JP5374787B2 - Steel ball marking recognition system - Google Patents

Description

As is well known, pachinko steel balls (hereinafter referred to as pachinko balls) are stamped as identification cards to indicate that they are used in their own store (own or in-house). . The present invention recognizes stamps that are mainly rolled into pachinko balls, and distinguishes pachinko balls stamped by other stores (stores of other people or other companies) from pachinko balls stamped by their own stores. It is related with the marking recognition system of steel balls , such as a pachinko ball.

  Until now, pachinko balls have been stamped by rolling as an identification card for their own store. This was done to prevent users from bringing pachinko balls from other stores to play, and to prevent unauthorized use of cash value in 1-yen and 4-yen pachinko balls. It is. However, at present, the game machines related to pachinko have only the function of counting the number of pachinko balls, and the imprint is merely a meaning as a mental warning to the user.

  Under such circumstances, the use of pachinko balls from other stores and the illegal use of the cash value of pachinko balls have become life-threatening problems at pachinko machine operators. In addition, manufacturers of this type of pachinko game machine have a strong need to recognize the markings on the pachinko ball on the device side, and to add or mount warning and play stop functions to existing game machines.

  By the way, the pachinko ball is engraved on a spherical surface, so when this engraving is taken with a camera, the engraving is taken from a slightly inclined direction, and the engraved letter or figure is curved along the spherical surface. Become. A method of extracting and correcting such a curved character string or graphic string has been proposed (for example, see Non-Patent Document 1).

  In this method, extraction is performed using general characteristics of a character string. First, in a series of character strings where the brightness difference between the character and its background is high, the character color is the same, and the local binarization is performed on an edge basis. Character strings are extracted by concatenating these regions based on the characteristics of “equal”, “short character spacing”, and “local linearity.” Next, approximate the curvature of the string with a second-order polynomial. The character string area is obtained, and the curvature is corrected linearly by using the last obtained curvature model, so that even a character string curved by oblique photographing can be extracted, and the correction is also a good result. The extraction accuracy was 60.5% for the average precision and 62.7% for the average recall.

Articles from Nagoya University http://ci.nii.ac.jp/naid/110004662969/ (Extraction and correction of character strings written on curved surfaces) (Theme session (1), character recognition and document understanding)

  As described above, this research and development issue is the recognition of the image on the steel ball where the start and end points of the image cannot be limited, and there is specular reflection because it is an imprint on the pachinko ball. This is an advanced recognition technology that can only be recognized as an image.

  On the other hand, the character recognition method described in Non-Patent Document 1 recognizes paint characters and does not recognize stamps. That is, in the case of the marking, since the image is reflected and lifted as shown in FIGS. 4 and 5, it is necessary to irradiate light so that the marking is shaded. Even using the method, it is difficult to identify the pachinko sphere imprint. Further, in this method, in the case of paint characters, the extraction accuracy is 60.5% in average precision and 62.7% in average reproduction rate, so the accuracy of 70% to 80% or more is required. This is not applicable to the pachinko ball marking recognition system.

  The present invention has been made in view of the above points. Mainly, a pachinko ball of the own store and a pachinko ball of another store can be identified with a probability of 70% to 80% or more. It is intended to provide a pachinko ball marking recognition system that can be installed in equipment for store main halls (aggregators, general management devices).

To achieve the above object, a pachinko ball marking recognition system according to claim 1 of the present invention is provided.
(1) means for moving the pachinko ball in the passage of the pachinko machine center device while changing the direction of the pachinko ball, and imaging the pachinko ball with a camera from a plurality of different directions;
(2) means for cutting out and labeling the color stamped image of the pachinko sphere;
(3) The labeled color image of the pachinko sphere is converted into a gray scale, and the two-dimensional Fourier transform (2DFFT) is applied to the rotation-invariant information (the power spectrum of the two-dimensional Fourier transform (root of coefficient square)). RBF (radial based function, for example, Gaussian function) instead of the normal (conventional) sigmoid function as the output function of each cell of the neural network (NN), and the two-dimensional Fourier power as the rotation invariant information Means for inputting a spectrum into an input layer of a three-layer neural network (NN);
(4) The pachinko ball moving in the passage is photographed with a camera a plurality of times, the color stamped image is converted into a gray image, and the two-dimensional Fourier power spectrum is recognized NN a plurality of times (for example, 10 times). Based on the majority rule, more than half of them (for example, 6 times or more) is provided with means for eliminating pachinko balls only when it is determined by NN recognition that the pachinko balls of other stores are stamped. To do.

  Here, by using the RBF function as the output function, the response value of the output cell of the NN appears on the surface of the Gaussian function as an example of the rotational histogram during the learning of the NN, and the output function is configured with the conventional sigmoid function. In addition to the formation of the curved curved identification surface as in the case of the case, the value in the height direction is output three-dimensionally. For the input value to the NN near the center of the store pattern, A high reaction value is exhibited. On the other hand, a low reaction value is shown for an input value to the NN far from the center of the own store pattern. Thereby, a certain kind of pseudo absolute recognition is realized. The color stamped image of the pachinko sphere means a stamped image including the image of the pachinko sphere. Further, the gray scale of the color image of the pachinko sphere may be performed before labeling.

