CN103913468A - Multi-vision defect detecting equipment and method for large-size LCD glass substrate in production line - Google Patents

Abstract

The invention discloses a multi-vision defect detecting equipment and a multi-vision defect detecting method for a large-size LCD glass substrate in a production line. For the equipment, a linear scanning imaging detection system is adopted for acquiring the high-definition gray image of an LCD glass substrate to be detected, and the equipment is simple in structure and convenient to operate; according to the method disclosed by the invention, the acquired image is processed, a kmeans clustering method is adopted to carry out defect existence judgment on the processed image of the LCD glass substrate, and by making the defect regions, the defect category is judged by a classification method of a support vector machine (SVM), and the defect quantity is counted. Processed image and detection results are transmitted to an office in real time by a filed bus control system for operators on duty to check, and the field production parameters are transmitted, so that remote monitoring of the system is realized.

Description

Many defects of vision checkout equipment and the method for large-scale LCD glass substrate on production line

Technical field

The invention belongs to visual detection equipment field on electronics manufacturing line, particularly many defects of vision checkout equipment and the method for large-scale LCD glass substrate on a kind of production line.

Background technology

Along with the development of electronic technology, the price of consumption electronic product is more and more lower, and people constantly expand the demand of electronic product, and this has also driven the manufacturing development of LCD glass substrate.But for pursuing more perfect seeing and hearing enjoyment, the main flow size of liquid crystal display, LCD TV and mobile terminal, all in continuous expansion, even there will be the lcd screen of some oversizes.Therefore, large-sized LCD glass substrate is produced and will be the turning point of seeking following industry tremendous development.But the quality detection technology of large-scale LCD glass substrate is also a bottleneck of restriction industry development.

Although most of production run of LCD glass substrate is all to complete in the dust free room the inside of high-cleanness, high, but on LCD glass substrate, still inevitably there will be some defects, these defects will cause the integrated circuit printing on substrate normally to work, and cause the flaw of LCD display.Cause the reason of these defects to have a lot, all may defective generation at the links of its manufacture.These defects mainly comprise cut, scraping, hole, particulate, bubble, snotter etc.Some defects are very trickle, and human eye is difficult to observe.

In high-speed automated electronics manufacturing line, how LCD glass substrate is carried out to quality testing fast and accurately, be an important technology difficult problem that is directly connected to product quality.The quality testing of most production line mainly relies on manual method, i.e. manual observation detects defect.The defect of manual detection is: 1, detection speed is slow, and inefficiency cannot meet the demand of high-speed automated production line; 2, accuracy of detection is low, detects quality and is affected by human factors, and the undetected probability of flase drop is higher; 3, labor strength is large, poor working environment; 4, the preservation and the inquiry inconvenience that detect data, be not easy to management.

Summary of the invention

For the deficiency of existing detection technique, the invention provides many defects of vision checkout equipment and the method for large-scale LCD glass substrate on a kind of production line.

On production line, many defects of vision checkout equipment of large-scale LCD glass substrate, comprises image-forming detecting system and bus control system;

Wherein, described image-forming detecting system at least comprises two line scan camera 11, one or more LED line source 12, transmission control device 13 and detection computations machines 14 of installing side by side;

Described transmission control device 13 comprises the mutual display unit 134 of travelling belt 131, feeler 132, PLC133 and PLC;

The mutual display unit 134 of described feeler 132, PLC133 and PLC is connected successively; Described travelling belt is controlled by PLC; Described feeler is installed on the fixed frame of travelling belt one side, and in the time that LCD glass substrate arrives predeterminated position with travelling belt, feeler sends a signal to PLC;

Described line scan camera 11 is connected with described detection computations machine, and is controlled by PLC133; Described line scan camera 11 is under the triggering control of PLC, obtain the gray level image of LCD glass substrate, the optical axis of line scan camera is installed perpendicular to the plane of movement of LCD glass substrate, and the angle theta of the light plane that LED line source sends and the normal plane of travelling belt is 5 °-10 °;

Described bus control system comprises supervisory control comuter 25, Industrial Ethernet bus system 21 and PROFIBUS field bus system 22;

PLC133 drives travelling belt to move with uniform velocity to fixed-direction as controller control servomotor, thereby drive LCD glass substrate uniform motion, form stable relative motion with line scan camera 11, line scan camera obtains the gray level image of LCD glass substrate, and detection computations machine 12 detects identification to the gray level image of LCD glass substrate;

The processing result image that described detection computations machine 14 is exported is sent to the supervisory control comuter 21 that is positioned at workstation by Industrial Ethernet bus system;

Describedly by PROFIBUS field bus system 22, the travelling speed of travelling belt and detected glass substrate position parameter are sent to workstation supervisory control comuter 25 for the PLC133 that controls conveyer belt.

