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

Multi-visual defect detection equipment and method for large-size LCD glass substrate on production line

technical field

The invention belongs to the field of visual inspection equipment on an electronic manufacturing production line, and in particular relates to a multi-visual defect inspection equipment and method for a large-size LCD glass substrate on a production line.

Background technique

With the development of electronic technology, the price of consumer electronic products is getting lower and lower, and people's demand for electronic products continues to expand, which also drives the development of LCD glass substrate manufacturing. However, in order to pursue a more perfect audio-visual enjoyment, the mainstream sizes of LCD monitors, LCD TVs and mobile terminals are constantly expanding, and even some super-sized LCD screens will appear. Therefore, the production of large-size LCD glass substrates will be a turning point for the rapid development of the industry in the future. However, the quality inspection technology of large-size LCD glass substrates is also a bottleneck restricting the development of the industry.

Although most of the production process of LCD glass substrates is completed in a clean room with high cleanliness, some defects will inevitably appear on the LCD glass substrates, which will cause the integrated circuits printed on the substrates to fail. work properly, resulting in blemishes on the LCD display. There are many reasons for these defects, and defects may occur in all aspects of its manufacture. These defects mainly include scratches, scratches, holes, particles, air bubbles, inclusions, etc. Some defects are so subtle that it is difficult for the human eye to observe them.

In the high-speed automated electronic manufacturing production line, how to quickly and accurately inspect the quality of LCD glass substrates is an important technical problem directly related to product quality. At present, the quality inspection of most production lines mainly relies on manual methods, that is, manual inspection and detection of defects. The defects of manual detection are: 1. The detection speed is slow, the efficiency is low, and it cannot meet the needs of high-speed automatic production lines; 2. The detection accuracy is low, the detection quality is affected by human factors, and the probability of false detection and missed detection is high; 3. The labor intensity of workers 4. It is inconvenient to save and query the test data, and it is not easy to manage.

Contents of the invention

Aiming at the deficiencies of the existing detection technology, the invention provides a multi-visual defect detection device and method for large-size LCD glass substrates on a production line.

A multi-vision defect detection equipment for large-size LCD glass substrates on a production line, including an imaging detection system and a bus control system;

Wherein, the imaging detection system includes at least two line scan cameras 11 installed side by side, one or more LED line light sources 12, a transmission control device 13 and a detection computer 14;

The transmission control device 13 includes a conveyor belt 131, a contact sensor 132, a PLC 133 and a PLC interactive display unit 134;

The touch sensor 132, the PLC 133 and the PLC interactive display unit 134 are connected in sequence; the conveyor belt is controlled by the PLC; the touch sensor is installed on a fixed frame on one side of the conveyor belt, and when the LCD glass substrate arrives at a preset position with the conveyor belt, The contact sensor sends a signal to the PLC;

The line scan camera 11 is connected with the detection computer and is controlled by PLC133; the line scan camera 11 obtains the grayscale image of the LCD glass substrate under the trigger control of the PLC, and the optical axis of the line scan camera is perpendicular to the LCD The movement plane of the glass substrate is installed, and the angle θ between the light plane emitted by the LED line light source and the normal plane of the conveyor belt is 5°-10°;

The bus control system includes a monitoring computer 25, an industrial Ethernet bus system 21 and a PROFIBUS field bus system 22;

PLC133 as a controller controls the servo motor to drive the conveyor belt to move at a constant speed in a fixed direction, thereby driving the LCD glass substrate to move at a constant speed, and forms a stable relative motion with the line scan camera 11. The line scan camera obtains the grayscale image of the LCD glass substrate, and the detection computer 12 Detect and recognize the grayscale image of the LCD glass substrate;

The image processing result output by the detection computer 14 is transmitted to the monitoring computer 21 located at the workstation through the industrial Ethernet bus system;

The PLC 133 for controlling the movement of the conveyor belt transmits the running speed of the conveyor belt and the position parameters of the detected glass substrate to the monitoring computer 25 of the workstation through the PROFIBUS field bus system 22 .

The line CCD sensor in the line scan camera has 7450 pixels.

The distance between the lens of the line scan camera and the glass substrate to be detected is 400mm.

