TW202248628A - Inspection device, inspection method, manufacturing method of glass plates, inspection program, and computer program product - Google Patents

Abstract

The present invention strikes a balance between a reduction in personnel costs of detecting the size of a flaw generated in a glass plate and maintenance of detection accuracy. An inspection device (1) includes: an image acquisition unit (101) that obtains an acquired image of a glass plate; and a size detection unit (103) that inputs the acquired image to a size detection model (112), which has learned so as to infer the position and a range of a flaw generated in the glass plate, and that detects the size of the range in an obtained inference result, as the size of the flaw.

Description

Inspection device, inspection method, glass plate manufacturing method, inspection program, and computer program product

The present invention relates to an inspection device and the like for inspecting defects of a glass plate based on an image obtained by taking a glass plate.

Conventionally, defect inspection of glass sheets in the manufacture of glass sheets, especially size inspection of defects generated in glass sheets, was performed by visual inspection, and thus labor costs were inevitably incurred. In view of such a problem, Patent Document 1 discloses a technique of performing image processing on an image obtained by capturing an end surface of a glass plate to detect the size of a defect. [Prior art literature] [Patent Document]

[Patent Document 1] International Publication No. 2004/079352

[Problem to be Solved by the Invention] In the prior art as described above, the inventors of the present application found a problem that the size and the like of a defect may not be detected with sufficient accuracy. One of the causes of the above-mentioned problems is considered to be that the depth of the periphery of the defect is often unclear in the image obtained by capturing the glass plate.

The object of the present invention is to provide an inspection device, etc., which can balance the reduction of labor costs and the maintenance of detection accuracy in the size detection of defects generated in glass plates. [Means to solve the problem]

In order to solve the above-mentioned problems, an inspection device according to an aspect of the present invention includes: an image acquisition unit that acquires a captured image of a glass plate; In the inference result, the size of the range is taken as the size of the defect generated on the glass plate, and the learned model calculates the position of the defect generated on the glass plate and the range. and scope were studied.

In order to solve the above-mentioned problems, an inspection method according to an aspect of the present invention is performed by an inspection device, and includes: an image acquisition step of acquiring a captured image obtained by photographing a glass plate; and a size detection step of detecting a learned model. The size of the range in the inference result obtained by the captured image is input as the size of the defect generated on the glass plate captured, and the learned model is used to infer the position of the defect generated on the glass plate and the The location and range are learned in a range-based manner.

In order to solve the above-mentioned problems, a method for manufacturing a glass plate according to an aspect of the present invention includes a step of molding an original glass plate into a glass plate of a predetermined size, and an inspection step of the glass plate performed by an inspection device, and the The inspection step includes: an image acquisition step of acquiring a photographed image of the glass plate; and a size detection step of detecting the size of a range in an inference result obtained by inputting the photographed image to the learned model as The size of the defect generated on the glass plate is imaged, and the learned model learns the position and range of the defect generated on the glass plate so as to infer the position and range of the defect generated on the glass plate. [Effect of the invention]

According to an aspect of the present invention, in the dimension detection of the defect generated in the glass plate, both the reduction of labor cost and the maintenance of detection accuracy can be taken into account.

[Embodiment 1] (check system 100) Hereinafter, an embodiment of the present invention will be described in detail. FIG. 1 is a block diagram showing an overview of an inspection system 100 according to the present embodiment and an example of a configuration of main parts of an inspection device 1 included in the inspection system 100 .

The inspection system 100 is a system for detecting the size of a defect generated on an end surface of a glass sheet. In addition, in this embodiment, although the glass plate is demonstrated as a rectangular shape, the shape of a glass plate is not limited to this example. The so-called defects are, for example, cracks, chipping, etc. generated on the end surfaces of the glass plate. The inspection system 100 includes an inspection device 1 , an imaging device 2 , an image memory device 3 and a test result memory device 4 .

The inspection apparatus 1 is an apparatus which detects the size of the defect which arises in the end surface of a glass plate. For the end face of a glass plate, the size inspection of defects has been performed visually, and in the size inspection using image processing, it may be difficult to perform accurate size inspection because the depth of the periphery of the defect in the captured image is not clear. Although the details will be described later, the inspection device 1 detects, as the size of the defect, the size of a range in the inference result of the defect obtained by inputting the captured image to the learned model that infers the defect generated on the end surface. The location and scope of the method are studied. That is, since human beings are not involved in the dimensional inspection of the defect on the end surface, the labor cost required for this inspection can be reduced. Furthermore, in the learning of the learned model, by using the images of the blurred end faces around the defects as teaching materials for many times of learning, the images of the dark and dark end faces around the defects can also be highly improved. Detect defects with precision. Therefore, in the size detection of the defect generated on the end surface of the glass plate, both the reduction of labor cost and the maintenance of detection accuracy can be taken into account.