Linear CCD sensor in described line scan camera has 7450 pixels.

The detected glass substrate distance of distance of camera lens of described line scan camera is 400mm.

Many defects of vision detection method of large-scale LCD glass substrate on production line, on the electronics manufacturing line described in adopting, many defects of vision checkout equipment of large-scale LCD glass substrate, comprises the steps:

Step 1: the image of LCD glass substrate on Real-time Collection production line;

The signal that the glass substrate that feeler receives reaches predeterminated position is sent to PLC133, and PLC133 triggers line scan camera and carries out image acquisition;

Step 2: the LCD glass substrate image on the production line gathering is carried out to the pre-service of denoising and sharpening, improve picture quality;

Step 3: adopt kmeans clustering method to carry out the judgement of defect existence to pretreated LCD glass substrate image, if there is defect area in present image, enter step 4, otherwise, finish this defects detection, return to step 1 lower piece image is carried out to defects detection;

Step 4: defect area is carried out to mark, and extract defect characteristic;

Step 5: adopt the sorting technique of support vector machines, carry out the identification of defect classification according to the defect characteristic extracting, complete defects detection;

The process that described step 3 adopts kmeans clustering method to carry out the judgement of defect existence to pretreated LCD glass substrate image comprises following concrete steps:

1) first pretreated LCD glass substrate image is divided into several 8*8 rectangle sub-blocks g (x, y), press following two-dimensional dct transform formula respectively to each rectangle sub-block g (x, y) carry out discrete cosine transform, obtain each rectangle sub-block g (x, y) DCT coefficient C (u, v);

$C(u,v)=a\left(u\right)a\left(v\right)\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}g(x,y)\cos \left[\frac{(2x+1)\mathrm{u\π}}{2N}\right]\cos \left[\frac{(2y+1)\mathrm{v\π}}{2N}\right]$

Wherein, (x, y) represents the pixel coordinate in each rectangle sub-block, N=8, and u and v all get the integer between [0,7], $a\left(u\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (u=0)\\ \sqrt{2/N},& (u=\mathrm{1,2}...N-1)\end{array}\right.,$ $a\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (v=0)\\ \sqrt{2/N},& (v=\mathrm{1,2}...N-1)\end{array}\right.;$

2) calculate high frequency coefficient mean value mean_high and the low frequency coefficient mean value mean_low of each rectangle sub-block;

Mean_high=avg{C (u _h, v _h) | u _h, v _h∈ [0,2] }, mean_low=avg{C (u _l, v _l) | u _l, v _l∈ [3,7] }, wherein, agv represents the computing of averaging;

3) calculate high frequency coefficient mean value and the low frequency coefficient mean value ratio r atio of each rectangle sub-block in DCT territory: $\mathrm{ratio}=\frac{\mathrm{mean}\_\mathrm{high}}{\mathrm{mean}\_\mathrm{low}}\×100;$

4) the three-dimensional feature amount of utilizing kmeans clustering method to form each rectangle sub-block high frequency coefficient mean value mean_high and low frequency coefficient mean value mean_low and the ratio r atio of the two is classified, and concrete steps are as follows:

Step a: each rectangle sub-block high frequency coefficient mean value mean_high and low frequency coefficient mean value mean_low and the ratio r atio of the two are normalized respectively, obtain the three-dimensional feature vector after the normalization of each rectangle sub-block;

Step b: by the ratio histogram sequence after the normalization of all rectangle sub-blocks of sequential build from left to right, from top to bottom, and histogram sequence is carried out to first order difference calculating, to be less than ratio value that the difference value of the first setting threshold T1 is corresponding as histogram peak, calculate the ratio histogram peak number hist_peak after normalization; T1 gets the integer between 15-25;

Step c: cluster classification is set and counts cluster=hist_peak, and choose at random three-dimensional feature vector after the normalization of the rectangle sub-block initial value as cluster centre;

Steps d: utilize k-means clustering method that all rectangle sub-blocks are divided into cluster class, obtain cluster cluster centre simultaneously;

Step e: from cluster cluster centre, delete two minimum and maximum cluster centres of ratio value, using classification represented remaining cluster centre as defect classification;

Step f: calculate all total number count of rectangle sub-block that belong to each defect classification;

If count is less than Second Threshold T2, judge and in present image, do not have defect, otherwise, judge in present image and have defect; The value of T2 is the integer between 50 to 200;

In described step 4, defect area is carried out to mark, and extract defect characteristic, detailed process is as follows:

1) according to the cluster result of step 3, the pixel value that is judged as the each pixel in the rectangle sub-block of defect is all set to 255, the pixel value that is judged to each pixel in the rectangle sub-block of non-defect is all set to 0, obtain bianry image;

2) utilize opencv profile to search algorithm and extract defect rectangular area profile, and with the minimum boundary rectangle marking of defects region of each defect area;

3) obtain the outline position (Px, Py) of each defect area, area S, average gray L and circularity E;

Wherein, the outline position (Px, Py) of defect is the lower left corner point coordinate of defect part boundary rectangle;

Area S is with non-vanishing number of pixels in the minimum boundary rectangle of defect area;

Average gray L is the average gray of all pixels in the minimum circumscribed rectangular region of defect area;

Circularity E, p is the girth of the minimum boundary rectangle of defect area.

