TW200844429A - An automatic optical inspection approach for detecting and classifying the surface defects on coating brightness enhancement film - Google Patents

Abstract

This patent develops an automatic optical inspection (AOI) approach to inspect the surface defects such as Mura, bubble, streak, and contamination on coating Brightness Enhancement Film (BEF). First of all, histogram equalization and local statistics are used to enhance the contours and inner parts of surface defects. Second, using the Otsu threshold method to segment the areas of defects and obtain their locations and shapes. Once the location and shape of a surface defect are addressed, best fitting ellipse algorithm is utilized to extract the geometric features of a defect such as number, area, length of major axis, length of minor axis, and the ratio calculated by the lengths of the major and minor axes. Finally, classifier K-Nearest Neighbor (KNN) classifier is implemented to classify the defective types. Totally 120 testing images including 30 Mura, 30 bubbles, 30 streaks, and 30 contamination defective images are practically verified. Experimental results have shown that the proposed method is able to achieve 100% detection rate and 100% identification rate by appropriate parameters for KNN classifiers.

Description

200844429 IX. Invention: [Technical field of invention] In recent years, the government has actively developed the image display industry and Chinese manufacturers have actively invested in the manufacture of thin film electric liquid crystal display (TFT-LCD) panels, and in the information industry in 2004. The recovery of the economy and the increase in sales of LCD TVs and PDPTVs have driven the growth of the flat panel display market and the growth of TFTLCD «the industry, and thus (4) the business opportunities of raw materials on the upstream side. The global market for flat panel display materials has grown from $260 million in 2004 to $14.14 billion in 2.5 years. Demand for flat panel display materials will grow again in 2008, driven by demand for LCD TVs and plasma TV panels. 50%, reaching $21.3 billion. Among them, the ratio of liquid crystal display materials to flat display materials exceeds 87〇/. , reaching 12.33 billion US dollars. In the domestic part of the domestic TFT-LCD materials market reached NT$29.2 billion in 2004, it is forecast to grow to NT$135.1 billion in 2005. Due to the huge market opportunities in the liquid crystal display materials market, domestic and foreign materials manufacturers are looking at In this business opportunity, they have set up factories in Taiwan alone. Although the material self-sufficiency rate has increased, the core technology is still in the hands of foreign businessmen. Domestic material manufacturers are interested in rushing to eat this market. However, there is still a gap between domestic material manufacturers' R&D technology and advanced foreign manufacturers' advanced technologies. In addition, the small scale of domestic manufacturers makes production and quality unable to reach a competitive market economy. It is the difficulty encountered in the development of flat panel display materials in China (1). Domestic demand for liquid crystal display materials is huge. It is predicted that it will reach NT$131.5 billion in 2.5 years. Among them, the glass substrate, brightness enhancement film / enamel film, liquid crystal, cold cathode fluorescent tube are the top five products in the market. The cost of the two materials of brightness enhancement film / enamel film accounts for 14% of all costs, second only to 49% of the glass substrate. It can be seen that if China can develop its own high-quality and low-cost brightness-enhancing film and enamel film, it can greatly reduce the cost of manufacturing LCD 200844429 display, and improve the competitiveness of China's flat-panel display industry. In order to provide a high-quality brightness enhancement film, it is necessary to have excellent recording capabilities. The brightening film is an optical film. The most important inspection item is the product of the product. The current detection method is artificial and financial, and is still dominated by adopted personnel. China's TFT-LCD industry's optical inspection equipment is mainly focused on the detection and finishing equipment of lcd. For the detection of brightness-enhancing film or other optical film, it is necessary for the quantification of the determination and the saving of human-powered optical inspection. . Therefore, this specializes in the development of the coated brightening film-automatic optical method, which can effectively detect and classify the surface of the thinned surface. The coating type brightness-increasing film is produced by using a roll-type (secret t〇_R〇n) and using a precision coating method to insert a high-quality female fabric with a reading birefringence. After repeated coating, it is formed into an optical film having a polymer layer of coffee layer and a thickness of only 132 Å. The color of the process cap, the extension and the surface of the finished product may be colored by the foreign matter mixed into the silk or stained. Degree, shadow _ optical film's degree of polarization, transmittance and twist, so that the optical film's product f butterfly [3M5 chest. In order to assist manufacturers to detect defects and determine their types, the quality information of the products is immediately fed back to the process unit for process improvement. At the same time, in order to respond to the goal of improving the self-sufficiency rate of domestic equipment by the government to reach 50% in 2008, this patent The invention develops an effective coating-type brightness-enhancing film occlusion automatic optical detection method, and performs cockroache detection for the four common cockroaches which are commonly used in the precision coating process [7]. 1 · Large-scale uneven coating or uneven brightness (Mura or Unevenness) 200844429