The imaging device 2 is a device for imaging the end surface of the glass plate. Here, the term "end surface" refers to a side surface part when the widest both surfaces of the glass plate are set as an upper surface and a lower surface, respectively, and may also be referred to as an outer peripheral part. Hereinafter, the four end surfaces of the glass plate may be respectively expressed as an X1 end surface, an X2 end surface, a Y1 end surface, and a Y2 end surface. Furthermore, let the X1 end surface and the X2 end surface be mutually parallel end surfaces, and the Y1 end surface and the Y2 end surface be mutually parallel end surfaces.

As an example, the imaging device 2 is arranged along a conveying path of a conveying device (not shown) that conveys a glass plate, and continuously images one end surface of a glass plate being conveyed by the conveying device at predetermined time intervals. In this manner, by taking multiple shots of one end face, a plurality of captured images in which one end face is divided into a plurality can be obtained, and the resolution of the end face of each captured image can be sufficiently high for inspection. Of course, if a captured image with sufficient resolution can be obtained even if the entire end surface is captured by one capture, the number of captures may be one.

Only one camera device 2 is shown in FIG. 1 , but the camera device 2 can also be arranged for each end face of the glass plate. For example, two imaging devices 2 may be arranged so as to face each other across the conveyance path, and simultaneously image parallel end faces (for example, X1 end face and X2 end face, or Y1 end face and Y2 end face). Moreover, when the conveyance path branches, the imaging device 2 may be arrange|positioned for every branch.

Moreover, although the illustration is omitted, the imaging device 2 may be accompanied by an illumination device and an information processing device. On the one hand, the illumination device irradiates light to the end surface of the glass plate, and on the other hand, the imaging device 2 takes pictures of the end surface.

The information processing device is a device that adds additional information to the captured image captured by the imaging device 2 and stores it in the image memory device 3 . The additional information includes glass identification information indicating the glass plate captured in the captured image, and glass position information indicating which part of the glass plate is captured in the captured image.

The glass identification information may be, for example, an identification number given to each glass plate. The glass position information only needs to be information indicating what part of the glass plate is. For example, when the conveying speed of the glass plate is constant and the number of photographs taken at each position of the end surface is also constant, the information indicating which photographed image was obtained may be used as the glass position information.

The image memory device 3 is a memory device for storing captured images captured by the imaging device 2 . Furthermore, the detection result storage device 4 is a storage device that stores the detection result detected by the inspection device 1 , that is, information indicating the size of a defect. Furthermore, multiple image memory devices 3 and test result memory devices 4 may be provided respectively. For example, an image memory device 3 and a test result memory device 4 may also be provided for each end surface of the glass plate. Furthermore, the image storage device 3 and the detection result storage device 4 may also be omitted. In this case, it is only necessary to cause the imaging device 2 to generate a captured image to the inspection device 1 and cause the inspection device 1 to store the detection result in the storage unit 11 .

(check device 1) As shown in FIG. 1 , the inspection device 1 includes a control unit 10 , a memory unit 11 and a communication unit 12 . The control unit 10 comprehensively controls each unit of the inspection device 1 . The storage unit 11 stores various data used by the inspection device 1 . The communication unit 12 is used for the inspection device 1 to communicate with other devices. Typical examples of these other devices are the image memory device 3 and the detection result memory device 4 .

As shown in FIG. 1 , the control unit 10 includes an image acquisition unit 101 , a defect determination unit 102 , and a size detection unit 103 . The memory unit 11 memorizes a defect determination model 111 and a size detection model 112 .

The image acquisition part 101 acquires the captured image which captured the glass plate. In the present embodiment, the captured image acquired by the image acquiring unit 101 is a captured image obtained by imaging the end surface of the glass plate by the imaging device 2 as described above. As an example, the image acquisition unit 101 receives captured images from the image memory device 3 via the communication unit 12 .

The defect determination unit 102 determines whether or not there is a defect in the end surface captured in the captured image. Specifically, the defect determination unit 102 inputs the captured image acquired from the image acquisition unit 101 to the defect determination model 111 , and determines the presence or absence of a defect based on the inference result output from the defect determination model 111 .

Here, the defect determination model 111 will be described. The defect judgment model 111 is a learned model that has been learned so as to infer whether there is a defect in the captured image obtained by capturing the end surface of the glass plate. Such a defect determination model 111 is constructed by machine learning using a plurality of captured images whose presence or absence of defects is known as teaching data. The machine learning algorithm is not particularly limited as long as it can generate the defect determination model 111 capable of classifying captured images into two types of defective and non-defective. For example, a deep learning convolutional neural network (convolution neural network) with high image classification accuracy is suitable, but it is not limited to this example.

In addition, the defect determination unit 102 may determine the type of defect. When the determination is also performed for the type of defect, it is only necessary to prepare teaching data for each type of defect and perform machine learning using the teaching data. Examples of types of defects include cracks, chips, and the like described above.

The dimension detection part 103 detects the dimension of the range of a defect in the inference result which inputted the captured image to the dimension detection model 112, as the dimension of the defect which arises in the captured glass plate.