  According to the pachinko ball marking recognition system of the present invention having the above-described configuration, the pachinko ball moving in the passage is photographed by a camera from a plurality of directions to form a color stamp image, and the pachinko ball is oriented. Since the image is taken a plurality of times while changing, the marking on the pachinko ball is surely taken as an image. Further, the captured color stamp image is divided into RGB and equally averaged to be converted into a gray scale, and the gray image is converted into a Fourier power spectrum (root of coefficient square) by two-dimensional Fourier transform. This Fourier power spectrum is rotation invariant information, and this information is input to the three-layer NN.

  On the other hand, as described above, the NN uses an RBF (for example, a Gaussian function) such as a rotational histogram as a component of the NN, and outputs a value in the three-dimensional height direction. When the input to the NN is a two-dimensional Fourier power spectrum of the own store imprinted image, it reacts only with the cell indicating the own store imprinted image, and when the input to the NN is the two-dimensional Fourier power spectrum of the other store imprinted image, the other store Reacts only to cells showing engraved images. Then, RBF (for example, Gaussian function) is used as an output function when a signal is input to the next cell by adding a weight to the cell in the NN. By using a Gaussian function as an example of the rotation type histogram as an output function, the response value of the output cell of the NN appears on the surface of the Gaussian function as a rotation type histogram during the learning process of the NN, and the conventional sigmoid function In addition to forming a curved plane identification curved surface as in the case of configuring an output function, a value in a three-dimensional height direction is output. That is, a high reaction value (close to 1) is shown for an input value to the NN near the center of the own shop pattern. On the other hand, a low reaction value (close to 0) is shown for an input value to the NN far from the center of the own store pattern. As a result, a certain kind of pseudo absolute recognition is realized, and even when the engraved image is not only a complete image but also a partial image, the pachinko ball of the own store and the pachinko ball of another store can be distinguished. . As a result, it is possible to recognize and determine the stamps of the own store and other stores with high accuracy.

  In this way, as shown in FIG. 6, the own store stamp image and the other store stamp image are distinguished, and at the same time, the complete image and the partial image are also distinguished. For example, when the two-dimensional Fourier power spectrum of the self-engraved complete image of the own store is input, the response value of the output cell that reacts only to the NN own store is close to 1, but the two-dimensional Fourier power of the other store engraved image When the spectrum is input, the response value of the output cell that reacts only to other stores of the NN is close to zero. On the contrary, when the 2D Fourier power spectrum of the imprinted complete image of another store is input, the response value of the output cell that reacts only to the other store of NN is close to 1, but the 2D Fourier power of the own store imprinted image When the spectrum is input, the response value of the output cell that reacts only to the own store of the NN is close to zero. The pachinko balls moving through the passage in the center hall of the pachinko gaming machine are picked up by a plurality of cameras from different directions while changing the direction, and the color stamped images of the respective pachinko balls are displayed. A two-dimensional Fourier power spectrum converted into a gray image and finally Fourier-transformed is input to the NN input layer a plurality of times (for example, 10 times) with the same pachinko sphere having different orientations. As a result, the response value of the output cell of the NN to the predetermined stamped image of the own store is compared with the response value of the output cell of the NN to the stamped image of the other store, and the pachinko ball of the own store or the pachinko ball of the other store is compared. Is recognized by NN. Thus, when the NN recognizes more than half (for example, six times or more) based on the principle of majority with other store pachinko balls, the pachinko balls are excluded from the passage.

To achieve the above object, a pachinko ball marking recognition system according to claim 2 of the present invention is provided.
(1) A pachinko ball is introduced into the passage of each pachinko machine while changing its orientation, and a predetermined number of pachinko balls are lined up and moved in series, and a plurality of the pachinko balls are grouped by a plurality of cameras from a plurality of different directions. Means for imaging once;
(2) means for cutting out and labeling each color stamped image of the pachinko ball group;
(3) Two-dimensional Fourier power spectrum (the root of the square of the coefficient) as the information that is invariant to rotation by converting each color stamped image of the pachinko sphere group thus labeled to gray scale, and performing two-dimensional Fourier transform (2DFFT) RBF (radial based function, for example, Gaussian function) is used for the output function of each cell of the neural network (NN) instead of the conventional sigmoid function, and the two-dimensional Fourier power spectrum is converted into a three-layer neural network ( NN) input layer,
(4 ′) A plurality of (for example, 10) pachinko ball groups that flow into the passage are photographed multiple times with a camera, and the two-dimensional Fourier is obtained by converting the color stamped images into gray. The power spectrum is input to the NN multiple times, and more than half (for example, 6 or more) of the pachinko spheres of other stores are more than half of the input color stamped images of the pachinko ball group based on the majority rule. And a means for excluding all (for example, 10) pachinko balls only when it is determined by NN recognition.

