Spherical Appearance Inspection System and Object Appearance Inspection System

Abstract

To enable, in production engineering and visual inspection region fields, inspection of the appearance of a sphere having a spherical surface such as a ball bearing, steel ball, glass ball, and spherical products in the primary industry (orange, melon, and water melon).SOLUTION: An inspection system is obtained by integrating means that rotates a sphere (ceramic ball) at a uniform speed in front of a camera to pick up an image of the total spherical surface with the camera, and hardware and software that determine the presence or absence of defects in the surface such as scratches from the movement of the defects in the image, in inspection of the appearance of the total spherical surface of a sphere. In particular, the inspection system takes two images with a temporal delay as images before and after the movement, and accurately determines a change in shading of the same portion in the image using a neural network. This allows detection of even ultra-micro defects.SELECTED DRAWING: Figure 8

Description

Technical Field

[0001] The present invention relates to an appearance inspection system for spheres or rotatable objects using motion analysis by AI application. In particular, it is used in the fields of production technology, visual inspection, ball bearings, steel balls, glass balls, and identification of spherical products in the primary industry (such as oranges, melons, watermelons, mangoes, etc.).

Background Art

[0002] Conventionally, pachinko balls (steel balls for pachinko) have been marked as recognition tickets in order to indicate that they are pachinko balls used in the store itself. There is a known steel ball marking recognition system that recognizes the marking and discriminates and excludes pachinko balls marked by other stores from the pachinko balls marked by the store itself (see, for example, Patent Document 1).

[0003] On the other hand, defective products of ceramic balls (hereinafter abbreviated as ceramic balls) used for ball bearings used in precision mechanical devices and the like are picked out by human visual inspection. That is, bearings and the like are imaged in a specific conveyance direction with a camera or the like, and the presence or absence of defects is inspected by pattern matching or the like. However, at present, minute surface defects are inspected visually.

[0004] Incidentally, as shown in FIG. 1, there are four types of scratches that are currently confirmed to be extremely minute surface defects. FIG. 1(a) is a crack, which is like a scratch on the surface of a ceramic ball. FIG. 1(b) is a hole. FIG. 1(c) is a chip, which is like a depression on the ball surface. FIG. 1(d) is unevenness, which is uneven firing in the process of making a ceramic ball.

[0005] However, since the ceramic balls are very small with a diameter of about 5 mm, if the inspection takes a long time, the burden on the operator, the non-uniformity of the inspection, and the lack of reproducibility become problems. In addition, since the inspection object is a sphere, highlights are likely to occur.

Prior Art Documents

Patent Documents

[0006] [Patent Document 1] Japanese Patent No. 5374787 [Summary of the Invention] [Problems to be Solved by the Invention]

[0007] As described above, when imaging a specific transport direction with a camera or the like and inspecting the presence or absence of defects by pattern matching or the like, it is necessary to set a plurality of transport directions. In addition, precise image processing for spherical images cannot always be realized with high precision, and the determination accuracy is insufficient. Further, when performing the inspection visually, there are problems such as the fact that the highly accurate determination of the inspection operator cannot be maintained for a long time and the accuracy is not stable.

[0008] Generally, at present, it is difficult to detect such extremely minute surface defects even by carefully observing a still image. As described above, experts always roll a sphere during visual inspection to detect the movement of surface defects and determine whether there are surface defects.

[0009] Therefore, the inventor conceived of realizing this person's visual operation by motion analysis and realizing the determination by AI (neural network), which is a non-linear discrimination more flexible than linear discrimination, and improving its performance to high precision. Specifically, two images with a time delay are treated as still images before and after movement, and the density variation of the same part in the images is determined with high precision by a neural network, leading to the present invention for discriminating whether it is a defective product or not.

[0010] That is, for a curved surface image, it is difficult to perform sufficient determination processing by image processing, but by using a neural network that realizes motion analysis and non-linear discrimination in combination, it is possible to maintain the determination accuracy even for noise such as variation in irradiation amount and reflection.

[0011] The present invention aims to solve the above problems by devising an imaging method and performing inspections using a neural network, which is a type of AI (Artificial Intelligence).