Circularity E value more approaches 1, represents that body form approaches circle; Its value is larger, represents that body form is more elongated;

In described step 5, adopt the sorting technique of support vector machines, carry out the identification of defect classification according to the defect characteristic extracting, complete defects detection, detailed process is as follows:

First gather the sample data that contains defect area, in sample data, comprise cut, scrape scrape along hole;

Calculate respectively the area S of the defect area in each sample data, average gray L and circularity E;

Utilize the area S of sample data, tri-proper vectors of average gray L and circularity E build two support vector machines sorters, first sorter is distinguished a class of hole one class and cut and scraping composition, and second sorter distinguished cut and scraping;

Utilize two sorters that build to carry out Classification and Identification to the defect characteristic of extract real-time, complete defects detection.

In described step 2, the LCD glass primary image on the production line gathering is carried out the pre-service of denoising and sharpening, refer to and adopt the window of 3*3 to carry out medium filtering denoising, obtain denoising image; Then the image after adopting Laplce's sharpening operator to denoising carries out sharpening processing.

Beneficial effect

Compared with prior art, the invention has the advantages that:

Simple and compact equipment structure of the present invention, with low cost, easy and simple to handle, degree of accuracy is high, the advantages such as detection speed is fast, realized LCD glass substrate automatic production line and detected in real time for the online of substrate quality, and there are 3.8 meters of above detection speeds of per minute, each line scan camera can detection width be 20 centimetres.Its detection efficiency is far above manual detection, and in testing process, can not introduce other pollutions.In addition, the present invention also adopts distributed control concept, respectively detects between station and does not interdepend, and has realized the real-time communication of each detection station computing machine and monitoring host computer simultaneously.Detection algorithm of the present invention can, in defects detection of the same type, have versatility widely, and transformation can be used for the quality testing of product of the same type a little.

Brief description of the drawings

Fig. 1 is that in the present invention, imaging detects machine system architecture schematic diagram;

Fig. 2 is image-forming principle schematic diagram;

Fig. 3 is apparatus of the present invention image-forming detecting system workflow diagram;

Fig. 4 is this true picture that installs accessed checked object;

Fig. 5 is field bus control system structural representation;

Fig. 6 is the workflow schematic diagram of LCD glass substrate defect detecting system;

Fig. 7 is the image treatment scheme schematic diagram of LCD glass substrate defect inspection method in the present invention;

Fig. 8 is dct transform 8*8 sub-block block plan;

Fig. 9 is the histogram of ratio;

Figure 10 is image segmentation figure;

Figure 11 is the detection effect schematic diagram of detection method of the present invention;

Label declaration:

11-line scan camera, 12-LED line source, 13-telecontrol equipment, 14-detection computations machine, 131-travelling belt, 132-feeler, 133-PLC, the mutual display unit of 134-PLC, 21-supervisory control comuter, 22-Industrial Ethernet bus system, 23-PROFIBUS field bus system.

Embodiment

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

For the deficiency of existing detection technique, the invention provides many defects of vision checkout equipment and the method for large-scale LCD glass substrate on a kind of production line.

On production line, many defects of vision checkout equipment of large-scale LCD glass substrate, comprises image-forming detecting system and bus control system;

Wherein, described image-forming detecting system at least comprises two line scan camera 11, one or more LED line source 12, transmission control device 13 and detection computations machines 14 of installing side by side;

Described transmission control device 13 comprises the mutual display unit 134 of travelling belt 131, feeler 132, PLC133 and PLC;

The mutual display unit 134 of described feeler 132, PLC133 and PLC is connected successively; Described travelling belt is controlled by PLC; Described feeler is installed on the fixed frame of travelling belt one side, and in the time that LCD glass substrate arrives predeterminated position with travelling belt, feeler sends a signal to PLC;

Described line scan camera 11 is connected with described detection computations machine, and is controlled by PLC133; Described line scan camera 11 is under the triggering control of PLC, obtain the gray level image of LCD glass substrate, the optical axis of line scan camera is installed perpendicular to the plane of movement of LCD glass substrate, and the angle theta of the light plane that LED line source sends and the normal plane of travelling belt is 5 °-10 °;