  Here, by using an RBF (for example, a Gaussian function) as an output function in the NN, the response value of the output cell of the NN appears on the surface of a Gaussian function as an example of a rotational histogram during the learning of the NN. It is not just the formation of a curved plane identification curved surface as in the case where the output function is configured with a conventional sigmoid function, but the value in the height direction is output three-dimensionally and is close to the center of the own store pattern It shows a high response value (close to 1) for the input value to NN. On the other hand, a low reaction value (close to 0) is shown for an input value to the NN far from the center of the own store pattern. Thereby, a certain kind of pseudo absolute recognition is realized. In addition, each color image of the pachinko sphere group can be gray scaled, cut out, and labeled. Further, the color stamped image of the pachinko sphere means a stamped image including the image of the pachinko sphere.

  According to the pachinko ball marking recognition system according to claim 2 of the present invention having the above-described configuration, the moving pachinko ball group (for example, ten pachinko balls arranged in a line) in the passage from a plurality of directions. Color stamped images (including the entire image of the pachinko sphere, in other words, the pachinko sphere and the stamped image above it) are captured by multiple cameras and the pachinko sphere group is captured multiple times from different directions. Is done. Also, each color stamped image taken is divided into RGB and averaged equally and grayscaled to become a grayscale image. This grayscale image information is converted into a Fourier power spectrum (root of the square of the coefficient) by two-dimensional Fourier transform. This Fourier power spectrum is information that is invariant to rotation, and this information is a three-layer NN power spectrum. Is input. On the other hand, as described above, the NN uses an RBF (for example, a Gaussian function) such as a rotational histogram as a component of the NN, and outputs a value in the three-dimensional height direction. When the input to the NN is a two-dimensional Fourier power spectrum of the own store imprinted image, it reacts only with the cell indicating the own store imprinted image, and when the input to the NN is the two-dimensional Fourier power spectrum of the other store imprinted image, the other store Reacts only to cells showing engraved images. Then, an RBF (for example, a Gaussian function) is used as an output function when a cell is weighted and added in the NN and a signal is input to the next cell, and a Gaussian function as this rotational histogram is used as an output function. Thus, the response value of the output cell of the NN appears on the surface of the Gaussian function such as a rotation type histogram during the learning of the NN, and is simply identified in a curved plane as in the case where the output function is configured with a conventional sigmoid function. The value in the height direction is output three-dimensionally, not limited to the formation of the curved surface. That is, a high reaction value (close to 1) is shown for an input value to the NN near the center of the own shop pattern. On the other hand, a low reaction value (close to 0) is shown for an input value to the NN far from the center of the own store pattern. As a result, some kind of pseudo absolute recognition is realized, and even when the engraved image is not only a complete image but also a partial image, the pachinko ball of the own store and the pachinko ball of another store can be distinguished. It becomes like this. As a result, it is possible to recognize and determine the stamps of the own store and other stores with high accuracy.

  In this way, as shown in FIG. 6, the own store stamp image and the other store stamp image are distinguished, and at the same time, the complete image and the partial image are also distinguished. For example, when the two-dimensional Fourier power spectrum of the self-engraved complete image of the own store is input, the reaction value of the cell indicating the self-stored image of the NN is close to 1, but the two-dimensional Fourier power spectrum of the other store engraved image is obtained. The cell response value indicating the other store image of the NN when input is close to zero. Conversely, when the 2D Fourier power spectrum of the imprinted complete image of another store is input, the response value of the cell indicating the other store image of NN is close to 1, but the 2D Fourier power spectrum of the own store imprinted image is The response value of the cell indicating the NN's own store image when input is close to zero.

  The pachinko balls moving in the path of each pachinko gaming machine are picked up by a plurality of cameras from different directions while changing the direction, and each color stamp image of the pachinko balls is Fourier transformed. The two-dimensional power spectrum is input to the NN input layer a plurality of times (for example, five times) by the input means corresponding to the stamped images having different directions of a plurality of (for example, ten) pachinko balls. On the other hand, as described above, the NN uses the RBF function (Gaussian function) as a component of the NN, and outputs a value in the height direction three-dimensionally. When the input to the NN is a two-dimensional Fourier power spectrum of the own store imprinted image, it reacts only with the cell indicating the own store imprinted image, and when the input to the NN is the two-dimensional Fourier power spectrum of the other store imprinted image, the other store Reacts only to cells showing engraved images. In addition, the response value of the NN output cell to a predetermined stamped image of the own store is compared with the response value of the NN output cell to the stamped image of another store, and the pachinko ball of the own store or the pachinko ball of another store is compared. Is recognized by NN. Thereby, it is possible to recognize and determine the stamps of the own store and other stores with high accuracy.

