CN105158272A - Textile defect detection method - Google Patents

Abstract

The invention relates to a textile defect detection method. The method comprise the steps that firstly, a to-be-detected textile image with the to-be-cut edge characteristics is collected, and a series of image preprocessing operations are performed on the collected to-be-detected image; secondly, sawteeth of the image are eliminated, and textile edges are roughly determined through an LSD straight line detection algorithm; thirdly, the accurate edge positions are obtained by transversely cutting a longitudinal average gray value variation diagram, real-time edge reference data are determined according to relevant parameters of an acquisition system, and then extracted edge information and the edge reference data are compared to determine the edges and count; defect distribution is determined by combining standard diagram characteristic parameters and ignoring the defects around the edges, all defect communicating regions are obtained by adopting a region growing method, and then the number of a textile where the detects are located is judged by combining the edge positions. According to the textile defect detection method, information reference determination is repeatedly performed on the textile edges, and therefore the precision of on-line edge counting is improved; fine and scattered defects are regrown and communicated, and therefore the precision of on-line defect detection is improved.

Description

A kind of method for detecting textile defect

Technical field

The invention belongs to industrial automation, relate to the defect inspection method of a kind of product or workpiece.

Background technology

Machine vision is with the development of computer technology and field bus technique, and technology is increasingly mature, has become the product that the enterprise in increasing design automation field is indispensable.In some traditional forms of enterprises, the defect of workpiece or product still dependence official can detect, introduce machine vision and replace classic method, can greatly improve detection efficiency and automaticity, especially some high-risk working environments, artificial vision has been difficult to meet industrial automation enterprise to high-level efficiency, high-precision demand.In some streamline on-line checkingi, one Analysis of Nested Design is agreed with, it is complete to apply, the machine vision automatic checkout system of 24 hours steady operations, not only greatly reduce the manpower, material resources and financial resources of enterprise, can also the most detailed information feed back be obtained to make decisions the soonest.For improving enterprise competitiveness further, need more innovative technologies based on machine vision, the automatic online defect detecting system that development more agrees with enterprise demand is very urgent.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, provide the method for detecting textile defect of tool edge characteristics to be cut on a kind of streamline.

The technical scheme that technical solution problem of the present invention is taked is:

Step (1). gather image and Image semantic classification.

Step (2). use LSD line detection algorithm to extract textile side information in image.

Step (3). confirm real-time edge reference data according to acquisition system correlation parameter.

Step (4). the side information that contrast is extracted and edge reference data confirm edge and counting

Step (5). combined standard figure characteristic parameter and the distribution of edge center point location confirmation flaw.

Step (6). adopt region-growing method differentiate flaw location and split.

Beneficial effect of the present invention:

(1) the repeatedly information reference at textile edge is confirmed, improve the degree of accuracy of online edge counting.

(2) the thin loose class regrowth of textile flaw is communicated with, improves the degree of accuracy of on-line checkingi flaw.

Accompanying drawing explanation

Fig. 1 towel label defect detecting system process flow diagram;

Embodiment

Below in conjunction with Figure of description, the present invention is described further.

According to Figure of description (1), implementation step is described in detail:

Step (1). gather image and Image semantic classification.

For ensureing that image acquisition is continuous, stable, need be fixed on the production line of stabilized speed by industrial camera, illumination is stablized, and image acquisition interval can increase a little, but ensure that gatherer process does not have information gaps and omissions, namely continuous two image input contents can repeat but can not breakthrough.To the pre-service of image, mainly comprise and denoising is carried out to image, several processing procedure such as gradation conversion and image binaryzation.

Step (2). use LSD line detection algorithm to extract textile side information in image.

LSD algorithm target is the straight profile of local in detected image, extracts textile side information than being more suitable for.

The first step: eliminate sawtooth, the LSD algorithm first step is former is that input picture is reduced into 80% of original size, this object reduced is to weaken the crenellated phenomena occurred in even removal of images, and the image reduced lost some characteristic informations, fuzzy operation is carried out to image and equally can reach similar effect, but the interference of white noise can be increased.The present invention proposes a kind of modified LSD line detection algorithm, rim detection is carried out to removing pretreated figure, carry out simple Hough transformation straight line to the figure after process again to search, get a pretreated figure again, the thick line similar with edge block by color describes in rear figure, eliminate sawtooth with this, hold up straight edge, reduce line interruption.

