JP2023008662A - Tire inspection device and inspection method, identification model generation method, and generation device as well as program - Google Patents

Abstract

An object of the present invention is to improve the detection accuracy of defects occurring in the internal structure of a tire. An image data acquiring unit (11a) acquires image data of an X-ray image of a tire to be inspected. The defect detection unit 11v is a discrimination model generated by machine learning using image data of a defective region, which is a region containing a defect in the internal structure of the tire, and image data of a normal region, which is a region different from the defective region, as teacher data. Part of the image data of the tire to be inspected is input to M1 and M2 to detect defects in the internal structure of the tire to be inspected. [Selection drawing] Fig. 5

Description

The present disclosure relates to a tire inspection device and inspection method, a discriminative model generation method and generation device, and a program.

Patent Literature 1 listed below discloses an inspection method for inspecting the internal structure of a tire. Some tires have steel carcass cords. In tires of this type, defects called cross cords, where adjacent carcass cords intersect, may occur. Patent Document 1 discloses an inspection method for detecting this across code. In the inspection disclosed in Patent Document 1, the inner surface of the tire is irradiated with X-rays from the inside of the tire while the tire is rotated in the circumferential direction, and an X-ray image is acquired by an imaging device arranged outside the tire. . Based on the area of the carcass code image and the presence or absence of branching of the carcass code image, it is determined whether or not a defect called the above-described cross code has occurred.

JP 2008-309644 A

In the inspection method disclosed in Patent Document 1, over-detection may occur in which a normal carcass cord is determined to be defective, and improvement in detection accuracy has been desired.

(1) A tire inspection apparatus proposed in the present disclosure includes image data acquisition means for acquiring image data of an X-ray image of a tire to be inspected; A part or all of the image data of the tire to be inspected is input to a discrimination model generated by machine learning using image data of a normal region, which is a region different from the defective region, as teacher data, and the tire to be inspected and defect detection means for detecting defects in the internal structure of the. According to this inspection device, it is possible to improve the accuracy of tire inspection.

(2) The tire inspection apparatus described in (1) performs defect area candidate extraction for extracting, from the image data of the tire to be inspected, image data of defect area candidates, which are candidates for areas containing defects in the internal structure. It may further include a means. The defect detection means may input the image data of the defect area candidate to the identification model as part of the image data of the tire to be inspected, and determine whether or not the defect area candidate includes a defect. . According to this inspection device, the tire inspection accuracy can be further improved.

(3) The internal structure of the tire may have a plurality of linear members arranged in one direction, and the defect of the internal structure of the tire may include intersection of the linear members. In the tire inspection apparatus described in (2), the defect area candidate extraction means extracts image data of the defect area candidates based on at least one of an area and a shape of each linear member appearing in the image of the tire to be inspected. can be extracted.

(4) The tire internal structural defect may include a first type and a second type. In the tire inspection apparatus described in (1), the identification model includes image data of the defect area including the first type defect and image data of the defect area including the second type defect. , and the image data of the normal region as the teacher data. The defect detection means may input the part or all of the image data of the tire to be inspected into the identification model to specify the type of defect in the internal structure of the tire to be inspected.

(5) The tire to be inspected includes a belt and a carcass, and the carcass has a carcass central portion that is a portion covered by the belt and carcass side portions that are not covered by the belt. It's okay. The tire inspection apparatus described in (1) includes carcass center image extracting means for extracting image data of the carcass center from the tire image data, and carcass side image extracting means from the tire image data. carcass side image extraction means for extracting data; The discriminative models may comprise a first discriminative model for detecting defects in the carcass center and a second discriminative model for detecting defects in the carcass sides. The defect detection means inputs part or all of the image data of the carcass central portion into the first identification model, and inputs part or all of the image data of the carcass side portion into the second identification model. You can

(6) The tire to be inspected may have a carcass layer, a belt layer, and a finishing layer as the internal structure. In the tire inspection device described in (1), the defect detection means may detect defects in the carcass layer, the belt layer, or the finishing layer.

(7) A tire inspection method proposed in the present disclosure includes an image data acquisition step of acquiring image data of an X-ray image of a tire to be inspected; A part or all of the image data of the tire to be inspected is input to a discrimination model generated by machine learning using image data of a normal region, which is a region different from the defective region, as teacher data, and the tire to be inspected and a defect detection step of detecting defects in the internal structure of the. According to this inspection method, the tire inspection accuracy can be improved.

(8) The program proposed in the present disclosure includes image data acquisition means for acquiring image data of an X-ray image of a tire to be inspected, and image data of a defect area that is an area containing a defect in the internal structure of the tire and the defect area. A part or all of the image data of the tire to be inspected is input to a discrimination model generated by machine learning using image data of a normal region, which is a different region, as teacher data, and the inside of the tire to be inspected The computer functions as defect detection means for detecting structural defects. According to this program, the tire inspection accuracy can be improved.

(9) The identification model generation method proposed in the present disclosure includes an image data acquisition step of acquiring image data of an X-ray image of a tire, and image data of a defect region that is a region containing a defect of the internal structure of the tire. a defective region extracting step of extracting from the image data; a normal region extracting step of extracting image data of a normal region, which is a region different from the defective region, from the image data; and a model generation step of generating a discriminative model for detecting defects in the internal structure of the tire to be inspected, using the image data of the region as teacher data. Using the discriminative model generated by this method can improve the accuracy of tire inspection.

(10) In the generation method described in (9), the defect area extracting step extracts image data of defect area candidates, which are areas that may contain defects in the internal structure of the tire. The method may include a candidate extracting step of extracting from the image data and a determination result receiving step of receiving a worker's determination as to whether or not a defect is drawn in the defect area candidate. In the model generation step, the image data of the defect area candidate determined by the operator to depict a defect may be used as the image data of the defect area. Using the discriminative model generated by this method can further improve the tire inspection accuracy.

(11) The tire includes a belt and a carcass, and the carcass has a central portion of the carcass that is covered by the belt and side portions of the carcass that are not covered by the belt. . In the generating method described in (9), in the defective area extracting step, image data of the defective area is extracted from the image data of the carcass central portion, and image data of the defective area is extracted from the image data of the carcass side portion. is extracted, and in the model generation step, a first identification model is generated using the image data of the defect region extracted from the image data of the carcass central portion, and the first identification model extracted from the image data of the carcass side portion A second discriminant model may be generated using the image data of the defect region.

(12) The identification model generation device proposed in the present disclosure includes image data acquisition means for irradiating a tire with X-rays and acquiring image data of the tire, and defect, which is an area including a defect of the internal structure of the tire. defective area extracting means for extracting image data of an area from the image data; normal area extracting means for extracting image data of a normal area, which is an area different from the defective area, from the image data; and and a model generating means for generating a discriminative model for detecting defects in the internal structure of the tire to be inspected, using the image data and the image data of the normal region as teacher data. By using the discriminative model generated by this device, the accuracy of tire inspection can be improved.

(13) The program proposed in the present disclosure includes image data acquisition means for acquiring image data of an X-ray image of a tire, and image data of a defect area, which is an area containing a defect in the internal structure of the tire, from the image data. defective area extracting means for extracting; normal area extracting means for extracting image data of a normal area, which is an area different from the defective area, from the image data; and image data of the defective area and image data of the normal area. The computer is made to function as model generating means for generating a discriminative model for detecting defects in the internal structure of the tire to be inspected using the training data. Using discriminative models with this program can improve the accuracy of tire inspection.