Mura is defined as the phenomenon of irregular brightness or uneven color (4) on the surface with uniformity of color and color. The cause of this thief is very high. The dye output may be poorly controlled during the coating process. Vibration problems, foreign matter in the coating process, gaps in the mold, accumulation of paint in the mold, and damage to the mold or scraper, etc. Any small error may lead to large-scale production, the first picture (a) That is, coating unevenness on the brightness enhancement film. 2. Bubbles Bubbles can produce different types of fatigue due to different reasons, such as dust on the substrate, dyes that are not full of nafe, coated gas, and _Teng# original _ will lead to different forms Bubbles are generated. Although there are many kinds of bubbles, their shapes are similar to the circular geometry. Therefore, it is easy to judge the bubble fatigue by machine vision. The first is the bubble fatigue on the party film. 3. Streaks Streaks are similar to Μ (10), but the difference between the two is that the stripes only appear - the unevenness of the single-line coating, while the Mwa is unevenly applied over a wide range of lines. . This type of ruthenium is one of the most common types of enamel in the production of brightening film. The first picture (c) is the 瑕疵 on the brightness enhancement film. 4. Spots or Contamination In the process of coating, foreign matter 上 appears on the brightness-enhancing film because paint, tape or dust adheres to the back pressure roller or the roller does not clean regularly. Such enamel can be avoided if the substrate is first inspected and the roller is cleaned at regular intervals. Figure _(3) shows the image of 瑕 200844429 /疵. [Prior Art] There are two main types of brightness enhancing films on the market today, one is an optical film called a enamel film, and the other is a reflective optical film having a multilayer film structure. Both of these are independently developed and produced by American 3M, which has patent protection and cannot be produced by other manufacturers. In addition, Nitto Denko has also developed a non-multilayer film reflective polarizer PCF (pdarizati()n Conversion Film) [10] that utilizes a cholesterol liquid crystal mechanism to recover light. Today's brightness enhancement film is mainly used in thin film transistor liquid crystal display (TFT-LCD) thin film transistor liquid crystal display (TFT_LCD). The process is mainly OLED process, panel process and module process. The process requires a large number of components required to form a complete liquid crystal display module, including a TFT-Array, a Color filter, and a Liquid Crystal 'glass substrate. , Polarizer, Alignment film, Electrode, Thin Film Transistor (TFT), Color Filter (c〇i〇r Red), Gap (Chong (10) Sink), Storage Capacitor (c), a prism sheet/brightness ehancement film, and a driver IC (Driver Integrated 1C). There have been many researches on the appearance and appearance of TFT-LCD products in the month, which can be mainly divided into the appearance and detection of finished products assembled by TFT_LCD and related components of TFT-LCD such as color filters. Polarizers, glass substrates, etc. are two parts of the appearance inspection. The following are discussed: Surface detection of TFT-LCD

Kim et al [9] used the machine vision method to check the brightness unevenness of the lcd panel. The 200844429 test uses the gradient of the 瑕疵 image in the spatial domain due to the change in brightness to capture the contour η of the image party degree change. And according to this high-definition silk _ thief's location. Finally, the test results are compared with the results of manual identification. The results show that this method is consistent with the results of manual identification. MGri and riding on the LCD Μ·瑕 瑕疵 瑕疵 , , , , , , LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD LCD MG LCD MG MG MG MG MG MG Xin [11] uses machine vision on the surface of the TFT panel for plague detection, for the TFT-LCD panel towel, the human eye is not easy to microscopic thieves, including picking, dust, alignment film holes and ship's scales' _Fu Liye County (FQurief she ―) and anti-Fourier imagery, 魄 , , 配合 , , , 棘 棘 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县 县Newkinpe Ship, and (4) Statistical Control Boundary Law highlights the position of thieves and thieves on the side of the silk. The county uses the Hough transform method to divide the _ marks and _ _ _ holes and dust in the Fu-reversion image, and distinguishes them by the morphological study. h et al. (10) developed the characteristics of bribe CD giant crane I. [CD panel detection green' divides TFT_LCD __ into several dump blocks, and uses a luminance meter to measure the brightness in each block, and the sugar __ block Gray_exclusive, and using the variance analysis and the exponential weighted average method as the criterion for judging, find out the blocks with uneven brightness in the module, and the effect of the __ touch towel is abundance. Chen 200844429 Zhizhong [13] is aimed at the giant image of TFT_LCD, such as Yuan ^, Jue, etc., the TFT-LCD module panel is divided into several blocks, using ccd to obtain images of various regions and calculate each series The average brightness, and then the highest brightness value and the lowest brightness value are compared by the average brightness of each block, and the ratio of the highest brightness value to the lowest brightness value is used to determine whether the module has a plague of brightness unevenness . ^ et al (4) using a linear camera with a gray-scale to capture the surface of the TFT_LCD, and the image obtained by the ride is separated into two fresh, with grayscale saturation and directional and periodic and edge images. Next, Lapladan ef L〇G, according to the wave, is on the list. The image is processed by the image, and the processed image is combined and binarized to obtain the desired image. All of the above technologies are used for the appearance inspection of the finished TFT-LCD. Careful observation of the finished product of the TFT-LCD reveals that a background pattern with regular lines on the panel appears, as if the system is neatly arranged as a matrix. If the inspection function is a small-scale inspection on the panel, the hybrid method is _ fresh_method such as Fourier transform or wavelet transform, etc., and the thief face of Lin Weilu (10) is observed; however, if it is for a large area The plague, such as Mum, is measured by the brightness meter, and then the image processing method in the spatial domain is used to highlight the position, thereby detecting the position and shape of the cockroach. ^ Surface detection of TFT_LCD related components

Lhn et al. (10) used the machine vision method to identify the fatigue of the glass substrate of the LCD. The main development - the identification system is used to select some fine particles or the 200844429 does not affect the quality of the product. In this way, without affecting the quality of the product, avoid excessive resources to detect and repair some minor flaws that are not potentially threatening, and improve the output rate of the product. Zhao Xinmin [16] used ITO conductive glass used in touch panels to detect surface plague by nonlinear diffusion, which uses spatial domain (Spatiald_called nonlinear diffusion for non-homogeneous random texture surface) Detection, according to the gray level of the image (Gradient) information for diffusion (Diffosi〇n) processing, smoothing in the weak gradient region, and in the strong gradient region, the suppression smoothing process. In images with random texture, the edge of the ridge usually has a higher gradient than the background texture. Therefore, the nonlinear diffusion is applied to the detection of the random grain surface, which can accurately smooth the area of the image towel with random grain background. Processing, and the Ministry will destroy the edge of the fatigue and achieve the purpose of suppressing the texture and highlighting the fatigue. The effect of using nonlinear diffusion to process the image for the random texture will be related to the DiffUsivity functions. Diffusivity parameters and the number of iterations. Qiu Xueyuan [17] uses machine vision technology Applied to the light guide plate, the flaw detection, the object to be detected includes scratches, flow marks, bright spots and black spots, and it is expected to utilize machine vision technology to improve the loss caused by the false detection caused by manual detection. First, Fourier transform is used to filter out the regular texture background composed of dots on the light guide plate and use binarization to highlight the flaws in the image, and then use the neural network to classify the ripples. Four kinds of eigenvalues such as area, slope variation, aspect ratio and least square error are used as input of percepti〇n neural network, and combined with 12 200844429 grayscale threshold Four sputum samples with bright spots and seventeen samples with flow marks are used to train the perceptron. Cai et al. [18] applied image processing technology and neural network theory to conduct Polarizing Film.瑕疵Detection and classification, for the common color difference (strip, cloud, dot), scratches and poor fit of polarizing film, etc. The image mask-wave noise interference filter, in addition, then the region growth method (Regk) n &