Second step: gradient calculation and angle differentiate, set point X (m, n) represents the gray-scale value at pixel (m, n) place on gray level image, and image gradient is calculated by formula below:

$f_m(m,n)=\frac{X(m+1,n)+X(m+1,n+1)-X(m,n)-X(m,n+1)}{2}$

$f_n(m,n)=\frac{X(m,n+1)+X(m+1,n+1)-X(m,n)-X(m+1,n)}{2}$

The angle of this level-line is calculated by formula below:

$\mathrm{arc}tan\left(\frac{f_m(m,n)}{-f_n(m,n)}\right)$

Need to carry out gradient magnitude sequence and Grads threshold after the gray-scale map of non-binaryzation carries out LSD algorithm, and then determine the saltus step border of the degree of achieving the goal.The present invention directly carries out LSD algorithm to image after binaryzation, and saltus step is obvious, only needs simple angle to differentiate the position that just roughly can confirm straight region.

3rd step: two ends, left and right, edge judge, given set point X (m, n) represents the gray-scale value at pixel (m, n) place on gray level image, and picture altitude is H, and width is W, get the longitudinal average gray of image, calculated by formula below:

$f_{average}\left(m\right)=\frac{\underset{i=0}{\overset{H}{\Σ}}\left(X\right(m,i\left)\right)}{H},0\≤m\≤W$

After by each point (m, f _average(m)) retouch and make average gray variation diagram in a rectangular coordinate system, get peak (m _m, f _average(m _m) _max), if transversal ratio p (0<p<1), make straight line at this coordinate system, y=p*f _average(m _m) _max, transversal variation diagram, makes variation diagram and straight line intersection two point (m _l, f _average(m _l)) and (m _r, f _average(m _r)).After the process that there is edge in image, ideally there is one piece of constant width, highly equal the white rectangle of picture height, and other regions are all black, then, in longitudinal average gray variation diagram, about edge center point, have one piece of obvious projection.If m _r-m _lbe not similar to edge width, then reset p, otherwise then obtain exact edge center position ((m _r+ m _l)/2, H/2).

Step (3). confirm real-time edge reference data according to acquisition system correlation parameter.

Parameter is selected, number of image frames n per second, at the uniform velocity crawler belt motion fixed speed v from right to left, wall scroll towel towel face length degree l, edge width b, video picture actual width k.Whole piece towel video picture cycle T=(b+l)/v, frame number z=t*n, if the instantaneous towel edge of t=t1 enters image, during t=b/2v+t1, after b*n/2v frame, towel face enters field of view, during t=(b/2+k)/v+t1, edge is from picture drop-out, during t=(b/2+l)/v+t1 and t=T+t1-b/2v, edge occurs once again, then the edge time of occurrence of m bar of towel is (mT+t1-b/2v, mT+t1+k/v+b/2v) (m>1), with image lower frame for one-dimensional coordinate system, rightest point is initial point, x-axis direction is identical with crawler belt, from right to left, edge center point position is x=(t-t.mod (T) * T-t1) * v, x belongs to (-2/b, k+2/b) edge is had in the picture time.

Step (4). the side information that contrast is extracted and edge reference data confirm edge and counting

If this centerline and reference edge centrally to put distance very far away, or there is edge at reference edge along not time of occurrence section, the bulk flaw being similar to edge may be occurred, now centrally put with reference edge and be as the criterion.The edge center point gray scale judged in step 2 second step is too low maybe cannot find edge about or reference edge be boundless edge along not time of occurrence; Exact edge central point figure left side and this central point and reference edge centrally put apart from close, then edge is half side on an image left side; In like manner judge on right side; If distance is very far away, centrally puts with reference edge and be as the criterion.In the image of continuous acquisition, same edge there will be repeatedly, here only edge center point in a upper figure in left side, in this figure, right side is to edge counting number.

Step (5). combined standard figure characteristic parameter and the distribution of edge center point location confirmation flaw.

Carry out triple channel pixel to standard coloured picture to average, then get the cromogram after the denoising that simply gathers, get rid of edge center point position fixed range, generate self-defined binaryzation gray-scale map with each passage aberration contrast, the distribution of display flaw.

Step (6). adopt region-growing method differentiate flaw location and split.