1 is a block diagram showing hardware of a tire inspection device proposed in the present disclosure; FIG. It is a figure for explaining arrangement of an X-ray irradiation part and an image pick-up part, and an internal structure of a tire. It is a figure which shows the example of an X-ray image. FIG. 4 is a diagram showing an example of a carcass cord containing defects; FIG. 4 is a diagram showing an example of a carcass cord containing defects; FIG. 4 is a diagram showing an example of a carcass cord containing defects; 3 is a block diagram showing functions of a control unit according to the first embodiment; FIG. It is a figure for demonstrating the process of a carcass center part image extraction part. It is a figure for demonstrating the process of a carcass center part image extraction part. It is a figure for demonstrating the process of a carcass center part image extraction part. It is a figure for demonstrating the process of a carcass side part image extraction part. It is a figure for demonstrating the process of a candidate extraction part. It is a figure for demonstrating the process of a candidate extraction part. It is a figure for demonstrating the process of a candidate extraction part. 4 is a flow chart showing an example of processing executed in generating a discriminative model; 4 is a flow chart showing an example of processing executed in generating a discriminative model; 4 is a flow chart showing an example of processing executed in tire inspection. FIG. 9 is a block diagram showing functions of a control unit according to the second embodiment; FIG. 11 is a diagram showing an example of an image extracted by the belt image extraction unit shown in FIG. 10; FIG.

Hereinafter, a tire inspection apparatus and inspection method proposed in the present disclosure, and a discriminative model generation method and generation apparatus will be described.

FIG. 1 is a block diagram showing hardware of a tire inspection device 10 proposed in the present disclosure. As shown in FIG. 1 , the tire inspection device 10 has a control section 11 , a display section 13 , an operation section 14 , an imaging section 15 , an X-ray irradiation section 16 and a support section 17 . Elements of the tire inspection device 10 , such as the control unit 11 , function as an identification model generator used by the tire inspection device 10 .

FIG. 2 is a diagram for explaining the arrangement of the X-ray irradiation unit 16 and imaging unit 15 shown in FIG. 1 and the internal structure of the tire. Reference numeral 90 shown in FIG. 2 indicates both the tire to be inspected and the tire used to generate the discriminative model.

[Tire internal structure]
As shown in FIG. 2 , tire 90 has carcass layer 91 , belt layer 92 and finishing layer 93 . Each of these three layers 91, 92, 93 is composed of a plurality of linear members (cords) arranged in one direction.

As shown in FIG. 2 , the carcass layer 91 has a plurality of carcass cords 91 a arranged in the circumferential direction of the tire 90 . Each carcass cord 91a extends from one sidewall portion 90R of the tire 90 to the other sidewall portion 90L.

As shown in FIG. 2, the belt layer 92 is arranged inside the tread portion 90T and covers the central portion 91A of the carcass layer 91. As shown in FIG. The tire 90 may have multiple belt layers 92 . Each belt layer 92 has a plurality of belt cords 92 a arranged in the circumferential direction of the tire 90 . Each belt cord 92 a is arranged obliquely with respect to the circumferential direction of the tire 90 .

As shown in FIG. 2, the finishing layer 93 is disposed on the edges of the sidewall portions 90R and 90L and covers the bead 94. As shown in FIG. The finishing layer 93 is composed of a plurality of finishing cords 93 a arranged in the circumferential direction of the tire 90 .

The linear members (carcass cords 91a, belt cords 92a, and finishing cords 93a) forming the three layers 91, 92, and 93 are made of metal (specifically, steel).

[Hardware of inspection equipment]
The X-ray irradiation unit 16 is an X-ray tube that radially emits X-rays. As shown in FIG. 2, the X-ray irradiation unit 16 is arranged, for example, inside the tire 90 and irradiates the tire 90 with X-rays. The imaging unit 15 is arranged outside the tire 90 and outputs image data of X-rays that have passed through the tire 90 . Since X-rays pass through the rubber portion of the tire 90, an X-ray image showing the internal structure of the tire 90 is obtained.

As shown in FIG. 2 , the imaging unit 15 is arranged, for example, above the tire 90 (outside in the radial direction). In addition to or in place of the imaging unit 15, the tire inspection apparatus 10 is provided with the image pickup unit 15 or in place of the image pickup unit 15. It may have an imaging unit arranged in the upper left direction. Further, the tire inspection device 10 may have imaging units arranged on the right and left sides of the sidewall portions 90R and 90L of the tire 90, respectively.

The tire 90 is rotatably supported by the support portion 17 (FIG. 1). The support unit 17 rotates the tire 90 at a predetermined speed during tire inspection and generation of image data (teacher data) for machine learning. The imaging unit 15 continuously images the tire 90 at a frequency corresponding to the rotation speed of the tire 90 and outputs image data for one round of the tire 90 . The imaging unit 15 is, for example, a line sensor camera, but may be an area camera.

The control unit 11 has arithmetic units such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). Also, the control unit 11 has a storage unit 12 . The storage unit 12 includes a main storage device composed of RAM (random access memory) and an auxiliary storage device composed of ROM (read only memory), SSD (solid state drive), HDD (hard disk drive), and the like. . The storage unit 12 stores programs executed by an arithmetic device such as a CPU, image data output by the imaging unit 15, generated identification models, and the like.

The display unit 13 is a display device such as a liquid crystal display, and displays various images according to instructions from the control unit 11 .

The operation unit 14 is a user interface such as a keyboard and a mouse, and receives user operation input and outputs a signal indicating the content of the input to the control unit 11 .

FIG. 3 is a diagram showing an example of an X-ray image acquired by the imaging unit 15. As shown in FIG. In FIG. 3 , the vertical direction (Y direction) of the image corresponds to the circumferential direction of the tire 90 . The carcass layer 91 and the belt layer 92 appear in the central portion of the image (the area indicated by reference numeral 91A in the drawing). The belt layer 92 has a plurality of belt cords 92a arranged in the circumferential direction of the tire 90, as described above. Each belt cord 92 a is arranged obliquely with respect to the circumferential direction of the tire 90 . A plurality of carcass cords 91 a are arranged in the circumferential direction of the tire 90 . Each carcass cord 91a extends laterally in this image. Carcass layer 91 has sides 91B that are not covered by belt layer 92 . In this specification, the central portion 91A of the carcass layer 91 that is covered by the belt layer 92 in the carcass layer 91 is referred to as the "carcass central portion", and the side portion 91B of the carcass layer 91 that is not covered by the belt layer 92 is referred to as the " "Carcass side".

[Defects in tire internal structure]
The tire inspection device 10 is, for example, a device that inspects the carcass cords 91a. The plurality of carcass cords 91a are arranged at equal intervals in the circumferential direction of the tire 90. As shown in FIG. However, some of the carcass cords 91a are arranged obliquely, and a defect that intersects the adjacent carcass cords 91a may occur. Such defects are called across codes.

For example, as shown in FIG. 4A, two carcass cords 91a may be in line contact while intersecting defects. Also, as shown in FIG. 4B, two carcass cords 91a may intersect in an X shape. Also, as shown in FIG. 4C, one carcass cord 91a may be bent in a U-shape and cross the adjacent carcass cord 91a at two positions. The tire inspection device 10 detects such defects by the processing described later. Such defects are hereinafter referred to as "code defects".

[Processing of the control unit in the first embodiment]
The functions of the control unit 11 will be described below. FIG. 5 is a block diagram showing the functions of the control section 11. As shown in FIG. The control unit 11 has, as its functions, an image data acquisition unit 11a, a preprocessing unit 11b, a teacher data generation unit 11e, a model generation unit 11j, a defect area candidate extraction unit 11u, and a defect detection unit 11v. These functions are implemented by the control unit 11 executing programs stored in the storage unit 12 . The control unit 11 also has identification models M1 and M2. The identification models M1 and M2 are stored in the storage unit 12. FIG.