  To repeat, in the NN recognition in the pachinko sphere marking recognition system of the present invention, the color stamped image is grayscaled, and these are two-dimensional Fourier transformed (2DFFT) to perform two-dimensional Fourier transformation as rotation invariant information. Rotate the output function of each cell instead of the normal sigmoid function so that the power spectrum (the root of the square of the coefficient) is input to the NN and the partial image missing the part of the pachinko sphere engraved image can be recognized. RBF (radial based function, for example, Gaussian function) like a type histogram is used. In addition, if the stamped image is a complete image, it is possible to accurately compare the stamp of the own store or the stamp of the other store, so it is desirable to change the orientation of the pachinko ball stamp so that a complete image of the stamp can be captured. It should be taken repeatedly with the camera. Note that those engraved images are image information in a low frequency region.

  Therefore, in the present invention, as described in claim 4 to be described later, the floor portion and the guide of the passage (pachinko hall equipment or pachinko game machine) through which the pachinko ball moves are formed in a spiral shape, for example, thereby forming a pachinko. The sphere moves while changing the direction of the stamp. In addition, a plurality of cameras are installed so that a plurality of images can be taken from a plurality of directions. Thereby, there is a high possibility that the engraved image on the pachinko ball is captured as a complete image or an image close to the complete image.

In addition, the color stamped image is converted to gray scale and subjected to two-dimensional Fourier transform, and a two-dimensional Fourier spectrum as rotation invariant information is input to NN. On the other hand, in the NN, RBF (radial based function, for example, Gaussian function) such as a rotation histogram is used for the output function of each cell instead of the conventional sigmoid function, and specified as a bell-shaped rotation histogram (for example, Gaussian function). Therefore, as shown in FIG. 6, the self-stored stamp image and the other store stamped image are distinguished, and at the same time the complete image and the partial image are also distinguished. For example, when the value σ of the own store stamp image is 1 or between 1 and 0, the other store stamp image is 0, and conversely, the value σ of the other store stamp image is between 1 or 1 and 0 The self-engraved image is 0. These numerical values are registered in advance in the NN so as to correspond to the stamped image of the other store. Further, the teaching method is different from the conventional method and includes the following modes.
Case 1.
Full image of own store; teacher value T = 1.0
Teacher value to other stores T = 0.0
Case 2
Partial image of own store; teacher value T = 0.7
Teacher value to other stores T = 0.0
Case 3.
Partial image of own store; teacher value T = 0.3
Teacher value to other stores T = 0.0
Case A.
Complete image to another store; teacher value T = 1.0
Teacher value to own store T = 0.0
Case B.
Partial image to another store; teacher value T = 0.7
Teacher value to own store T = 0.0
Case C.
Partial image to another store; teacher value T = 0.3
Teacher value to own store T = 0.0
The pachinko balls moving through the passages of each pachinko machine change the direction, and a plurality of (for example, 10) pachinko ball groups are imaged multiple times (for example, 10 times) from a plurality of different directions. A two-dimensional Fourier power spectrum obtained by converting individual engraved images with different sphere group orientations into gray scales and two-dimensional Fourier transform is input to the input layer a plurality of times (ten times) by the input means. Then, the response value of the output cell when the two-dimensional Fourier power spectrum of the self-engraved image inputted in advance is inputted to NN and the two-dimensional Fourier power spectrum of the imprinted image of another store are inputted to NN. The response value of the output cell is compared and contrasted, and recognition is performed based on whether or not the pachinko sphere in the store is judged to be more than 50% (for example, 6 or more) based on the majority rule. Thus, for example, when it is recognized that the number of pachinko balls in its own store is more than 50% based on the majority rule (for example, 6 or more), it is not excluded from the passage, but the number of pachinko balls in other stores is based on the majority rule. When it is recognized that the number is more than 50% (for example, 6 or more), all (10) pachinko balls are excluded from the passage.

  As described above, the stamped image recognition system of the present invention realizes recognition of a stamped image of a pachinko ball in its own shop using the associative memory capability of the neural network (NN). Here, the NN associative memory means that in this example, a complete image of a pachinko ball engraving is learned, and a part of the image (partial image) is input to the NN with respect to the evaluation image after learning is completed. , To restore a complete image. In the present invention, as described above, the teacher value: t = 1.0 for the complete image of the pachinko ball engraved at the store, and the degree of omission of the image for the partial image of the pachinko ball engraved at the store. For example, teacher learning is performed according to how the image is viewed in an analog manner with a teacher value: t = 0.7 or t = 0.3. On the other hand, by giving an analog teacher value also to a typical pachinko ball stamp image of another store, it is finer than the conventional NN learning by 1 and 0 teacher values, (Including partial images) can be recognized.

  In addition, in NN learning, as shown in FIG. 6, by using an RBF (radial based function) such as a rotation histogram as an output function, the response value of an output cell of NN is as if it is a rotation type during NN learning. Appears on the surface of a Gaussian function as a histogram, and outputs a value in the height direction three-dimensionally, not just the formation of a curved plane identification curved surface as in the case where the output function is configured with a conventional sigmoid function. In other words, a high response value (close to 1) is shown for the input value to the NN near the center of the own store pattern. On the other hand, a low reaction value (close to 0) is shown for an input value to the NN far from the center of the own store pattern. Thereby, a certain kind of pseudo absolute recognition is realized.