A multi-visual defect detection method for large-size LCD glass substrates on a production line, using the multi-visual defect detection equipment for large-size LCD glass substrates on the electronic manufacturing production line, comprising the following steps:

Step 1: Collect images of LCD glass substrates on the production line in real time;

The signal received by the contact sensor that the glass substrate has reached the preset position is transmitted to PLC133, and PLC133 triggers the line scan camera for image acquisition;

Step 2: Perform denoising and sharpening preprocessing on the collected LCD glass substrate images on the production line to improve image quality;

Step 3: Use the kmeans clustering method to judge the existence of defects on the preprocessed LCD glass substrate image. If there is a defect area in the current image, go to step 4. Otherwise, end this defect detection and return to step 1 for the next image. Image for defect detection;

Step 4: mark the defect area and extract defect features;

Step 5: Use the classification method of support vector machine SVM to identify the defect category according to the extracted defect features, and complete the defect detection;

In step 3, the process of using the kmeans clustering method to judge the existence of defects on the preprocessed LCD glass substrate image includes the following specific steps:

1) First, divide the preprocessed LCD glass substrate image into several 8*8 rectangular sub-blocks g(x, y), and discretize each rectangular sub-block g(x, y) according to the following two-dimensional DCT transformation formula Cosine transform DCT to obtain the DCT coefficient C(u,v) of each rectangular sub-block g(x,y);

$\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; cos</span>}<span\; class="notranslate">\; [</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u\&pi;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ]</span>\mathrm{<span\; class="notranslate">\; cos</span>}<span\; class="notranslate">\; [</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; v\&pi;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ]</span>$

Among them, (x, y) represents the pixel point coordinates in each rectangular sub-block, N=8, u and v are integers between [0,7], $a\left(u\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (u=0)\\ \sqrt{2/N},& (u=\mathrm{1,2}...N-1)\end{array}\right.,$ $a\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (v=0)\\ \sqrt{2/N},& (v=\mathrm{1,2}...N-1)\end{array}\right.;$

2) Calculate the mean_high of high-frequency coefficients and mean_low of low-frequency coefficients of each rectangular 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 average operation;

3) Calculate the ratio ratio of the average value of high-frequency coefficients to the average value of low-frequency coefficients of each rectangular sub-block in the DCT domain: $\mathrm{ratio}=\frac{\mathrm{mean}\_\mathrm{high}}{\mathrm{mean}\_\mathrm{low}}\&times;100;$

4) Use the kmeans clustering method to classify the three-dimensional feature quantity composed of the average high-frequency coefficient mean_high and low-frequency coefficient mean_low of each rectangular sub-block and the ratio ratio of the two. The specific steps are as follows:

Step a: Perform normalization processing on the average value mean_high of high-frequency coefficients of each rectangular sub-block, the average value mean_low of low-frequency coefficients and the ratio ratio of the two, and obtain the normalized three-dimensional feature vector of each rectangular sub-block;

Step b: Construct the normalized ratio histogram sequence of all rectangular sub-blocks in order from left to right and from top to bottom, and perform a first-order difference calculation on the histogram sequence to make it smaller than the first set threshold T1 The ratio value corresponding to the difference value is used as the peak value of the histogram, and the normalized ratio histogram peak number hist_peak is calculated; T1 takes an integer between 15-25;

Step c: Set the number of clustering categories cluster=hist_peak, and randomly select a normalized three-dimensional feature vector of a rectangular sub-block as the initial value of the cluster center;

Step d: use the k-means clustering method to divide all rectangular sub-blocks into cluster classes, and obtain cluster cluster centers at the same time;

Step e: from the cluster cluster centers, delete the two cluster centers with the largest and minimum ratio values, and use the category represented by the remaining cluster centers as the defect category;

Step f: Calculate the total number count of all rectangular sub-blocks belonging to each defect category;

If count is less than the second threshold T2, it is determined that there is no defect in the current image, otherwise, it is determined that there is a defect in the current image; the value of T2 is an integer between 50 and 200;

In the step 4, the defect area is marked, and the defect feature is extracted, and the specific process is as follows:

1) According to the clustering result of step 3, set the pixel value of each pixel in the rectangular sub-block judged as defective to 255, and set the pixel value of each pixel in the non-defective rectangular sub-block The values are all set to 0 to obtain a binary image;

2) Use the opencv contour search algorithm to extract the contour of the defect rectangular area, and mark the defect area with the smallest circumscribed rectangle of each defect area;

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

Among them, the contour position (Px, Py) of the defect is the coordinate of the lower left corner point of the circumscribed rectangle of the defect part;

The area S is the number of non-zero pixels in the smallest circumscribed rectangle of the defective area;

The average gray level L is the average gray level of all pixels in the smallest circumscribed rectangular area of the defective area;

circularity E, P is the perimeter of the smallest circumscribed rectangle of the defect area.