The image data acquisition unit 11a, the preprocessing unit 11b, the teacher data generation unit 11e, and the model generation unit 11j generate (learn) the discrimination models M1 and M2. The internal structure of the tire 90 is inspected using the identification models M1 and M2 by the image data acquisition section 11a, the preprocessing section 11b, the defect area candidate extraction section 11u, and the defect detection section 11v.

The control unit 11 may be composed of a plurality of personal computers. Alternatively, the control unit 11 may consist of one or more personal computers and one or more server computers. In this case, some functions of the control unit 11 (for example, the defect area candidate extraction unit 11u and the defect detection unit 11v) are executed by the personal computer, and other functions of the control unit 11 ( For example, the model generator 11j) may be executed on another personal computer or server computer.

The carcass central portion 91A and the carcass side portion 91B do not have the same form of crossing of the carcass cords 91a to be determined as cord defects. For example, there is a form of intersection that should be considered normal if it occurs in the carcass side 91B, whereas it should be considered a code defect if it occurs in the carcass center 91A. There are also types of intersections that are likely to occur in the carcass central portion 91A but are less likely to occur in the carcass side portions 91B. Therefore, the tire inspection apparatus 10 separately has a first identification model M1 used for inspection of the carcass center portion 91A and a second identification model M2 used for inspection of the carcass side portion 91B.

[Image data acquisition unit]
The image data acquisition unit 11 a acquires image data of an X-ray image of the tire 90 from the imaging unit 15 . Alternatively, the image data acquisition unit 11 a may acquire image data of the X-ray image of the tire 90 stored in the storage unit 12 . During inspection, the image data acquisition unit 11a acquires image data of the tire 90 to be inspected. At the time of learning (during generation) of the discrimination models M1 and M2, the image data acquisition unit 11a acquires image data for learning (image data for generating teacher data). The image data acquisition unit 11a may acquire image data of a plurality of tires during learning. Each image data may be image data for one round of the tire.

[Pretreatment part]
As shown in FIG. 5, the preprocessing section 11b has a carcass center image extraction section 11c and a carcass side image extraction section 11d.

[Carcass Central Image Extraction Unit]
The carcass central portion 91A is covered with a belt layer 92, and carcass cords 91a and belt cords 92a appear in the area of the carcass central portion 91A as shown in FIG. The carcass central portion image extracting section 11c extracts an image of the carcass cords 91a of the carcass central portion 91A from the image data of the tire 90 acquired by the image data acquiring section 11a. This extraction can be performed, for example, by the following processing.

The average density value in the region of the carcass central portion 91A is darker than the average density value in the region of the carcass side portion 91B. Therefore, the carcass central portion image extraction unit 11c generates an image by smoothing the density values of a plurality of pixels arranged in the vertical direction (the circumferential direction of the tire 90) of the image, as shown in FIG. 6A, for example. . In this image, an average density value of a plurality of pixels arranged in the vertical direction is given. In this image, the boundary (edges E1 and E2 in the example of FIG. 6A) between the area of the carcass central portion 91A and the area of the carcass side portion 91B appears. Then, the carcass central portion image extracting unit 11c performs edge detection on this image, and the area between the right edge E1 and the left edge E2 is extracted from the image data of the tire acquired by the image data acquiring unit 11a. Extract from Thereby, as shown in FIG. 6B, the image data of the area of the carcass central portion 91A is extracted.

In the image of the area of the carcass central portion 91A, both the carcass cords 91a and the belt cords 92a appear, as shown in FIG. 6B. The carcass central portion image extracting section 11c executes processing for sharpening the carcass cords 91a in the carcass central portion 91A.

Each belt cord 92a is obliquely arranged with respect to the circumferential direction of the tire 90, and the plurality of belt cords 92a are arranged in the circumferential direction of the tire 90 at regular intervals. Therefore, in the image of the tire 90, the belt cords 92a periodically appear as oblique lines. Therefore, the carcass central portion image extracting section 11c performs, for example, a two-dimensional Fourier transform process on the extracted image data of the carcass central portion 91A (image including the carcass layer 91 and the belt layer 92). Then, the carcass central portion image extracting section 11c removes the image data of the frequency component corresponding to the belt cord 92a. By doing so, as shown in FIG. 6C, an image is obtained in which the carcass cords 91a in the carcass central portion 91A appear clearly. The carcass central portion image extracting section 11c may perform a binarization process on the image thus obtained. This binarization processing may be dynamic binarization processing. Hereinafter, the image (FIG. 6C) extracted by the carcass center image extraction unit 11c is referred to as "carcass center image".

[Carcass Side Image Extraction Unit]
The carcass side image extraction unit 11d extracts image data of the carcass cord 91a of the carcass side portion 91B from the tire image data acquired by the image data acquisition unit 11a. This extraction can utilize part of the processing of the carcass central portion image extraction section 11c. That is, the carcass side image extraction unit 11d generates an image (see FIG. 6A) in which the density values of a plurality of pixels arranged in the longitudinal direction of the image (the circumferential direction of the tire 90) are smoothed, and the generated image Perform edge detection on . Then, the carcass side image extracting unit 11d extracts a further right region of the right edge E1 (see FIG. 6A) and a further left region of the left edge E2 (see FIG. 6A). As a result, image data of the left and right carcass side portions 91B are obtained as shown in FIG. 6D. The carcass side image extraction unit 11d may perform binarization processing on these image data. This binarization processing may also be dynamic binarization processing. Hereinafter, the image (FIG. 6D) extracted by the carcass side image extraction unit 11d is referred to as "carcass side image".

[Generation of discriminative model]
Processing for generating the identification models M1 and M2 using the learning image data (carcass center image and carcass side image) extracted by the processing of the preprocessing unit 11b will be described. The teacher data generating unit 11e and the model generating unit 11j shown in FIG. 5 execute processing for generating the discriminative models M1 and M2.

[Training data generator]
As shown in FIG. 5, the teacher data generator 11e has a defective area extractor 11f and a normal area extractor 11i. The defect area extractor 11f has a candidate extractor 11g and a determination result receiver 11h.

[Candidate extraction unit]
The candidate extracting unit 11g extracts a region in which a cord defect (for example, defects shown in FIGS. 4A to 4C) may occur from each of the learning carcass center image and the learning carcass side image. Extract. Hereinafter, a region where a code defect may occur is referred to as a "defective region candidate". The candidate extraction unit 11g extracts image data of defect area candidates based on at least one of the area and shape of the carcass cords 91a appearing in the carcass center image and the carcass side image. This processing is executed, for example, as follows.

FIG. 7A is an example of an enlarged carcass center image. Although ten carcass cords 91a are shown in this figure, some of the carcass cords 91a (the carcass cords within the area indicated by the two-dot chain line L) are overlapped. When a plurality of carcass cords 91a are arranged normally, each of the plurality of carcass cords 91a appears in the binarized image as a connected region in which pixels having a density value of 1 are connected. However, when a plurality of carcass cords 91a intersect, the entirety of the plurality of carcass cords 91a appears as one connected region. For example, in the example of FIG. 7A, three carcass cords 91a within the area indicated by the two-dot chain line L form one connecting area. Therefore, the area of the intersecting carcass cords 91a (the number of pixels in the connecting region) is larger than the area of the other carcass cords 91a.