Method), mark and segment the plague image part, and finally use the Lapladan son to recognize the thief side, and the thief _ 特性 特性 特性 瑕 瑕 瑕 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性 特性Four characteristic values, such as the average area of the vocabulary and the perimeter of the ', as the eigenvalue of the thief, which uses the three-layer reverse-transfer-like neural network to classify the five types of money. Take eight copies of the book, a total of 4 瑕 silk book as a sample of the inverted transmission network, after the network training is completed, use another 40 瑕疵 samples to test the network identification.

Naka Shima [19] detects macro-defects and micro-defects of color filters, using image subtraction and optical Fourier filters to measure fatigue, for microscopic The five types of 瑕疵 in the 瑕疵, including SIR0-NUKE > Black Matrix Hole ^ Particle ^ Black Matrix Pattern Defect ^ Color Filter Defect for image detection; and the other three types of plague (dust, conductive glass, particles ) is to make ^He_Ne laser with Fourier filter to produce a good UX 丨 丨 thief. These two methods have their own rules. The hetero-phase subtraction technique needs to consider the reference of the hybrid (Goldenimage) to generate the mosquitoes (Regis (10) (5), etc. 13 200844429 questions' and if the reference sample selection is not accurate, the misjudgment Opportunities will greatly improve the old He Chang Lei _ Newgger's Wei control, only to receive the addition of impurities. There are also many studies on TFT_LCD __ transfer into the pure recording, but the main research object is only for a few parts such as IT〇 Conductive glass, light guide plate, polarizing film and color filter, etc., for the detection of enthalpy of the passenger film. In the domestic and foreign patents, the patent is called the annual film 瑕巍 for the detection object; Innovative and sturdy roads 1. Ye Yangzhe, “Flat Display Materials Industry”, Industrial Materials, No. 219, 2005, pp. 74-83. — 2. Garvin, L. Hugh, Pinnow, and A. Douglas et. al.9 "High Selectivity Thin Film

Polarizer ", United States Patent 4289381, 1981. 3. Kelimann and Fritz, Polarizer for Infrared Radiation, ', United States Patent 5177635 1993. 4 4. Song Dacheng, "Introduction to Polarized Film", Industrial Materials, No. 140, 1998, Pp. 118-126. 5. Huang Chaoyi, “Introduction to Polarizing Films”, Optoelectronics Industry and Technical Information, No. 20, 1999, pp. 40-47. 6. Liu Yijun, “Principles and Applications of Polarizing Films”, Industrial Materials, No. 188, 2002, pp. 153-159. 7. Wen Shuheng, “Coating Defect Elimination Method – Turning Decay into Magic”, Industrial Materials, No. 205, 2004, pp. 98-105. 8. Zhang Yonghan, “Introduction to the Application of Precision Coating Technology in Optical Films for Flat Panel Displays”, Industrial Materials, No. 211, 2004, pp. 127-136. 9. JH Kim and BA Barsky, uHuman Vision Based Detection of Non-Uniform Brightness on LCD Panels", Proc. Of SPIE4S&T Electronic Imaging, 6070, 2006? 60700P. 10· Y· Mori, K. Tanahashi, and S. Tsuji, “Quantitative Evaluation of Mura in Liquid Crystal Display, 95 Optical Engineering, 43(11), 2004, pp. 2696-2700. 11·Zeng Yanxin, Application of Machine Vision to Surface Detection and Classification of TFT Panels, Master thesis, Yuan Department of Industrial Engineering and Management, Chiku University, China, 2003. 12. BC Jiang, CC Wang, and HC Liu, ''Liquid Crystal Display Surface Uniformity Defect Inspection using Analysis of Variable Exponentially Weighted Moving Average Techniques9,9 International Journal of Production Research, 43,2005, pp. 67-80. 13. Chen Zhizhong, Automated Detection of Pixel Point Defects and Relative Brightness Uniformity of Liquid Crystal Display, Master Thesis, Institute of Mechanical Engineering, Chung Yuan Christian University, Zhongli, 2000. 14. KB Lee , MS Κο, Τ. Μ. Κοο, and Κ. Η. Park et. al.? "Defect Detection Method for TFT-LCD panel based on Saliency Map Model", 2004 IEEE Region 10 Conference, A? 2004, pp. 21-24. 14 200844429 15* D· C. Lim, D. G· Seo, D· Η· Jeong, “ Defect Classification for the Inspection of TFT-LCD Glass99, Proceedings ofSPIE Optomechatronic Machine Vision, 6051, 2005, 6051 OF. 16. Zhao Xinmin, Application of Non-homogeneous Diffusion in Non-homogeneous Textures of IT Conductive Glass Surface Inspection, Master Thesis, Yuan Department of Industrial Engineering and Management, Chiba University, China, 2003. 17. Qiu Xueyuan, Development of Automatic Light Detection System for Light Guide Plate, Master's Thesis, Department of Mechanical and Automation Engineering, Kaohsiung First University of Science and Technology, Kaohsiung, 2004. 18·Cai Yingnan, Applied Image Processing and Neural Network for Polarized Film Recognition, Master Thesis, Department of Polymer Engineering, National Taiwan University of Technology, Taipei, 2003. 19·K· Nakashima,, 'Hibrid Inspection System for LCD Color Filter Panels,,, Instrumentation and Measurement Technology Conference Proceedings, IMTC/94, 1994 ? pp. 689-692. 20. RC Gonzalez and RE Woods, Digital Image Processings New York: Addison-Wesley, 1992.