  Furthermore, in the present invention, complete or missing images taken by the camera while the pachinko sphere rolls are imaged and recognized NN multiple times (for example, 10 times) at regular intervals, and more or less (for example, 1 to 2 times). When the ball engraved image and the NN recognize, the final decision is made that the store is a pachinko ball. On the other hand, only when more than half (for example, six or more) pachinko balls are recognized as NN images of the pachinko ball stamps of other stores, the final judgment is made as an image of the other store pachinko balls. As a result, the number of pachinko balls having in-store stamps is reduced as much as possible, and pachinko balls in other stores are excluded only when the pachinko balls clearly have other store stamps.

  As described in claim 3, the complete image of the own store and several types of missing images are used as learning data for the marking pattern of the pachinko ball of the own store, and the complete image of the other store and several types of missing images are used as the pachinko of another store. The learning data of the marking pattern of the sphere is used, and the defect images of the own store and the other stores are 1> σ> 0 while the complete image is σ = 1.

  In this way, when the complete image of the own store is 1, the partial image has an intermediate value between 1 and 0 (for example, 0.7), and the other stores have 0. On the other hand, when the complete image of the other store is 1, the partial image has an intermediate value between 1 and 0 (for example, 0.3), and the own store becomes 0. In other words, as shown in FIG. 6, σ is a bell-shaped Gaussian function, and not only the pachinko ball of the own store and other stores can be distinguished, but also a complete image or an incomplete (partial) image can be learned. Even in the case of partial images, it is possible to identify the inscriptions of pachinko balls of the own store and other stores.

  According to a fourth aspect of the present invention, it is preferable that a floor surface of the pachinko ball passage is formed in a spiral shape and a guide portion of the pachinko ball is provided along both sides of the passage.

  In this way, when the pachinko ball travels (moves) in the passage, the pachinko ball travels while changing the direction of the marking provided on the pachinko ball. When shooting the image, a complete image of a pachinko ball engraving or a partial image close thereto can be captured at a minimum number of times. Even if the system of the present invention cannot capture the complete image of the pachinko ball on the pachinko ball, it can recognize the stamp of the pachinko ball of the own store and the other store, but the stamp of the own store and the other store can be compared between the complete images. It goes without saying that it can be recognized reliably if possible.

The pachinko ball marking recognition system according to the present invention has the following excellent effects.
・ When the pachinko ball moves in the aisle, the direction of the stamp is changed as much as possible, and it is shot with multiple cameras from multiple directions, so the stamp is only provided on a part of the pachinko ball. There is a high possibility that the image is captured as a complete image or an image close to the complete image.
・ Shooting pachinko balls from multiple directions with multiple cameras, each time the captured imprinted image is converted to gray scale, and two-dimensional Fourier transform is performed to input a two-dimensional Fourier power spectrum as rotation invariant information. Since it is determined whether it is own store or other store based on a numerical value between 1 and 0 in an analog manner according to the degree of image loss, it is NN recognition even if it is a partial image Can be judged accurately.
Because RBF (radial based function) is used for the output function of NN, and it is made to appear on the bell-shaped rotator histogram of closed curves in the area of the own store and other stores, as shown in FIG. Whether it is a pachinko ball at your own store or a pachinko ball at another store can be reliably identified regardless of whether it is a complete image or a partial image.
・ Photograph multiple pachinko spheres located in the aisle with a camera multiple times at a time from multiple directions, cut out and label each pachinko sphere in gray scale, label, and 2D Fourier for each image Either converted and recognized by NN, or cut out and labeled individual pachinko sphere images, gray scaled for each image and two-dimensional Fourier transformed to be recognized by NN. It is possible to distinguish between own store and other stores with high efficiency due to high speed.
Since the color images of individual pachinko spheres are gray scaled, and these are two-dimensionally Fourier transformed and input to the NN, it is possible to recognize the NN by inputting information that is invariant to rotation.
-In the past, only 1 and 0 teachers were given by giving only complete self-store images and complete other store images to the NN, but this did not show the expected response in the partial images of the own store and other stores. . Therefore, the complete image of the own store and several types of partial images are used as learning data of the own store pattern, and the complete image of the other store and several types of partial images are learned and registered in the NN as learning data of the other store pattern. When the partial image of NN is given to NN, the teacher value of NN proportional to the missing state of the image is given at the time of learning. A shop can be identified by NN recognition.
・ Pushing ball images taken by the camera while rolling, make NN recognition multiple times, and remove it only when it is determined to be a pachinko ball from another store more than a predetermined number of times (50%) based on the majority rule. Therefore, it is not possible to mistakenly exclude pachinko balls from your own store.
-When the recognition system according to the present invention is installed for each pachinko machine (pachinko machine), it recognizes a plurality of pachinko ball marks multiple times at a time, and more than half of them are based on the majority rule. Only when a pachinko ball is recognized as an NN, it is a batch batch process that eliminates all pachinko balls or issues an alarm. , And high-speed processing is realized.
・ When the recognition system according to the present invention is installed in a general management system that manages a plurality of pachinko machines in a pachinko hall, the pachinko ball is scrutinized several times for each pachinko ball, and a single pachinko ball is based on the majority rule. Since it is a sequential process that eliminates pachinko balls or triggers an alarm only when it is recognized as a pachinko ball of another store in more than half of the times, it is highly probable (for example, 90% or more) The pachinko sphere can be identified.