Here, the size detection model 112 will be described. The size inspection model 112 is a learned model that has learned the position and range of a defect generated in a glass plate so as to infer the position and range. In this way, the size detection model 112 can be constructed by machine learning using teaching data that correlates the position and range of the defect as positive solution data with respect to the captured image of the defect. . The position and range of a defect can also be represented by, for example, a rectangle surrounding the defect. In this case, the position of the rectangle indicates the position of the defect, and the width and height of the rectangle indicate the size of the defect.

As the teaching data, it is preferable to use a plurality of captured images in which the shades of the periphery of the defect are not clear. Thereby, the size of the defect can be detected with high precision even from the photographed image with blurred shades around the defect. The algorithm of machine learning is not particularly limited in the same way as the defect determination model 111 .

Furthermore, when the defect judging unit 102 judges the type of defect, the size detection model 112 may be prepared in advance for each type of defect, and the size detection unit 103 may use the size detection model 112 corresponding to the type judged by the defect judging unit 102. for size inspection. Thereby, the detection accuracy of a dimension can be improved. For example, it is also possible to use the photographed image obtained by photographing the end surface of the glass plate with cracks as teaching data, construct the size detection model 112 for crack defects, and use the photographed image obtained by photographing the end surface of the glass plate with defects as a demonstration material. Teaching materials, constructing the size detection model 112 for defect defects. In this case, it is only necessary to perform size detection using the size detection model 112 for the crack defect with respect to the captured image determined by the defect judging unit 102 to have a crack defect. On the other hand, what is necessary is just to perform size detection using the size detection model 112 for chip|tip defect with respect to the captured image judged by the defect determination part 102 that a chip defect has arisen.

(inference example) FIG. 2 is a diagram showing an example of inference of the position and range of a defect using the dimension inspection model 112 . In this example, the image obtained by imaging the end surface of the glass plate by the imaging device 2 is the captured image 21 . The black portion extending laterally in the central portion of the captured image 21 is the end face of the glass plate, and the white portion photographed near the central portion of the end face of the glass plate is a defect. As described above, the captured image 21 is stored in the image memory device 3 .

The image acquisition unit 101 of the inspection device 1 acquires the captured image 21 from the image storage device 3 . When the captured image 21 is acquired, first, the defect determination unit 102 inputs the captured image 21 to the defect determination model 111 , and determines the presence or absence of a defect based on an inference result output from the defect determination model 111 . Since a defect is captured in the captured image 21 as described above, the defect determination unit 102 determines that there is a defect. Furthermore, as described above, the defect judging unit 102 may also judge the type of defect, but here it only judges the presence or absence of a defect.

The size detection unit 103 detects the size of the captured image 21 determined to be defective by the defect determination unit 102 . That is, the size detection unit 103 inputs the captured image 21 to the size detection model 112 and acquires the inference result output from the size detection model 112 . The inference result indicates the position and range of the defect in the captured image 21 .

The size detection unit 103 stores the range indicated by the above-mentioned deduction result in the detection result storage device 4 as size information indicating the defect size. For example, when the range is a rectangle, the size detection unit 103 only needs to set the representative coordinates (for example, the coordinates of the upper left corner of the rectangle) representing the position of the rectangle in the captured image 21, and the coordinates representing the width and height of the rectangle. The size information can be memorized.

In the image 31 shown in FIG. 2 , the inference result of the size detection model 112 is indicated by a rectangle 41 . As shown in the figure, the width and height of the rectangle 41 are equal to the width and height of the defect, and thus it can be seen that an accurate inference has been made.

Such rectangles may also be called annotations. By displaying the comment, the user of the inspection device 1 can visually confirm the inference result. By using the size detection model 112 constructed by machine learning using the captured image of the defect as teaching data, it is possible to detect the defective portion so accurately.

(Inspection in manufacturing steps of glass plate) The inspection by the inspection apparatus 1 can also be performed as a part of the manufacturing process of a glass plate. At this time, an example of a manufacturing procedure of a glass plate accompanied by inspection by the inspection device 1 will be described based on FIG. 3 . FIG. 3 is a flowchart showing an example of a manufacturing procedure of a glass plate accompanied by inspection by the inspection device 1 .

In S101, the original glass plate is molded into a glass plate of a predetermined size. The original glass plate is a glass plate of a larger size than the glass plate to be a product, and is manufactured by a glass original plate manufacturing device. In S101, the original glass plate is adjusted to a glass plate of a predetermined size by, for example, a cutting device that cuts the original glass plate into a predetermined size, and a processing device that processes the end surface of the cut glass plate.

In S102, the inspection device 1 executes an inspection step. The glass plate judged to be a good product by this inspection process becomes a product. The details of the inspection procedure will be described below based on FIG. 4 .

