TWI771594B - Training method for defect and system using same, and determent method for defect and system using same - Google Patents

Abstract

The training method for defect includes the following steps. Firstly, an image of an optical film is captured, the image includes a defect area. Then, a plurality of coordinate values of a plurality of boundary points of a defect boundary of the defect area along a direction is obtained. Then, an average coordinate value of the coordinate values is obtained. Then, the difference between each coordinate value and the average coordinate value is obtained. Then, based on the largest one of the differences, the area in the image as the defect determent machine learning is determined.

Description

Defect judgment training method and its application system and defect judgment method Laws and systems applying them

The present invention relates to a training method and a system using the same, a judging method and a system using the same, and more particularly, a training method for judging defects and a system using the same, and a method for judging defects and a system using the same.

The conventional optical film must be manually observed to observe whether the optical film has defects and the types of defects after the production is completed. However, the artificial naked eye observation method is prone to misjudgment. Therefore, it is one of the goals of those skilled in the art to propose a new defect judgment technology that can increase the judgment accuracy.

Embodiments of the present invention provide a defect judgment training method and a system using the same, as well as a defect judgment method and a system using the same, which can improve the above problems.

An embodiment of the present invention provides a defect judgment training method. The defect judgment training method includes the following steps. Capture an image of an optical film, the image includes a defect area; obtain a plurality of coordinate values along a direction of a plurality of boundary points of a defect boundary of the defect area; obtain One of these coordinate values is an average coordinate value; a difference between each coordinate value and the average coordinate value is obtained; and, according to the largest of these difference values, an area in the image to be used for defect judgment machine learning is determined.

Another embodiment of the present invention provides a defect judgment training system. The defect judgment training system includes a camera and a defect judgment machine learner. The camera is used for capturing an image of an optical film, and the image includes a defect area. The defect judgment machine learner is used to: obtain a plurality of coordinate values along a direction of a plurality of boundary points of a defect boundary of a defect area; obtain an average coordinate value of these coordinate values; obtain a difference between each coordinate value and the average coordinate value value; and, according to the largest of these differences, determine the region in the image to be used for defect judgment machine learning.

An embodiment of the present invention provides a defect judgment method. The defect judgment method includes the following steps. Capture an image of an optical film, the image has a defect area to be determined; analyze the defect area to be determined, and generate a similarity score of the defect area to be determined relative to each of several defect types; determine the similarity Whether the defect type corresponding to the highest degree score belongs to a specific one of these defect types; when the defect type corresponding to the highest degree belongs to a specific one, determine whether the highest one is greater than a preset value; when the highest one is greater than the preset value , determine that the defect area to be determined belongs to a specific one; and, when the highest value is not greater than the preset value, determine that the defect area to be determined belongs to another of these defect types.

An embodiment of the present invention provides a defect judgment system. The defect judging system includes a camera and a defect judging device. The camera is used for capturing an image of an optical film, and the image has a defect area to be determined. The defect judger is used to: analyze the defect area to be judged, and generate a similarity score of the defect area to be judged relative to each of several defect types; judge this Whether the defect type corresponding to the highest similarity score belongs to one of these defect types; when the defect type corresponding to the highest belongs to a specific type, determine whether the highest value is greater than a preset value; Set the value to determine that the defect area to be determined belongs to a specific one; and, when the highest value is not greater than the preset value, determine that the defect area to be determined belongs to another of these defect types.

In order to have a better understanding of the above-mentioned and other aspects of the present invention, the following specific examples are given and described in detail in conjunction with the accompanying drawings as follows:

10: Optical film

100: Defect Judgment Training System

110, 210: Camera

120: Image cropper

130: Defect Judgment Machine Learner

200: Defect Judgment System

220: Defect Judger

A, B, C: boundary points

B _U : upper boundary

B _L : lower boundary

D1: Defective area

L: Boundary reference line

M: Defect Judgment Model

P:image

P1: local area

P _f : Optical film image area

P _b : background area

S: where the grayscale value is different

S110~S153H, S210~S270: Steps

Y _A , Y _B , Y _C : Coordinate values

Y _AV : Average coordinate value

Y _p : image height

△Y _AV : Difference

△Y _L : Lower offset

△Y: Upper offset

σ _L : lower limit of standard deviation

σ: standard deviation value

σ _U : standard deviation upper limit

FIG. 1 is a functional block diagram of a defect judgment training system according to an embodiment of the present invention.

