Glass bottle inspection method and glass bottle inspection device

Abstract

The present invention provides a glass bottle inspection method that allows a determination to be automatically made as to whether there is a defect with high inspection accuracy even in a region in which a joint line is present.　One aspect of the glass bottle inspection method comprises: an image acquisition step for acquiring an image of a glass bottle; a line setting step for generating a line display image in which an identification line is disposed in the image in alignment with an uneven portion resulting from the mold shape for the glass bottle; a segmentation step for segmenting the image into a plurality of minute segments; an extraction step for extracting an image of interest by performing image processing on each of the minute segments; an image creation step for creating, if an image of interest is extracted in the extraction step, an inspection image by setting the coordinates of the image of interest in the line display image; and a determination step for inputting the inspection image to a trained model so as to determine whether there is a defect. The trained model is characterized by performing machine learning by using training images without a defect, training images with a defect, and training images with an identification line.

Description

Glass bottle inspection method and glass bottle inspection device

[0001] The present invention relates to a glass bottle inspection method and a glass bottle inspection device.

[0002] A method for inspecting glass bottles with uneven engraving on the surface is proposed, for example, in Patent Document 1. The invention of Patent Document 1 proposes a process for masking the engraved area and a process for determining whether or not the area where the seam line appears is a defect.

[0003] In the invention of Patent Document 1, in order to distinguish between the detected object and the seam in the image, the influence of the seam is reduced by measuring the horizontal width and by image processing that eliminates vertical shadows.

[0004] Patent No. 7220128

[0005] However, in the invention of Patent Document 1, the inspection area set in accordance with the seam line and wider than the seam line inevitably has lower inspection accuracy than other inspection areas.

[0006] Therefore, the present invention provides a glass bottle inspection method and glass bottle inspection device that automatically determines the presence or absence of defects with high inspection accuracy even in areas where seams exist.

[0007] The present invention has been made to solve at least some of the above-mentioned problems, and can be realized as the following aspects or application examples.

[0008] [1] One aspect of the glass bottle inspection method of the present invention is a glass bottle inspection method comprising: an image acquisition step of acquiring an image of the glass bottle; a line setting step of generating a line display image in which an identification line is arranged in the image to match the uneven portions resulting from the mold shape of the glass bottle; a division step of dividing the image into a plurality of minute sections; an extraction step of performing image processing for each minute section to extract an image of interest; an image creation step of, if the image of interest is found in the extraction step, setting the coordinates of the image of interest in the line display image to create an inspection image; and a determination step of inputting the inspection image into a trained model and determining whether or not there is a defect, wherein the trained model is machine learned using training images without defects, training images with defects, and training images with the identification line.

[0009] According to one aspect of the above-mentioned glass bottle inspection method, the presence or absence of defects can be determined using a trained model that uses training images with identification lines, making it possible to automatically determine the presence or absence of defects with high inspection accuracy even in areas with uneven parts such as seam lines.

[0010] [2] In one aspect of the above-mentioned glass bottle inspection method, the extraction process includes an integration process that integrates adjacent micro-sections to combine the target image, and the image creation process can set the coordinates of the target image after the integration process to the line display image.

[0011] According to one aspect of the glass bottle inspection method, even if an image of interest spans multiple small sections, the presence or absence of defects can be determined by integrating the image as a single image of interest.

[0012] [3] In one aspect of the glass bottle inspection method, the learning images having the identification lines may include learning images having a plurality of types of identification lines with different thicknesses.

[0013] According to one aspect of the above-mentioned glass bottle inspection method, by using identification lines of different thicknesses, uneven parts can be judged as non-defective even when the judgment process is carried out using enlarged and / or reduced images.

[0014] [4] One aspect of the glass bottle inspection device of the present invention is a glass bottle inspection device comprising: a line setting unit that places an identification line in an image of the glass bottle in accordance with uneven portions resulting from the mold shape; an image of interest extraction unit that divides the image into a plurality of minute sections and performs image processing for each of the minute sections to extract an image of interest; an image creation unit that sets the coordinates of the image of interest in the image in which the identification line is placed to create an inspection image; and a judgment unit that inputs the inspection image into a trained model and judges whether or not there is a defect, wherein the trained model performs machine learning using training images without defects, training images with defects, and training images with the identification line.