21. Zhong Guoliang, “Image Processing and Computer Vision”, Taipei: Donghua Book Company, 2002. 22. N. Otsu, A threshold selection method from gray-level histogramM? IEEE Transaction on System, Man, and cybernetics, 9(1), 1979, pp. 62-66. 23. J. Weickert5, fA Real-Time Algorithm For Assessing Inhomogeneities in Fabrics'1, Real-Time Imaging, 5, 1999, pp. 15-12. 24. RM Haralick and LG Shapiro, Computer and Robot Vision, New York: Addison-Wesley, 1992, pp. 639-658 25. A. Khotanzad, H. Banerjee and MD Srinath, UA vision system for inspection of ball bonds and 2-D profile of bonding wires in integrated circuits 99, IEEE Transactions on Semiconductor Manufacturing, no. 4, vol.7? 1994, Pp. 413-422. 26· Η·K· Lee and S·I· Yoo, “A method for inspection of ball bonds in integrated circuits”, Systems, Man, and Cybernetics, IEEE SMC, 99 Conference Proceedings, 2 (12 -15), 1999, pp. 975-980. ' 27. L. Higgins, LG Bahler, and JE Porter, MVoice Identification using Nearest-Neighbor Distance Measure." Proc. ICASSP, 1993, pp. 375-378. 28 Guo Jianting, application shadow Processing technology in the research of a quality management system in order to detect, for example a credit card, master's thesis, Institute of Management Evergreen School of Management, Tainan, 2001. [Description of the Invention] The patented invention includes a method for pre-processing a brightness enhancement film image; that is, an image enhancement and a binary image processing method, and a method for capturing the geometric feature of the image, and finally different types of images. It is classified by probability neural network and K nearest neighbor neighbors. 15 200844429 ι·Image pre-treatment is the most difficult to increase the pain _ _ Jun cloth uneven ripples (heart a) and stripes, this record visual appearance ± color _ smaller, rectification bubble or foreign body When the silk is easy to distinguish, it is necessary to make the (10) special image ageing method to deal with the county thief image before it can highlight the thief's miscellaneous and position. Therefore, the “Rectangle® Equalization” and “Regional Statistical Parameter Processing Method” of this patent revolving space domain, image enhancement by riding a bright film image. ... 1·1 Histogram Equalization 70 Innocent increase In the bright film image, the gray scale value of the whole image is concentrated in a small range, and the gray scale value variation degree is very small. In the image with 瑕 _ brightening film, the gray point value of the pixel in the area where the sputum is located is compared. Others have a large change in the grayscale value, so the grayscale value of the pixel in the area where the shelter is located is smaller than that of the non-sickness-enhanced image. However, the grayscale image of the brightening film image has a glimpse or no madness. The value changes are very small, so in order to facilitate the subsequent image enhancement processing, the original brightening film plague image of the patented invention is ordered to be equalized. The physical meaning of the histogram scale method is to set the original gray scale value. The grayscale values of the distribution that are too concentrated are redistributed to the grayscale values 〇~255. The images that have been processed by the histogram equalization have a wide range of grayscale values and can more easily detect the distribution of the 瑕疵. The second picture is the histogram of the image of the brightness enhancement film Mura瑕疵 image after the histogram equalization method and the distribution of the gray level value, and the third picture is the result of equalization of the histograms of the first image by the histogram. 2 Image enhancement - regional statistical parameter processing method (L〇cal 16 200844429