Figures 2A to 2F illustrate images of several types of actual damage defects that may occur in optical films according to embodiments of the present invention.

Figures 2G to 2H illustrate images of several types of non-defective defects that may occur in optical films according to embodiments of the present invention.

FIGS. 3A to 3B are flowcharts of the defect judgment method of the defect judgment training system of FIG. 1 .

FIG. 4 is a functional block diagram of a defect judgment system according to an embodiment of the present invention.

FIG. 5 is a flowchart showing a defect judgment method of the defect judgment system of FIG. 4 .

For a better understanding of the above and other aspects of the present invention, the following Examples are given and described in detail with the accompanying drawings as follows.

Please refer to FIGS. 1 and 2A~2H. FIG. 1 shows a functional block diagram of a defect judgment training system 100 according to an embodiment of the present invention, and FIGS. 2A to 2F show possible occurrences of an optical film according to an embodiment of the present invention. 2G~2H are images of several types of non-physical defects that may occur in the optical film according to the embodiment of the present invention.

The defect judgment training system 100 includes a camera 110 , an image cutter 120 and a defect judgment machine learner 130 . The camera 110 is used to capture the image P of the optical film, and the image P includes the defect area D1. The image cutter 120 is used for analyzing the characteristics of the defect area D1 of the image P, so as to determine the image P as an input area of the defect judgment machine learning device 130 , wherein the area includes the defect area D1 . The defect judging machine learning device 130 is used to analyze the determined area to learn the characteristics of the defect area D1. Since the defect type of the defect area D1 is known, the defect judgment machine learning device 130 can improve the judgment of the defect type of the defect area D1 by analyzing the characteristics of the defect area D1.

The image cutter 120 and the defect judgment machine learner 130 are circuit structures formed by, for example, semiconductor manufacturing processes. In one embodiment, the image cutter 120 and the defect judgment machine learning device 130 may be integrated into a single component or integrated into a processor.

The optical film 10 can be a single-layer or multi-layer film, including optical gain, alignment, compensation, steering, orthogonality, diffusion, protection, anti-sticking, scratch resistance, anti-glare, reflection suppression, high refractive index, etc. The film, for example, can be a polarizing film, a release film, a wide viewing angle film, a brightness enhancement film, a reflective film, a protective film, a Alignment liquid crystal film, hard coat film, anti-reflection film, anti-film, diffusion film, anti-glare film and other surface-treated films with compensation or birefringence and other characteristics, or a combination of the above, but not limited to this.

FIGS. 3A to 3B are flowcharts of the defect judgment method of the defect judgment training system of FIG. 1 .

In step S110, the camera 110 captures an image P of the side surface of the optical film, and the image P includes the defect area D1. The defect area D1 is a known defect type judged manually. In terms of actual damage defects, the defect area D1 of the image P shown in Figure 2A is a convex film type defect, and the defect area D1 of the image P shown in Figure 2B is a dander type defect (for example, it is based on cutting the edge of the optical film. generated dander), the defect area D1 of the image P shown in Figure 2C is a laser starting point type defect (for example, a cutting starting point defect based on the requirement of using a laser to cut optical films), Figure 2D The defect area D1 of the image P shown is a bubble-type defect (for example, based on the bubbles generated at the bonding interface of the multilayer optical film), the defect area D1 of the image P shown in Fig. 2E is a crack-type defect, and the defect area shown in Fig. 2F. The defect area D1 of the image P is shown as a contamination type defect. In terms of non-physical defects, the defect area D1 of the image P shown in FIG. 2G is a mark-type defect (a mark deliberately formed on the optical film 10 based on process requirements), while the defect of the image P shown in FIG. 2H is a mark-type defect. Region D1 is a bond-type defect (bonded feature based on the need for bonding two rolls of optical film).