Means for Solving the Problems

[0012] An appearance inspection system according to one aspect of the present invention is a sphere appearance inspection system for inspecting the presence or absence of defects on a sphere, and includes a rotating means for rotating the sphere at a specific position, an imaging means for imaging the surface of the sphere rotating at the specific position, a signal receiving means for receiving a signal from the imaging means, capturing a plurality of still images at intervals of time with respect to the surface of the sphere, and determining whether there are surface defects such as scratches on the surface of the sphere The determination is performed based on the shading variation of the plurality of still images. and a defect discrimination means, and is characterized by this. The presence or absence of the shading variation of the plurality of still images is determined by dividing each of the still images into a plurality of regions and generating a high-order discrimination curve of non-linear discrimination by a neural network for the shading variation of each region in the plurality of still images captured with a time delay.

[0013] Here, "capturing still images" includes a case where the imaging means first captures an image as a moving image and then selects and captures a plurality of still images at intervals of time from the moving image, and a case where a plurality of still images are sequentially captured and captured at intervals of time. In addition, the "sphere" of the present invention includes not only a clean sphere such as a pachinko ball but also those having a substantially spherical shape such as a watermelon or a tomato.

[0014] In this way, for the surface of the sphere, a plurality of still images are captured at intervals of time while rotating the sphere, and it is determined whether there are surface defects based on those still images, so that it is possible to determine whether a product is defective without visual inspection by an expert.

[0015] Also, a sphere There are various colors of products in the form of spheres, and there are also various shapes and colors of defects such as scratches. By determining based on the fluctuations in the shading of a plurality of images, it is possible to accurately determine the presence or absence of defects in products of any color and the shape of defective parts with high accuracy.

[0016] Also, a product It is possible to more accurately determine the presence or absence of defects.​

[0017] In addition, in this appearance inspection system, the rotating means has an upper plate that houses the sphere with a clearance, and a lower plate that is disposed below the upper plate and is movable relative to the upper plate in at least a first direction. By moving the lower plate in the first direction, the Inside the hole sphere housed therein is rotated in the first direction. In this way, the sphere can rotate at least in the first direction, and it is possible to image all surfaces except for at least the upper and lower vertices. Inside the hole In this way, the sphere can rotate at least in the first direction, and it is possible to image all surfaces except for at least the upper and lower vertices. Inside the hole In this way, by rotating at least in the first direction, it is possible to image all surfaces except for at least the upper and lower vertices.

[0018] In addition, in this appearance inspection system, the rotating means has an upper plate that houses the sphere with a clearance, and a lower plate that is disposed below the upper plate and is movable relative to the upper plate in at least a first direction and a second direction orthogonal to the first direction. By moving the lower plate in the first or second direction, the Inside the hole sphere housed therein is rotated in the first or second direction. It is desirable that this is the case. Inside the hole In this way, by rotating the sphere at least in the first direction and a second direction orthogonal to the first direction, it is possible to image the entire surface of the sphere.

[0019] In this way, by rotating the sphere at least in the first direction and a second direction orthogonal to the first direction, it is possible to image the entire surface of the sphere. Inside the hole In this way, by rotating at least in the first direction and a second direction orthogonal to the first direction, it is possible to image the entire surface of the sphere. It is possible.

[0020] In addition, in this appearance inspection system, the rotating means has a lower plate that is movable relative to the upper plate in a third direction different from the first direction and the second direction. By moving the lower plate in the third direction, the Inside the hole sphere housed therein is rotated in the third direction. It is more desirable that this is the case. In this way, it becomes possible to rotate the sphere in the first, second, and third directions, and it is possible to image the entire surface of the sphere from various directions, so the discrimination accuracy of defective portions is improved.

[0021] In addition, in this appearance inspection system, the defect discrimination means has a magnifying means for magnifying the still image, and theHole image It has correction means for making the image center of the Hole image image coincide with the image center of the still image of the sphere and correcting minute displacements of the image center due to enlargement. By doing so, since the image centers of the still images of the sphere coincide, the discrimination accuracy of defective portions is improved.