Therefore, the candidate extraction unit 11g extracts a defect region candidate based on the area of the carcass code 91a appearing in the binarized carcass center image. For example, the candidate extraction unit 11g calculates the average value of the areas of the carcass cords 91a, and detects images of the carcass cords 91a having areas equal to or larger than the average value by a predetermined magnification (for example, 1.5 times). The candidate extraction unit 11g extracts the area including the detected image of the carcass cord 91a as a defect area candidate. For example, as indicated by a two-dot chain line L in FIGS. 7A and 7B, a circumscribing rectangle of the carcass cord 91a having an area equal to or larger than a predetermined magnification with respect to the average value is extracted as a defect region candidate.

The candidate extracting section 11g may extract the defect area candidate based on the shape of the carcass cord 91a appearing in the carcass center image. For example, the candidate extraction unit 11g may search the carcass center image for the branch shape of the carcass code 91a. Then, when a branch shape is found, a region including the branch shape may be extracted as a defect region candidate.

For example, the candidate extraction unit 11g may perform the thinning process on the area extracted based on the area (the image illustrated in FIG. 7B). By this thinning process, the line width of the image of each carcass cord 91a is gradually reduced, and each carcass cord 91a is represented as a line having a width of, for example, one pixel (see FIG. 7C). Then, the candidate extraction unit 11g may search for branch shapes in the thinned image of the carcass cord 91a. The search for the branch shape can be performed, for example, by scanning the thinned carcass cord 91a image in the lateral direction of the image. Then, when a branch shape exists, the candidate extraction unit 11g may extract a circumscribed rectangle of a connected region (the region shown in FIG. 7B) including the branch shape as a defect region candidate. By performing such processing, for example, when two adjacent carcass cords 91a are close to each other but do not intersect, the image extracted based on the area is immediately extracted as a defect region candidate. can be prevented.

As described above, the candidate extraction unit 11g extracts image data of defect area candidates based on the area and shape (whether or not there is a branch shape) of the carcass cord 91a appearing in the carcass center image, for example. . Alternatively, the candidate extracting section 11g may extract the image data of the defect region candidate based only on the area of the carcass cord 91a appearing in the carcass center image. As still another example, the candidate extracting section 11g may extract the image data of the defect area candidate based only on the shape (whether or not there is a branch shape) of the carcass code 91a appearing in the carcass central portion image.

The candidate extraction unit 11g performs the same processing as that for the carcass center image on the carcass side image. That is, the candidate extraction unit 11g may extract the defect area candidate from the carcass side image based on the area of the carcass code 91a appearing in the binarized carcass side image. Further, the candidate extracting section 11g may extract defect area candidates from the carcass side image based on the shape of the carcass cord 91a appearing in the carcass side image.

[Judgment result reception part]
The determination result reception unit 11h displays on the display unit 13 images of the defect area candidates extracted by the candidate extraction unit 11g. Then, the judgment result reception unit 11h receives the operator's judgment result as to whether or not the displayed image actually includes a code defect. The operator can input the determination result through the operation unit 14 . When a plurality of defect area candidates are stored in the storage unit 12, the determination result receiving unit 11h sequentially displays the plurality of defect area candidates on the display unit 13, and receives the worker's determination result for each defect area candidate. .

The determination result reception unit 11h assigns an identification label to each defect area candidate based on the input determination result, and stores the identification label in the storage unit 12 . Specifically, when a determination result indicating that a code defect appears in a defect area candidate is input, the determination result reception unit 11h assigns the identification label "defect" to the defect area candidate. Then, the determination result reception unit 11h stores the image data of the defect area candidate in the storage unit 12 as the defect area. On the other hand, when the determination result indicating that the code defect does not appear in the defect area candidate is input, the determination result reception unit 11h assigns the identification label "normal" to the defect area candidate, and returns the image data of the defect area candidate to the normal area. , and stored in the storage unit 12 . The defective area and the normal area are stored in the storage unit 12 together with information on the part (the carcass central part 91A or the carcass side part 91B) where they existed.

If a code defect appears in the defect area candidates, the operator may enter the type of code defect. Types of code defects include, for example, the types shown in FIGS. 4A to 4C. For example, when a shape similar to the code defect shown in FIG. 4A appears in the defect area candidate, the determination result reception unit 11h assigns the identification label "first type defect" to this defect area candidate. Then, the determination result receiving unit 11h stores the image data of the defect area candidate in the storage unit 12 as the defect area of the "first type defect". Similarly, when a code defect having a shape similar to the code defect shown in FIG. The image data of this defect area candidate is stored in the storage unit 12 as the defect area of the "second type defect". The same is true when a code defect having a shape similar to the code defect shown in FIG. 4C appears in the defect area candidates. The number of defect types is not limited to this. The number of defect types may be two or more than three.

[Normal region extraction unit]
The normal area extractor 11i extracts a part of the area that was not extracted as a defect area candidate in the process of the candidate extractor 11g as a normal area. The normal region extracting unit 11 i assigns an identification label “normal” to the image extracted as the normal region, and stores the image data of this normal region in the storage unit 12 . The normal region extraction unit 11i extracts normal regions from each of the carcass center image and the carcass side image. The image data of the normal region thus extracted are stored in the storage unit 12 together with the information of the portion (the carcass central portion 91A or the carcass side portion 91B) where they existed.

The normal region extraction process can be performed, for example, as follows. The normal region extraction unit 11i randomly extracts an image having a predetermined number of pixels and a predetermined aspect ratio from the carcass center image. Then, if the randomly extracted image does not have a portion that overlaps with the image extracted as the defect area candidate, the normal area extractor 11i may set this extracted image as the normal area. Here, the predetermined number of pixels and the predetermined aspect ratio may be, for example, the average number of pixels and the average aspect ratio of the defect area candidate.

The teacher data generation unit 11e converts each of the image data of the plurality of defect areas obtained by the above-described processing into image data having a predetermined number of pixels and a predetermined aspect ratio. (Hereinafter, this number of pixels will be referred to as the "model input pixel number", and this aspect ratio will be referred to as the "model input aspect ratio".) The image data of the area is converted into the model input pixel number and the model input aspect ratio. In other words, the teacher data generator 11e unifies the dimensions of the teacher data.

[Model generator]
The model generation unit 11j inputs the image data of the defect region and the identification label (that is, the correct label) to the pre-learning identification model as teacher data. Also, the model generation unit 11j inputs the image data of the normal region and the identification label (that is, the correct label) to the same identification model as teacher data. By doing so, the model generator 11j generates the identification models M1 and M2 for detecting defects in the internal structure of the tire 90. FIG.

A neural network, for example, may be used as the discrimination models M1 and M2. A convolutional neural network (CNN) may be used as the discriminative models M1 and M2. Alternatively, support vector machines (SVM), random forests, etc. may be used as discriminative models M1 and M2.

As described above, the teacher data generator 11e extracts the defective region and the normal region from the carcass center image and the carcass side image, respectively. The model generator 11j uses the defective region and the normal region extracted from the carcass center image to generate a first identification model M1 for detecting cord defects in the carcass center 91A. The model generator 11j also uses the defective region and the normal region extracted from the carcass side image to generate a second identification model M2 for detecting cord defects in the carcass side 91B.

As described above, the carcass central portion 91A and the carcass side portion 91B do not have the same form of crossing of the carcass cords 91a that should be determined to be defective. For example, there is a form of intersection that should be considered normal if it occurs in the carcass side 91B, whereas it should be considered a code defect if it occurs in the carcass center 91A. There are also types of intersections that are likely to occur in the carcass central portion 91A but are less likely to occur in the carcass side portions 91B. In this embodiment, two identification models M1 and M2 different from each other are generated for two portions of the carcass layer 91 (the carcass central portion 91A and the carcass side portions 91B). Therefore, it is possible to improve the code defect detection accuracy using the discrimination models M1 and M2.

