energy value and energy variance, determine boundary value, determine and may contain primitive defective, boundary value is asked for as follows:

T11＝(1-λ)*min(EYY _n)，T12＝(1+λ)*max(ENY _n)

T21＝(1-λ)*min(VAR)，T11＝(1+λ)*max(VAR _n)

Wherein T11, T12, T21 and T22 represent that respectively in Two dimensional Distribution, energy lower limit, upper energy limit, variance roll off the production line and the variance upper limit, and λ is for limiting coefficient.

4. a kind of online visible detection method of pattern Fabric Defect according to claim 1, is characterized in that: online Detection and Extraction primitive feature to be detected with Fuzzy Classifier classification in described step (3), comprises the following steps:

The first step, extract primitive feature to be detected, in stripe region to be detected, choose successively from left to right the image block sequence of primitive cycle size in order to detect, note M _i ^rbe i image block, 1≤i≤NUM, NUM is separable graph picture piece number; Ask for M ₁' be offset Sequence of Primitive Elements L with standard form _stdbetween energy poor, note standard form skew Sequence of Primitive Elements L _std(a, b) individual element is M _ab=[f _ab(c, d)], primitive to be detected is M ^r ₁=[f ' 1 (c, d)], wherein m and n are primitive level and vertical cycle, f _ab(c, d) and f _i ^r(c, d) is element in primitive, 1≤c≤m, 1≤d≤n.The energy between two primitives is designated as:

$u_1(a,b)=\left|\right|M_{\mathrm{ab}}-{M^r}_1\left|\right|/N=\left(\underset{c=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{d=1}{\overset{n}{\mathrm{\&Sigma;}}}\right|f_{\mathrm{ab}}(c,d)-{f_1}^r(c,d)\left|\right)/N$ N=m * n wherein

To each primitive M ' to be detected ₁all can obtain m * n energy value;

The average energy value of this primitive to be detected and standard form skew Sequence of Primitive Elements is:

The average energy value of this primitive to be detected and standard form skew Sequence of Primitive Elements

Step gained u in second step, judgement ₁and v ₁whether all in the Two dimensional Distribution space of the training feature of extracting, if T11≤u ₁≤ T12 and T21≤v ₁≤ T22, to any i all set up the Sequence of Primitive Elements that detects all indefectible, otherwise the Sequence of Primitive Elements that detects may have flaw; Note u=[u ₁, u ₂..., u _num], v=[v ₁, v ₂..., v _num], may contain flaw primitive index and be: Logicl=u≤T11||u>=T12||v≤T21||v>=T22;

If sum (Logicl)=0, illustrates that detected Sequence of Primitive Elements is all indefectible, return to step 1 and obtain next frame image detection, if sum (Logicl) ≠ 0 illustrates that detected Sequence of Primitive Elements is all indefectible, continue step 4 and do precise classification.

5. a kind of online visible detection method of pattern Fabric Defect according to claim 1, is characterized in that: in described step (4), build precise classification device, comprise the following steps:

In the first step, calculation procedure two in gained stripe region to be detected all with standard form skew Sequence of Primitive Elements L _stdin each element gained energy matrix u _i(a, b) sum, obtaining big or small is the stripe region energy matrix u of m * n: $U=\underset{i=1}{\overset{\mathrm{num}}{\mathrm{\&Sigma;}}}{u}_1(a,b);$

Second step, ask for minimum value place coordinate (x in matrix U ₀, y ₀), (x ₀-1, y ₀-1) be primitive to be detected and standard primitive position offset, i.e. M ^rwith texture is identical, i primitive M to be detected _i ^rwith

between difference can further react M ' ₁whether for there being flaw; I primitive M ' to be detected ₁with

between difference matrix D IF ₁and DIF ₁middle all elements variance yields VAR _ifor:

${\mathrm{DIF}}_1=\mathrm{abs}(M_1^{\&prime;}-M_{(x_0-1)(v_0-1)}),$ ${\mathrm{VAR}}_1=\underset{c=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{d=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{DIF}}_1(c,d)-u_1(a,b))$

Remember ten thousand difference Room row VAR=[VAR ₁, VAR ₂..., VAR _num], ask for two-way difference and obtain VAR_differ, relatively VAR_differ and its mean value relationship obtain: Logic2=VAR_differ > 2*mean (VAR_differ);

Note Logic=Logicl & Logic2, if sum (Logic) ≠ 0 exists flaw, marks defect areas, stopping alarm; Otherwise enter step 1 and detect next frame image.

Images