[0015] According to one aspect of the above-mentioned glass bottle inspection device, the presence or absence of defects can be determined using a trained model that uses training images with identification lines, making it possible to automatically determine the presence or absence of defects with high inspection accuracy even in areas with uneven parts such as seam lines.

[0016] According to one aspect of the glass bottle inspection method and glass bottle inspection device of the present invention, it is possible to automatically determine the presence or absence of defects with high inspection accuracy even in areas with uneven parts such as seam lines.

[0017] Fig. 1 is a front view schematically showing an inspection device according to this embodiment. Fig. 2 is a diagram schematically showing an image of a glass bottle. Fig. 3 is a diagram schematically showing an image in which an identification line is arranged. Fig. 4 is a diagram explaining minute compartments. Fig. 5 is a diagram showing a display unit that displays inspection results. Fig. 6 is a flowchart of an inspection method according to this embodiment.

[0018] Preferred embodiments of the present invention will be described in detail below with reference to the drawings. Note that the embodiments described below do not unduly limit the content of the present invention as defined in the claims. Furthermore, not all of the configurations described below are necessarily essential components of the present invention.

[0019] The glass bottle inspection method of this embodiment is a glass bottle inspection method that includes an image acquisition process for acquiring an image of the glass bottle, a line setting process for generating a line display image in which identification lines are arranged in the image to match the uneven portions resulting from the mold shape of the glass bottle, a division process for dividing the image into a plurality of tiny sections, an extraction process for performing image processing for each of the tiny sections to extract an image of interest, an image creation process for, if the image of interest is found in the extraction process, setting the coordinates of the image of interest in the line display image to create an inspection image, and a determination process for inputting the inspection image into a trained model and determining whether or not there is a defect, wherein the trained model is characterized by performing machine learning using training images without defects, training images with defects, and training images with the identification lines.

[0020] The glass bottle inspection device of this embodiment is a glass bottle inspection device that includes a line setting unit that places an identification line on an image of the glass bottle in accordance with the uneven parts resulting from the mold shape; an image of interest extraction unit that divides the image into a plurality of tiny sections and performs image processing for each of the tiny sections to extract an image of interest; an image creation unit that sets the coordinates of the image of interest on the image in which the identification line is placed to create an inspection image; and a judgment unit that inputs the inspection image into a trained model and judges whether or not there is a defect, and is characterized in that the trained model performs machine learning using training images without defects, training images with defects, and training images with the identification line.

[0021] 1. Inspection Device An inspection device 100 for glass bottles 10 will be described in detail using Figures 1 to 5. Figure 1 is a front view schematically showing the inspection device 100 according to this embodiment, Figure 2 is a diagram schematically showing a first image 110 of the glass bottle 10, Figure 3 is a diagram schematically showing a second image 112 in which an identification line 18a is arranged, Figure 4 is a diagram explaining the microsection A3, and Figure 5 is a diagram showing a display unit 80 that displays the inspection results.

[0022] The inspection device 100 shown in Figure 1 is an inspection device 100 for glass bottles 10. The inspection device 100 is incorporated as part of a manufacturing line for glass bottles 10 (not shown), and after molding, the glass bottles 10 that have been slowly cooled are transported to the inspection device 100, and after inspection, the glass bottles 10 are transported to the next process. The inspection device 100 may be equipped with multiple inspection stages, or may be installed on a conveyor.

[0023] The inspection device 100 includes a light-emitting unit 20 having a light-emitting surface 20a that irradiates light onto the glass bottle 10, an imaging unit 50 arranged opposite the light-emitting unit 20 across the glass bottle 10, and a control device 60 equipped with a judgment unit 64 that judges whether or not there are defects based on, for example, a first image 110 (Figure 2) of the glass bottle 10 captured by the imaging unit 50.

[0024] As shown in Figure 1, the glass bottle 10 is inspected in an upright position, i.e., with its central axis 11 aligned vertically. The vertical direction is the direction of gravity, and the horizontal direction is the direction perpendicular to the vertical direction. Note that the central axis 11, indicated by the dashed line, is an imaginary line.