[0022] Further, in this appearance inspection system, the defect discrimination means assumes the sphere as a cube, captures a plurality of still images for each face of the cube, divides the still images into a plurality of regions for each face, and determines the density variation for each region by forming a high-order discrimination curve of non-linear discrimination with a neural network and performing the determination a plurality of times. If it is determined that there is a surface defect even once among the multiple determinations for each face, it is discriminated that the cube has a surface defect. By doing so, the discrimination accuracy of defective portions is improved.

[0023] The above appearance inspection system is an appearance inspection system for a sphere, but the present invention can also discriminate surface defects of an object by a system having a similar configuration for a rotatable object other than a sphere. Rotatable objects include, in addition to a sphere, for example, rotating bodies such as an ellipsoid and a cone, a rugby ball, and substantially ellipsoidal objects such as a cucumber, an eggplant, and a mango, and a substantially spherical polyhedron formed by assembling a large number of polygons such as a soccer ball. That is, if an image of each surface can be captured by rotating such a rotatable object, the appearance inspection system of the present invention can be applied.

[0024] An appearance inspection system according to one aspect of the present invention is an appearance inspection system for inspecting the presence or absence of defects in a rotatable object, and includes rotating means for rotating the object at a specific position, imaging means for imaging the surface of the object rotating at the specific position, receiving a signal from the imaging means, capturing a plurality of still images of the surface of the object at time intervals, and determining whether there are surface defects such as scratches on the surface of the object The determination is performed based on the shading variation of the plurality of still images. and defect discrimination means. The presence or absence of the shading variation of the plurality of still images is determined by dividing each of the still images into a plurality of regions and generating a high-order discrimination curve of non-linear discrimination by a neural network for the shading variation of each region in the plurality of still images captured with a time delay. It is characterized by this.

[0027] Further, in this appearance inspection system, the rotating means holds a clearance for the objectInside the hole an upper plate for accommodating the Inside the hole , and a lower plate disposed below the upper plate and movable relative to the upper plate at least in a first direction. By moving the lower plate in the first direction, the Inside the hole object accommodated in the Inside the hole is rotated in the first direction. For example, in the case of the above-described ellipsoid or substantially ellipsoid, it is possible to image the entire surface of these objects only by rotating in the first direction.

[0028] In addition, in this appearance inspection system, the rotating means accommodates the object with a clearance Inside the hole an upper plate for accommodating the Inside the hole , and a lower plate disposed below the upper plate and movable relative to the upper plate at least in a first direction and a second direction orthogonal thereto. By moving the lower plate in the first or second direction, the Inside the hole object accommodated in the Inside the hole is rotated in the first or second direction. For example, in the case of the above-described ellipsoid or substantially ellipsoid, it is possible to accommodate the object in the Inside the hole with a predetermined clearance. Inside the hole

[0029] In addition, in this appearance inspection system, the rotating means is a lower plate in which the lower plate is movable relative to the upper plate in a third direction different from the first direction and the second direction. By moving the lower plate in the third direction, the Inside the hole object accommodated in the Inside the hole is rotated in the third direction.

[0030] In addition, in this appearance inspection system, the defect determination means includes a magnifying means for magnifying the still image, and Hole image a correction means for aligning the image center of the Hole image of the upper plate with the image center of the still image of the object and correcting a minute deviation of the image center due to magnification.

[0031] In addition, in this appearance inspection system, the defect determination means assumes the object as a rectangular parallelepiped, captures a plurality of still images for each surface of the rectangular parallelepiped, divides the still images into a plurality of regions for each surface, and determines the shading variation for each region multiple times by forming a high-order discrimination curve of non-linear discrimination with a neural network. If it is determined that there is a surface defect even once among the multiple determinations for each surface, the rectangular parallelepiped is discriminated as having a surface defect.

Advantages of the Invention

[0032] The present invention enables the appearance inspection of the surfaces of rotatable objects or spheres having curved surfaces, such as ball bearings, steel balls, glass balls, and spherical products in the primary industry (oranges, melons, watermelons), in the production technology and visual inspection fields.