The embodiment of the present invention is not limited to the aforementioned defect types. In another embodiment, the defect judgment training system can handle more defect types, but also fewer defect types.

In step S120 , taking FIG. 2A as an example, the image cutter 120 obtains several coordinate values Y _A , Y of several boundary points A, B and C of the defect boundary of the defect area D1 along a direction Y _B and Y _C , wherein the direction Y is, for example, perpendicular to the boundary reference line L of the optical film image area P _f of the image P. FIG. The defect area D1 protrudes from the boundary reference line L (eg, FIG. 2A ) or sinks in (eg, FIG. 2C ).

The image cropper 120 may binarize the image P. After binarization, each pixel of the optical film image area P _f of the image P has the first gray-scale value (shown by the oblique cross-section in Figure 2A), and each pixel of the background area P _b of the image P has the first grayscale value. Two grayscale values (represented by a dotted section in FIG. 2A ), wherein the first grayscale value is different from the second grayscale value. In this way, the image cutter 120 can distinguish the optical film image area P _f and the background area P _b through the difference in grayscale values, and determine the defect boundary of the defect area D1 . The image cutter 120 extends along the direction Y with the direction X where the relative change "relatively gentle" in the gray-scale value difference between the optical film image area P _f and the background area P _b as the boundary reference line L. The aforementioned definition of “softer” complies with the requirements: S=|the first grayscale value of P _f -the second grayscale value of _Pb |, where S>the second grayscale value of _Pb , and S <The first gray-scale value of P _f , and S<(the first gray-scale value of P _f + the second gray-scale value of P _b )/2. The boundary reference line L is, for example, the extension of the contour line image of the outer surface of the optical film 10 , that is, the boundary line between the optical film image area P _f and the background area P _b . When the image P has no convex film-type defects (such as Fig. 2A) and dander-type defects (such as Fig. 2B), the outer surface contour line of the optical film image area P _f is, for example, a straight line, such as the image in Fig. 2C P is a horizontal line.

The number of boundary points in the embodiment of the present invention is described by taking three as an example, which is not intended to limit the embodiment of the present invention. In another embodiment, the number of boundary points may be less or more than three.

In step S130 , the image cutter 120 calculates the coordinate values Y _A , Y _B and Y _C to obtain the average coordinate value Y _AV of the coordinate values Y _A , Y _B and Y _C .

In step S140, the image cutter 120 obtains the difference value ΔY _AV between each coordinate value Y _A , Y _B and Y _C and the average coordinate value Y _AV . This difference ΔY _AV is, for example, an absolute value.

In step S150 , the image cutter 120 determines the learning area (learning object) in the image P as the learning area (learning object) for the defect judgment machine learning according to the largest of these differences ΔY _AV . Step S150 can be completed by the following steps.

In step S151 , the image cutter 120 determines whether the largest of the difference values ΔY _AV falls within a preset range; if not, the process proceeds to step S152 ; if so, the process proceeds to step S153A.

Taking Fig. 2A as an example, the difference ΔY _AV between the coordinate value Y _B of the image P and the average coordinate value Y _AV (ie, Y _B -Y _AV =ΔY _AV ) is the maximum value of all the difference values ΔY _AV Otherwise, it exceeds the preset range, wherein the preset range is the range between the upper boundary _BU and the lower boundary _BL in FIG. 2A . The upper boundary B _U is the sum of the average coordinate value Y _AV and the upper offset ΔY _U (ie, Y _AV +ΔY _U ), and the lower boundary _BL is the average coordinate value Y _AV and the lower offset Δ The difference in Y _L (ie, Y _AV - ΔY _U ). As shown in FIG. 2A , since the maximum difference ΔY _AV falls within the preset range, it indicates that the size of the defect area D1 is large enough to not affect the accuracy of the machine learning, so the process proceeds to step S152 .