The closer the circularity E value is to 1, the closer the object shape is to a circle; the larger the value, the thinner the object shape;

In the step 5, the support vector machine (SVM) classification method is adopted, and the defect category is identified according to the extracted defect features, and the defect detection is completed. The specific process is as follows:

First collect the sample data of the defect area, the sample data includes scratches, scratches and holes;

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

Using the area S of the sample data, the average gray level L and the circularity E three eigenvectors to construct two support vector machine SVM classifiers, the first classifier is used to classify holes and scratches and scrapes. distinction, the second classifier differentiates between scratches and scratches;

The two classifiers built are used to classify and identify the defect features extracted in real time to complete the defect detection.

In the step 2, the preprocessing of denoising and sharpening the LCD glass basic image collected on the production line refers to the use of a 3*3 window to perform median filter denoising to obtain a denoising image; then use Lapla The sharpening operator sharpens the denoised image.

Beneficial effect

Compared with the prior art, the present invention has the advantages of:

The equipment of the present invention has the advantages of simple and compact structure, low cost, easy operation, high precision, fast detection speed, etc., and realizes the online real-time detection of substrate quality in the LCD glass substrate automatic production line, and has a detection speed of more than 3.8 meters per minute. , the detectable width of each line scan camera is 20 cm. Its detection efficiency is much higher than that of manual detection, and no other pollution will be introduced during the detection process. In addition, the present invention also adopts the concept of distributed control, and each detection station does not depend on each other, and at the same time realizes the real-time communication between the computers of each detection station and the monitoring host. The detection algorithm of the invention can be used in the same type of defect detection, has wide versatility, and can be used for the quality detection of the same type of products with a little modification.

Description of drawings

Fig. 1 is a schematic structural diagram of the imaging detection machine system in the present invention;

Figure 2 is a schematic diagram of the imaging principle;

Fig. 3 is a work flow chart of the imaging detection system of the device of the present invention;

Fig. 4 is the real image of the detected object obtained by the device;

Fig. 5 is a structural schematic diagram of the fieldbus control system;

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

Fig. 7 is a schematic diagram of the image processing flow of the LCD glass substrate defect detection method in the present invention;

Fig. 8 is a DCT transform 8*8 sub-block partition diagram;

Figure 9 is a histogram of ratio;

Figure 10 is an image segmentation effect diagram;

Fig. 11 is a schematic diagram of the detection effect of the detection method of the present invention;

Label description:

11-line scan camera, 12-LED line light source, 13-motion device, 14-detection computer, 131-conveyor belt, 132-contact sensor, 133-PLC, 134-PLC interactive display unit, 21-monitoring computer, 22-industry Ethernet bus system, 23-PROFIBUS field bus system.

Detailed ways

The present invention will be described in further detail below in conjunction with specific examples and accompanying drawings.

Aiming at the deficiencies of the existing detection technology, the invention provides a multi-visual defect detection device and method for large-size LCD glass substrates on a production line.

A multi-vision defect detection equipment for large-size LCD glass substrates on a production line, including an imaging detection system and a bus control system;

Wherein, the imaging detection system includes at least two line scan cameras 11 installed side by side, one or more LED line light sources 12, a transmission control device 13 and a detection computer 14;

The transmission control device 13 includes a conveyor belt 131, a contact sensor 132, a PLC 133 and a PLC interactive display unit 134;

The touch sensor 132, the PLC 133 and the PLC interactive display unit 134 are connected in sequence; the conveyor belt is controlled by the PLC; the touch sensor is installed on a fixed frame on one side of the conveyor belt, and when the LCD glass substrate arrives at a preset position with the conveyor belt, The contact sensor sends a signal to the PLC;

The line scan camera 11 is connected with the detection computer and is controlled by PLC133; the line scan camera 11 obtains the grayscale image of the LCD glass substrate under the trigger control of the PLC, and the optical axis of the line scan camera is perpendicular to the LCD The movement plane of the glass substrate is installed, and the angle θ between the light plane emitted by the LED line light source and the normal plane of the conveyor belt is 5°-10°;

The bus control system includes a monitoring computer 21, an industrial Ethernet bus system 23 and a PROFIBUS field bus system 22;

PLC133 as a controller controls the servo motor to drive the conveyor belt to move at a constant speed in a fixed direction, thereby driving the LCD glass substrate to move at a constant speed, and forms a stable relative motion with the line scan camera 11. The line scan camera obtains the grayscale image of the LCD glass substrate, and the detection computer 12 Detect and recognize the grayscale image of the LCD glass substrate;

The image processing result output by the detection computer 14 is transmitted to the monitoring computer 21 located at the workstation through the industrial Ethernet bus system;

The PLC 133 transmits the running speed of the conveyor belt and the position parameters of the detected glass substrate to the workstation monitoring computer 21 through the PROFIBUS field bus system 22 .