Brief Description of the Drawings

[0033]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Embodiments for Carrying Out the Invention

[0034] Hereinafter, an embodiment according to the present invention will be described in detail with reference to the drawings. Note that the present invention is not limited by this embodiment. For example, in the following embodiment, an example in which the object to be inspected for appearance is a spherical ball bearing will be described. However, the present invention is not limited to a sphere such as a ball bearing, and can also be applied to an object having a substantially elliptical shape such as a rugby ball, a cucumber, an eggplant, or a mango.

[0035] The appearance inspection system of this embodiment is a spherical appearance inspection system premised on inspecting the entire spherical surface of a sphere. For example, it can highly accurately detect extremely minute surface defects of a ceramic ball bearing with a diameter of about 5 mm.

[0036] In order to inspect the entire spherical surface of the sphere without omission, as shown in FIG. 2, the ceramic ball 1 (sphere) to be inspected is assumed to be a six-sided virtual cube 2 and rotated to capture images of the entire spherical surface. That is, it is realized by rotating (rotating on its own axis) the ceramic ball 1 assuming the spherical surface as a cube and surely imaging each of the six surfaces.

[0037] Then, directly below the camera, the ceramic ball 1 to be rotated is arranged toward the camera central axis, and by rotating (rotating on its own axis), all six surfaces can be imaged within a certain period of time. In this case, as shown in FIG. 3, by rotating the ceramic ball in the horizontal direction and then rotating it in the vertical direction orthogonal to the horizontal direction, all six surfaces can be imaged within a certain period of time.

[0038] That is, while continuously imaging the entire spherical surface while rotating the ceramic ball 1 vertically and horizontally, static images corresponding to each surface of the cube are captured at intervals of time from the captured video, and they are compared for each surface to detect extremely minute surface defects (for example, cracks, chips, dirt, spots, etc.). Since it rotates at a uniform speed and is assumed to be a six-sided virtual cube 2, static images corresponding to each surface of the cube can be easily captured, and the entire spherical surface can be imaged without omission.

[0039] As an example of an imaging means, the imaging device 11 includes a ball storage unit 12 and first and second buttons 13, 14. The ball storage unit 12 can accommodate approximately 3000 ceramic balls. Also, by pressing the first button 13, the gear 15 inside the ball storage unit 12 rotates, and one ceramic ball 1 is fed into the internal rotating unit 16 one by one.

[0040] As shown in FIG. 5, the rotating unit 16 has an upper plate 21 that accommodates the ceramic ball 1 in a square hole 21a with a clearance, and a lower rubber plate 22 that is disposed below the upper plate 21 and moves in a first direction S1 and a second direction S2 orthogonal to the first direction S1 with respect to the upper plate 21.

[0041] By moving the lower rubber plate 22 in the first or second direction S1, S2, the ceramic ball 1 rotatably accommodated in the square hole 21a is rotated at a constant speed in the first or second direction S1, S2. The ceramic ball 1 in the state of being accommodated in the square hole 21a is arranged directly below the camera for imaging and toward the camera central axis.

[0042] By long-pressing the second button 14, the ceramic ball 1 held in the square hole 21a of the upper plate 21 inside moves vertically and horizontally as the lower rubber plate 22 in contact with the ceramic ball 1 from below moves, and the ceramic ball 1 rotates. Finally, it falls through a through hole (not shown) provided in the lower rubber plate 22 and is sorted into defective or non-defective products.

[0043] And while one ceramic ball is rotating, imaging is performed a plurality of times, for example, 3 times, a total of 18 images, for one surface so that the entire spherical surface can be photographed. Appearance inspection is performed in pairs of 2 images, and determination is made, for example, 2 times, a total of 12 times, for each surface. This number of determinations can be changed as appropriate. Regarding the determination, optical flow is adopted in which two still images are captured from the video and the presence or absence of surface defects is determined based on the density change (see FIG. 6).

[0044] The two still images are set as still image 1 (before movement) and still image 2 (after movement), and the in-block addition average values (grayscale images) obtained by block-dividing them are created for still image 1 (before movement) and still image 2 (after movement) respectively, and these are used as input values to the neural network. Here, instead of determining the grayscale variation by a threshold in motion analysis, a high-order discrimination curve of non-linear discrimination is formed by the neural network for strict and flexible discrimination.