(Flow of checking steps) FIG. 4 is a flowchart showing an example of the inspection procedure shown in FIG. 3 . Furthermore, it is assumed that before the inspection step starts, the imaging device 2 photographs the end face of the glass plate to be inspected, and stores the photographed image in the image memory device 3 . Furthermore, the timing for performing the inspection step shown in FIG. 4 is not particularly limited. For example, it can be carried out every time a new image is taken and stored in the image memory device 3, it can also be carried out every time the shooting of one end surface of a glass plate is completed, and it can also be carried out on all end surfaces of all glass plates of the inspection object. Carried out after shooting.

In S1 (image acquisition step), the image acquisition unit 101 acquires a captured image from the image memory device 3 . When a plurality of captured images are stored in the image memory device 3, it is only necessary to acquire captured images that have not yet been submitted to the inspection step. Then, the image acquisition unit 101 outputs the acquired captured image to the defect determination unit 102 .

In S2, the defect determination part 102 inputs the captured image acquired in S1 to the defect determination model 111, and determines the presence or absence of a defect based on the inference result output from the defect determination model 111. In addition, it is assumed here that the defect determination unit 102 also determines the type of the defect.

In S3, the dimension detection part 103 decides to use the dimension detection model 112 corresponding to the kind of defect judged in S2 among the some dimension detection models memorize|stored in the storage part 11 beforehand, for dimension detection. In addition, when it is determined in S2 that there is no defect, the process after S3 is not performed, and the inspection procedure for the captured image acquired in S1 ends.

In S4 (size detection step), the size detection unit 103 inputs the captured image acquired in S1 to the size detection model 112 determined in S3 . Next, the size detection unit 103 detects the size of the range (in other words, the size of the range indicated by the inference result) in the inference result output from the size detection model 112 as the size of the defect. For example, when the said range is a rectangle, the size detection part 103 detects the width and height of this rectangle as the width and height of a defect. Furthermore, the size detecting unit 103 may also convert the detected width and height into actual sizes.

In S5 , the size detection unit 103 stores the size detected in S4 in the detection result storage device 4 . Specifically, the dimension detection unit 103 transmits and stores dimension information indicating the width and height of a defect to the detection result storage device 4 . As above, the inspection procedure for the captured image acquired in S1 ends.

Furthermore, although not shown in FIG. 4 , the inspection step in FIG. 4 is repeated until the inspection of the captured images of all end surfaces of at least one glass plate to be inspected is completed, and then returns to the manufacturing step in FIG. 3 .

Furthermore, the inspection step in FIG. 4 may include a step of determining whether the glass plate is a good product or a defective glass plate for which the presence or absence of defects has been determined for all end faces. The glass plate judged to be a good product in this step becomes a product. The criteria for judging a good product and a defective product may be appropriately determined. For example, a glass plate in which a defect larger than a predetermined size is detected may be regarded as a defective product, and a glass plate in which no defect is detected or a detected defect smaller than a predetermined size may be regarded as a good product. This determination may be performed by the inspection device 1 , or may be performed by another information processing device different from the inspection device 1 .

(Effect) As mentioned above, the inspection apparatus 1 of this embodiment includes the image acquisition part 101 which acquires the image|captured image which image|photographed the end surface of a glass plate. Furthermore, the inspection device 1 includes a size detection unit 103 that detects the size of the above-mentioned range in the inference result obtained by inputting the captured image to the size detection model 112 as the size of the defect generated in the captured glass plate. Here, the dimension inspection model 112 is a learned model that has learned the position and range of the defect generated on the end surface so as to infer the position and range.

Furthermore, the inspection method of this embodiment includes, as above, an image acquisition step (S1) of acquiring a photographed image obtained by photographing a glass plate; The size of the range in the inference result obtained from the image is taken as the size of the defect generated on the glass plate, and the size detection model 112 is used to deduce the position of the defect generated on the glass plate and the size of the range The method has learned the position and the range, and the learned model is completed.

Furthermore, the method for manufacturing a glass plate according to the present embodiment includes the step of molding the original glass plate into a glass plate of a predetermined size ( S101 ), and the step of inspecting the glass plate by the inspection device 1 ( S102 ) as described above. , and the inspection step includes: an image acquisition step (S1) of acquiring a photographed image obtained by photographing the glass plate; In the deduction result, the size of the range is taken as the size of the defect generated in the glass plate, and the size detection model 112 is used to deduce the position and the range of the defect generated in the glass plate. The position and the range are learned, and the learned model is completed.

According to these configurations, human beings are not involved in the size detection of defects, and thus the labor cost for the detection can be reduced. In addition, in the learning of the size inspection model 112, by learning a large number of images with unclear shades around defects as teaching data, defects can be detected with high accuracy even from images with unclear shades around defects. . Therefore, the reduction of labor costs and the maintenance of detection accuracy can be taken into account during the size detection of the defects generated in the glass plate.

[Embodiment 2] Other embodiments of the present invention will be described below. In addition, for the convenience of description, the same code|symbol is attached|subjected to the member which has the same function as the member demonstrated in the said embodiment, and the description is not repeated. This is also the case in the third and subsequent embodiments.