It is explanatory drawing which shows the outline | summary of the NN recognition system used for the marking recognition system of the pachinko sphere of this invention. FIG. 1A is a perspective view showing an apparatus for a pachinko gaming machine, and FIG. 2B is a view for a center hall, showing an embodiment of a pachinko ball marking recognition system according to the present invention. FIG. It is a perspective view of a pachinko ball. Figures (a) and (b) are captured images of pachinko spheres using a non-reflective light source, and the stamps being rolled can be confirmed. Figures (a) to (c) are captured images of pachinko spheres using a non-reflective light source, and the stamps being rolled can be confirmed. In the pachinko ball marking recognition system of the present invention, it is an explanatory diagram illustrating the outline for distinguishing between a store and another store regardless of whether the stamped image of the pachinko ball is a complete image or an incomplete image. It is a learning screen which shows the simulator of NN. It is a learning drawing showing the line image and the Gaussian function (analog) showing the stamp of the other store on a pachinko ball.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below with reference to the drawings showing examples of a pachinko ball marking recognition system. FIG. 2A is a schematic diagram showing a state in which a part of the NN recognition system is installed in the pachinko ball passage in the pachinko hall, and FIG. 2B is a schematic diagram showing another mode.

  As shown in FIG. 2A, the floor surface 2a of the passage 2 is installed so as to be inclined downward from the proximal end toward the distal end toward the container 3 for storing the pachinko balls P. Bar-shaped guides 4 and 4 are arranged in parallel on both sides of the floor surface 2a of the passage 2, and the pachinko balls P move while rolling. A gap is provided between the tip of the floor surface 2a and the container 3, and a rod-shaped solenoid body 5 capable of removing pachinko balls P 'from other stores is disposed above and below the gap so as to be movable up and down. The pachinko ball P ′ to be rejected is projected by the solenoid body 5 at the tip of the passage 2 and falls downward. On the floor surface 2a of the passage 2, for example, a spiral groove (not shown) is formed from the base end to the tip, and the pachinko ball P moves while rotating so as to change the direction of the marking. A plurality of cameras (digital cameras) 6 are provided along the passage 2. In FIG. 2 (a), two cameras 6 are provided above the passage. However, as shown in FIG. 2 (b), the cameras 6 are also arranged on both sides of the passage 2, and the pachinko balls move along the passage 2. It is preferable that P can be photographed from at least three directions of the upper side and the left-right direction.

  In this way, the pachinko sphere P moves while rolling, and at the same time, the pachinko sphere is photographed with the camera 6 from above or from three directions, so that the markings formed on a part of the pachinko sphere can be reliably and as few times as possible. So that it can be taken with a pachinko ball. Each camera 6 is a color digital camera, and each camera 6 is connected to a personal computer 7. A non-reflective illuminator (non-reflective light source) 8 is mounted around the lens of each camera 6. Although a detailed drawing is omitted, a dome type illuminator or a flat dome illuminator in which a plurality of LEDs, which are a kind of indirect illumination, are arranged inside a semi-spherical ball is used for the illuminator 8. Thereby, when the pachinko sphere P as shown in FIG. 3 is photographed by the camera 6, the reflection of the surface of the pachinko sphere P is reduced, and the rolled stamp is taken as an image as shown in FIGS. can do.

  The pachinko sphere P is a silver or gold steel ball, and reflects the illumination light entirely as shown in FIG. 3, so the non-reflective illuminator 8 is used. Moreover, since the marking on the pachinko sphere is located on a part of the spherical surface, the camera 6 is photographed from a plurality of directions with different directions, and is photographed from a slightly inclined direction so that the marking becomes a shadow.

  Here, the camera 6 takes a picture while irradiating a plurality of pachinko balls P moving while rolling on the floor surface 2a with the non-reflective illuminator 8, and takes in the personal computer 7 as a color image (bitmap image). The color image is converted into a gray scale, the edge portion is extracted, cut out, and labeled.

The stamped image of the labeled pachinko sphere is input to the NN input layer as shown in FIG. The input to the NN is a two-dimensional Fourier transform of the imprinted image on the gray scale pachinko sphere as shown in the following formula 1, and a two-dimensional Fourier power spectrum (rotation root of the coefficient) as rotation invariant information. ) Is done.

  The NN shown in FIG. 1 has a three-layer structure in which cells indicating the own store and other stores are set as output cells. When the own store image is input, a response close to 1 is shown in the output cell indicating the own store. However, when a partial image of the own store is input, a reaction smaller than 1 is shown. In either case, the output cell indicating another store shows a reaction close to zero.