Described bus control system comprises supervisory control comuter 21, Industrial Ethernet bus system 23 and PROFIBUS field bus system 22;

PLC133 drives travelling belt to move with uniform velocity to fixed-direction as controller control servomotor, thereby drive LCD glass substrate uniform motion, form stable relative motion with line scan camera 11, line scan camera obtains the gray level image of LCD glass substrate, and detection computations machine 12 detects identification to the gray level image of LCD glass substrate;

The processing result image that described detection computations machine 14 is exported is sent to the supervisory control comuter 21 that is positioned at workstation by Industrial Ethernet bus system;

Described PLC133 is sent to workstation supervisory control comuter 21 by PROFIBUS field bus system 22 by the travelling speed of travelling belt and detected glass substrate position parameter.

Linear CCD sensor in described line scan camera has 7450 pixels.

The detected glass substrate distance of distance of camera lens of described line scan camera is 400mm.

In this example, adopt two line scan cameras that linear transducer camera lens is housed, be connected with computing machine by image pick-up card; Adopt PLC logic controller to export the trigger pip of linear scanning camera, control the speed of travelling belt, the parameters such as acceleration.

Referring to Fig. 1, be that apparatus of the present invention imaging detects machine system 1 structural representation, Fig. 2 is image-forming principle schematic diagram of the present invention, lens focus is 50mm, the tested substrate 400mm of distance of camera lens, horizontal field of view is 240mm.Line-scan digital camera is the linear CCD sensor that contains 7450 pixels, and travelling belt drives detected glass substrate can make image-forming detecting system obtain the minimum image that distorts in direction of motion with the speed operation of 3.8 meters of per minutes.

Its image-forming principle is as follows: in the time that directional light irradiates glass to glass surface with certain incident angle, if glass surface is smooth, zero defect, mirror-reflection occurs, and the camera directly over being positioned at receives less than reflection ray, and captured image-region should be black; In the time that light incident place defectiveness exists, will there is diffuse reflection, the camera directly over being positioned at can receive by the light of defect reflection, thereby forms the defect image that gray-scale value is higher under dark background.Line scan camera be arranged on detected glass substrate directly over, it is vertical with glass substrate direction of motion that it takes direction.Linear array LED light source is between camera and glass substrate, and LED line source light becomes the small angle theta of one 5 to 10 degree to install with plane of movement normal plane, and concrete setting angle is taking the picture rich in detail that can obtain object to be detected as criterion.In native system, below detected glass substrate, in 1 meter, cannot there is any object that blocks, no side can affect imaging effect.

Apparatus of the present invention image-forming detecting system workflow diagram as shown in Figure 3, transmission control device drives detected glass substrate towards fixing direction of motion uniform motion under the control of PLC, when conveyer belt during to fixing trigger position PLC trigger image-forming detecting system and start working, multiple line scan cameras carry out image acquisition to tested glass substrate, Site Detection computing machine carries out a series of processing realization to defective locations to image, the calculating of area and type, and testing result is passed to supervisory control comuter and add up, preserve and show.

The true picture of checked object glass substrate of the present invention as shown in Figure 4, comprising cut, scraping and three kinds of defects of hole.Wherein 2,4,5 is cut, and 3 is scraping, and 1,6 is hole.Cut is the elongated texture defect that glass substrate is higher and more sharp-pointed with some hardness in process of production object is produced as screw, nail etc. contact; Scraping be in production run glass substrate with surface with some shaggy objects as the defect being produced that rubs of the device in dust friction or clean link on travelling belt, hole is the defect because materials inequality or bubble produce in production technology.

Be the structural representation of field bus control system in the present invention as shown in Figure 5, based on the field bus control system of Industrial Ethernet and PROFIBUS bus, its Computer is divided by its function as Site Detection computing machine and workstation supervisory control comuter.Detection computations machine is processed the image of glass substrate at the scene, obtains defective data and image after treatment, by Industrial Ethernet, view data and result is sent to the supervisory control comuter that is positioned at workstation.By PROFIBUS bus, the parameters of the travelling speed of on-the-spot motor and spot sensor is sent to workstation simultaneously.The staff of workstation can browsing histories defective data and is watched realtime graphic and make on-the-spot Long-distance Control decision-making.

Shown in workflow Fig. 6 of LCD glass substrate defect detecting system, travelling belt drives detected glass substrate towards fixing direction of motion uniform motion, PLC (133) triggers image-forming detecting system in conveyer belt during to fixing trigger position and starts working, image-forming detecting system acquisition process cell picture passes to Site Detection computing machine, on-the-spot industrial computer passes to workstation detection computations machine by result by Industrial Ethernet bus system after image is processed, and staff's navigation process result of workstation also makes a policy.