The line CCD sensor in the line scan camera has 7450 pixels.

The distance between the lens of the line scan camera and the glass substrate to be detected is 400 mm.

In this example, two line-scan cameras equipped with linear sensor lenses are used, which are connected to the computer through an image acquisition card; a PLC logic controller is used to output the trigger signal of the line-scan cameras to control the speed, acceleration and other parameters of the conveyor belt.

Referring to Fig. 1, it is a structural schematic diagram of the imaging inspection machine system 1 of the present invention, and Fig. 2 is a schematic diagram of the imaging principle of the present invention, the focal length of the lens is 50 mm, the distance between the lens and the substrate to be inspected is 400 mm, and the horizontal field of view is 240 mm. The line-scan camera is a line-scan CCD sensor with 7450 pixels, and the conveyor belt drives the inspected glass substrate to run at a speed of 3.8 meters per minute, so that the imaging inspection system can obtain images with minimal distortion in the direction of motion.

The imaging principle is as follows: when parallel light irradiates the glass surface at a certain incident angle, if the glass surface is smooth and free of defects, specular reflection occurs, and the camera directly above cannot receive the reflected light, and the captured image area It should be black; when there is a defect where the light is incident, diffuse reflection will occur, and the camera directly above will receive the light reflected by the defect, thus forming a defect image with a high gray value under a dark background. The line scan camera is installed directly above the glass substrate to be inspected, and its shooting direction is perpendicular to the moving direction of the glass substrate. The linear array LED light source is located between the camera and the glass substrate. The light from the LED line light source and the normal plane of the motion plane are installed at a small angle θ of 5 to 10 degrees. The specific installation angle is based on the ability to obtain a clear image of the detected object. In this system, there must not be any blocking objects within 1 meter below the detected glass substrate, otherwise it will affect the imaging effect.

As shown in Figure 3 is the work flow chart of the imaging detection system of the present invention. The transmission control device drives the detected glass substrate to move at a constant speed in a fixed direction of motion under the control of the PLC. When the conveyor belt moves to a fixed trigger position, the PLC triggers the imaging. The inspection system starts to work, multiple line scan cameras collect images of the inspected glass substrate, and the on-site inspection computer performs a series of processing on the images to realize the calculation of defect location, area and type, and transmits the inspection results to the monitoring computer for statistics , save and display.

As shown in FIG. 4 , it is a real image of the inspected glass substrate of the present invention, which includes three kinds of defects: scratches, scratches and holes. Among them, 2,4,5 are scratches, 3 are scratches, and 1,6 are holes. Scratches are elongated texture defects caused by the contact between the glass substrate and some hard and sharp objects such as screws and nails during the production process; scratches are the glass substrate and the surface and some rough objects during the production process. Such as defects caused by dust friction on the conveyor belt or device friction in the cleaning process, holes are defects caused by uneven materials or air bubbles in the production process.

As shown in Figure 5, it is a structural representation of the field bus control system in the present invention, based on the field bus control system of industrial Ethernet and PROFIBUS bus, wherein the computer is divided into a field detection computer and a workstation monitoring computer according to its function. The inspection computer processes the image of the glass substrate on site to obtain defect data and processed images, and transmits the image data and processing results to the monitoring computer at the workstation through the industrial Ethernet. At the same time, the running speed of the on-site motor and various parameters of the on-site sensors are transmitted to the workstation through the PROFIBUS bus. Workstation staff can browse historical defect data and watch real-time images and make remote control decisions on site.

The working flow diagram of the LCD glass substrate defect detection system is shown in 6. The conveyor belt drives the detected glass substrate to move at a constant speed in a fixed direction of motion. When the conveyor belt moves to a fixed trigger position, the PLC (133) triggers the imaging detection system to start working. The detection system collects and processes the image of the processing unit and transmits it to the on-site detection computer. After the on-site industrial computer processes the image, the processing result is transmitted to the workstation detection computer through the industrial Ethernet bus system. The staff at the workstation browses the processing results and makes decisions.