In step S152, since the size of the defect area D1 in the picture 2A is large enough to affect the accuracy of the machine learning, it is not necessary to cut the image P in the picture 2A, and the image cutter 120 directly uses the entire The image P is input to the defect judgment machine learner 130 . In order to speed up the processing speed, the defect judgment machine learning device 130 may first reduce the resolution (size) of the input image, and then perform the training of defect judgment. Since the size of the defect area D1 in FIG. 2A is large enough, even if the defect judgment machine learning device 130 is reducing the size of the entire image P Defect judgment training after inch, still does not affect the accuracy of machine learning.

In addition, the defect judgment machine learner 130 may use machine learning technology to perform defect judgment training. Specific machine learning techniques such as: deep neural network (Deep Neural Networks, DNN), support vector machine (SVM), decision tree (decision tree), ensemble (ensemble), K-neighbor method (K-NN), linear regression ( Linear regression), Bayesian probability, neural network (neural network), Logis regression, perceptron (perceptron) or correlation vector machine (relevance vector machine, RVM) and other algorithms. In some embodiments, for example, Deep Neural Networks (DNN) or Support Vector Machines (SVM). The feature classification technology has the advantages of fast operation speed and low technical cost, while the machine learning classification technology has a very high recognition accuracy rate and can continuously (on the quality inspection line) correct the machine learning model to further improve the recognition accuracy rate.

In step S151 , when the largest of these differences ΔY _AV does not fall within the predetermined range, it means that the size of the defect area D1 may be too small, and further enlargement processing is required. Taking Fig. 2D as an example, the maximum difference ΔY _AV falls within the preset range (the range between the upper boundary _BU and the lower boundary _BL ), indicating that the size of the defect area D1 may be too small, so the flow enters the step S153A, the image trimmer 120 further determines whether to trim the image P and the size of the trimmed local area (if it is to be trimmed).

In step S153A, the image cutter 120 obtains the standard deviation value σ of the coordinate values Y _A , Y _B and Y _C .

In step S153B, the image cutter 120 determines whether the standard deviation value σ is equal to or smaller than the standard deviation lower limit value σ _L . If yes, the flow goes to step S153C; if not, the flow goes to step S153E. The aforementioned lower limit value σ _L of the standard deviation is, for example, 1% of the image height Y _p of the image P, but the embodiment of the present invention is not limited to this.

In step S153C, the image trimmer 120 trims a local area of the image P, wherein the ratio of the area of the local area to the area of the image P is within the first cropping scale range, wherein the first cropping scale range is between between 30% and 60%.

Taking the 2D image as an example, since the maximum difference ΔY _AV falls within the preset range, the image cutter 120 cuts the local area P1 of the image P in the 2D image as the machine learning object ( Input defect judgment machine learner 130). The local area P1 is within the first crop ratio range R1 of the entire area of the image P, wherein the defect area D1 is entirely located within the local area P1. The proportion of the defective area D1 in the local area P1 is greater than the proportion of the defective area D1 in the image P, which highlights the characteristics of the defective area D1 in the local area P1. In other words, compared with the whole image P, the defect area D1 in the partial area P1 seems to be enlarged, which highlights the characteristics of the defect area D1. In an embodiment, the local area P1 may be the area with the average coordinate value Y _AV as the center line up, down, or both up and down, wherein in the case of simultaneous down and down, the up ratio may be approximately It is equal to the proportion below, but it can also be different.

In step S153D, the image cropper 120 inputs the cropped partial region P1 to the defect judgment machine learner 130. When the cropped local area P1 is used as the learning object (input to the defect judgment machine learner 130), even if the defect judgment machine learner 130 performs defect judgment training after reducing the size of the local area P1, it does not affect the increase in machine learning. accuracy.