[Tire inspection]
Processing for inspecting a tire using the carcass center image and the carcass side image of the tire to be inspected will be described. The defect area candidate extraction unit 11u and the defect detection unit 11v execute processing for inspection.

[Defective Area Candidate Extraction Unit]
The defect area candidate extraction unit 11u extracts defect area candidates (see FIG. 7B) from the image data of the tire 90 to be inspected extracted by the preprocessing unit 11b, that is, from the carcass center image and the carcass side image. Extract.

The processing of the defect area candidate extraction unit 11u may be the same as the processing of the candidate extraction unit 11g of the teacher data generation unit 11e described above, for example. That is, the defect area candidate extraction unit 11u extracts the image data of the defect area candidate based on the areas of the carcass cords 91a appearing in the carcass center image and the carcass side image, for example. Instead of or in addition to the area of the carcass cord 91a, the defect area candidate extraction unit 11u extracts image data of the defect area candidate based on the shape of the carcass cord 91a appearing in the carcass center image and the carcass side image. may be extracted. For example, the defect area candidate extracting unit 11u may search for branch shapes of the carcass code 91a in the carcass center image and the carcass side image, and extract an area including the branch shape as a defect area candidate.

[Defect detector]
The defect detection unit 11v inputs the image data of the defect area candidates extracted by the defect area candidate extraction unit 11u to the identification models M1 and M2. The defect detection unit 11v inputs the defect area candidate extracted from the carcass center image to the first discrimination model M1, and classifies the defect area candidate into one of a plurality of classes. Further, the defect detection unit 11v inputs the defect area candidate extracted from the carcass side image to the second discrimination model M2, and classifies the defect area candidate into one of a plurality of classes.

Here, the plurality of classes are, for example, a class in which a code defect appears in a defect area candidate (a class given the identification label "defect"), or a class in which a code defect does not appear in a defect area candidate (identification label " normal” is given). Such processing by the defect detection unit 11v can detect defects in the internal structure of the tire 90 to be inspected. The output of the defect detector 11v is, for example, the probability that the defect area candidate corresponds to each class. Alternatively, the output of the defect detector 11v may be the identification label of the class to which the defect area candidate has the highest probability.

As described above, when the identification models M1 and M2 are generated, the image data of the defect area to which the identification label representing the type of defect (for example, "type 1 defect" or "type 2 defect") is assigned is used as a teacher. It may be used as data. In this case, the plurality of classes classified by the defect detector 11v may include a plurality of classes respectively corresponding to the types of code defects appearing in the defect area candidates. For example, the plurality of classes may be a class corresponding to the first type code defect (see FIG. 4A) (class assigned with the identification label "first type defect") or a second type code defect (see FIG. 4B). (a class to which the identification label "Type 2 defect" is assigned) corresponding to . This processing by the defect detector 11v can detect not only the presence or absence of defects but also the types of defects.

The defect detection unit 11v displays the output on the display unit 13 or stores it in the storage unit 12 together with the image of the defect area candidate, for example. When a plurality of defect area candidates are stored in the storage unit 12 for one tire 90, the defect detection unit 11v sequentially inputs all of the plurality of defect area candidates to the identification models M1 and M2, and outputs each may be displayed on the display unit 13 or stored in the storage unit 12 .

[Learning flow]
8A and 8B are diagrams showing an example of processing executed by the control unit 11 to generate the discriminative models M1 and M2.

The control unit 11 drives the X-ray irradiation unit 16 to irradiate the tires 90 for generating learning data with X-rays. The image data acquisition unit 11a acquires an X-ray image of the tire 90 from the imaging unit 15 (S101). The image data acquisition unit 11 a may acquire X-ray images stored in the storage unit 12 . The carcass center image extraction unit 11c extracts the carcass center image (see FIG. 6C) from the image data acquired in S101. Further, the carcass side image extracting unit 11d extracts the carcass side image (see FIG. 6D) from the image data acquired in S101 (S102).

Next, the teacher data generation unit 11e executes processing for generating teacher data (S103). Specifically, as shown in FIG. 8B, the defective area extracting unit 11f (candidate extracting unit 11g) extracted an area (defective area candidate, FIG. 7B) that may cause a code defect in S102. It is extracted from the carcass center image (S201). The defective area extraction unit 11f (determination result reception unit 11h) displays the extracted defect area candidates on the display unit 13. FIG. Then, the defect area extraction unit 11f (determination result reception unit 11h) receives the operator's determination result as to whether or not the displayed defect area candidate includes a code defect, and an identification label ( correct label) is given to the defect area candidate (S202). Identification labels are, for example, "normal" and "defective". Image data to which the identification label "defect" is assigned is stored in the storage unit 12 as a defective area, and image data to which the identification label "normal" is assigned is stored in the storage unit 12 as a normal area. As the identification label, a label indicating the type of code defect (for example, "type 1 defect" or "type 2 defect") may be added.

The normal region extraction unit 11i extracts a part of the region that was not extracted (selected) as a defect region candidate in S201 as a normal region from the carcass center image, and assigns the identification label “normal” to the image data of the normal region. (S203). This image data with the identification label "normal" is stored in the storage unit 12 as a normal region.

Next, the teacher data generator 11e executes the same processes as S201 to S203 for the carcass side portion 91B. Specifically, the defect area extraction unit 11f (candidate extraction unit 11g) extracts an area (defect area candidate) that may cause a code defect from the carcass side image extracted in S102 (S204). . The defective area extraction unit 11f (determination result receiving unit 11h) displays the defective area candidates on the display unit 13. FIG. Then, the defect area extraction unit 11f (determination result reception unit 11h) receives the operator's determination result as to whether or not the displayed defect area candidate includes a code defect, and an identification label ( correct label) is given to the defect area candidate (S205). Identification labels are, for example, "normal" and "defective". Image data to which the identification label "defect" is assigned is stored in the storage unit 12 as a defective area, and image data to which the identification label "normal" is assigned is stored in the storage unit 12 as a normal area. As the identification label, a label indicating the type of code defect (for example, "type 1 defect" or "type 2 defect") may be assigned instead of "defect".

The normal region extraction unit 11i extracts a part of the region that was not extracted (selected) as a defect region candidate in S204 as a normal region from the carcass side image, and assigns the identification label “normal” to the image data of the normal region. (S206). This image data with the identification label "normal" is stored in the storage unit 12 as a normal region.

The teacher data generation unit 11e determines whether or not a predetermined number of image data sets (a predetermined number of teacher data sets) have been prepared for each of the plurality of classes through the processing of S201 to S206 (S207). Here, the plurality of classes are, for example, a class in which a defect area candidate has a code defect (a class given the identification label "defect"), a class in which a defect area candidate does not have a code defect (a class given the identification label "normal"). class). In this case, the teacher data generation unit 11e generates a defective region extracted from the carcass central image, a normal region extracted from the carcass central image, a defective region extracted from the carcass side image, and a normal region extracted from the carcass side image. has reached a predetermined number. Here, the predetermined number is the number deemed necessary to generate the discriminative models M1 and M2. If the number of pieces of image data for each area has not reached the predetermined number, the teacher data generator 11e returns to the process of S201 and executes the subsequent processes. A plurality of classes into which defect area candidates are classified may include a plurality of classes corresponding to types of code defects, such as "type 1 defect" and "type 2 defect".