Be the treatment scheme schematic diagram of described Site Detection computing machine image that described image-forming detecting system is collected as shown in Figure 7, wherein mainly comprise collection image, image pre-service, judgement fast, feature extraction, pattern-recognition five steps.Be described below respectively:

Step 1: the image of LCD glass substrate on Real-time Collection production line;

Travelling belt drives detected glass substrate towards fixing direction of motion uniform motion, form metastable relative motion with the first scanning camera being positioned at directly over it, the light plane that line source sends becomes a small angle theta to irradiate with plane of movement normal plane, utilize minute surface transmitting and irreflexive principle to glass substrate imaging.PLC triggers image-forming detecting system in conveyer belt during to fixing trigger position and starts working, and image-forming detecting system is adopted processing unit image and passed to on-the-spot industrial computer.

Step 2: image pre-service;

The pretreated object of image is to improve the quality of image, for follow-up image algorithm provides better input picture.Pre-service comprises denoising and two steps of sharpening.

Adopt the window of 3*3 to carry out medium filtering.Realize as follows, wherein f (t, f) represents after medium filtering the pixel value of position (i, j) in image, Z _kbe that in the 3*3 window centered by (i, j), source image data pixel value is pressed the array after order arrangement from small to large, Med represents to ask median operation.

f(i,j)=Med{Z _k|k=1,2,3,4,5,6,7,8,9}

And then adopt Laplce's sharpening operator and image array to carry out convolution, and result of calculation and f (t, f) to be added and subtracted and obtained sharpening result g (x, y) mutually, implementation procedure is as follows:

First use Laplace operator to carry out convolution algorithm to image,

${\▿}^{2}f(i,j)=[f(x+1,y)+f(x-1,y)+f(x,y+f)+f(x,y-1)]-4f(x,y)$

Then the result after convolution algorithm is combined with original digital image data,

After sharpening, can obtain the better image of visual effect.For defect characteristic below provides good source images.

Step 3: defect judges fast;

Native system has been realized the uninterrupted vision-based detection to glass substrate, but in reality, glass substrate is not that each section all can defectiveness, if each processing unit image that image-forming detecting system is gathered carry out same processing can; caused the waste to resource, this also there is no need.Therefore, the present invention has designed the method that whether has defect in the image that a quick judgement processing.

In image, the difference of defect part and non-defect background parts is mainly gray scale and grey scale change.Because various defects all do not have specific texture features, it is a kind of random texture image.So no matter defect is hole, cut or particle, aspect gray scale and grey scale change, all there is the same characteristic.And in discrete cosine transform coefficient, low frequency coefficient has just represented the gray average in image, HFS has represented the grey scale change in image.Reason based on such just, the proper vector that this programme adopts discrete cosine transform to calculate characterizes defect.Thereby feature space analysis is obtained to the conclusion whether defect exists.The method also has good effect for some fine defects in image.

This programme adopts and image is carried out to discrete cosine transform judges in image, whether there is trickle defect.First image is divided into the rectangular block of several 8*8 sizes, respectively it is carried out to discrete cosine transform by following two-dimensional dct transform formula, obtain DCT coefficient C (u, v):

$C(u,v)=a\left(u\right)a\left(v\right)\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}g(x,y)\cos \left[\frac{(2x+1)\mathrm{u\π}}{2N}\right]\cos \left[\frac{(2y+1)\mathrm{v\π}}{2N}\right]$

Wherein, (x, y) represents the pixel coordinate in each rectangle sub-block, N=8, and the span of u and v is all the integers between [0,7], $a\left(u\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (u=0)\\ \sqrt{2/N},& (u=\mathrm{1,2}...N-1)\end{array}\right.,$ $a\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (v=0)\\ \sqrt{2/N},& (v=\mathrm{1,2}...N-1)\end{array}\right.;$

For the rectangle sub-block of 8*8, this programme is divided into two regions, high-frequency region and low frequency region, and as shown in Figure 8, diagonal line hatches region is wherein low frequency region, vertical line shadow region is high-frequency region.

Definition high frequency coefficient mean value mean_high=agv{C (u _h, v _h) | u _h, v _h∈ [0,2] }, low frequency coefficient average mean_low=agv{C (u _v, u _t) | u _vv _l∈ [3,7] }, agv represents the computing of averaging.

Calculate high frequency coefficient mean value and low frequency coefficient mean value and the ratio r atio of the two in DCT territory, calculate as follows: $\mathrm{ratio}=\frac{\mathrm{mean}\_\mathrm{high}}{\mathrm{mean}\_\mathrm{low}}\×100;$

Respectively mean_high, mean_low and ratio are normalized with following formula, wherein y is the result after normalization, and x is the data before normalization.Max and min represent respectively maximal value and the minimum value of the data before normalization.