Statistics);,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,

Deviatio_ images are processed. In terms of image enhancement, this method develops with the concept of mean and standard deviation, where the mean represents the measurement of the average grayscale value in an image, while the cardiac standard represents the average contrast, which is the regional average. The difference between the difference and the difference is related to the image characteristics in a predefined area near each pixel in the image. The square space of the mask is defined. The center position of the mask is in the pixel (four) point-by-point crane. The fresh axis mode is from left to right, and the image mother pixels are processed sequentially from top to bottom. After the image is processed, as shown in the fourth figure, the size of the whole image is 16G X 12G pixels, and the mask size is set to 5 χ 5. When the mask is processed, it will be occupied on the image. A blank boundary of 2 pixels is generated. The patented invention fills the blank boundary around the image with the gray scale value of the adjacent pixel. If the pixel (A value is added), the gray level statistics in the mask area, such as the regional average and _ standard deviation, are calculated at each pixel point (4). How to calculate the regional mean and regional standard deviation as follows [20]: First let the 4-off-board machine Wei, which represents the shed in the [GH]_ residual order, and let the ice be the normalization of the corresponding value of r The histogram component, in which ice) is an estimate of the probability of grayscale occurrence. Then get the "step difference of the mean of its r (she defined as Tao: 17 (1) (1) 200844429 where 00 = Σ (Γ / - 尸 /) ο where w is the average of r (r The average gray scale value): Ll ° (2) The second-order motion difference (π Shi _) can be obtained from (1) and (7) two equations (L) ΜΟ = Σλη - m)2p(rj " ( 3) The second moment is the 4th variance; called the standard deviation is defined as the square root of the variance, which is the square root of the second moment; called σ (0. Then define a fixed-size rectangle as the execution region Histogram statistical processing of the mask. The size of the mask is very important 'fresh too small can not _ the purpose of the thief; conversely too large will produce too blurred image and greatly increase the processing of the program, so 3 X 3 or 5 X 5 as the size of the mask, the patent invention uses 5 X 5 as the mask = small (x, _y) is the coordinate of one pixel in one image, so that ^ represents the image pixel in any image The grayscale value on the coordinate center, let \ denote the mask centered on the coordinate (four) (the sub-image Η fourth_show, from the formula (7) can calculate ^ all pixels in the ^ mask The average grayscale value of the point ^^, ''Si is) (4) where ~ is the grayscale value at the coordinate (〇) in the mask, and 〆~) is the grayscale value at the coordinate CM) ~ in the ~ mask The probability of occurrence. At this time, if the average grayscale value is replaced by the original k', the effect is equivalent to the smoothing (Sm00thing) processing of the average filter (Meanfllter), and the purpose of the smoothing processing is to eliminate the noise on the original image. (5) (5) 200844429 Similarly, the formula (3) can be used to find the pixel grayscale value in the sail mask '] 2 〆 ~)) 2 Therefore, the entire image of each _ pixel () In its own right, ^^, at this time, it can be judged that the entire image is a pixel (X, W's A includes "Nove, 隹* " value ~_ and minimum value Amln, and the following & Normalized gray-scale conversion,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,, On the edge of the bubble or the foreign body, the edge of the female will cause a large I'. If there is a gas in the image, the foreign ship and the M., the pixel in the Mura瑕 domain (4) is normalized to the grayscale value conversion. After processing, the grayscale value will become very small. In order to overcome this special practice, if ~"over_ in formula (6) she becomes 1G' and the formula (6) towel is super-speaking and its setting is (7). Then, the plague images of the third figure are smoothed by the mask of the average mask of the formula (4), and then the smoothing process is performed by the formula (5), and then each pixel (four) is smoothed, and finally "marked I The difference normalized gray-scale value conversion, and the processed new gray-scale value melon force image. As can be seen from the fifth figure, the above-mentioned smoothing processing and the function of "fresh difference normalized gray-scale value conversion" processing and general edge preparation The effects of the devices such as Sobel and prewitt # are similar, but the effects are more blurred than that of s〇bel and prewitt, and the range that is highlighted is not only the outline of the madness, but also the area of the scorpion. Image enhancement is based on this, and the advantage is that the enhanced part is not only for the edge of the ,, but the area enhanced by 19 200844429 also includes the interior of the 。. Then judge a pixel (x, 3 〇 is the edge pixel, this The method is mainly to compare the regional standard deviation α of the treatment, the flattening β processing and the standard deviation normalized gray-scale value conversion.

The mean value of the gray scale standard deviation of the Sxy overall image ~—, defines the average value of all mask standard deviations of σ_4 °Χν, ie Μ Ν ΣΣσ, ^ ^ average = 土〇 0

MxN (7) Eight middle Μ and N are the number of pixels included in the scale width and height. The images of the present invention are all _ pixels X 12G pixels. If a bite is greater than or equal to, the pixel point (4) is regarded as a pixel point that needs grayscale value conversion, and the Μ is a constant greater than 0 and not more than 3, in order to detect the pixel edge of the silk edge, it is necessary to Set ~ in the range of 〇 to 3, which is based on the concept that the hang (4) is within the three fresh errors of the average number, and contains about 99.7% of all possible pixel grayscale values. set up. Let / (to represent the image pixel point at any image coordinate (the above grayscale value, and let the heart) represent the new grayscale value after conversion on the coordinate (4), ie, ;;). £ χ / (χ, less) if σ ~ 2 夂 σ, _ chat / 0, less) Other (8) where the parameter is corrected, the catch frequency is greater than 255, then 255, then move the neighbor 'to the phase Hao's pixel position, and repeat the above steps until all the image_processing_ on the entire image is recorded. The sixth is to take the conversion result of her image by formula (8), which is tender. The parameters are set to £=3 and moxibustion respectively. =1. , 20 200844429 1.3 Binary image segmentation - value of the main purpose of the gradual grayscale image of the theme or target of interest, the thief order value of the change (four) its convex _ to make the glory of the touch The grayscale value of the value and the background are completely reduced in black and white. In order to make the image of the fatigue after binarization effectively highlight the remaining position and the corridor, the selection of the binary selection threshold is very important. The patented invention adopts the value of 〇 网 网 ☆ ☆ 界 H 。 。 。 。 。 。 。 。 。 。 。 。 。 。 The principle of this search for binarized A i values is 4 to find a threshold so that the weighted total sum of the variances in each cluster is the smallest. Suppose there is a grayscale value in the image, which is the grayscale value (four) the number of pixel occurrences'. The total pixel in the image, the secret of each grayscale value is: N (9) If you want to divide the grayscale of the image into two Group c. With ς, where c. The ash ° gradation possessed by the population is 〇~k; the grayscale values possessed by the same captured population are within. The probability of these two groups is w. And called··

K == ^(k) /=〇 (11) μ = =1-male) /=^+1 The mean value of the two groups ".和Μ为^〇-Σ~ /=0 冰〇21 (12) 200844429 i=k+\ (13) The variances %2 and %2 of the two groups are: σ〇=Σ(ζ·-"〇) 2立/=0 >^〇(14)