[0025] The inspection device 100 captures images of glass bottles 10 moving on a conveying path 40, such as a top chain conveyor, and inspects them sequentially. Glass bottles 10 are conveyed on the conveying path 40 with a gap between them and the glass bottles 10 in front and behind them. While the inspection device 100 is shown with one set of light-emitting unit 20 and image-capturing unit 50, multiple sets of light-emitting unit 20 and image-capturing unit 50 (not shown) may be provided so that the entire circumference of the glass bottle 10 can be imaged without omission. The inspection device 100 may also be configured to capture the entire circumference of the glass bottle 10 while rotating it around its central axis 11. In this case, a mounting table rotated by an electric motor may be provided instead of the conveying path 40.

[0026] The glass bottle 10 is, for example, transparent or translucent. Translucency refers to a degree of transparency that allows light from the light-emitting element 20 passing through the glass bottle 10 to detect defects 19a, 19b, such as surface bubbles, on the body 14 of the glass bottle 10. The glass bottle 10 has, for example, a mouth 12 with a circular cross section, a body 14, and a bottom 16. The body 14 includes a neck extending downward from the mouth 12 and a shoulder that gradually widens in diameter downward from the neck. The cross-sectional shape of the glass bottle 10 may also be polygonal. The glass bottle 10 has uneven portions resulting from the mold shape. Examples of uneven portions include a seam line 18 formed at the joint of the mold, a recess defining the position for attaching a label, and patterns, letters, or other engravings carved into the mold.

[0027] The light-emitting unit 20 is a light source that illuminates the glass bottle 10. The light-emitting unit 20 has a light-emitting surface 20a on the side facing the glass bottle 10. The light-emitting unit 20 is a surface light source that can illuminate the glass bottle 10 from the opposite side of the imaging unit 50. The light-emitting surface 20a is, for example, rectangular in shape, and emits light from almost its entire surface. The light-emitting surface 20a is positioned approximately directly facing the glass bottle 10 and the imaging unit 50 so that light transmitted through the glass bottle 10 reaches the imaging unit 50. The light source for the light-emitting unit 20 can be a known light source, such as an LED or organic EL. The light-emitting unit 20 provides diffuse lighting. When an LED is used, a light-restricting unit 30 can be laminated on the light-restricting surface 20a to irradiate the glass bottle 10 with uniform light from the entire light-restricting surface 20a. The light-restricting unit 30 can be a known unit that suppresses the diffusion angle of light from a light source such as an LED and emits directional light to the outside. The light-limiting unit 30 suppresses the diffusion angle of light, making it possible to capture images of, for example, wrinkles or streak shadows with a step depth of approximately 0.02 mm. The light-limiting unit 30 can be formed by stacking a film that suppresses the diffusion angle in the horizontal direction and a film that suppresses the diffusion angle in the vertical direction. The light-limiting unit 30 may also be formed by stacking multiple light control films, for example. When stacking films that suppress the diffusion angle in the same direction (e.g., the horizontal direction), it is preferable to stack films with different diffusion angles. Commercially available light control films can be used. Furthermore, any film that can suppress the diffusion angle of light can be used for the light-limiting unit 30, and is not limited to light control films.

[0028] The imaging unit 50 is positioned opposite the light-emitting unit 20 across the glass bottle 10. The imaging unit 50 is positioned so as to capture an image of the surface of the glass bottle 10 on an extension of the central axis 11. The imaging unit 50 preferably has a shallow depth of field so as not to capture the seam line 18 on the light-emitting unit 20 side. The imaging unit 50 may include multiple imaging units, such as a first imaging unit 51a and a second imaging unit 51b arranged vertically and spaced apart as shown in FIG. 1. The first imaging unit 51a and the second imaging unit 51b can capture at least the inspection target portion of the glass bottle 10 designated for each imaging unit. For example, the first imaging unit 51a captures at least the upper region of the mouth 12 and the body 14 within its field of view. For example, the second imaging unit 51b captures at least the lower region of the body 14 within its field of view. The first and second image capturing units 51a and 51b can capture a first image 110 ( FIG. 2 ) including defects 19a and 19b using light from the light emitting unit 20 that has passed through the glass bottle 10. The first and second image capturing units 51a and 51b can be, for example, a known area sensor camera, or a line sensor camera that captures images of the glass bottle 10 by rotating it. The image capturing unit 50 transmits the captured image data to the control device 60.