$y=\frac{255}{\max -\min }(x-\min );$

After normalization, obtain the three-dimensional feature vector (mean_high that several become to characterize 8*8 rectangle sub-block, mean_low, ratio), in this application, by the histogram analysis to ratio, the histogram of ratio as shown in Figure 9, from figure, can obtain cluster class counts cluster and should get hist_peak=3, by calculating the first order difference of histogram sequence of ratio, in this example, the first order difference value that histogram peak position is found in analysis by some examples more than 90% all below 20, get thus T1=20, difference value is less than to T1 pixel as histogram peak, thereby obtaining histogrammic peak value number is hist_peak=3, clustering algorithm classification is set and counts cluster=count_peak.Characterize the three-dimensional feature vector of 8*8 rectangle sub-block with mean_high, mean_low and ratio composition, choose at random cluster initial cluster center, utilize k-means clustering method that rectangle sub-block is divided into cluster class, in cluster result, the distribution of raito presents interval property, different classes of raito value is distributed in respectively (0,0.4), (0.4,3) and in (3 ,+∞) three disjoint intervals.In cluster result ratio value size in the middle of being positioned at the classification of (0.4,3) belong to defect area.Calculating belongs to the number count of the 8*8 rectangle sub-block of defect area.Think in image not have defect if count is greater than a certain threshold value T2, and then be for further processing, otherwise, judge in current rectangle blockage and have defect, will not carry out follow-up processing; The value of T2 is the integer between 50 to 200.

Step 4: target area feature extraction;

Through step 3 after above step 3, the value that is classified as 8*8 pixel in each rectangle sub-block of defect area in image is all set to 255, then image is carried out to inverse processing, the pixel value of defect part image becomes 255, and the pixel value of background parts becomes 0.Thereby obtain the bianry image after cutting apart, segmentation effect as shown in figure 10, in figure, defect profile is not used as an entirety preservation, so next used profile to search algorithm, each profile is done to as a whole preservation, so that follow-up calculation of parameter.Use opencv profile to search algorithm and carry out defect profile and search, by a series of summits composition profile sequence that represents profile.On original image, defect profile is marked with the boundary rectangle frame of profile, and according to from left to right, order is from top to bottom arranged label.If Figure 11 is the final effect of flaw labeling.

Calculate respectively the position (Px, Py) of each defect profile, area S, average gray L, circularity E parameter.

Wherein, the outline position (Px, Py) of defect is the lower left corner point coordinate of defect part boundary rectangle; Area S represents with number of pixels non-vanishing in bianry image, and average gray L is with the gray scale draw value representation of pixel that rectangular area is comprised, and circularity E calculates as follows: E=P ²/ 4 π S;

Wherein, P is girth, is obtained by the distance accumulation calculating between vertex sequence.S is area, represents with defect image pixel count.Circularity E value more approaches 1, represents that body form approaches circle; Its value is larger, represents that body form is more elongated.

Step 5: Classifcation of flaws;

Defect is divided into cut by the present invention, scraping, hole three classes.First to the area S calculating before, circularity E, tri-characteristic quantities of average gray L are normalized as follows, and wherein y is the result after normalization, x is the data before normalization, and max and min represent respectively maximal value and the minimum value of the data before normalization:

$y=\frac{x-\min }{\max -\min };$

Three-dimensional feature vector after normalization is classified as input feature vector amount.Here adopt the sorting technique of support vector machine (SVM).Construct 2 two classification svm classifier devices, realize three classes to data are divided.Definition cut, the label of scraping is respectively 1 and 2, and the label of hole classification is 3; Sorter C1 realizes the classification of 1 and 2 pair 3, and sorter C2 realizes the classification of 1 pair 2.

SVM method is to pass through function sample in the input space is mapped in high-dimensional feature space, and constructs optimal classification face in this feature space.And when construct optimum lineoid in feature space time, training algorithm is the dot product in applicable characteristic space only, so, if can find a function K to make like this, in higher dimensional space, in fact only need carry out inner product operation, even must not know conversion form.

Through a series of conversion, SVM method is following optimization problem.

$\min \frac{1}{2}{\mathrm{\α}}^T\mathrm{Q\α}-e^T\mathrm{\α}$

Y satisfies condition ^tα=0,0≤α _i≤ C, wherein i=1 ... n

In formula, C is penalty coefficient, and e is unit matrix, Q be n*n positive semidefinite matrix Q ( ⁱ, j)=yiy _jk (xi, x _j) this example employing RBF kernel function:

K (x _i, x _j)=exp (γ || x _i-x _j|| ²), γ >0, wherein γ is called radial basis radius.

Structure svm classifier device determines the process of C and γ.