In step S153E, the image cutter 120 determines whether the standard deviation value σ is equal to or greater than the standard deviation upper limit value σ _U . If it is, it means that the height fluctuation of the defect area D1 is too large, it is judged as a large size defect, and the image P does not need to be cropped, so the flow goes to step S153F. In step S153F, the image cutter 120 inputs the entire image P to the defect judgment machine learner 130. The aforementioned upper limit value σ _U of the standard deviation is, for example, 2% of the image height Y _p , but the embodiment of the present invention is not limited to this.

If the image cropper 120 determines whether the standard deviation value σ is smaller than the standard deviation upper limit value σ _U , it means that the standard deviation value σ is between the standard deviation lower limit value σ _L and the standard deviation upper limit value σ _U . The image cutter 120 determines that the defect area D1 has very small fluctuations in height and is a very small size defect, so the process proceeds to step S153G.

In step S153G, the image trimmer 120 trims a local area of the image P.

Taking FIG. 2F as an example, since the standard deviation value σ of the defect area D1 is between the lower standard deviation value σ _L and the upper standard deviation value σ _U , the image cutter 120 trims the image of FIG. 2F . The local area P1 like P is used as a machine learning object (input defect judgment machine learner 130). The partial area P1 is within the second crop ratio range R2 of the entire area of the image P, wherein the defect area D1 is entirely located within the partial area P1. The second cutting ratio range R2 is smaller than the aforementioned first cutting ratio range R1, for example, the second cutting ratio range is between 6% and 20%. In an embodiment, the local area P1 of the 2F diagram may be the area with the average coordinate value Y _AV as the center line up, down, or both up and down, wherein in the case of simultaneous down and down, up can be roughly equal to the scale down.

In step S153H, the image cropper 120 inputs the cropped partial region P1 to the defect judgment machine learner 130. When the cropped local area P1 is used as the learning object (input to the defect judgment machine learner 130), even if the defect judgment machine learner 130 reduces the size of the local area P1, it does not affect the accuracy of the machine learning. Spend.

To sum up, after analyzing the defect area D1 of the image P in Figures 2A to 2H using the aforementioned steps, when the size of the defect area D1 is large enough (the defect judgment machine learner 130 will not negatively affect the machine learning after reducing the image size). accuracy), the image cutter 120 performs defect judgment training on the entire image P (eg, steps S152 and S153F). When the size of the defect area D1 is not large enough (the defect judgment machine learner 130 may negatively affect the accuracy of machine learning after reducing the image size), the image cutter 120 cuts a local area of the image P, so that the The defect area D1 seems to be enlarged. Even if the defect judgment machine learner 130 performs defect judgment training after reducing the size of the image, the accuracy of machine learning will not be adversely affected. In addition, depending on the standard deviation of several coordinate values of the defect area D1, the local area may be within the first cropping scale range R1 (step S153C) or the second cropping scale range R2 (step S153C) of the entire area of the image P S153G).

Please refer to FIGS. 4 and 5. FIG. 4 shows a functional block diagram of a defect judgment system 200 according to an embodiment of the present invention, and FIG. 5 shows a flowchart of a defect judgment method of the defect judgment system 200 of FIG. 4. .

As shown in FIG. 4 , the defect judgment system 200 includes a camera 210 and a defect judgment unit 220 . The defect judger 220 is, for example, a circuit structure formed by a semiconductor process. In one embodiment, the defect determiner 220 may be integrated into a processor.

The camera 210 is used for capturing an image of the optical film 10 at the process site, such as an automated optical inspection (Automated Optical Inspection, AOI) system. The defect judging machine learner 130 uses several images P to judge the defects through the above process. After the training is completed, the defect judgment machine learner 130 generates a defect judgment model M (shown in FIG. 1 ). The defect judging unit 220 can judge the defect type of the image captured by the camera 210 according to the defect judging model M. The judgment criteria of the defect judger 220 are shown in Table 1 below, and the following is a further example to illustrate the flow chart of FIG. 5 .