As shown in Fig. 7, it is a schematic diagram of the processing flow of the image collected by the imaging detection system by the on-site inspection computer, which mainly includes five steps of image acquisition, image preprocessing, quick judgment, feature extraction, and pattern recognition. They are described as follows:

Step 1: Collect images of LCD glass substrates on the production line in real time;

The conveyor belt drives the glass substrate to be inspected to move at a constant speed in a fixed direction, and forms a relatively stable relative motion with the first-scanning camera directly above it. The principles of specular emission and diffuse reflection image glass substrates. When the conveyor belt moves to a fixed trigger position, the PLC triggers the imaging detection system to start working, and the imaging detection system collects the image of the processing unit and transmits it to the on-site industrial computer.

Step 2: Image preprocessing;

The purpose of image preprocessing is to improve the quality of images and provide better input images for subsequent image algorithms. Preprocessing includes two steps of denoising and sharpening.

Use a 3*3 window for median filtering. Realize according to the following formula, where f(t, f) represents the pixel value of position (i, j) in the image after median filtering, and Z _k is the source image data pixel in the 3*3 window centered on (i, j) An array of values arranged in ascending order, and Med represents the median operation.

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

Then, the Laplacian sharpening operator is used to convolve with the image matrix, and the calculation result is added and subtracted with f(t, f) to obtain the sharpening result g(x, y). The implementation process is as follows:

First, the image is convolved using the Laplacian operator,

${<span\; class="notranslate">\; \&dtri;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 4</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>$

Then combine the result of the convolution operation with the original image data,

Images with better visual effects can be obtained after sharpening. Provide good source images for subsequent defect feature extraction.

Step 3: Defect quick judgment;

This system realizes the uninterrupted visual inspection of the glass substrate, but in reality, not every section of the glass substrate will have defects. If the image of each processing unit collected by the imaging inspection system is processed in the same way, it will cause a waste of resources. waste, which is also unnecessary. Therefore, the present invention designs a method for quickly judging whether there is a defect in the image being processed.

The difference between the defect part and the non-defect background part in the image mainly lies in the gray scale and the gray scale change. Since various defects have no specific texture characteristics, it is an irregular texture image. So whether the defect is a hole, a scratch, or a particle, it has the same characteristics in terms of grayscale and grayscale variation. In the discrete cosine transform coefficients, the low-frequency coefficients represent the gray mean value in the image, and the high-frequency part represents the gray-scale change in the image. It is for this reason that this program uses the eigenvectors calculated by the discrete cosine transform to characterize the defects. The feature space is analyzed to get the conclusion of whether the defect exists. This method also has a good effect on some subtle defects in the image.

This scheme adopts discrete cosine transform DCT to the image to judge whether there are subtle defects in the image. First, the image is divided into several 8*8 rectangular blocks, and the discrete cosine transform DCT is performed on them according to the following two-dimensional DCT transformation formula to obtain the DCT coefficient C(u,v):

$\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; cos</span>}<span\; class="notranslate">\; [</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u\&pi;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ]</span>\mathrm{<span\; class="notranslate">\; cos</span>}<span\; class="notranslate">\; [</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; v\&pi;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ]</span>$

Among them, (x, y) represents the pixel coordinates in each rectangular sub-block, N=8, and the value ranges of u and v are 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 8*8 rectangular sub-block, this scheme divides it into two areas, the high-frequency area and the low-frequency area, as shown in Figure 8, where the diagonally shaded area is the low-frequency area, and the vertical line shaded area is the high-frequency area .

Define the average value of high-frequency coefficients mean_high=agv{C(u _h , v _h )|u _h , v _h ∈[0,2]}, the mean value of low-frequency coefficients mean_low=agv{C(u _v , u _t )|u _V v _l ∈ [3, 7]}, agv represents the mean value operation.

Calculate the average value of the high-frequency coefficients and the average value of the low-frequency coefficients in the DCT domain and the ratio of the two, and calculate according to the following formula: $\mathrm{ratio}=\frac{\mathrm{mean}\_\mathrm{high}}{\mathrm{mean}\_\mathrm{low}}\&times;100;$

Use the following formulas to normalize mean_high, mean_low, and ratio respectively, where y is the result after normalization, and x is the data before normalization. max and min represent the maximum and minimum values of the data before normalization, respectively.