Experiment sample is cut, and each 20 of scraping and hole three class defects, remember that three sample sets are respectively S1, and S2 and S3 respectively choose at random the sample set S0 that 10 sample composition labels are 0 from cut and scraping sample.Altogether four groups of experiment samples, choose in every group of sample 14 at random as training set, and 6 samples are as test set.Application the method for the invention is trained, and sample training data are as shown in the table:

Table 1 sample training data

Adopt the method for test to determine parameters C and γ.With sample set S0 and sample set S3 training classifier C1, sample set S1 and S2 training classifier C2.By test repeatedly, determine classifier parameters, the parameter area of experiment is C ∈ (60,300), parameters C=120 of γ ∈ (0.2,3.0) sorter C1, the result that γ=0.4 o'clock obtains is best, and test accuracy now can reach 100%; Parameters C=180 of sorter C2, the result that γ=1.2 o'clock obtain is best, test accuracy now can reach 83.33%.

The final process result example of this example detection scheme as shown in figure 11, wherein defect 2,4,5 are divided into cut, and 3 are divided into scraping, and 1,6 is divided into hole.

Claims (6)

1. many defects of vision checkout equipment of large-scale LCD glass substrate on production line, is characterized in that, comprises image-forming detecting system and bus control system;

Wherein, described image-forming detecting system at least comprises a set of imaging detection device, and described imaging detection device at least comprises two line scan camera (11), one or more LED line source (12), transmission control device (13) and the detection computations machines (14) installed side by side;

Described transmission control device (13) comprises travelling belt (131), feeler (132), PLC (133) and the mutual display unit of PLC (134);

Described feeler (132), PLC (133) and the mutual display unit of PLC (134) are connected successively; Described travelling belt is controlled by PLC; Described feeler is installed on the fixed frame of travelling belt one side, and in the time that LCD glass substrate arrives predeterminated position with travelling belt, feeler sends a signal to PLC;

Described line scan camera (11) is connected with described detection computations machine, and is controlled by PLC (133); Described line scan camera (11) is under the triggering control of PLC, obtain the gray level image of LCD glass substrate, the optical axis of line scan camera is installed perpendicular to the plane of movement of LCD glass substrate, and the angle theta of the light plane that LED line source sends and the normal plane of travelling belt is 5 °-10 °;

Described bus control system comprises supervisory control comuter (21), Industrial Ethernet bus system (23) and PROFIBUS field bus system (22);

The processing result image of described detection computations machine (14) output is sent to the supervisory control comuter (21) that is positioned at workstation by Industrial Ethernet bus system;

The described supervisory control comuter (21) that by PROFIBUS field bus system (22), the travelling speed of travelling belt and detected glass substrate position parameter is sent to workstation for controlling the PLC (133) of conveyer belt.

2. many defects of vision checkout equipment of large-scale LCD glass substrate on production line according to claim 1, is characterized in that, the linear CCD sensor in described line scan camera has 7450 pixels.

3. many defects of vision checkout equipment of large-scale LCD glass substrate on production line according to claim 1 and 2, is characterized in that, the detected glass substrate distance of distance of camera lens of described line scan camera is 400mm.

4. many defects of vision detection method of large-scale LCD glass substrate on production line, is characterized in that, adopts many defects of vision checkout equipment of large-scale LCD glass substrate on the electronics manufacturing line described in claim 1 or 2, comprises the steps:

Step 1: the image of LCD glass substrate on Real-time Collection production line;

The signal that the glass substrate that feeler receives reaches predeterminated position is sent to PLC (133), and PLC (133) triggers line scan camera and carries out image acquisition;

Step 2: the LCD glass substrate image on the production line gathering is carried out to the pre-service of denoising and sharpening;

Step 3: adopt kmeans clustering method to carry out the judgement of defect existence to pretreated LCD glass substrate image, if there is defect area in present image, enter step 4, otherwise, finish this defects detection, return to step 1 lower piece image is carried out to defects detection;

Step 4: defect area is carried out to mark, and extract defect characteristic;

Step 5: adopt the sorting technique of support vector machines, carry out the identification of defect classification according to the defect characteristic extracting, complete defects detection;

The process that described step 3 adopts kmeans clustering method to carry out the judgement of defect existence to pretreated LCD glass substrate image comprises following concrete steps:

1) first pretreated LCD glass substrate image is divided into several 8*8 rectangle sub-blocks g (x, y), press following two-dimensional dct transform formula respectively to each rectangle sub-block g (x, y) carry out discrete cosine transform, obtain each rectangle sub-block g (x, y) DCT coefficient C (u, v);

$C(u,v)=a\left(u\right)a\left(v\right)\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}g(x,y)\cos \left[\frac{(2x+1)\mathrm{u\π}}{2N}\right]\cos \left[\frac{(2y+1)\mathrm{v\π}}{2N}\right]$