In step S210, the camera 210 captures an image of the optical film 10 at the process site, and the image has a defect area to be determined.

In step S220, the defect determiner 220 analyzes the defect area to be determined, and generates a similarity score of the defect area to be determined with respect to each of several defect types. In this embodiment, the similarity score of the defect area to be judged relative to the mark-type defect is used as all the similarity scores. For example, the highest number.

In step S230, the defect determiner 220 determines whether the defect type corresponding to the highest similarity score belongs to a specific one of the defect types. If yes, the flow goes to step S240; if no, the defect determiner 220 directly determines that the defect area to be determined belongs to the defect type corresponding to the highest similarity score.

Taking the specific one as a marker-type defect for example, the defect type corresponding to the highest similarity score of the defect area to be determined also belongs to the marker-type defect, and the process proceeds to step S240.

In another embodiment, a specific one is a marking type defect and the defect type corresponding to the highest similarity score is a convex film type defect. For example, since the convex film type defect does not belong to the specific one (marking type defect), therefore The flow proceeds to step S270. In step S270, according to the above Table 1, the defect judger 220 directly judges that the defect area to be judged belongs to the defect type corresponding to the highest similarity score, which is a convex film type defect in this example.

In step S240, the defect determiner 220 determines whether the highest similarity score is greater than a predetermined value. Take a specific person as a marker defect and the defect type corresponding to the highest similarity score is also a marker defect. For example, when the highest similarity score (such as 0.8) is greater than the preset value (Table 1 takes 0.4 as an example), the process Going to step S250, the defect judger 220 judges that the defect area to be judged belongs to the specific one, that is, it belongs to the mark type defect.

When the highest similarity score (such as 0.8) is not greater than the preset value (0.4 is used as an example in Table 1), the process proceeds to step S260, and the defect determiner 220 determines that the defect area to be determined belongs to the other of these defect types. The other is similar to the defect characteristics of that particular one. For example, since the characteristics of mark-type defects and contamination-type defects are close, In terms of type defects, in addition to the highest similarity score relative to marked type defects, the similarity score must be higher than 0.4, and the defect judgment machine learner 130 will determine that the defect area to be judged belongs to non-actually harmful marked type defects; If the similarity score relative to the marked defect is the highest, but the score is less than 0.4, the defect judgment machine learner 130 judges that the defect area to be judged belongs to the dirty defect. Similarly, as shown in Table 1, the characteristics of bonding-type defects and bubble-type defects are similar, so it can also be determined that the defect area to be judged belongs to bonding-type defects or bubble-type defects by a similar method.

To sum up, it can be seen that the conditions for judging that the to-be-determined defect area of the image belongs to the non-actual defect type (such as mark-type defects and joint-type defects) are relatively strict. For example, in addition to the highest similarity score relative to the non-actual defect type, the similarity score must be higher than the preset value (the specific value is not limited by the above table 1), which can prevent the actual defect type from being misjudged as non-actually harmful. Defect type.

To sum up, although the present invention has been disclosed by the above embodiments, it is not intended to limit the present invention. Those skilled in the art to which the present invention pertains can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the scope of the appended patent application.

S110~S153D: Steps

Claims (27)