  In the conventional NN, when the own store image is input to the NN at the time of learning, 1 is given to the output cell showing the own store, 0 is given to the output cell showing the other store, these are taught, and the learning weight is set to As shown in Equation 2, the backpropagation method was used for adjustment.

However, in the pachinko sphere marking recognition system 1 according to the present invention, the pachinko sphere stamp image of both the own store and the other store cannot be captured as a complete image in many cases. Consideration must be given to both the partial image from which the imprinted image of the part is missing. For this reason, not only the integer of 1 and 0 but also the teacher value for the partial image is taken into consideration in the manner of the teacher, and a teacher value (intermediate value between 1 and 0) proportional to the image loss state is set. This is called analog teacher learning. That is, the teacher method is different from the conventional method and has the following aspects.
Case 1.
Full image of own store; teacher value T = 1.0
Teacher value to other stores T = 0.0
Case 2
Partial image of own store; teacher value T = 0.7
Teacher value to other stores T = 0.0
Case 3.
Partial image of own store; teacher value T = 0.3
Teacher value to other stores T = 0.0
Case A.
Complete image to another store; teacher value T = 1.0
Teacher value to own store T = 0.0
Case B.
Partial image to another store; teacher value T = 0.7
Teacher value to own store T = 0.0
Case C.
Partial image to another store; teacher value T = 0.3
Teacher value to own store T = 0.0
In addition, the output function of each cell in the NN has conventionally used a sigmoid function. However, since this is a high-order curve with an open identification region, it is difficult to identify partial images. Therefore, RBF (radial based function, for example, Gaussian function) shown in the following formula 3 is adopted as the output function of each cell of NN.

  That is, during the learning of NN, the response value of the output cell of NN appears on the surface of the Gaussian function as a rotational histogram, and is simply a curved plane discrimination surface as in the case where the output function is configured with a conventional sigmoid function. The value in the height direction is output three-dimensionally. That is, in this example, the bell-shaped rotating body histogram (Gauss function) D · E formed by the closed curve in FIG. 6 is used. Further, as shown in FIG. 6, the upper area A is the pachinko sphere P of the own store and the lower area B is the pachinko sphere P ′ of the other store with the curve C inclined from the lower left to the upper right in the vertical direction as a boundary. Is shown. Although the value of NN appears on the surfaces of the upper and lower rotating body histograms (Gaussian functions) D and E, when the store image is input to NN, it is 1 if it is a complete image. Further, since the data is located near the center of the rotation histogram (Gaussian function) D, in other words, at the top, a large response close to 1 is shown. In the case of a partial image of the own store, the data is located away from the center, in other words, at the bottom, and therefore the reaction is smaller than 1. However, both are located on the rotating body histogram (Gauss function) D of the own store. On the other hand, when a stamped image of another store is input to NN, it is positioned on the lower rotating body histogram (Gauss function) E.

  In addition, when a stamped image of a pachinko ball from another store is input to the NN on the NN recognition screen of the display in the personal computer 7, the response of the other store is 0.992564 as an NN recognition determination as shown in FIG. A value close to 1 is shown. Further, the evaluation image is almost complete as can be seen from the engraved image on the pachinko sphere depicted in the diagram, and the NN recognition result is understood to be another store. The screen for the own store is not shown, but the same applies to the case of the own store.

  First, the pachinko ball marking recognition system 1 according to the first embodiment, which is provided in the overall management device for pachinko gaming machines in the pachinko hall, will be described.

  As shown in FIG. 2A, the pachinko ball P moves toward the storage container 3 while changing the direction of the marking in the passage 2. At this time, while the plurality of cameras 6 irradiate the non-reflective light source 8 from a plurality of directions, the moving pachinko spheres P are photographed one by several times (for example, 15 times). The engraved image of the captured pachinko sphere P (including the image of the pachinko sphere P) is taken into the PC 7 as data. In the PC 7, the color pachinko sphere (engraved) image is converted into a gray scale, the edge portion is extracted, cut out, and labeled. The engraved image of the labeled pachinko sphere is input to an input layer of a three-layer NN as shown in FIG. The input to the NN can be divided into RGB and averaged into a gray scale after the color stamped image of the pachinko sphere is cut out and labeled. Then, the gray image is subjected to two-dimensional Fourier transform as shown in the above equation 1, and the two-dimensional Fourier power spectrum (the root of the coefficient square) as rotation invariant information is performed. On the other hand, for the NN, by using an RBF (for example, a Gaussian function) such as a rotational histogram shown in FIG. 6 as an output function, the response value of the output cell of the NN is as if on the surface of the Gaussian function as a rotational histogram. Appears and is not limited to the formation of a curved plane identification curved surface as in the case where the output function is configured with a conventional sigmoid function, and the value in the height direction is output three-dimensionally. A high response value (close to 1) is shown for an input value to NN close to. On the other hand, a low reaction value (close to 0) is shown for an input value to the NN far from the center of the own store pattern. Then, in the NN recognition, the pachinko ball P of the own store or the pachinko ball P ′ of another store is output from the output layer through the intermediate layer. NN recognition is performed a plurality of times (15 times) for the same pachinko ball, and if more than half (8 times or more) is recognized as the own store based on the principle of majority decision, it moves in the passage 2 and is accommodated in the container 3. On the other hand, if more than half (8 times or more) are recognized as other stores based on the majority rule, the alarm sound is generated, and at the same time, the solenoid body 5 descends and the pachinko ball P ′ to be removed is solenoided at the tip of the passage 2. It is struck by the body 5, falls downward from the gap between the tip of the floor surface 2a and the container 3, and is accommodated in another container (not shown). When more than half (eight times or more) are recognized as the store, the pachinko ball P passes through the gap and is accommodated in the container 3.