Region-growing method is grown by the iteration of Seed Points, looks for the method for closed connected region.To the gray-scale map region growing after binaryzation, avoid cavity and over-segmentation that noise and gray scale heterogeneity may produce, but its maximum inferior position to be calculation cost large, the modified region-growing method that the present invention proposes effectively can overcome this point.

The first step: initial point selection, the pixel found during image sequence scans first carries out ownership inquiry, as seed after confirmation does not belong to.And ownership inquiry has larger calculated amount, only getting edges of regions pixel here enters to belong to sequence, and namely this pixel 4 neighborhood territory pixel meets growth criterion, then at intra-zone instead of edge, do not add ownership sequence.

Second step: growth, formulates growth criterion, centered by Seed Points, consider that its 4 neighborhood territory pixel meets preliminary growth criterion and not in Seed Sequences, then resets as Seed Points and add Seed Sequences, continue as central point, iteration grows, and the point meeting edge criterion is added ownership sequence.If former seed point location is (m, n), growth criterion is:

X (m-1, n)+X (m+1, n)+X (m, n-1)+X (m, n+1) >V, V=255, X (m, n) pixel value during the marking without Seed Sequences of representative point (m, n), if having, is 0.

Edge criterion is:

Y (m-1, n)+Y (m+1, n)+Y (m, n-1)+Y (m, n+1) <W, the pixel value of W=255*4, Y (m, n) representative point (m, n).

3rd step: stop, until ownership sequence joins end to end or the growth termination in Seed Sequences of each point in this region.

4th step: region merging technique, has marked off the region Q of m ' n for one, its barycenter (m ₀, n ₀) be:

$m_0=\frac{\underset{(m,n)\&Element;Q}{\&Sigma;}mX(m,n)}{\underset{(m,n)\&Element;Q}{\&Sigma;}X(m,n)},n_0=\frac{\underset{(m,n)\&Element;Q}{\&Sigma;}nX(m,n)}{\underset{(m,n)\&Element;Q}{\&Sigma;}X(m,n)}.$

In formula, X (m, n) is the gray-scale value of this point, sets a threshold value Y, includes the rectangular area Q in this region according to Q barycenter place system one minimum area _new, Q basis makes width and length all expand Y to Shi Yuan region, this region, judges which joint area becomes one piece by following two kinds of situations:

(m _p0, n _p0) ∈ Q _new, namely in the Q of the barycenter of region P after expansion,

$\frac{\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s+i,n_s)+\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s+i,n_s+Y)}{2(m+Y)}+\frac{\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s,n_s+i)+\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s+m+Y,n_s+i)}{2(n+Y)}>0,$

And (m _p, n _p) ∈ Q _new, (m _s, n _s) be Q _newupper left starting point coordinate, namely at Q _newlimit, rectangular area exists the pixel in other regions, belong to region P, then P and Q connects into one piece.Textile flaw is roughly divided into bulk, shot shape, strip, staggered-line by shape.For the easy differentiation of the obvious Connectivity Properties of tool such as bulk, strip, staggered-line, the region of shot shape distribution is then easily appeared multiaspect by differentiation and is amassed less zonule, is gone in size differentiates by sieve.For this situation, be necessary that the region less to area merges, region of being joined together in shot shape flaw face.All set merging rear region positions are differentiated, in conjunction with edge center point position, judges flaw place fabric numbering.

Claims (3)

1. a method for detecting textile defect, is characterized in that the method comprises the following steps:

Step (1). gather image and Image semantic classification;

Step (2). use LSD line detection algorithm to extract textile side information in image;

Step (3). confirm real-time edge reference data according to acquisition system correlation parameter;

Step (4). the side information that contrast is extracted and edge reference data confirm edge and counting;

Step (5). combined standard figure characteristic parameter and the distribution of edge center point location confirmation flaw;

Step (6). adopt region-growing method differentiate flaw location and split.