[0045] Regarding illumination, mini power dome illumination is used, but it can also be coaxial epi-illumination where the minimum RGB can be independently adjusted for irradiation. As the imaging means, a digital microscope (with a magnification of 1000 times) is used. In order to avoid the deviation of the image center due to magnification affecting the determination, the image center of the square hole image on the upper side plate 21 is made to coincide with the image centers of the two still images, and it is corrected to correct the minute deviation of the image center due to magnification (deviation correction means).

[0046] Thus, this system is based on motion analysis and uses two still images before and after movement to determine whether there is a variation in grayscale by a neural network. This method is equivalent to the conventional inspection that discriminates the presence or absence of surface defects (for example, scratches) by taking advantage of the fact that the position of the surface defect changes while the inspection operator visually moves the ceramic ball by hand.

[0047] Fig. 7 shows an example of the main menu screen of the spherical appearance inspection system. As described above, two still images are extracted from the video during the rotation of the ceramic ball for each surface, block-divided, and the grayscale variation of the addition average value of the in-block pixels is input to be determined by the neural network. That is, the bearing is assumed to be a cube, and the grayscale variation of the still images before and after movement due to rotation on each surface is used. In this example, the block division of each surface is 38 divisions, but it is not limited to this.

[0048] When there is no surface defect (scratch) on the ceramic ball, as shown in Fig. 8(a), there is no shading variation even when it is moved, so the input value vector of the neural network is flat. However, when there is a surface defect (scratch), as shown in Fig. 8(b), there is a shading variation when it is moved, so the input value vector of the neural network is not flat and changes occur.

[0049] Therefore, if the number of determinations is, for example, twice for the front and back sides with respect to the six sides, the total number of determinations is 6×2 = 12 times. If even one determination results in a determination of a defect by the neural network, it means that there is a defective part, that is, a surface defect, in the ceramic ball. Here, although the neural network is activated 12 times for the determination, the same learning result is used for all of them. The reason for making two determinations for each side is to probabilistically improve the accuracy of the determination, but the number of determinations can be changed according to the accuracy of the determination.

[0050] As described above, the preferred embodiments of the present invention have been described with reference to the drawings. However, various additions, changes, or deletions are possible without departing from the spirit of the present invention. For example, although mini power dome lighting is adopted, if halation occurs near the center of the spherical image due to excessive illuminance, it becomes difficult to detect surface defects. In such a case, the irradiation amount can be reduced with a dimming filter or the like.

[0051] Also, in the above embodiment, the lower rubber plate moves in the first direction and the second direction orthogonal to each other with respect to the upper plate, and the spherical body rotates in the directions of the first and second directions, respectively. However, it can also be configured to rotate only in the first direction, or can also be configured to rotate in an oblique direction (third direction) with respect to the first and second directions.

[0052] Also, the holes provided in the upper plate are not limited to square holes, and may be rectangular holes such as rectangular holes or diamond-shaped holes, as long as they have a shape with corners. Also, the corners may be acute angles or rounded. The presence of two sides forming the corners can suppress accidental displacement when the spherical body rotates.

[0053] In addition, in the above-described embodiment, the object to be inspected is a sphere, but the present invention is not limited thereto. For example, it is also possible to target a rotatable object such as an ellipsoid or a substantially ellipsoidal object. For example, in the case of an ellipsoid or a substantially ellipsoidal object, the lower rubber plate may be movable only in one of the first or second directions, and the object may be configured to be rotatable in one direction. If the object has an ellipsoidal or substantially ellipsoidal shape, it is possible to image the entire surface by rotating it in only one rotatable direction.

Description of Reference Numerals

[0054] 1 Ceramic ball 2 Virtual cube 11 Imaging device 12 Ball storage section 13, 14 Buttons 15 Gear 16 Rotating section 21 Upper plate 22 Lower rubber plate (lower plate)

Claims

1. A spherical appearance inspection system for inspecting the presence or absence of defects in a sphere, comprising: rotating means for rotating the sphere at a specific position; imaging means for imaging the surface of the sphere rotating at the specific position; defect discrimination means for receiving a signal from the imaging means, capturing a plurality of still images of the surface of the sphere at time intervals, and determining whether there are surface defects such as scratches on the surface of the sphere based on the density variations of the plurality of still images; The spherical appearance inspection system is characterized in that the presence or absence of density variations in the plurality of still images is determined by dividing each of the still images into a plurality of regions and generating a high-order discrimination curve for non-linear discrimination by a neural network for discriminating the density variations for each region in the plurality of still images captured with a time delay.