[0029] The control device 60 includes a line setting unit 61, a target image extraction unit 62, an image creation unit 63, a determination unit 64, and an output unit 65. The control device 60 may further include a storage device (not shown). The control device 60 is configured with, for example, a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), storage devices such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), a ROM (Read-Only Memory), and a RAM (Random Access Memory), input devices such as a keyboard, a mouse, and a touchpad, and digital input / output boards such as an I / O board. The control device 100 may also include a programmable logic controller (PLC), various network devices (e.g., switches, hubs, routers, etc.), an image input board for transmitting captured image data, etc. The inspection device 100 may further include a display unit 80, such as a liquid crystal display or an organic electroluminescence (EL) display, that displays output from the control device 60. The control device 60 acquires image data from the imaging unit 50 and executes a process for inspecting the glass bottles 10. The process of transporting the glass bottles 10 at a predetermined speed along the conveying path 40 may be executed by a control unit separate from the control device 60, or may be configured to be executed by the control device 60. The control device 60 can determine the timing of image capture based on signals from, for example, a transparent object detection sensor, a rotary encoder on the conveying path 40, a solid-state relay, etc.

[0030] As shown in FIGS. 2 and 3 , for example, the line setting unit 61 detects a glass bottle 10 in the first image 110 captured by the second imaging unit 51b and cuts out the inside of a first region A1, which is pre-prepared to correspond to, for example, the maximum width of the glass bottle 10, from the first image 110. The process of cutting out the first region A1 may be performed by a processing unit other than the line setting unit 61. The glass bottle 10 is detected by, for example, edge processing, detecting the shadow that appears at the boundary between the glass bottle 10 and the background. Since the glass bottle 10 to be inspected is predetermined, the size of the first region A1 can be stored in a storage device to match the outer shape of the glass bottle 10. By cutting out the first region A1, the background is reduced, as in the second image 112 of FIG. 3 , thereby reducing the burden on the inspection device 100 in each process described below. Note that the same process as for the first image 110 can also be performed on an image captured by the first imaging unit 51a.

[0031] As shown in FIG. 3 , the line setting unit 61 places an identification line 18a in the second image 112 of the glass bottle 10, aligned with uneven portions resulting from the mold shape, such as the seam line 18. Note that in FIG. 3 , the glass bottle 10 is depicted in black to show the identification line 18a as a white line. The seam line 18 is formed on the surface of the glass bottle 10 as a step along the mold seam, and a portion of it may appear as a dark line in the first image 110 ( FIG. 2 ). For example, if a portion of the seam line 18 can be detected in the straight portion of the body 14, the entire seam line 18 can be predicted based on shape information about the glass bottle 10, such as the body diameter. Because the seam line 18 appears as a straight line in a planar view, it is easy to predict the entire seam line 18. The identification line 18a is generated along the coordinates of the predicted seam line 18. The seam line 18 can be detected, for example, by setting a rectangular edge detection area in the center of the straight portion of the body 14 in the first image 110 and performing edge detection processing to find vertical lines. Information about the body diameter of glass bottle 10 can be detected, for example, by edge detection processing on both the left and right sides of first image 110. In this embodiment, the seam line 18 is described, but similar processing can be used for other uneven portions. In the case of uneven portions other than seam line 18 that result from the mold shape, for example, an identification line 18a may be artificially created in advance and stored in a memory device by estimating or acquiring coordinate information for the uneven portion of glass bottle 10, and then arranged in second image 112 in accordance with the coordinate information.

[0032] As shown in FIG. 4 , the target image extraction unit 62 may set a second area A2 (surrounded by a dashed line) to be inspected within the second image 112. The second area A2 in the second image 112 captured by the second imaging unit 51b is set, for example, from the center to the lower half of the body 14 of the glass bottle 10. In this embodiment, since there are images captured by the first imaging unit 51a and images captured by the second imaging unit 51b, the target area can be set for each image according to the range where high inspection accuracy is desired. The second area A2 is shown on the right side of FIG. 4 to illustrate the size of the second area A2. The target image extraction unit 62 divides the second image 112 into multiple micro-sections A3 and performs image processing for each micro-section A3 to extract the target image 114. The micro-sections A3 are sections surrounded by dotted lines in FIG. 3 and can be set, for example, within the second area A2. To illustrate each micro-section A3, one unit of the micro-section A3 is shown in the upper left corner of FIG. 4. 4, the minute sections A3 divide the second region A2 into a grid pattern, but this is not limiting, and the minute sections A3 may be divided so that they partially overlap with adjacent minute sections A3. The size of the minute sections A3 can be set according to the size of the target defect, and can be set, for example, to be slightly larger than the target defect.