A defect judgment training method, comprising: capturing an image of a side surface of an optical film, the image including a defect area and a boundary reference line, the defect area protruding or indented relative to the boundary reference line; obtaining the a plurality of coordinate values along a direction of a plurality of boundary points of a defect boundary of a defective area; obtain an average coordinate value of the coordinate values; obtain a difference between each of the coordinate values and the average coordinate value; according to the differences The largest value determines the area in the image that is used for defect judgment machine learning. The defect judgment training method as described in claim 1, wherein the step of determining the region in the image to be learned by the defect judgment machine according to the maximum value comprises: judging whether the maximum value of the differences is less than a a preset range; when the maximum is smaller than the preset range, crop a local area of the image, wherein the defective area is located in the local area; and input the local area into a defect judgment machine learner. According to the defect judgment training method described in item 1 of the scope of application, the step of determining the region in the image as the defect judgment machine learning area according to the maximum value includes: judging whether the maximum value of the difference values is less than a predetermined value setting a range; and when the maximum is not less than the preset range, inputting the entire image into a defect judgment machine learner. The defect judgment training method described in claim 2 or 3, wherein in the step of judging whether the largest of the differences is smaller than the predetermined range, the predetermined range is between an upper boundary and a lower boundary In between, the upper boundary is the sum of the average coordinate value and an upper offset, and the lower boundary is the difference between the average coordinate and a lower offset. The defect judgment training method as described in claim 2, wherein the step of cropping the local area of the image includes: obtaining a standard deviation value of the coordinate values; judging whether the standard deviation value is equal to or less than a lower standard deviation value; and when the standard deviation value is equal to or less than the lower standard deviation value, crop the local area of the image, wherein the ratio of the area of the local area to the area of the image is between a and the defect area is located in the local area. The defect judgment training method described in item 5 of the scope of the patent application, wherein the range of the first cropping ratio is between 30% and 60%. The defect judgment training method as described in claim 2, wherein the step of cropping the local area of the image comprises: obtaining a standard deviation value of the coordinate values; judging whether the standard deviation value is between a between a lower standard deviation value and an upper standard deviation value; and when the standard deviation value is between the lower standard deviation value and the upper standard deviation value, capturing the local area of the image, wherein The ratio of the area of the local area to the area of the image is between a second cropping ratio range, and the defect area is located in the local area. The defect judgment training method as described in item 7 of the scope of the application, wherein the range of the second cropping ratio is between 6% and 20%. The defect judgment training method described in claim 7, wherein the lower standard deviation value is 1% of the height of an image of the image; and the upper standard deviation value is the image of the image 2% of height. The defect judgment training method described in item 1 of the claimed scope further comprises: binarizing the image so that each pixel in the optical film image area of the image has a first grayscale value, and the image Each pixel of a background area has a second grayscale value; taking the absolute value of the difference between the first grayscale value and the second grayscale value as a grayscale value difference; and according to the grayscale value The extension of the relative change in the difference is the extension of the boundary reference line. The defect judgment training method as described in item 10 of the scope of the patent application, wherein the moderate point satisfies the following requirements: the difference between the grayscale values is greater than the first grayscale value; the difference between the grayscale values is smaller than the second grayscale value and the difference between the gray level values is smaller than the average value of the first gray level value and the second gray level value. The defect judgment training method as described in claim 2, wherein the step of cropping the local area of the image comprises: obtaining a standard deviation value of the direction coordinate values; judging whether the standard deviation value is greater than or is equal to an upper standard deviation value; and When the standard deviation value is between a lower standard deviation value and an upper standard deviation value, the entire image is input into the defect judgment machine learner. A defect judgment training system, comprising: a camera for capturing an image of a side surface of an optical film, the image including a defect area and a boundary reference line, the defect area protruding relative to the boundary reference line or indentation; and a defect judging machine learning device for: obtaining a plurality of coordinate values along a direction of a plurality of boundary points of a defect boundary of the defect area; obtaining an average coordinate value of the coordinate values; obtaining each A difference between the coordinate value and the average coordinate value; and according to the largest of the difference values, determine an area in the image to be learned by the defect judgment machine. The defect judgment training system as described in claim 13, wherein in the step of determining the area in the image to be learned by the defect judgment machine according to the largest one, the defect judgment machine learner is further used for: judging the whether the largest of these differences is less than a predetermined