The teacher data generation unit 11e converts the number of pixels and the aspect ratio of each image data (defective region and normal region) stored in the storage unit 12 into a predetermined number of pixels (the "number of model input pixels" described above) and a predetermined aspect ratio. ("model input aspect ratio" described above) (S208).

Note that the order of processing performed by the training data generation unit 11e is not limited to the example shown in FIG. 8B. For example, the process of extracting the defective area and the normal area from the carcass side image (S204 to S206) may be performed before the process of extracting the defective area and the normal area from the carcass side image (S201 to S203). . Further, the defect area extracting section 11f may extract a plurality of defect area candidates from the carcass center image (or carcass side image). Then, the defect area extraction unit 11f may display the plurality of defect area candidates on the display unit 13 at once.

Returning to FIG. 8A, the model generation unit 11j inputs the image data (teacher data) extracted in S201 and S202 to the pre-learning discrimination model M1, and extracts the discrimination model M1 for detecting cord defects in the carcass central portion 91A. is generated (learned) (S104). Further, the model generation unit 11j inputs the image data (teacher data) extracted in S204 and S206 to the discrimination model M2 before learning, and generates the discrimination model M2 for detecting the cord defect in the carcass side part 91B ( learning) (S105). The above is an example of the processing executed by the control unit 11 to generate the discriminative models M1 and M2.

[Inspection flow]
Next, an example of processing executed by the control unit 11 to detect defects in the internal structure of the tire will be described with reference to FIG.

The control unit 11 drives the X-ray irradiation unit 16 to irradiate the tire 90 to be inspected with X-rays. The image data acquisition unit 11a acquires the image data of the X-ray image of the tire 90 from the imaging unit 15 (S301). The preprocessing unit 11b (carcass center image extraction unit 11c) extracts a carcass center image (see FIG. 6C) from the image data acquired in S301 (S302). The preprocessing unit 11b (carcass side image extraction unit 11d) extracts a carcass side image (S302).

The defect area candidate extraction unit 11u extracts areas (defect area candidates) in which code defects may occur from the carcass center image extracted in S302 and the carcass side image extracted in S302 ( S303). The defect area candidate extraction unit 11u converts the number of pixels and the aspect ratio of the extracted defect area candidate into the number of model input pixels and the model input aspect ratio (S304).

The defect detection unit 11v inputs the image data of the defect area candidate extracted from the carcass center image to the first identification model M1 (S305). Then, the defect detector 11v classifies the defect area candidates into a plurality of classes (S306). The plurality of classes are, for example, a class in which a code defect appears in a defect area candidate (a class with an identification label "defect"), or a class in which a code defect does not appear in a defect area candidate (a class with an identification label "normal"). given class). A plurality of classes classified by the defect detection unit 11v may include a plurality of classes corresponding to types of code defects. This processing by the defect detection unit 11v makes it possible to detect not only the presence or absence of defects in the internal structure of the tire 90 to be inspected, but also the types of defects. The defect detection unit 11v determines whether or not all defect area candidates extracted from the carcass center image in S303 have been classified (S307). If unclassified defect area candidates remain, the defect detection unit 11v returns to S305 and performs subsequent processing for the unclassified defect area candidates.

Next, the defect detection unit 11v performs the same processing as S305 to S307 on the defect area candidate extracted from the carcass side image. That is, the defect detection unit 11v inputs the image data of the defect area candidate extracted from the carcass side image to the second identification model M2 (S308). Then, the defect detector 11v classifies the defect area candidates into a plurality of classes (S309). The defect detection unit 11v determines whether or not all defect area candidates extracted from the carcass side image in S303 have been classified (S310). If unclassified defect area candidates remain, the defect detection unit 11v returns to S308 and performs subsequent processing for the unclassified defect area candidates.

If it is determined in S310 that all defect area candidates extracted from the carcass side image have been classified, the control section 11 ends the process. That is, the control unit 11 ends the inspection of the internal structure of the tire 90 .

[Processing of the control unit in the second embodiment]
Further, in the tire inspection apparatus 10 described above, cord defects occurring in the carcass layer 91 are detected. However, code defects may also occur in the belt layer 92 and the finishing layer 93 . For example, two adjacent belt cords 92a may cross, or two adjacent finishing cords 93a may cross. The tire inspection device may detect cord defects occurring in the belt layer 92 and the finishing layer 93 . When detecting a cord defect in the finishing layer 93, the tire inspection device 10 faces the sidewall portions 90L and 90R of the tire 90 in addition to the imaging portion 15 positioned radially outwardly of the tire 90. It may have an imaging unit.

FIG. 10 is a block diagram showing functions of the control section 111 included in such an inspection apparatus. An example of detecting a cord defect occurring in the belt layer 92 and a cord defect occurring in the carcass layer 91 will be described below with reference to FIG.

The control unit 111 has an image data acquisition unit 11a, a preprocessing unit 111b, a teacher data generation unit 111e, a model generation unit 111j, a defect area candidate extraction unit 111u, and a defect detection unit 111v. The preprocessing section 111b has a belt image extraction section 111p and a carcass image extraction section 111q.

The image data acquisition unit 11a acquires the image data of the X-ray image of the tire through the imaging unit 15, like the image data acquisition unit 11a of the control unit 11 described above. During inspection, the image data acquisition unit 11a acquires image data of the tire 90 to be inspected. During learning, the image data acquisition unit 11a acquires image data of the tire 90 for generating image data (teacher data) for learning.

The belt image extraction unit 111p extracts image data of the belt cords 92a from the tire image data acquired by the image data acquisition unit 11a. This extraction can be performed, for example, by the following processing.

The belt layer 92 covers the carcass central portion 91A, and the average of the density values in the region where the belt layer 92 appears is the average of the density values in the region where only the carcass layer 91 appears (region of the carcass side portion 91B). becomes darker than Therefore, the belt image extraction unit 111p, like the carcass center image extraction unit 11c described above, smoothes the density values of a plurality of pixels arranged in the vertical direction (circumferential direction of the tire 90) of the image (Fig. 6A ). In this image, the boundaries (edges E1 and E2 in the example of FIG. 6A) between the area of the belt layer 92 (the area of the carcass central portion 91A) and the area of the carcass side portions 91B appear. Then, the belt image extraction unit 111p performs edge detection on this image, and extracts an area between the right edge E1 and the left edge E2 from the image data acquired by the image data acquisition unit 11a. As a result, the image data (see FIG. 6B) of the belt layer 92 and the carcass central portion 91A are extracted.

The belt image extraction unit 111p executes processing for sharpening the belt cord 92a. For example, the belt image extraction unit 111p executes a two-dimensional Fourier transform process on the extracted image (image including the carcass cords 91a and the belt cords 92a). Then, the belt image extraction unit 111p removes the image data of the frequency component corresponding to the carcass code 91a. By doing so, as shown in FIG. 11, an image in which the belt cord 92a clearly appears can be obtained. The belt image extraction unit 111p may perform binarization processing on the image thus obtained. This process may be a dynamic binarization process. An image obtained by these processes of the belt image extraction unit 111p is hereinafter referred to as a "belt image".