$\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 255</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

After normalization, several three-dimensional feature vectors (mean_high, mean_low, ratio) representing 8*8 rectangular sub-blocks are obtained. In this application, through the analysis of the histogram of ratio, the histogram of ratio is shown in Figure 9. From In the figure, the number of clusters can be obtained. Cluster should take hist_peak=3. By calculating the first-order difference of the histogram sequence of ratio, in this example, through the analysis of several examples, it is found that the first-order difference value of the peak position of the histogram is more than 90% All are below 20, so T1=20 is taken, and the difference value is less than T1 pixels as the peak of the histogram, so that the number of peaks in the histogram is hist_peak=3, and the number of clustering algorithm categories is set to cluster=count_peak. Use mean_high, mean_low and ratio to form a three-dimensional feature vector representing 8*8 rectangular sub-blocks, randomly select cluster initial cluster centers, use the k-means clustering method to divide rectangular sub-blocks into clusters, and the raito in the clustering results The distribution is interval, and the raito values of different categories are distributed in three disjoint intervals (0,0.4), (0.4,3) and (3,+∞). In the clustering results, the category whose ratio value is in the middle (0.4,3) belongs to the defect area. Calculate the number count of 8*8 rectangular sub-blocks belonging to the defect area. If the count is greater than a certain threshold T2, it is considered that there is no defect in the image, and then proceed to the next step of processing; otherwise, it is determined that there is a defect in the current small rectangular square, and no subsequent processing will be performed; the value of T2 is between 50 and 200 an integer of .

Step 4: Target area feature extraction;

After step 3 and the above step 3, all the values of 8*8 pixels in each rectangular sub-block classified as defective areas in the image are set to 255, and then the image is reversed, that is, the defective part The pixel value of the image becomes 255, and the pixel value of the background part becomes 0. Thus, the segmented binary image is obtained. The segmentation effect is shown in Figure 10. The defect contours in the figure are not saved as a whole, so the contour search algorithm is used to save each contour as a whole, so that the subsequent parameter calculation. Use the opencv contour search algorithm to find the defect contour, and use a series of vertices representing the contour to form a contour sequence. Use the circumscribed rectangular frame of the contour to mark the defect contour on the original image, and arrange the labels in the order from left to right and from top to bottom. Figure 11 is the final effect of the defect mark.

The position (Px, Py), area S, average gray level L, and circularity E parameters of each defect contour are calculated separately.

Among them, the contour position (Px, Py) of the defect is the coordinates of the lower left corner of the rectangle circumscribing the defect; the area S is represented by the number of pixels that are not zero in the binary image, and the average gray level L is represented by the gray level of the pixels contained in the rectangular area The tie value indicates that the circularity E is calculated according to the following formula: E=P ² /4πS;

Among them, P is the perimeter, which is calculated by accumulating the distances between the vertex sequences. S is the area, represented by the number of pixels in the defect image. The closer the value of circularity E is to 1, the closer the shape of the object is to a circle; the larger the value, the thinner the shape of the object.

Step 5: Defect type identification;

The present invention classifies defects into three types: scratches, scratches, and holes. First, normalize the previously calculated three feature quantities of area S, circularity E, and average gray level L according to the following formula, where y is the result after normalization, x is the data before normalization, and max and min represent the maximum and minimum values of the data before normalization, respectively:

$\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; ;</span>$

The normalized three-dimensional feature vector is used as the input feature quantity for classification. The classification method of support vector machine (SVM) is adopted here. Construct two two-category SVM classifiers to realize the three-category division of the data. Define scratches, the labels of the scratches are 1 and 2, and the label of the hole category is 3; the classifier C1 realizes the classification of 1 and 2 to 3, and the classifier C2 realizes the classification of 1 to 2.

The SVM method is through the function The samples in the input space are mapped to the high-dimensional feature space, and the optimal classification surface is constructed in the feature space. When constructing the optimal hyperplane in the feature space, the training algorithm only applies the dot product in the feature space, that is Therefore, if we can find a function K such that In this way, in a high-dimensional space, only the inner product operation is actually needed, and the transformation does not even need to be known form.

After a series of transformations, the SVM method is the following optimization problem.

$\mathrm{<span\; class="notranslate">\; min</span>}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; Q\&alpha;</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}$

Satisfy the condition y ^T α=0,0≤α _i ≤C, where i=1,…n

In the formula, C is the penalty coefficient, e is the identity matrix, and Q is the positive semi-definite matrix of n*n Q( ⁱ ,j)=yiy _j K(xi,x _j ) In this example, the RBF kernel function is used:

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

Constructing an SVM classifier is the process of determining C and γ.