[0033] The target image extraction unit 62 can reduce processing time by performing image processing for each minute section A3 using parallel processing. Image processing can be performed using a combination of multiple image processing algorithms and parameters used for each algorithm, which can detect the target defects 19a and 19b as points of interest 115. The points of interest 115 are singular points detected by image processing, and include the defects 19a and 19b. The minute section A3 containing the points of interest 115 is the target image 114. An example of the target image 114 after image processing is shown on the left side of Figure 4. Image processing algorithms include, for example, various filter processes for noise removal, correction processes such as brightness correction, binarization processes, edge detection processes, frequency filter processes, arithmetic operations such as arithmetic operations, morphology processes, detection target determination processes, branching processes, etc.

[0034] Image processing for each micro-section A3 can be optimized using, for example, evolutionary computation using the GNP (Genetic Network Programming) method. The GNP method can be applied to, for example, the method disclosed in Japanese Patent Application Laid-Open No. 2022-89430, which obtains a detector by optimizing the points of interest 115 in an image using a discrete optimization algorithm. As the discrete optimization algorithm, evolutionary computation, which uses genetic operations for optimization, is preferably used, as an optimization method that mathematically simulates the evolution of living organisms. However, other algorithms may also be used as long as they achieve similar effects. Image processing for each micro-section A3 can be performed by a detector that has been optimized in advance using training information including a training image divided in the same manner as the micro-section A3 and individual information (image processing algorithm). The training image includes, for example, the shape, number, position, range, etc. of multiple types of points of interest 115. For example, an image in which the points of interest 115 are depicted in white on a black background can be used. Optimization can be performed by repeating generations through, for example, crossover, mating, and mutation, where an input image is input to individual information, and individual information that produces an evaluation value close to that of the teacher image remains, thereby generating an optimized detector. By performing image processing for each micro-section A3 using the optimized detector, it is possible to detect and extract an image of interest 114 that includes a desired point of interest 115.

[0035] The image-of-interest extraction unit 62 can execute a merging process to merge adjacent minute sections A3 to combine the image of interest 114. For example, if one defect 19a shown in Figure 4 is divided into two minute sections A3, the merging process can be executed to merge them into one merged section A4, and the original image of interest 114 can be extracted.

[0036] The image creation unit 63 sets the coordinates of the image of interest 114 in the second image 112 ( FIG. 3 ) on which the identification line 18a is arranged, to create the test image 113 ( FIG. 5 ). The identification line 18a in FIG. 5 is depicted as a black line for ease of explanation, but it may also be a white line. When multiple micro-sections A3 are integrated into an integrated section A4, the image creation unit 63 can set the coordinates of the image of interest 114 after the integration process in the second image 112. The image creation unit 63 may also set information such as the shape, size, and brightness of the point of interest 115 in addition to the coordinates in the image of interest 114 of the test image 113. Note that FIG. 5 shows the test image 113 with the indication unit 90 and result display unit 92 added and displayed on the display unit 80.

[0037] The judgment unit 64 inputs the inspection image 113 into the trained model to judge the presence or absence of defects 19a and 19b. For example, the coordinates of multiple target images 114 are set in the inspection image 113, and judgment processing is performed using the trained model for the image with those coordinates. The judgment processing may be performed in parallel for the coordinates of multiple target images 114. The trained model is machine-learned using training images without defects, training images with defects, and training images with the discrimination line 18a. Training images with the discrimination line 18a can include images with and without defects, and training images without defects but with the discrimination line 18a can be trained as non-defective products. The discrimination line 18a is preferably thinner than the target defects 19a and 19b so that the target defects 19a and 19b are not hidden by the discrimination line 18a. Using training images with the discrimination line 18a makes it less likely that the trained model will erroneously judge the discrimination line 18a as a defect. The training image having the discrimination line 18a may include a training image in which a plurality of types of discrimination lines 18a with different thicknesses are set. By training with training images in which discrimination lines 18a with different thicknesses are set, the judgment unit 64 is less likely to judge the discrimination line 18a as a defect even when it cuts out the focus point 115 of the test image 113 and enlarges or reduces it for judgment. The trained model can be machine-learned using a neural network. As the neural network, it is preferable to use a convolutional neural network (CNN) having a convolutional layer. Furthermore, for example, an artificial image may be generated using a generative adversarial network that applies GAN (Generative Adversarial Nets) disclosed in Japanese Patent Laid-Open No. 2021-89219, and the generated image may be used as a training image. The types of defects in the learning images are, for example, streaks, wrinkles, transparent stones, cat scratches, bubbles, and surface bubbles, but other types of defects may be added depending on the inspection requirements.