The tire 90 may have a plurality of belt layers 92 with belt cords 92a having different angles. For example, the tire 90 has a first belt layer 92 composed of belt cords 92a arranged at a first angle (+45 degrees with respect to the circumferential direction of the tire 90) and a second angle (circumferential direction of the tire 90). and a second belt layer 92 consisting of belt cords 92a arranged at -45 degrees to the direction. In this case, the belt image extraction unit 111p removes the carcass cords 91a and the belt cords 92a of the second belt layer 92 by two-dimensional Fourier transform, and extracts the belt image from which the belt cords 92a of the first belt layer 92 clearly appear. Generate an image (see FIG. 11). Further, the belt image extracting unit 111p removes the carcass cords 91a and the belt cords 92a of the first belt layer 92 by two-dimensional Fourier transform, and extracts a belt image in which the belt cords 92a of the second belt layer 92 clearly appear. also generates

The carcass image extraction unit 111q performs two-dimensional Fourier transform processing on the image acquired by the image data acquisition unit 11a, and removes frequency components corresponding to the belt cords 92a. Also, the carcass image extraction unit 111q executes a binarization process to generate an image in which the carcass code 91a appears clearly. Below, the image obtained by this process of the carcass image extraction unit 111q is referred to as a "carcass image". Here, the carcass image may be, for example, an image including both the carcass center portion 91A and the carcass side portions 91B described above.

[Generation of discriminative model]
Processing for generating the discriminative models M3 and M4 using the learning belt image and carcass image will be described. The teacher data generating unit 111e and the model generating unit 111j execute processing for generating the discriminative models M3 and M4. The teacher data generation unit 111e has a candidate extraction unit 111g and a determination result reception unit 111h.

[Candidate extraction part]
The candidate extraction unit 111g extracts an area (defective area candidate) in which a code defect may occur from the belt image. When a plurality of belt cords 92a intersect, the area of the intersecting belt cords 92a (the number of pixels in the connecting area) is larger than the area of each of the other belt cords 92a. Therefore, the candidate extraction unit 111g extracts the image data of the defect area candidates, for example, based on the area of the belt cord 92a appearing in the belt image. For example, the candidate extracting unit 111g detects an image of the belt cord 92a having an area equal to or larger than a predetermined magnification (for example, 1.5 times) with respect to the average value of the areas of the belt cords 92a, and extracts the circumscribed rectangle of the belt cord 92a. is a defect region candidate.

The candidate extraction unit 111g may extract the image data of the defect area candidates based on the shape of the belt cord 92a appearing in the belt image. For example, the candidate extraction unit 111g may search for branch shapes of the belt cords 92a in the belt image, and extract regions including the branch shapes as defect region candidates.

The candidate extraction unit 111g also extracts defect area candidates from the carcass image. This process may be similar to the process of extracting defect area candidates from the belt image. That is, the candidate extraction unit 111g detects, for example, an image of the carcass cord 91a having an area equal to or greater than a predetermined magnification (for example, 1.5 times) of the average value of the areas of the carcass cords 91a, and extracts the image of the carcass cord 91a. A circumscribing rectangle is set as a defect area candidate. The candidate extraction unit 111g may search for branch shapes of the carcass code 91a in the carcass image, and extract regions including the branch shapes as defect region candidates.

[Judgment result reception part]
The determination result reception unit 111h displays on the display unit 13 the defect area candidates extracted from the carcass image and the defect area candidates extracted from the belt image. Further, the judgment result reception unit 11h receives the operator's judgment result as to whether or not each of these defect area candidates includes a defect. Then, the determination result receiving unit 11h assigns an identification label (“defective” or “normal”) according to the determination result to the image data of the defect area candidate. The image data with the identification label "defect" is stored in the storage unit 12 as a defect area. Image data to which the identification label "normal" is assigned is stored in the storage unit 12 as a normal region.

If a code defect appears in the defect area candidates, the operator may input the type of code defect as an identification label through the operation unit 14 . For example, when a shape similar to the code defect shown in FIG. 4A appears in the defect area candidate, the determination result reception unit 11h assigns the identification label "first type defect" to this defect area candidate. Similarly, for other types of code defects, the determination result reception unit 11h assigns identification labels such as "second type defect" and "third type defect" to each defect area candidate. The number of defect types is not limited to this. The number of defect types may be two or more than three. The types of cord defects may be specified for each of defects occurring in the carcass cord 91a and defects occurring in the belt cord 92a.

[Normal region extraction part]
The normal region extraction unit 111i extracts, as normal regions, portions of regions that were not selected as defective region candidates in the process of the candidate extraction unit 111g from the carcass image and the belt image. For example, the normal region extraction unit 111i randomly extracts an image having a predetermined number of pixels and a predetermined aspect ratio from each of the carcass image and the belt image. Then, when the randomly extracted image does not have a portion that overlaps with the defect area candidate, the normal area extraction unit 111i determines the randomly extracted image as a normal area. The normal region extraction unit 111 i assigns the identification label “normal” to the extracted image and stores it in the storage unit 12 .

[Model generator]
The model generation unit 111j uses the image data (the image data of the defective region and the image data of the normal region) generated by the training data generation unit 111e as training data to detect defects in the internal structure of the tire. Generate discriminative models M3 and M4.

The model generation unit 111j generates the first discriminant model M3 using teacher data obtained from the belt image. Also, the model generation unit 111j generates the second discriminant model M4 using teacher data obtained from the carcass image.

A neural network, for example, may be used as the discrimination models M3 and M4. A convolutional neural network (CNN) may be used as the discrimination models M3 and M4. Alternatively, a support vector machine (SVM), a random forest, or the like may be used as the discriminative models M3 and M4.

[Tire inspection]
Processing for inspecting a tire using a belt image and a carcass image of the tire to be inspected will be described. The defect area candidate extraction unit 111u and the defect detection unit 111v execute processing for inspection.

[Defective Area Candidate Extraction Unit]
The defect area candidate extraction unit 111u extracts defect area candidates from the image data of the tire 90 to be inspected (that is, the carcass image and the belt image) obtained by the preprocessing unit 11b.

Similar to the candidate extraction unit 111g, the defect area candidate extraction unit 111u extracts the image data of the defect area candidates from the belt image, for example, based on the area of the belt cord 92a appearing in the belt image. Further, the defect area candidate extracting section 111u may search for the branch shape of the belt cord 92a in the belt image and extract the area including the branch shape as the defect area candidate.

The defect area candidate extraction unit 111u also extracts defect area candidates from the carcass image. For example, the defect area candidate extraction unit 111u extracts image data of defect area candidates from the carcass image based on the area of the carcass code 91a appearing in the carcass image. Further, the defect area candidate extraction unit 111u may search for a branch shape of the carcass code 91a in the carcass image and extract an area including the branch shape as a defect area candidate.

[Defect detector]
The defect detection unit 111v inputs the defect area candidates obtained from the belt image to the first discrimination model M3, and classifies the defect area candidates into a plurality of classes. Further, the defect detection unit 111v inputs the defect area candidates extracted from the carcass image to the second identification model M4, and classifies the defect area candidates into a plurality of classes.

Here, the plurality of classes are, for example, a class in which a code defect appears in a defect area candidate (a class given the identification label "defect"), or a class in which a code defect does not appear in a defect area candidate (identification label " normal” is given). Through such processing by the defect detection section 111v, defects in the belt layer 92 and defects in the carcass layer 91 can be detected. The output of the defect detection unit 111v is, for example, the probability that the defect area candidate corresponds to each class. Alternatively, the output of the defect detector 111v may be the identification label of the class to which the defect area candidate has the highest probability.

As described above, when the identification models M3 and M4 are generated, the image data to which identification labels indicating the types of defects (for example, "type 1 defect" or "type 2 defect") are added are used as training data. may be In this case, the plurality of classes classified by the defect detection section 111v may be a plurality of classes corresponding to the types of code defects appearing in the defect area candidates.