The experimental samples are 20 scratches, 20 scratches and 20 holes. The three sample sets are respectively S1, S2 and S3. 10 samples are randomly selected from the scratches and scratches to form a sample with a label of 0. Set S0. There are four sets of experimental samples in total, 14 samples in each set are randomly selected as the training set, and 6 samples are used as the test set. Apply the method described in the present invention to train, and sample training data is as shown in the following table:

Table 1 Sample training data

Experimental methods are used to determine the parameters C and γ. The classifier C1 is trained with the sample set S0 and the sample set S3, and the classifier C2 is trained with the sample sets S1 and S2. Through repeated tests, the classifier parameters are determined. The experimental parameter range is C ∈ (60, 300), γ ∈ (0.2, 3.0). The parameter C=120, γ=0.4 of the classifier C1 is the best result. , the test accuracy rate can reach 100% at this time; the result obtained when the parameter C=180 and γ=1.2 of the classifier C2 is the best, and the test accuracy rate at this time can reach 83.33%.

As shown in Figure 11 is an example of the final processing results of the detection scheme in this example, in which defects 2, 4, and 5 are classified as scratches, 3 are classified as scratches, and defects 1 and 6 are classified as holes.

Claims (6)

1. A multi-visual defect detection device for large-size LCD glass substrates on a production line, characterized in that it includes an imaging detection system and a bus control system; Wherein, the imaging detection system includes at least one set of imaging detection device, and the imaging detection device includes at least two line scan cameras (11) installed side by side, one or more LED line light sources (12), a transmission control device (13 ) and detection computer (14); The transmission control device (13) includes a conveyor belt (131), a contact sensor (132), a PLC (133) and a PLC interactive display unit (134); The touch sensor (132), PLC (133) and PLC interactive display unit (134) are connected in sequence; the conveyor belt is controlled by the PLC; the touch sensor is installed on the fixed frame on one side of the conveyor belt, and when the LCD glass substrate is When the conveyor belt reaches the preset position, the contact sensor sends a signal to the PLC; The line scan camera (11) is connected to the detection computer and is controlled by the PLC (133); the line scan camera (11) obtains the grayscale image of the LCD glass substrate under the trigger control of the PLC, and the line scan The optical axis of the camera is installed perpendicular to the moving plane of the LCD glass substrate, and the angle θ between the light plane emitted by the LED line light source and the normal plane of the conveyor belt is 5°-10°; The bus control system includes a monitoring computer (21), an industrial Ethernet bus system (23) and a PROFIBUS field bus system (22); The image processing result output by the detection computer (14) is transmitted to the monitoring computer (21) located in the workstation through the industrial Ethernet bus system; The PLC (133) for controlling the movement of the conveyor belt transmits the running speed of the conveyor belt and the position parameters of the detected glass substrate to the monitoring computer (21) of the workstation through the PROFIBUS field bus system (22). 2. The multi-visual defect detection equipment for large-size LCD glass substrates on the production line according to claim 1, wherein the line array CCD sensor in the line scan camera has 7450 pixels. 3. The multi-visual defect inspection equipment for large-size LCD glass substrates on the production line according to claim 1 or 2, wherein the distance between the lens of the line scan camera and the glass substrate to be inspected is 400 mm. 4. A multi-visual defect detection method for large-size LCD glass substrates on a production line, characterized in that the multi-visual defect detection equipment for large-size LCD glass substrates on the electronic manufacturing production line according to claim 1 or 2 comprises the following steps : Step 1: Collect images of LCD glass substrates on the production line in real time; The signal received by the contact sensor that the glass substrate has reached the preset position is transmitted to the PLC (133), and the PLC (133) triggers the line scan camera for image acquisition; Step 2: Preprocessing the image of the LCD glass substrate collected on the production line for denoising and sharpening; Step 3: Use the kmeans clustering method to judge the existence of defects on the preprocessed LCD glass substrate image. If there is a defect area in the current image, go to step 4. Otherwise, end this defect detection and return to step 1 for the next image. Image for defect detection; Step 4: mark the defect area and extract defect features; Step 5: Use the classification method of support vector machine SVM to identify the defect category according to the extracted defect features, and complete the defect detection; In step 3, the process of using the kmeans clustering method to judge the existence of defects on the preprocessed LCD glass substrate image includes the following specific steps: 1) First, divide the preprocessed LCD glass substrate image into several 8*8 rectangular sub-blocks g(x, y), and discretize each rectangular sub-block g(x, y) according to the following two-dimensional DCT transformation formula Cosine transform DCT to obtain the DCT coefficient C(u,v) of each rectangular sub-block