, = "1 wx then the weighted sum of the two groups of variances ~ 2 is: aw = w〇cr 〇 2 (k) + \νλσ χ 2 (k) Using the above method to weight all the gray scale values respectively After the sum is calculated, the weighted sum of all gray scale values can be obtained, wherein the weighted sum is the minimum gray P white value which is the desired binarization threshold [23]. After the processing by the above binarization method ~ As shown in the seventh figure, the towel can clearly see that the inclusions, bubbles, stripes, and foreign objects can highlight the position and shape of the fatigue through this binarization. Take an optimal ellipse method (15) (16)

The patented invention uses the most suitable elliptical method to sneak 6] to perform the surface feature extraction of the brightness enhancement film. (4) After the binarization treatment, there are many groups of silk fibroin spots on the silk image, and the whole thread can be utilized. The geometric coordinates of the optimal ellipse of each group of towels, such as the number of vehicles, the total length of the silk, the length of the short axis, and the ratio of the length of the long and short axis, are calculated according to the coordinates of the points, and the surface of the ship can be touched. ·, milk foam, strip foreign matter. Here I use the financial b pixel coordinate group to calculate the reticle centroid (CentroidM^^t^fi,:J^r/A (r,c)eR (17) 22 200844429

c= ^c/A (r,c)eR (18) Among them: X pixel coordinates X coordinate value C: 瑕疵 pixel point coordinate coordinates and: 瑕疵 pixel point set 4: 瑕疵 pixel points total optimal ellipse The main geometric features of the method are the length of the major axis of the sugar circle, the length of the minor axis and the area of the ellipse. A binarized image of MX n is calculated as follows: Second order X-axis motion difference: = Μ — = T, (^r ) 2/A r=\ (18) Second-order 7-axis motion difference: /^ N =Σ(.- c=l (19) Second-order mixed motion difference: /^ NM =ΣΣ(Bu·*—1 (20) where (r, c): Irregular pixel points, pixel grayscale value is 255 J): Centroid of irregular pixel point set: Total number of irregular pixel points After calculating these second-order motion differences, you can follow these Regular pixel points find the optimal ellipse, so the formula for ellipse can be redefined as [25]: 23 (21) =1200844429 dx d2 rd 1 d3 c or dxr2 + 2d2rc + d3c2 = 1 (22) next [24 ] [25] [26]: The above D matrix can be re-presented by the eigenvalues of the second-order motion difference such that D = d 1 d 2 _ 1 ~ μ: d2 d3 _ 4 (MrrMcc -MrC) μη (23) 4. After the values of d2 and d3 are calculated, the length of the long axis, the length of the short axis and the area of the ellipse can be derived. The calculation method is as follows [24]: 2V2 + ^3) ~ V(^1 '~^ y=+4df Long axis length (24) 2V2 Short axis length = VK +^3)+VK -^3)2 +4c /22 Area = 0 = W3 - β - W2 W3 - The eighth figure is an example of the seventh method of performing the optimum expansion measurement after the detection. The rafts are Mm, bubbles, stripes and foreign bodies. In the eighth figure, it can be clearly seen that the number of _Mum and the striped crepe is made larger than 丨, so that two similar elliptical features are captured. In order to facilitate the input of the subsequent classifier, the elliptical area taken up is summed, and the ratio of the length of the wire, the length of the 姊 and the length of the 姊 are taken as the average value. The ninth picture shows the measured values of the geometrical features of the mechanical images obtained by the most suitable method. (25) dl (26) 24 200844429 3· Brightening film plague classification - K nearest neighbor classification (4) Collision system called KNN) K^«_n(KNN) to classify brightening surface defects' And to explore the optimal setting of the parameters of each side of the branch (4) Nearest neighbor method (Nearestneighbordedsicmrule) determines the classification by the closest neighbor data to rely on the category of the sample to be classified. If there is currently a sample vector v ' of the unknown category, then we find the closest point to v in all the data points, and in the heart category, the mosquitoes should be returned to the human-class 1 mathematical shirt, and the age is satisfied. In the case of ship type, V is classified into the same class of Vk [20]: dist(y -vk, S dist(y - v), i = 1,2,···, η,i this k (9) Euclidean distance can be used (Euclidean distance) to measure the distance between unclassified samples and categories: also / (vh). Extending the nearest neighbor decision method, it can evolve into k nearest neighbor decision decision (k-nearest neighbor decision rule) It is the category of the voting of the nearest neighbors (Voting). It is assumed that Kj (k, x) represents the number of the jth class among the k neighbors closest to χ, and then satisfies (k, χ) . ^ (k, χ), according to the k-NN rule, x is classified into the jth class. 4. The process of brightness enhancement film detection and classification This patented invention uses image processing to capture the geometry of the image of fatigue The feature and Kg classifier are used to detect and classify the surface of the enamel film. The complete detection and classification flow chart is shown in the tenth figure. As described below: (1) Pre-image processing First, the original brightness enhancement film image is subjected to histogram equalization processing, so that the grayscale value of the whole image 25 200844429 image will be gathered from the county group in a small part of the grayscale value. Converted into scatter between ~2 points. ___record processing butterfly _ she image enhancement of the silk image after equalization of the histogram. The image enhancement part is carried out in a two-stage manner, first smoothing it Processing and "standard deviation normalized gray-scale value conversion, processing" and then formula (8) transformational mismatch. Area word mean processing has the effect of removing image noise. Regional fresh steel has the function of making image gray-scale value change area The shape and position of the 即可 瑕 瑕 上述 上述 上述 上述 上述 上述 上述 影像 影像 影像 影像 影像 与 与 与 与 与 与 与 与 与 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 〇 The geometrical feature values of the binarized 瑕/疵 images are extracted using the optimal ellipsometry, including the number of 瑕 scales, total 瑕 积, average filament length, average short axis length and average length. Axis ratio, etc. (3) Classification result of classifier The patented invention selects KNN classifier as a sorting tool for surface fatigue of four kinds of shell-forming films such as Mura, bubbles, stripes and foreign matter. [Embodiment] The personal computer device of the patented invention uses a computer compatible with IBM. The central processing unit (CPU) of the personal computer is an Intel Pentium 4 2.8G processor, and the memory (SDram) is 512 MB. The NIPCI-1409 image capture card and the Watec WAT_2〇2B grayscale/color camera are used. The maximum resolution is 752x582 pixels and the operating voltage is 12VDC. Light source design 26 200844429 LED backlights 'application software using Measurement & Automati〇n (MAX) software developed by National Instruments. The image processing and the extraction of the geometric features are written in the development environment of Borland C++ 6.0 using the C++ programming language. Subsequent 瑕疵 classification is carried out using Matlab software. The sample size of the sample is measured as a pixel, and the actual field of view is 14 mm x 10.5 mm, and the resolution of the image can reach 9 〇w pixels. In the production process, the brightness-enhancing film may be subjected to sub-records such as Mum, bubbles, offerings and foreign objects, etc.: Detection. The actual scale - 瑕 __ taken the thief sample is 3 (), all plague images two pictures (a) (b) and (c) a total of 120. The sample images of 30 secret recipes are shown in the tenth-figure (8); the histograms are equal, and the results are as follows: Xuan-_) hemp; the image processed by the bad processing and the "difference positive gray scale value conversion" 10)), the image of all Mwa thief sample images converted by the formula (8), the image after the 0tsu binarization and the optimal ellipse method are listed in the tenth same as the '3G gas nucleus sample image as the tenth The fourteenth image of the triacid is not. The 30 lines are noisy. The image is as shown in the fifteenth and sixteenth figures. (IV) The foreign matter sample image is shown in Figures 17 and 18. The rotten and round method has 30 geometric features of Mura, bubbles, stripes and foreign objects, including the number of thieves, the contact area, the average long axis length, the average short axis length and the average length and length axis ratio. The geometric feature data of #1 is listed in the 19th, 20th, and 20th-20th. However, due to the different locating of the acquired geometric features, and the different singularity of the singularity and the sufficiency in the same geometrical features, it is necessary to normalize the eigenvalues of the gamma. This formal 27 200844429 is the normalization of statistical methods. If the 12G thief image is regarded as a group, the population average and group standard deviation of the number of 瑕疵 groups of 120 images are calculated, and then according to formula (28), 30 Normalize the number of groups of 瑕疵 images such as Mum, bubbles, stripes, and foreign objects. The total shouting area, average silk length, average miscellaneous length and average length and length of the 12G Zhang Yigui image are also normalized. The average and standard deviation of various geometric features are shown in Figure 23. . The normalized Mum, bubble, line and foreign body fatigue characteristics are paid in the twentieth, twenty-fifth, twenty-sixth and twenty-seventh. After the normalization, the geometric feature value can reduce the training time of the 瑕疵 classification, and the faster _(10) state or the convergence effect of the health and the high accuracy. Where V is the normalized 瑕疵 geometric feature value A is the 平均 geometric feature value of the group mean σ is the group standard deviation of the 几何 fatigue geometric feature value, then the ΚΝΝ classifier classifies the surface of the brightness enhancement film, various 瑕The turtle randomly selected 10 groups of samples for network training, so the training samples totaled 4G groups; listed in the twenty-eighth figure. The remaining 2 groups of each plague sample were used as test samples for the shame test samples; they are listed in Figure 29. The thirtieth figure shows the average recognition rate of _ to the test sample under different κ values. The κ value of each-turn (κ=ι, 2, 3, and the cuts are repeated in accordance with the random selection training and test samples described above, and the average recognition rate of the order is used to represent the translation. The square wire determines the red value of the leg ^. It can be easily observed from the twentieth towel. When K=1, the maximum average recognition rate of the k-k is 28 200844429, so the patent road day jg knee bracket &&; 1 a