[0038] The control device 60 outputs the determination result of the determination unit 64 to the outside, and can display the inspection result together with the inspection image 113 on the display unit 80, for example, as shown in FIG. 5. The inspection device 100 may also reject glass bottles 10 determined to have defects, for example, on a line subsequent to a discharge unit (not shown). The display unit 80 can display a designation unit 90 on the portion of the inspection image 113 determined to have defects 19a, 19b, and can display a result display unit 92 that displays information about the defects 19a, 19b. The result display unit 92 may display the type of defect 19a and the probability that it is of that type, or it may simply display whether or not it is a defect 19a.

[0039] According to the inspection device 100, the presence or absence of defects 19a and 19b can be determined using a trained model that uses a training image with an identification line 18a, so that the presence or absence of defects 19a and 19b can be automatically determined with high inspection accuracy even in areas with uneven portions such as the seam line 18.

[0040] 2. Inspection Method The inspection method for glass bottles 10 according to this embodiment using the inspection device 100 in Fig. 1 will be described with reference to Fig. 1 to Fig. 6. Fig. 6 is a flowchart of the inspection method according to this embodiment.

[0041] As shown in Figure 6, the inspection method according to this embodiment is a method for inspecting glass bottles 10 and includes at least an image acquisition step S10, a line setting step S32, a division step S14, an extraction step S20, an image creation step S22, and a judgment step S24. The inspection method according to this embodiment may further include a trimming step S12 after step S10, and the extraction step S20 may further include an image processing step S16 and an integration processing step S18. It may also further include a step S26 of displaying the judgment results after step S24. Each step will be described below in order with reference to Figures 1 to 5. Note that the following description will omit portions that overlap with the above description of the inspection device 100.

[0042] S10: The control device 60 executes an image acquisition step to acquire a first image 110 (FIG. 2) of the glass bottle 10. Specifically, when the glass bottle 10 is transported to a predetermined position by the transport path 40, the control device 60 commands the image capture unit 50 to begin capturing an image, and acquires the first image 110 output from the image capture unit 50. The first image 110 is stored in a storage device (not shown) of the control device 60.

[0043] S12: The control device 60 executes a trimming process to cut out a first region A1 (shown by a dashed line in FIG. 2) from the first image 110, the first region A1 being adjusted to the maximum width of the glass bottle 10. The second image 112 cut out in S12 has a reduced background region, as shown in FIGS. 3 and 4, thereby reducing the burden of each process. The control device 60 can process the flows of S14, S16, S18, and S20, and the flows of S30 and S32, in parallel, for the second image 112 cut out in S12.

[0044] S14: The target image extraction unit 62 of the control device 60, for example, executes a division step of dividing the second image 112 into a plurality of minute sections A3 (FIG. 4). In FIG. 4, the second area A2 to be inspected is divided into a plurality of minute sections A3. Each minute section A3 has a predetermined size.

[0045] S16: For example, the target image extraction unit 62 of the control device 60 executes an image processing step in which image processing is performed for each minute section A3. Performing image processing for each minute section A3 enables high-speed processing by the control device 60, and each image processing can be performed in parallel. The image processing executes an image processing algorithm that clearly extracts the target point 115 within the minute section A3. The target image extraction unit 62 may execute image processing for each minute section A3 using the optimized detector described above as the image processing algorithm. Performing image processing for each minute section A3 reduces the impact of the external shape of the glass bottle 10 on the image processing. If the external shape of the glass bottle 10 has little impact, the detector that detects the target image 114 can be more generalized, and the same detector can be applied to glass bottles of other shapes.