2. defect inspection method according to claim 1, is characterized in that: extract textile side information in image by LSD line detection algorithm in step (2), it is specially:

1) sawtooth is eliminated: carry out rim detection to removing pretreated figure, carry out Hough transformation straight line to the figure after process again to search, get a pretreated figure, the thick line similar with edge block by color describes in rear figure, eliminate sawtooth with this, hold up straight edge, reduce line interruption;

2) gradient calculation and angle differentiate: set point X (m, n) represents the gray-scale value at pixel (m, n) place on gray level image, and image gradient is calculated by formula below:

$f_m(m,n)=\frac{X(m+1,n)+X(m+1,n+1)-X(m,n)-X(m,n+1)}{2}$

$f_n(m,n)=\frac{X(m,n+1)+X(m+1,n+1)-X(m,n)-X(m+1,n)}{2}$

The angle of this level-line is calculated by formula below:

$arctan\left(\frac{f_m(m,n)}{-f_n(m,n)}\right)$

3) two ends, left and right, edge judge: set point X (m, n) represents the gray-scale value at pixel (m, n) place on gray level image, and picture altitude is H, and width is W, get the longitudinal average gray of image, are calculated by formula below:

$f_{average}\left(m\right)=\frac{\underset{i=0}{\overset{H}{\&Sigma;}}\left(X\right(m,i\left)\right)}{H},0\&le;m\&le;W$

After by each point (m, f _average(m)) retouch and make average gray variation diagram in a rectangular coordinate system, get peak (m _m, f _average(m _m) _max), if transversal ratio p, 0<p<1; Straight line is made, y=p*f at this coordinate system _average(m _m) _max, transversal variation diagram, makes variation diagram and straight line intersection two point (m _l, f _average(m _l)) and (m _r, f _average(m _r)); After the process that there is edge in image, ideally there is one piece of constant width, highly equal the white rectangle of picture height, and other regions are all black, then, in longitudinal average gray variation diagram, about edge center point, have one piece of obvious projection; If m _r-m _lbe not similar to edge width, then reset p, otherwise then obtain exact edge center position ((m _r+ m _l)/2, H/2).

3. defect inspection method according to claim 1, is characterized in that: differentiate flaw location with region-growing method in step (6) and split, it is specially:

1) initial point selection: the pixel found during image sequence scans first carries out ownership inquiry, as seed after confirmation does not belong to; Get edges of regions pixel to enter to belong to sequence, namely this pixel 4 neighborhood territory pixel meets growth criterion, then at intra-zone instead of edge, do not add ownership sequence;

2) grow: formulate growth criterion, centered by Seed Points, consider that its 4 neighborhood territory pixel meets preliminary growth criterion and not in Seed Sequences, then reset as Seed Points and add Seed Sequences, continue as central point, iteration grows, and the point meeting edge criterion is added ownership sequence; If former seed point location is (m, n), growth criterion is:

X (m-1, n)+X (m+1, n)+X (m, n-1)+X (m, n+1) >V, V=255, X (m, n) pixel value during the marking without Seed Sequences of representative point (m, n), if having, is 0;

Edge criterion is:

Y (m-1, n)+Y (m+1, n)+Y (m, n-1)+Y (m, n+1) <W, the pixel value of W=255*4, Y (m, n) representative point (m, n);

3) stop: until ownership sequence joins end to end or the growth termination in Seed Sequences of each point in this region;

4) region merging technique: the region Q one having been marked off to m × n, its barycenter (m ₀, n ₀) be: $m_0=\frac{\underset{(m,n)\&Element;Q}{\&Sigma;}mX(m,n)}{\underset{(m,n)\&Element;Q}{\&Sigma;}X(m,n)},n_0=\frac{\underset{(m,n)\&Element;Q}{\&Sigma;}nX(m,n)}{\underset{(m,n)\&Element;Q}{\&Sigma;}X(m,n)};$

In formula, X (m, n) is the gray-scale value of this point, sets a threshold value Y, includes the rectangular area Q in this region according to Q barycenter place system one minimum area _new, Q basis makes width and length all expand Y to Shi Yuan region, this region, judges which joint area becomes one piece: (m by following two kinds of situations _p0, n _p0) ∈ Q _new, namely in the Q of the barycenter of region P after expansion,

$\frac{\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s+i,n_s)+\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s+i,n_s+n+Y)}{2(m+Y)}+\frac{\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s,n_s+i)+\underset{i=0}{\overset{m+Y}{\&Sigma;}}X(m_s+m+Y,n_s+i)}{2(n+Y)}>0,$ And (m _p, n _p) ∈ Q _new, (m _s, n _s) be Q _newupper left starting point coordinate, namely at Q _newlimit, rectangular area exists the pixel in other regions, belong to region P, then P and Q connects into one piece; All set merging rear region positions are differentiated, in conjunction with edge center point position, judges flaw place fabric numbering.