g(x,y); $\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; cos</span>}<span\; class="notranslate">\; [</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u\&pi;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ]</span>\mathrm{<span\; class="notranslate">\; cos</span>}<span\; class="notranslate">\; [</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; v\&pi;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ]</span>$ Among them, (x, y) represents the pixel point coordinates in each rectangular sub-block, N=8, u and v are integers between [0,7], $a\left(u\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (u=0)\\ \sqrt{2/N},& (u=\mathrm{1,2}...N-1)\end{array}\right.,$ $a\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N},& (v=0)\\ \sqrt{2/N},& (v=\mathrm{1,2}...N-1)\end{array}\right.;$ 2) Calculate the mean_high of high-frequency coefficients and mean_low of low-frequency coefficients of each rectangular 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 average operation; 3) Calculate the ratio ratio of the average value of high-frequency coefficients to the average value of low-frequency coefficients of each rectangular sub-block in the DCT domain: $\mathrm{ratio}=\frac{\mathrm{mean}\_\mathrm{high}}{\mathrm{mean}\_\mathrm{low}}\&times;100;$ 4) Use the kmeans clustering method to classify the three-dimensional feature quantity composed of the average high-frequency coefficient mean_high and low-frequency coefficient mean_low of each rectangular sub-block and the ratio ratio of the two. The specific steps are as follows: Step a: Perform normalization processing on the average value mean_high of high-frequency coefficients of each rectangular sub-block, the average value mean_low of low-frequency coefficients and the ratio ratio of the two, and obtain the normalized three-dimensional feature vector of each rectangular sub-block; Step b: Construct the normalized ratio histogram sequence of all rectangular sub-blocks in order from left to right and from top to bottom, and perform a first-order difference calculation on the histogram sequence to make it smaller than the first set threshold T1 The ratio value corresponding to the difference value is used as the peak value of the histogram, and the normalized ratio histogram peak number hist_peak is calculated; T1 takes an integer between 15-25; Step c: Set the number of clustering categories cluster=hist_peak, and randomly select a normalized three-dimensional feature vector of a rectangular sub-block as the initial value of the cluster center; Step d: use the k-means clustering method to divide all rectangular sub-blocks into cluster classes, and obtain cluster cluster centers at the same time; Step e: from the cluster cluster centers, delete the two cluster centers with the largest and minimum ratio values, and use the category represented by the remaining cluster centers as the defect category; Step f: Calculate the total number count of all rectangular sub-blocks belonging to each defect category; If count is less than the second threshold T2, it is determined that there is no defect in the current image, otherwise, it is determined that there is a defect in the current image; the value of T2 is an integer between 50 and 200; In the step 4, the defect area is marked, and the defect feature is extracted, and the specific process is as follows: 1) According to the clustering result of step 3, set the pixel value of each pixel in the rectangular sub-block judged as defective to 255, and set the pixel value of each pixel in the non-defective rectangular sub-block The values are all set to 0 to obtain a binary image; 2) Use the opencv contour search algorithm to extract the contour of the defect rectangular area, and mark the defect area with the smallest circumscribed rectangle of each defect area; 3) Obtain the contour position (Px, Py), area S, average gray level L and circularity E of each defect area; Among them, the contour position (Px, Py) of the defect is the coordinate of the lower left corner point of the circumscribed rectangle of the defect part; The area S is the number of non-zero pixels in the smallest circumscribed rectangle of the defective area; The average gray level L is the average gray level of all pixels in the smallest circumscribed rectangular area of the defective area; circularity E, P is the perimeter of the smallest circumscribed rectangle of the defect area. 5. The multi-visual defect detection method for large-size LCD glass substrates on the production line according to claim 4, characterized in that, in the step 5, a support vector machine (SVM) classification method is used to identify defect categories based on the extracted defect features , to complete defect detection, the specific process is as follows: First collect the sample data of the defect area, the sample data includes scratches, scratches and holes; Calculate the area S, average gray level L and circularity E of the defect area in each sample data respectively; Using the area S of the sample data, the average gray level L and the circularity E three eigenvectors to construct two support vector machine SVM classifiers, the first classifier is used to classify holes and scratches and scrapes. distinction, the second classifier differentiates between scratches and scratches; The two classifiers built are used to classify and identify the defect features extracted in real time to complete the defect detection. 6. The multi-visual defect detection method of large-size LCD glass substrates on the production line according to claim 5, characterized in that, in the step 2, the basic image of the LCD glass on the production line collected is pre-denoised and sharpened. Processing refers to using a 3*3 window to perform median filter denoising to obtain a denoised image; then use the Laplacian sharpening operator to perform sharpening processing on the denoised image.