[0046] S18: The control device 60, for example, the image of interest extraction unit 62, executes an integration process to integrate adjacent minute sections A3 to combine the image of interest 114. The integrated image of interest 114 becomes the single point of interest 115 that existed before the division. Even if the image of interest 114 spans multiple minute sections A3, the process of integrating them into a single integrated section A4 allows the presence or absence of defects 19a, 19b to be determined as a single image of interest 114 in S24, which will be described later.

[0047] S20: The control device 60, for example, the image-of-interest extraction unit 62, executes an extraction step to extract the image of interest 114. The image-of-interest extraction unit 62 determines whether or not there are image-processed and / or integrated images of interest 114 for all micro-sections A3. The extraction of the image of interest 114 can be performed, for example, by labeling and blob analysis. The image-of-interest extraction unit 62 labels the image of interest 114 and can output the coordinates, size, aspect ratio, etc. of the point of interest 115. After image processing, the image of interest 114 clearly expresses the point of interest 115 as a cluster. As shown on the left side of FIG. 4, the image of interest 114 may have, for example, a black background against which a cluster of white points of interest 115 is formed. If there are no images of interest 114 in all micro-sections A3 (NO), the control device 60 executes S26. If there are images of interest 114 (YES), the control device 60 executes S22.

[0048] S30: The line setting unit 61 of the control device 60, for example, determines whether or not uneven portions resulting from the mold shape of the glass bottle 10 have been detected in the second image 112. The uneven portions are detected, for example, by subjecting the second image 112 to a known method, such as edge detection processing. Uneven portions, such as the seam line 18, can be detected as lines extending vertically to the body 14. If no uneven portions are detected in the second image 112 (NO), the control device 60 executes S22. If uneven portions are detected in the second image 112 (YES), the line setting unit 61 executes S32.

[0049] S32: The line setting unit 61 of the control device 60, for example, executes a line setting process to generate a line display image in which identification lines 18a are arranged in the second image 112 to match the uneven portions. Figure 3 shows the second image 112 with the identification lines 18a arranged. Because the uneven portions are due to the shape of the mold, if a portion of the uneven portion can be detected, the entire uneven portion can be predicted based on other shape information of the glass bottle 10. If an uneven portion, such as a portion of the seam line 18, can be detected in the straight portion of the body 14, the entire seam line 18 can be predicted based on information about the outer diameter of the glass bottle 10, for example. Information about the outer diameter of the glass bottle 10 can be detected, for example, using edge detection processing. A line display image can be generated by arranging an identification line 18a in the second image 112 that matches the coordinates of the uneven portion detected in S30. The identification line 18a may be generated for each image to match the detected uneven portion, or an appropriate identification line 18a may be selected and arranged from multiple identification lines 18a previously generated and stored to match the detected uneven portion. The identification line 18a may be a white line. The identification line 18a is not a mask for hiding the shadow of the uneven portion, and therefore may be thinner than, for example, the seam line 18 that appears in the second image 112.

[0050] S22: For example, the image creation unit 63 of the control device 60 sets the coordinates of the image of interest 114 in the line display image (second image 112 in FIG. 3) to create the test image 113 (FIG. 5). Using the small section A3 allows the image of interest 114 to be extracted efficiently, but dividing the seam line 18 makes it difficult to detect the seam line 18. Therefore, it is preferable to execute S32 in a flow separate from S14 to S20. If there is an image of interest 114 integrated in S18, the image creation unit 63 can set the coordinates of the image of interest 114 after the integration process in the line display image.