In this embodiment, an example will be described in which the dimension inspection model 112 detects the dimension of a defect based on the reliability of the inference result output together with the inference result. More specifically, the size detection unit 103 of the present embodiment detects the size of a rectangle in the inference result whose reliability is equal to or greater than a predetermined threshold value, as the size of a defect generated in the captured glass plate.

The reliability is a value indicating the certainty of an inference result, and is, for example, a numerical value ranging from 0 to 1. The larger the numerical value of the reliability in this embodiment, the higher the possibility that the inference result is actually a defect.

The reliability may be calculated, for example, by numerically analyzing the pixel values of the pixels constituting the captured image. In the captured image, the pixel value (darkness) greatly changes between a defective area and a non-defective area. Therefore, the reliability can be lowered when the change amount of the pixel value at the boundary between the inside and outside of the range indicated by the inference result of the size detection model 112 is small, and the reliability can be improved when the change amount of the pixel value is large. degree, etc., and the reliability is calculated from the amount of change in the pixel value. Thereby, the reliability of the value corresponding to the change amount of the pixel value is calculated. Furthermore, for example, the size detection model 112 that outputs a numerical value indicating the reliability of the inference result together with the inference result may be used. In this case, processing such as numerical analysis is unnecessary.

(Example of dimension inspection based on reliability) FIG. 5 is a diagram showing an example of dimension detection of defects based on reliability. In the photographed image 32 shown in FIG. 5 , rectangles 42 to 44 representing the inference results of the size inspection model 112 and the reliability of these inference results are drawn in the photographed image obtained by photographing the end face of the glass plate with defects. The value of degree is 52~54.

As shown in the figure, in this example, the size detection model 112 deduces that the ranges indicated by the rectangles 42 to 44 are defects, and the reliability of these inferences are 0.95, 0.90 and 0.75 as shown by the values 52 to 54, respectively.

Here, for example, it is assumed that the threshold value of reliability is set to 0.80. In this case, the dimension detection part 103 detects the dimension of the rectangle 42 and the rectangle 43 whose reliability is 0.80 or more among the rectangle 42-rectangle 44, respectively, as the dimension of a defect. On the other hand, the size detection unit 103 does not detect the size of the rectangle 44 whose reliability is less than 0.80 as the size of the defect.

Therefore, when the final size detection result obtained by the size detection unit 103 is displayed on the captured image 32, only the rectangle 42 and the rectangle 43 are shown on the lower side of FIG. 5 . In this case, the size detection unit 103 transmits and stores the size information indicating the width and height of the rectangle 42 and the size information indicating the width and height of the rectangle 43 to the detection result storage device 4 .

(Flow of checking steps) Fig. 6 is a flow chart showing an example of the inspection procedure in this embodiment. In this flowchart, the same processes as those in the flowchart of FIG. 4 are assigned the same numbers as those in FIG. 4 , and the description thereof will not be repeated.

In S11, the size detection part 103 inputs the captured image 21 to the size detection model 112 determined in S3, and acquires the deduction result of the size of a defect. This inference result includes not only the position and size of the range inferred to be defective by the size detection model 112 but also the reliability of the inference.

In S12, the size detection part 103 detects the size of the range in the inference result in which the reliability output in the inference of S11 is more than a threshold value, as the size of a defect.

(Effect) As described above, in the inspection device 1 of the present embodiment, the dimension detection model 112 further outputs the reliability of the inference result. And the size detection part 103 detects the size of the range of a defect in the inference result whose reliability is more than a predetermined threshold value, as the size of the defect which arises in the imaged glass plate.

In this way, the size of the range in the inference result with high reliability is detected as the size of the defect, so even if the position and range are obtained as the inference result for a non-defective part, the size of the range can be erroneously detected as the size of the defect. possibility.

In particular, due to deviations in the conveyance direction of the glass plate, etc., the light irradiation method at the time of imaging may change, and chromatic aberration may also occur in a non-defective portion. With respect to such parts with color difference, there is a possibility that the size detection model 112 may mistakenly deduce that it is a defective part, and by properly setting the threshold value, it is possible to distinguish the color difference part that has no problem with the quality of the product from the real defect part , for proper detection.

[Embodiment 3] When a plurality of inference results are obtained, the size detection unit 103 of the present embodiment detects the size of a region including a plurality of ranges in the plurality of inference results as the size of a defect generated in the captured glass plate.

FIG. 7 is a diagram showing an overview of size detection of defects performed by the size detection unit 103 according to the present embodiment. In the photographed image 33 shown in FIG. 7 , rectangles 45 to 48 representing the inference results of the size detection model 112 are drawn in the photographed image obtained by photographing the end face of the glass plate including the defective portion.