2. The rotating means includes: an upper plate for accommodating the sphere in a hole with a clearance, and a lower plate disposed below the upper plate and movable relative to the upper plate in at least a first direction; By moving the lower plate in the first direction, the sphere accommodated in the hole is rotated in the first direction. The spherical appearance inspection system according to Claim 1.

3. The rotating means includes: an upper plate for accommodating the sphere in a hole with a clearance, and a lower plate disposed below the upper plate and movable relative to the upper plate in at least a first direction and a second direction orthogonal thereto; By moving the lower plate in the first or second direction, the sphere accommodated in the hole is rotated in the first or second direction. The spherical appearance inspection system according to Claim 1.

4. The rotating means includes: The lower plate is a lower plate movable relative to the upper plate in a third direction different from the first and second directions; By moving the lower plate in the third direction, the sphere accommodated in the hole is rotated in the third direction. The spherical appearance inspection system according to Claim 3.

5. The defect discrimination means includes: enlarging means for enlarging the still image; correction means for aligning the image center of the hole image of the upper plate with the image center of the still image of the sphere and correcting minute displacements of the image center due to enlargement. The spherical appearance inspection system according to Claim 2.

6. The defect discrimination means includes: Assume the sphere as a cube, capture a plurality of still images for each face of the cube, divide the still images into a plurality of regions for each face, and determine the shading variation for each region multiple times by forming a high-order discrimination curve for non-linear discrimination with a neural network. If it is determined that there is a surface defect even once among the multiple determinations for each face, the cube is discriminated as having a surface defect. The sphere appearance inspection system according to claim 1.

7. An object appearance inspection system for inspecting the presence or absence of defects in a rotatable object, comprising: Rotation means for rotating the object at a specific position; Imaging means for imaging the surface of the object rotating at the specific position; Defect discrimination means that receives a signal from the imaging means, captures a plurality of still images at intervals of time for the surface of the object, and determines whether there is a surface defect such as a scratch on the surface of the object based on the shading variation of the plurality of still images. The presence or absence of the shading variation of the plurality of still images is determined by dividing each of the still images into a plurality of regions and generating a high-order discrimination curve for non-linear discrimination with a neural network for the shading variation of each region in the plurality of still images captured with a time delay. The object appearance inspection system is characterized by this.

8. The rotation means Has an upper plate that houses the object in a hole with a clearance, and a lower plate that is disposed below the upper plate and is movable relative to the upper plate in at least a first direction. By moving the lower plate in the first direction, the object housed in the hole is rotated in the first direction. The object appearance inspection system according to claim 7.

9. The rotation means Has an upper plate that houses the object in a hole with a clearance, and a lower plate that is disposed below the upper plate and is movable relative to the upper plate in at least a first direction and a second direction orthogonal to this. By moving the lower plate in the first or second direction, the object housed in the hole is rotated in the first or second direction. The object appearance inspection system according to claim 7.

10. The rotation means The lower plate is a lower plate that is movable relative to the upper plate in a third direction different from the first direction and the second direction. By moving the lower plate in the third direction, the object housed in the hole is rotated in the third direction. The object appearance inspection system according to claim 9.

11. The defect determination means is an enlargement means for enlarging the still image, and a correction means for making the image center of the hole image of the upper side plate coincide with the image center of the still image of the object and correcting a minute deviation of the image center due to enlargement. The appearance inspection system for an object according to claim 9.

12. The defect determination means is assuming the object as a rectangular parallelepiped, capturing a plurality of still images for each surface of the rectangular parallelepiped, dividing the still image into a plurality of regions for each surface, and determining the density variation for each region a plurality of times by forming a high-order discrimination curve of non-linear discrimination with a neural network. If it is determined that there is a surface defect even once among the plurality of determinations for each surface, the rectangular parallelepiped is determined to have a surface defect. The appearance inspection system for an object according to claim 7.

Images