The defect detection unit 11v displays the outputs from the identification models M3 and M4 on the display unit 13 or stores them in the storage unit 12 together with, for example, images of defect area candidates. When a plurality of defect area candidates are stored in the storage unit 12 for one tire 90, the defect detection unit 111v sequentially inputs all of the plurality of defect area candidates to the identification models M3 and M4, and outputs each of them. may be displayed on the display unit 13 or stored in the storage unit 12 .

[Modification]
The tire inspection device proposed in the present disclosure is not limited to the above-described first embodiment and second embodiment.

For example, in the tire inspection apparatus described above, defect area candidates are extracted from the image of the tire 90 to be inspected, and only the image data of the defect area candidates are input to the identification model. Unlike this, for example, an image of the tire 90 to be inspected (carcass center image, carcass side image, carcass image, belt image, etc.) is divided into a plurality of parts, all of which are represented by the identification model M1. It may be input to M2, M3, and M4.

Further, in the first embodiment, the image data of the defect area candidate extracted from the carcass center image and the image data of the defect area candidate extracted from the carcass side image are input to the two identification models M1 and M2, respectively. there is Alternatively, these two types of image data may be input into a common discriminative model. Further, in the second embodiment, the image data of the defect area candidates extracted from the carcass image and the image data of the defect area candidates extracted from the belt image may be input to a common identification model.

10: tire inspection device, 11/111: control unit, 11a/111a: image data acquisition unit, 11b/111b: preprocessing unit, 11c: carcass center image extraction unit, 11d: carcass side image extraction unit, 11e. 111e: teacher data generating unit, 11f/111f: defective region extracting unit, 11g/111g: candidate extracting unit, 11h/111h: determination result receiving unit, 11i/111i: normal region extracting unit, 11j/111j: model generating unit, 11u and 111u: defect area candidate extraction unit, 11v and 111v: defect detection unit, 12: storage unit, 13: display unit, 14: operation unit, 15: imaging unit, 16: X-ray irradiation unit, 17: support unit, 90: tire, 90L/00R: sidewall portion, 90T: tread portion, 91: carcass layer, 91A: carcass center portion, 91B: carcass side portion, 91a: carcass cord, 92: belt layer, 92a: belt cord, 93 : finishing layer, 93a: finishing cord, 94: bead, 111p: belt image extractor, 111q: carcass image extractor.

Claims (13)

image data acquisition means for acquiring image data of an X-ray image of a tire to be inspected;
A discrimination model generated by machine learning using image data of a defective region, which is a region containing a defect in the internal structure of the tire, and image data of a normal region, which is a region different from the defective region, as teacher data, is added to the inspection object. and defect detection means for inputting part or all of the image data of the tire and detecting defects in the internal structure of the tire to be inspected. further comprising defect area candidate extraction means for extracting image data of a defect area candidate, which is a candidate for an area containing a defect of the internal structure, from the image data of the tire to be inspected;
The defect detection means inputs the image data of the candidate defect area to the identification model as a part of the image data of the tire to be inspected, and determines whether or not the candidate defect area includes a defect. 1. The tire inspection device described in 1. The internal structure of the tire has a plurality of linear members arranged in one direction,
Defects in the internal structure of the tire include intersections of the linear members,
The tire inspection according to claim 2, wherein the defect area candidate extracting means extracts the image data of the defect area candidate based on at least one of the area and shape of each linear member appearing in the image of the tire to be inspected. Device. the tire internal structural defect includes a first type and a second type;
The identification model includes image data of the defect region including the first type of defect, image data of the defect region including the second type of defect, and image data of the normal region as the teacher data. is generated using as
The tire according to claim 1, wherein the defect detection means inputs the part or all of the image data of the tire to be inspected into the identification model to specify the type of defect in the internal structure of the tire to be inspected. inspection equipment. The tire to be inspected includes a belt and a carcass, and the carcass has a carcass center portion that is a portion covered by the belt and a carcass side portion that is a portion not covered by the belt,
The tire inspection device includes: a carcass central portion image extracting means for extracting image data of the carcass central portion from the tire image data; and a carcass side portion extracting image data of the carcass side portion from the tire image data. and partial image extraction means,
The discriminative models include a first discriminative model for detecting defects in the carcass center and a second discriminative model for detecting defects in the carcass sides;
The defect detection means inputs part or all of the image data of the carcass central portion into the first identification model, and inputs part or all of the image data of the carcass side portion into the second identification model. The tire inspection device according to claim 1. The tire to be inspected has a carcass layer, a belt layer, and a finishing layer as the internal structure,
The tire inspection device according to claim 1, wherein the defect detection means detects defects in the carcass layer, the belt layer, or the finishing layer. an image data acquisition step of acquiring image data of an X-ray image of a tire to be inspected;
A discrimination model generated by machine learning using image data of a defective region, which is a region containing a defect in the internal structure of the tire, and image data of a normal region, which is a region different from the defective region, as teacher data, is added to the inspection object. and a defect detection step of inputting part or all of the image data of the tire and detecting defects in the internal structure of the tire to be inspected. Image data acquisition means for acquiring image data of an X-ray image of a tire to be inspected, and image data of a defective area that is an area containing defects in the internal structure of the tire and image data of a normal area that is an area different from the defective area. A computer as defect detection means for detecting defects in the internal structure of the tire to be inspected by inputting part or all of the image data of the tire to be inspected into a discrimination model generated by machine learning using as teaching data program that makes it work. an image data acquisition step of acquiring image data of an X-ray image of the tire;
a defect area extracting step of extracting image data of a defect area, which is an area containing a defect in the internal structure of the tire, from the image data;
A normal region extraction step of extracting image data of a normal region, which is a region different from the defective region, from the image data;
A method for generating a discriminative model, comprising: a model generating step of generating a discriminative model for detecting defects in the internal structure of a tire to be inspected, using the image data of the defective region and the image data of the normal region as teacher data. . The defect area extraction step includes a candidate extraction step of extracting image data of a defect area candidate, which is an area that may contain a defect in the internal structure of the tire, from the image data of the tire; a decision result receiving step of receiving the worker's decision as to whether or not a defect is drawn in the
10. The identification model generation method according to claim 9, wherein in the model generating step, the image data of the defect area candidate determined by the operator to depict a defect is used as the image data of the defect area. . The tire includes a belt and a carcass, the carcass having a central carcass portion that is covered by the belt and side carcass portions that are not covered by the belt;
In the defective area extracting step, image data of the defective area is extracted from the image data of the carcass central portion, image data of the defective area is extracted from the image data of the carcass side portion,
In the model generation step, a first identification model is generated using image data of the defect area extracted from the image data of the carcass central portion, and an image of the defect area extracted from the image data of the carcass side portion. 10. The method of generating a discriminative model according to claim 9, wherein the data is used to generate a second discriminative model. image data acquisition means for irradiating a tire with X-rays and acquiring image data of the tire;
Defect area extracting means for extracting image data of a defect area, which is an area containing a defect in the internal structure of the tire, from the image data;
normal region extracting means for extracting image data of a normal region, which is a region different from the defective region, from the image data;
model generating means for generating a discriminative model for detecting defects in the internal structure of a tire to be inspected using the image data of the defective region and the image data of the normal region as teacher data; . image data acquisition means for acquiring image data of an X-ray image of a tire;
Defect area extracting means for extracting image data of a defect area, which is an area containing a defect in the internal structure of the tire, from the image data;
normal region extracting means for extracting image data of a normal region, which is a region different from the defective region, from the image data; and inspection using the image data of the defective region and the image data of the normal region as teacher data A program that causes a computer to function as model generation means for generating a discriminative model for detecting defects in the internal structure of a target tire.

Images