1%% recognition rate. If sample numbers 22, 19, and 10 are selected as training samples, the κ value is set to i, and the remaining samples have a recognition rate of 100%. This patent is issued on the day of the tree riding and recording. The results of the test are as follows: (1) The line of the lang is the same as the method _, there may be __ membrane plague and detection = brightening The four-year-old Wei Jian is in the German, the degree of hygiesis can reach 9 b / pixel, the minimum resolution of the industry is 200w pixels. Processing _ 14xl05_ size, and IX two for 160x120 pixels of the crazy image about 〇 295 seconds. In the general roll-form coating of the private towel, it is found that the vertical direction of the coating direction is 1Qmm, and the width is as long as the position of the coffee is required to be 6.8 seconds. In the future, several Ca_ into the column are combined. Optimizing the test program is expected to significantly reduce the processing time of the test. (2) The general class of cockroaches defined by the patented invention, such as Mura, bubbles, lines and foreign bodies, has a detection accuracy rate and a 瑕疵 classification identification rate of up to 1%, which is quite suitable for the detection of artificial visual inspection. . (3) Using the optimal ellipse method, the position and shape of the sputum can be effectively found, and the geometrical values of the 瑕疵 part can be obtained: the number of sputum groups, the total madness area, the average long axis length, the average short axis length and the average length and length axis. Ratio and so on. The geometric feature values extracted can be used as the input data of the classifier to further identify the species. (4) In the KNN classification, the KNN can achieve the best recognition when the κ value is set to 1. 29 200844429 ^ The 々 recognition rate decreases as the K value increases. The time required for KNN is as short as 0.05 seconds. [Simple diagram description] Tian Tian Ping ^ field of view (F0V, delete Of View # (8) bubbles; (c) stripes; different t X 120 pixels: (8) light and dark uneven (Mura);