When the size detection unit 103 of this embodiment detects a range of a plurality of defective parts in one captured image in this way, it obtains the size of the smallest area including these ranges. For example, the size detecting unit 103 may also determine the uppermost side among the sides constituting the rectangle 45 to the rectangle 48, and the lowermost side among the rectangles 45 to 48, and set the distance between these sides to the desired The height of the area. Moreover, the size detecting unit 103 may also determine the leftmost side among the sides constituting the rectangle 45 to the rectangle 48 and the rightmost side among the rectangles 45 to 48, and set the distance between these sides as the desired area width.

The size detection unit 103 can determine the inclusion area 61 including the rectangle 45 to rectangle 48 through such processing. Next, the size detection unit 103 sends and stores size information indicating the width and height of the included region 61 to the detection result storage device 4 .

In addition, in the example of FIG. 7, the containing area 61 containing the mutually contacting rectangle 45 - the rectangle 48 is specified, but it is not limited to this. That is, the size detection unit 103 may detect the size of a region including a plurality of ranges not in contact with each other as the size of the defect.

This configuration is effective, for example, when only the two ends of the actual defect are inferred as defects, and the central portion of the defect is not inferred as a defect. In this case, the size detection unit 103 can detect the size of the inclusion area with a small error from the actual size of the defect as the size of the defect by specifying the inclusion area including the detected both ends.

(Flow of checking steps) Fig. 8 is a flow chart showing an example of the inspection procedure in this embodiment. In this flowchart, the same processes as those in the flowchart of FIG. 4 are assigned the same numbers as those in FIG. 4 . In addition, the same numbers as in FIG. 6 are assigned to the same processes as those in the flowchart of FIG. 6 . The description of the same processing will not be repeated.

In S21, the size detection unit 103 determines whether there are a plurality of inference results acquired in S11. When it is determined that there are a plurality of items (YES in S21 ), the checking procedure proceeds to S22 . On the other hand, when it is determined that there are not a plurality of cases (NO in S21), the checking procedure proceeds to S23.

In S22 , the size detection unit 103 obtains the size of a region including a plurality of ranges indicated by the inference result acquired in S11 , and detects the size as the size of a defect. The method of calculating the size of the included region is as described based on FIG. 7 .

In S23, the size detection unit 103 detects the size of the range in the acquired inference result as the size of the defect. That is, the size detection unit 103 detects the width and height indicated by one of the inference results acquired in S11 as the size of the defect.

(Effect) As described above, in the inspection device 1 according to the present embodiment, when a plurality of inference results are obtained, the size detection unit 103 detects the size of a region including a plurality of ranges among the plurality of inference results as the captured image. The size of the defect produced by the glass sheet.

According to the above configuration, the size of the included region is detected as the size of the defect, so even when a plurality of ranges with a large size error from the actual defect are obtained, the ranges can be corrected so that the size error from the actual defect is small of the included area. As a result, even when a plurality of ranges are obtained, the error between the size detection of the defect using the size detection model 112 and the actual size can be made smaller.

Furthermore, even if it is a glass plate in which a plurality of defects are detected, the size of each detected defect is small, and the inspection apparatus 1 may finally determine the glass plate as a good product. This judgment result is correct when the size of the defect is actually small, but it is an incorrect judgment when there is actually a large-sized defect and a part of it is detected as a defect fragmentarily. According to the configuration of the present embodiment, the size of the included region, which includes a range including a plurality of defects, can be detected, so the possibility of such misjudgment can be reduced.

(modified example) This embodiment can be combined with Embodiment 2. Specifically, when there are a plurality of inference results whose reliability is equal to or higher than a predetermined threshold, the size detection unit 103 may specify an included area including a range in the inferred result, and detect the size of the included area as the inferred area. The size of the defect produced by the photographed glass sheet.

[Embodiment 4] When the size detection unit 103 of the present embodiment detects that the size of the defect is out of the normal range, it performs numerical analysis on the pixels constituting each pixel of the captured image, and detects the size of the defect again. The normal range should just be predetermined based on the size of a glass plate, the size of a normal defect, etc., for example. The size of the defect outside the normal range may satisfy (1) the width of the rectangle as the inference result is outside the predetermined first numerical range, and (2) the height of the rectangle as the inference result is outside the predetermined second numerical range. Dimensions of at least either of the two conditions.

The size detection unit 103 of the present embodiment also detects a size based on the inference result output from the size detection model 112, similarly to the size detection unit 103 of each of the aforementioned embodiments. The size detection unit 103 of this embodiment is different from the size detection unit 103 of each of the above-mentioned embodiments in determining whether or not the detected size is within a normal range.

In addition, when the detected size is out of the normal range, the size detection unit 103 of this embodiment performs numerical analysis on the pixel values of each pixel constituting the captured image, and detects the size of the defect again. The size detection unit 103 of this embodiment is also different from the size detection unit 103 of each of the above-mentioned embodiments in this point. In addition, the processing block for performing numerical analysis may be a processing block different from that of the size detection unit 103 .