[0051] S24: For example, the determination unit 64 of the control device 60 executes a determination step in which the inspection image 113 is input into a trained model to determine the presence or absence of defects 19a and 19b. The trained model can be the trained model described for the inspection device 100. The trained model is machine-learned using training images without defects, training images with defects, and training images with the discrimination line 18a. Training images with the discrimination line 18a can include images with and without defects. By training training images without defects and with the discrimination line 18a as images without defects, even if the shadow of the seam line 18 is extracted as the focus point 115 in S20, if there is no focus point 115 other than the seam line 18, it can be determined that there is no defect at that coordinate. Furthermore, by including training images with the discrimination line 18a as images with and without defects, it can be determined that there is a defect if there is a defect 19a or 19b near the seam line 18. In this way, the presence or absence of defects 19a and 19b can be determined using a trained model that uses training images with discrimination lines 18a, so that the presence or absence of defects 19a and 19b can be automatically determined with high inspection accuracy even in areas with uneven portions such as seam lines 18. Training images with discrimination lines 18a used in the trained model can include training images in which multiple types of discrimination lines 18a with different thicknesses are set. By using discrimination lines 18a with different thicknesses, uneven portions can be determined to be non-defective even when S24 is performed on an enlarged and / or reduced image.

[0052] The control device 60 may additionally train the trained model using rejected images of non-defective products. The rejected images of non-defective products may be test images 113, which are glass bottles 10 that have been determined to have defects 19a and 19b in S24 and visually inspected by an operator and determined to be non-defective. Because the test images 113 contain information on the coordinates and size of the focus point 115, the control device 60 automatically extracts an image including the focus point 115 from the test image 113 and stores it in a training image folder as a defect-free training image. The additionally stored defect-free training image is then used to additionally train the trained model. This allows the trained model to accurately determine that a focus point 115, which is likely to be determined to have defects 19a and 19b, is a non-defective product.

[0053] S26: For example, the output unit 65 of the control device 60 can display the determination result of the determination unit 64 on the display unit 80 together with the test image 113 as shown in FIG.

[0054] The present invention is not limited to the above-described embodiments, and various modifications are possible, including configurations that are substantially identical to the configurations described in the embodiments. Here, "same configuration" means a configuration that has the same function, method, and result, or a configuration that has the same purpose and effect. The present invention also includes configurations in which non-essential parts of the configurations described in the embodiments are replaced. The present invention also includes configurations that achieve the same effects or purposes as the configurations described in the embodiments. The present invention also includes configurations in which publicly known technology is added to the configurations described in the embodiments.

[0055] 10...glass bottle, 11...center axis, 12...mouth, 14...body, 16...bottom, 18...seam line, 18a...identification line, 19a, 19b...defect, 20...light emitting section, 20a...light emitting surface, 30...light restricting section, 40...conveying path, 50...imaging section, 51a...first imaging section, 51b...second imaging section, 60...controller, 61...line setting section, 62...target image extraction section, 63...image creation section, 64...judgment section, 65...output section, 80...display section, 90...display section, 92...result display section, 100...inspection device, 110...first image, 112...second image, 113...inspection image, 114...target image, 115...target point, A1...first region, A2...second region, A3...micro section, A4...integrated section

Claims

1. A method for inspecting glass bottles, comprising: an image acquisition step for acquiring an image of the glass bottle; a line setting step for generating a line display image in which identification lines are arranged in the image to match the uneven portions resulting from the mold shape of the glass bottle; a division step for dividing the image into a plurality of minute sections; an extraction step for performing image processing for each minute section to extract an image of interest; an image creation step for, if an image of interest is found in the extraction step, setting the coordinates of the image of interest in the line display image to create an inspection image; and a determination step for inputting the inspection image into a trained model and determining whether or not there is a defect, wherein the trained model is trained through machine learning using training images without defects, training images with defects, and training images with the identification lines.

2. A glass bottle inspection method as described in claim 1, characterized in that the extraction step includes a merging process for merging adjacent micro-sections to combine the images of interest, and the image creation step includes setting the coordinates of the images of interest after the merging process in the line display image.

3. A glass bottle inspection method as set forth in claim 1 or claim 2, characterized in that the learning images with the identification lines include learning images with multiple types of identification lines of different thicknesses.

4. A glass bottle inspection device comprising: a line setting unit that places an identification line on an image of the glass bottle in accordance with uneven portions resulting from the shape of the mold; an image of interest extraction unit that divides the image into a plurality of minute sections and performs image processing for each of the minute sections to extract an image of interest; an image creation unit that sets the coordinates of the image of interest on the image on which the identification line is placed to create an image for inspection; and a judgment unit that inputs the image for inspection into a trained model and judges whether or not there is a defect, wherein the trained model is trained using training images without defects, training images with defects, and training images with the identification line.

Images