Grayscale paper (C) J image after processing; (8) Gray scale value scatter diagram of Fig. 4 (4) The results of the equalization of the simple images of the third image are: (8) unevenness of light and darkness (Mum), (b) air bubbles; (c) stripes; (4) foreign body four figures · regional statistical parameter method of the mask movement mode fifth picture · Ιΐί ^ each thief teacher _ New Zealand, hiding "fresh difference normalized gray level _ ι processing After the image: _ dark uneven _ra); (b) bubbles; (e) stripes; (d) foreign matter 1 The eighth picture · the fifth picture of each fatigue image converted by the formula (8) ^ = 3 and heart =丨): (8) light and dark unevenness (Mura); (b) air bubbles; (c) stripes; (4) foreign matter seventh picture (4) of the fine images of the 〇tsu binarization results: (8) light and dark unevenness (Mum) (b) Bubbles; (c) Stripes; (Examples of the optimum expansion of the image of each of the fatigue images of the eighth figure in the eighth figure of the φ foreign object: (a) unevenness of light and darkness (Mura); (b) air bubbles; (c) streaks; (d) foreign matter ninth figure 瑕 fatigue image obtained after the most it elliptic method analysis 瑕疵 geometry special 200844429 tenth figure this patent invention complete detection and classification flow chart eleventh figure 30 sheets Mura瑕疵Sample image··(8) (9) Histogram equalization; (c) smoothing and "standard deviation normalized gray-scale value conversion, processing twelfth picture 30 sheets of Mura瑕疵 sample image: (a) conversion result by formula (8); (6) After binarization; (c) Image after optimal ellipse analysis Twelfth image 30 bubbles 瑕疵 sample image: (a) original 瑕疵 image; (the histogram is processed; (c) smoothed and Standard deviation normalized gray-scale value conversion, processing the fourteenth bubble 30 瑕疵 sample image: (a) the conversion result of the formula (8) 〇 value processing; (c) the image after the optimal expansion method The fifteenth figure 30 stripe 瑕 / 疵 sample image: (a) original sputum image; (histogram equalization; (6) smoothed and "standard deviation normalized grayscale value conversion" processing 16th 30th Stripe 瑕疵 sample image: (a) converted by the formula (8) 〇 binarization; (c) image after optimal ellipse analysis, 17th image 30 images of foreign body madness sample: (a) original 瑕Weak image; made) histogram uniform processing; (c) smoothed processing and "standard deviation normalized gray Value conversion, processing 18th image of 30 foreign objects 瑕疵 sample image: (4) by formula ( υ binarization; (c) image after analysis of optimal ellipse method (9) a Utsu 19th image 30 sheets of Mura 瑕疵 sample image The geometric characteristics of the number of 瑕疵 groups, total 瑕疵 area, length, average short axis length and average length and length axis are also twentieth; 3G 瑕 瑕 本 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 3 3 Geometric characteristics such as the ratio of the length of the shaft to the average length and length of the axis 1 The geometric characteristics of the twentieth-- 30 敝瑕 本 本 影像 image, the total thief area length, the average short axis length and the average length and length axis ratio - Twelve ϋ 3G $ 异 职 、 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异 异Group standard 31 200844429 The fourteenth item 3 〇 瑕 瑕 样本 样本 sample image after normalization of various geometric characteristics of the value of the fifteenth Figure 3 〇 瑕疵 瑕疵 瑕疵 瑕疵 sample image normalization of various geometric characteristics of the value of the twentieth 》3 瑕 瑕 瑕 瑕 瑕 影像 影像 影像 影像 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规 正规Network Training 40 Group Training Sample Numbers 29th Figure 80 Group Test Sample Number [Main Component Symbol Description] 32

Claims (1)

200844429 X. Patent application scope: 1. A method using the regional statistical parameters (L〇cai statistics), Best fitting ellipse and nearest neighbor method (K_Nearest Neighb〇r, KNN) to detect and coat The method for automatic optical detection of 瑕/疵 on the surface of the film, including: image enhancement by regional statistical wheat number, and then determining the position of the surface flaw, and extracting the five geometric characteristic values of the surface flaw by the optimum ellipse method, the nearest neighbor The neighbor rule classifies the geometric feature values to identify the type of 瑕疵. r , 2 · The method of claim 1 of the patent scope includes: · Detection and classification of coated brightening film surface fatigue including Uneveness or Mura, Bubble, Streak Four kinds of defects such as Contamination. 3. The method of claim 1 of the scope of the patent application further includes: 〇 _ domain statistical parameters for image enhancement, firstly, the gray-scale value conversion is normalized by regional difference, that is, / (four) = 7~^ 255, S max min ) where /&,;0 represents a pixel point (by the new grayscale value of the transition, ~ is the regional grayscale standard deviation before the pixel is converted, and is, respectively, the whole sheet The maximum and minimum values of the regional gray scale standard deviation before the image is not converted. If % is more than 10, it is specified as 10, and % is more than 1〇, which is also set to 1〇. 33 200844429 4·If the patent application scope The method of item 3, comprising: a subsequent grayscale value of the subsequent order table /〇:,>〇, that is, g(x = right σ Sxy kk〇CT 〇yerage where σ~ represents pixel point b, &gt The regional gray scale h standard deviation calculated by the / (X, y) gray scale value is the average value of the / (U) regional gray scale standard deviation of each pixel of the overall image, that is, Μ Ν ΣΣ σ σ average = , MxN where μ and ν are the number of pixels in the image width and height, and £ and ^ are the specified parameters, if the value is large In 255, the order is 255. 5. The method of the first paragraph of the Shenqing patent scope further includes: The optimum ellipse method for extracting five kinds of geometric features of the surface 瑕疵 includes the number of 瑕疵 groups, the plague area, the length of the long axis, The ratio of the length of the short axis to the length of the short axis. 6. The method of claim 1, further comprising: setting the value of the nearest neighbor method to 1. 34.