Wherein, (x, y) represents the pixel coordinate in each rectangle sub-block, N=8, and u and v all get the integer between [0,7], $a\left(u\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (u=0)\\ \sqrt{2/N},& (u=\mathrm{1,2}...N-1)\end{array}\right.,$ $a\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (v=0)\\ \sqrt{2/N},& (v=\mathrm{1,2}...N-1)\end{array}\right.;$

2) calculate high frequency coefficient mean value mean_high and the low frequency coefficient mean value mean_low of each rectangle sub-block;

Mean_high=avg{C (u _h, v _h) | u _h, v _h∈ [0,2] }, mean_low=avg{C (u _l, v _l) | u _l, v _l∈ [3,7] }, wherein, agv represents the computing of averaging;

3) calculate high frequency coefficient mean value and the low frequency coefficient mean value ratio r atio of each rectangle sub-block in DCT territory: $\mathrm{ratio}=\frac{\mathrm{mean}\_\mathrm{high}}{\mathrm{mean}\_\mathrm{low}}\×100;$

4) the three-dimensional feature amount of utilizing kmeans clustering method to form each rectangle sub-block high frequency coefficient mean value mean_high and low frequency coefficient mean value mean_low and the ratio r atio of the two is classified, and concrete steps are as follows:

Step a: each rectangle sub-block high frequency coefficient mean value mean_high and low frequency coefficient mean value mean_low and the ratio r atio of the two are normalized respectively, obtain the three-dimensional feature vector after the normalization of each rectangle sub-block;

Step b: by the ratio histogram sequence after the normalization of all rectangle sub-blocks of sequential build from left to right, from top to bottom, and histogram sequence is carried out to first order difference calculating, to be less than ratio value that the difference value of the first setting threshold T1 is corresponding as histogram peak, calculate the ratio histogram peak number hist_peak after normalization; T1 gets the integer between 15-25;

Step c: cluster classification is set and counts cluster=hist_peak, and choose at random three-dimensional feature vector after the normalization of the rectangle sub-block initial value as cluster centre;

Steps d: utilize k-means clustering method that all rectangle sub-blocks are divided into cluster class, obtain cluster cluster centre simultaneously;

Step e: from cluster cluster centre, delete two minimum and maximum cluster centres of ratio value, using classification represented remaining cluster centre as defect classification;

Step f: calculate all total number count of rectangle sub-block that belong to each defect classification;

If count is less than Second Threshold T2, judge and in present image, do not have defect, otherwise, judge in present image and have defect; The value of T2 is the integer between 50 to 200;

In described step 4, defect area is carried out to mark, and extract defect characteristic, detailed process is as follows:

1) according to the cluster result of step 3, the pixel value that is judged as the each pixel in the rectangle sub-block of defect is all set to 255, the pixel value that is judged to each pixel in the rectangle sub-block of non-defect is all set to 0, obtain bianry image;

2) utilize opencv profile to search algorithm and extract defect rectangular area profile, and with the minimum boundary rectangle marking of defects region of each defect area;

3) obtain the outline position (Px, Py) of each defect area, area S, average gray L and circularity E;

Wherein, the outline position (Px, Py) of defect is the lower left corner point coordinate of defect part boundary rectangle;

Area S is with non-vanishing number of pixels in the minimum boundary rectangle of defect area;

Average gray L is the average gray of all pixels in the minimum circumscribed rectangular region of defect area;

Circularity E, p is the girth of the minimum boundary rectangle of defect area.

5. many defects of vision detection method of large-scale LCD glass substrate on production line according to claim 4, it is characterized in that, in described step 5, adopt the sorting technique of support vector machines, carry out the identification of defect classification according to the defect characteristic extracting, complete defects detection, detailed process is as follows:

First gather the sample data that contains defect area, in sample data, comprise cut, scrape scrape along hole;

Calculate respectively the area S of the defect area in each sample data, average gray L and circularity E;

Utilize the area S of sample data, tri-proper vectors of average gray L and circularity E build two support vector machines sorters, first sorter is distinguished a class of hole one class and cut and scraping composition, and second sorter distinguished cut and scraping;

Utilize two sorters that build to carry out Classification and Identification to the defect characteristic of extract real-time, complete defects detection.

6. many defects of vision detection method of large-scale LCD glass substrate on production line according to claim 5, it is characterized in that, in described step 2, the LCD glass primary image on the production line gathering is carried out the pre-service of denoising and sharpening, refer to and adopt the window of 3*3 to carry out medium filtering denoising, obtain denoising image; Then the image after adopting Laplce's sharpening operator to denoising carries out sharpening processing.