The size detection method of the defect by numerical analysis is not particularly limited, and various methods can be applied. For example, as shown in FIG. 2 and the like, in the captured image, the pixel value greatly changes at the boundary between the region where the end surface of the glass plate is captured and the background region. Therefore, the size detection unit 103 may firstly extract the region captured on the end surface of the glass plate based on the change in the pixel value.

Furthermore, as shown in FIG. 2 and the like, among the end faces of the glass plate captured in the captured image, the pixel values of the portion where the defect is captured are different from those of the portion without the defect. Therefore, as long as the region including the pixel value specific to the portion where the defect is imaged is included in the region captured as described above and the end surface of the glass plate is extracted, the size detection unit 103 only needs to detect the width and height of the region as The size of the defect is sufficient. Next, the size detection unit 103 sends and stores size information indicating the detected width and height to the detection result storage device 4 .

(Flow of checking steps) FIG. 9 is a flow chart showing an example of the inspection procedure in this embodiment. In this flowchart, the same processes as those in the flowchart of FIG. 4 are assigned the same numbers as those in FIG. 4 , and the description thereof will not be repeated.

In S31, the dimension detection part 103 judges whether the dimension of the defect detected in S4 is within a normal range. When it is determined that it is within the normal range (YES in S31 ), the checking step proceeds to S5 . When it is determined that it is not within the normal range, that is, outside the normal range (NO in S31 ), the checking step proceeds to S32 .

In S32, the size detection unit 103 discards the size detection result based on the inference result output by the size detection model 112, and in S33, the size detection unit 103 numerically analyzes the pixel values of the pixels constituting the captured image, The size of the defect is checked again. Then, the checking step goes to S5.

In S5 entered from S33 , the size detection unit 103 may store information indicating that the re-detection has been performed in association with the size information in the detection result memory device 4 . Thereby, the user of the inspection system 100 can identify the captured image for which re-inspection was performed.

Moreover, S32 can also be omitted. In this case, in S5 , the size detecting unit 103 stores both the size information indicating the detection result in S4 and the size information indicating the detection result in S33 in the detection result storage device 4 .

(Effect) As described above, in the inspection device 1 of the present embodiment, when the size of the detected defect is out of the normal range, the size detection unit 103 performs numerical analysis on the pixel values of each pixel constituting the captured image, and again The size of the defect is detected.

According to the above configuration, for a defect whose size is out of the normal range, the pixel value of each pixel constituting the captured image is numerically analyzed to re-inspect the size. That is, since the re-inspection of the size is performed by a method different from the size detection of the defect using the learned model 1, it can be corrected to an appropriate size.

〔Modification〕 In each of the above-described embodiments, an example in which the size of the defect generated on the end surface of the glass plate was detected was described, but the size of the defect generated on a portion other than the end surface of the glass plate may also be detected. Furthermore, in each of the above-described embodiments, an example of detecting the size of a defect generated in a glass sheet cut out from a glass original sheet has been described, but the target glass sheet is not limited to cutting out from a glass original sheet. For example, it is also possible to detect the size and the like of a defect generated in a glass plate portion of a product including a glass plate.

Furthermore, in each of the above-described embodiments, an example was shown in which both determination of the presence or absence of defects (and type) and detection of dimensions were performed by one inspection device, but these processes may be performed by separate devices. That is, the inspection method described in each of the above-mentioned embodiments may be executed by one inspection device, or may be executed by a plurality of inspection devices.

Furthermore, in each of the above-described embodiments, an example in which the defect determination model 111 and the dimension detection model 112 are respectively different models has been described, but one model that learns the presence or absence of defects and the positions and ranges of defects may be used. In this case, when a captured image is input to the one model, an inference result indicating the presence or absence of a defect is output. Furthermore, when there is a defect, an inference result showing its position and range is also output together. Furthermore, it is also possible to infer this one model with respect to the type of defect.

〔Example of realization by software〕 The function of the inspection device 1 (hereinafter referred to as "device") can be realized by the following program (inspection program) for making a computer function as the device and for controlling each of the devices as the computer. The blocks (in particular, each unit included in the control unit 10 ) function.

In this case, the device includes a computer having at least one control device (such as a processor) and at least one storage device (such as a memory) as hardware for executing the program. Each function described in each of the above-described embodiments is realized by executing the program using the control device and the storage device.

The program can also be recorded in one or more non-transitory and computer-readable recording media. The recording medium may or may not be included by the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

Moreover, part or all of the functions of the control blocks may be realized by logic circuits. For example, an integrated circuit in which a logic circuit functioning as each of the above-mentioned control blocks is formed is also included in the scope of the present invention. In addition, for example, the functions of the control blocks can also be realized by a quantum computer. Furthermore, computer program products including the following configurations are also included in the scope of the present invention. That is, the computer program product of one aspect of the present invention loads (loads) the inspection program via the computer, and executes the following commands: causing the processor to acquire a captured image obtained by capturing a glass plate; The model inputs the size of the range in the inference result obtained from the photographed image as the size of the defect generated on the glass plate captured, and the learned model is used to deduce the position of the defect generated on the glass plate and the size of the defect generated on the glass plate. The position and the range are studied in the manner of the above-mentioned range.