range; when the largest is less than the predetermined range, crop a partial area of the image, wherein the defective area is located in the partial area; and input the partial area in A defect judgment machine learner. The defect judgment training system as described in item 14 of the scope of the application, wherein in the step of determining the area in the image to be learned by the defect judgment machine according to the largest one, the defect judgment machine learner is further used for: judging these whether the largest of the differences is less than a predetermined range; and when the largest is not less than the predetermined range, input the entire image into the defect judgment machine learner. The defect judgment training system according to claim 14 or 15, wherein the predetermined range is between an upper boundary and a lower boundary, and the upper boundary is the sum of the average coordinate value and an upper offset , and the lower boundary is the difference between the average coordinate value and the lower offset. The defect judgment training system as described in claim 14, wherein in the step of cropping the local area of the image, the defect judgment machine learner is further configured to: obtain a standard deviation value of the coordinate values ; determine whether the standard deviation value is equal to or less than a standard deviation lower limit value; and when the standard deviation value is equal to or less than the standard deviation lower limit value, crop the local area of the image, wherein the area of the local area accounts for The ratio of the area of the image is between a first cropping ratio range, and the defect area is located in a local area. The defect judgment training system according to item 17 of the scope of application, wherein the range of the first cutting ratio is between 30% and 60%. The defect judgment training system as described in claim 14, wherein in the step of cropping the local area of the image, the defect judgment machine learner is further configured to: obtain a standard deviation value of the coordinate values ; determine whether the standard deviation value is between a lower standard deviation value and an upper standard deviation value; and when the standard deviation value is between the lower standard deviation value and the upper standard deviation value, retrieve The partial area of the image is taken, wherein the ratio of the area of the partial area to the area of the image is between a second cropping scale range, and the defective area is located in the partial area. The defect judgment training system according to item 19 of the scope of application, wherein the second cropping ratio ranges from 6% to 20%. The defect judgment training system as described in claim 14, wherein in the step of cropping the local area of the image, the defect judgment machine learner is further used for: obtaining a standard deviation of the direction coordinate values value; determine whether the standard deviation value is greater than or equal to an upper standard deviation value; and when the standard deviation value is between a lower standard deviation value and the upper standard deviation value, input the entire image into the defect Judge the machine learner. A defect judgment method, comprising: capturing an image of an optical film, the image having a defect area to be judged; and performing the following steps according to a defect judgment model: Analyze the defect area to be determined, and generate a similarity score of the defect area to be determined relative to each of a plurality of defect types; determine whether the defect type corresponding to the highest similarity score belongs to the defect types a specific one; when the defect type corresponding to the highest one belongs to the specific one, it is determined whether the highest one is greater than a preset value; when the highest one is greater than the preset value, it is determined that the defect area to be determined belongs to the specific one and when the highest value is not greater than the preset value, it is determined that the defect area to be determined belongs to the other of the defect types; wherein, the defect determination model adopts the defect determination as described in item 1 of the scope of the patent application generated by the training method. The defect judging method as described in item 22 of the patent application scope, wherein the specific one and the other have similar defect characteristics. The defect judging method described in item 22 of the scope of the application further comprises: when the defect type corresponding to the highest one does not belong to the specific one, directly determining that the defect area to be determined belongs to the defect type corresponding to the highest one . A defect judgment system, comprising: a camera for capturing an image of an optical film, the image having a defect area to be judged; and a defect judger for executing according to a defect judgment model: Analyze the defect area to be determined, and generate a similarity score of the defect area to be determined relative to each of a plurality of defect types; determine whether the defect type corresponding to the highest similarity score belongs to the defect types a specific one; when the defect type corresponding to the highest one belongs to the specific one, it is determined whether the highest one is greater than a preset value; when the highest one is greater than the preset value, it is determined that the defect area to be determined belongs to the specific one and when the highest value is not greater than the default value, it is determined that the defect area to be determined belongs to the other of the defect types; wherein, the defect determination model adopts the defect determination as described in item 1 of the scope of the patent application generated by the training method. The defect judging system as described in claim 25, wherein the specific one and the other one have similar defect characteristics. The defect judging system as described in item 25 of the scope of application, wherein the defect judging device is further used for: when the defect type corresponding to the highest one does not belong to the specific one, directly determine that the to-be-determined defective area belongs to the highest one The corresponding defect type.

Images