The present invention is not limited to the above-mentioned embodiments, and various changes can be made within the range indicated in the patent claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

1: Check the device 2: camera device 3: Image memory device 4: Test result memory device 10: Control Department 11: Memory Department 12: Department of Communications 21, 32, 33: Capture images 31: Image 41～48: rectangle 52～54: value 61: Include area 100: Check system 101: Image Acquisition Department 102:Defect Judgment Department 103: Size inspection department 111: Defect Judgment Model 112: Dimensional detection model S1～S5, S11, S12, S21～S23, S31～S33, S101, S102: steps

FIG. 1 is a block diagram showing an overview of an inspection system according to Embodiment 1 of the present invention and an example of a configuration of main parts of an inspection device included in the inspection system. FIG. 2 is a diagram showing an example of inference of the position and range of a defect using a dimension inspection model. Fig. 3 is a flow chart showing an example of a manufacturing procedure of a glass plate accompanied by inspection by the inspection device. FIG. 4 is a flowchart showing an example of an inspection step included in the manufacturing step. FIG. 5 is a diagram showing an example of dimension detection of defects based on reliability. Fig. 6 is a flow chart showing an example of an inspection procedure according to Embodiment 2 of the present invention. Fig. 7 is a diagram showing an overview of size detection of defects performed by a size detection unit according to Embodiment 3 of the present invention. Fig. 8 is a flow chart showing an example of an inspection procedure according to Embodiment 3 of the present invention. Fig. 9 is a flow chart showing an example of an inspection procedure according to Embodiment 4 of the present invention.

1: Check the device

2: camera device

3: Image memory device

4: Test result memory device

10: Control Department

11: Memory Department

12: Department of Communications

100: Check system

101: Image Acquisition Department

102:Defect Judgment Department

103: Size inspection department

111: Defect Judgment Model

112: Dimensional detection model

Claims (10)

An inspection device comprising: an image acquisition unit that acquires a photographed image obtained by photographing the glass plate; and A size detecting unit detects a size of a range in an inference result obtained by inputting the captured image to a learned model that infers the size of a defect generated in the glass plate as a size of the captured defect in the glass plate. The positions and ranges of the generated defects are learned, and the positions and ranges are learned. The inspection device according to claim 1, wherein the learned model further outputs the reliability of the inference result, The size detection unit detects the size of the range in the inference result in which the reliability is equal to or greater than a predetermined threshold value, as the size of the imaged defect generated in the glass plate. The inspection device according to claim 1 or claim 2, wherein the size detecting unit detects the size of a region including a plurality of ranges among the plurality of inference results when a plurality of the inference results are obtained, As the size of the defect produced by the glass plate as photographed. The inspection device according to claim 1 or claim 2, wherein when the size of the detected defect is out of the normal range, the size detection unit checks the pixel value of each pixel constituting the captured image Numerical analysis was carried out to check again the size of the defect. The inspection device according to claim 1 or claim 2, wherein the image acquisition unit acquires the photographed image obtained by photographing the end surface of the glass plate. The inspection device according to claim 1, wherein the size detection unit detects that in the inference result, the change amount of the pixel value at the boundary between the inside and the outside of the range indicated by the inference result is equal to or greater than a predetermined threshold value The size of the range is taken as the size of the defect generated on the glass plate. An inspection method, performed by an inspection device, comprising: An image acquisition step of acquiring a photographed image obtained by photographing the glass plate; and A size detection step of detecting the size of a range in the inference result obtained by inputting the photographed image to the learned model, as the size of the defect generated in the captured glass plate, the learned model deduces the size of the glass plate The positions and ranges of the generated defects are learned, and the positions and ranges are learned. A method of manufacturing a glass plate, comprising the steps of forming a glass raw plate into a glass plate of a predetermined size, and inspecting the glass plate by an inspection device, and The inspection steps include: an image acquiring step of acquiring a photographed image obtained by photographing the glass plate; and A size detection step of detecting, as the size of a defect generated in the captured glass plate, the size of a range in the inference result obtained by inputting the captured image to the learned model that deduces the size of the glass plate. The positions and ranges of the defects generated on the board were studied, and the positions and ranges were learned. An inspection program for causing a computer to function as the inspection device according to Claim 1, and for causing the computer to function as the image acquisition unit and the size detection unit. A computer program product that loads an inspection program via a computer and executes the following commands: causing the processor to acquire a captured image of the glass plate; and causing the processor to detect, as a size of a defect generated in the captured glass plate, a size of a range in an inference result obtained by inputting the captured image to a learned model that infers the size of the glass plate The positions and ranges of the generated defects are learned, and the positions and ranges are learned.

Images