  Next, a pachinko ball marking recognition system 1 'according to a second embodiment installed in each pachinko machine pachinko machine will be described.

  The pachinko ball P moves in the predetermined passage 2 while changing the direction of the marking. At this time, while the plurality of cameras 6 irradiate the non-reflective light source 8 from a plurality of directions, the moving pachinko spheres P are simultaneously photographed, for example, ten times several times (for example, five times). The ten stamped images (including images of the pachinko spheres P) of the ten captured pachinko spheres P are taken into the PC 7 as data. In the PC 7, each color imprinted image is converted into a gray scale to become a gray image, and an edge portion is extracted, cut out, and labeled. The labeled image of each pachinko sphere is subjected to two-dimensional Fourier transform as shown in the equation 1 above, and a two-dimensional Fourier power spectrum (root of coefficient square) as rotation invariant information is shown in FIG. Is input to the input layer of a three-layer type NN. In the NN, the pachinko ball P of the own store or the pachinko ball P ′ of another store is output from the output layer through the intermediate layer. In this way, in this example, NN recognition is performed a plurality of times (5 times) for 10 pachinko spheres, and more than half (6 or more) of pachinko spheres in other stores are based on the majority rule. If it is recognized, all 10 pachinko balls P ′ in the passage 2 are eliminated at the same time as an alarm sound is generated. On the contrary, if more than half (6 or more) are recognized as one's own store among a plurality of times, they pass through the passage 2.

  As described above, the pachinko ball marking recognition system 1 for the overall management device of the pachinko hall and the pachinko ball marking recognition system 1 ′ for each pachinko gaming machine have been described by way of examples, but the present invention is limited to the above examples. It goes without saying that it is not done.

Industrial application fields

  This invention is a pachinko-related technology such as a pachinko game machine and its management device, which automatically picks up the pachinko balls of its own store and other stores from the difference of the stamps by imaging the markings on steel balls, mainly pachinko balls. Related to the field. In addition, it can be used in the field of auditing “Yuzu, melon, apples, etc.” with spherical fruits and vegetables, and the field of inspection of the entire surface of drugs such as capsules.

1.1 'pachinko ball marking recognition system 2 passage 2a floor 3 storage container 4 guide 5 solenoid body 6 camera (digital camera)
7 Personal computer 8 Non-reflective illuminator (non-reflective light source)
P / P 'Pachinko ball

Claims (4)

(1) means for moving the pachinko ball in the passage of the pachinko machine center device while changing the direction of the pachinko ball, and imaging the pachinko ball with a camera from a plurality of different directions;
(2) means for cutting out and labeling the color stamped image of the pachinko sphere;
(3) The labeled color image of the pachinko sphere is converted into a gray scale, and the two-dimensional Fourier transform is performed to obtain a two-dimensional Fourier power spectrum (root of the coefficient square) as information that is invariant to rotation. Means for inputting the two-dimensional Fourier power spectrum to an input layer of a three-layer neural network (NN) using RBF (radial based function) as an output function of each cell of the network (NN);
(4) The pachinko ball moving in the passage is photographed multiple times by the camera, the color stamped image is converted into a gray image, and the two-dimensional Fourier power spectrum is recognized NN multiple times. A pachinko ball marking recognition system comprising means for removing pachinko balls only when it is determined by NN recognition that more than half of the pachinko balls are stamped by other stores. (1) A pachinko ball is introduced into the passage of each pachinko machine while changing its orientation, and a predetermined number of pachinko balls are lined up and moved in series, and a plurality of the pachinko balls are grouped by a plurality of cameras from a plurality of different directions. Means for imaging once
(2) means for cutting out and labeling each color stamped image of the pachinko ball group;
(3) Each labeled color image of the pachinko sphere group is converted to gray scale, and these are converted into two-dimensional Fourier transform to obtain a two-dimensional Fourier power spectrum (root of coefficient square) as rotation invariant information. Using RBF (radial based function) as an output function of each cell of the neural network (NN), and inputting the two-dimensional Fourier power spectrum to an input layer of the three-layer neural network (NN);
(4 ′) A plurality of (for example, 10) pachinko ball groups flowing into the passage are photographed in a plurality of times with a camera, and the color stamp image is converted into a gray image. The Fourier power spectrum is input to the NN multiple times, and it is determined by NN recognition that more than half of the pachinko spheres of the input pachinko ball group are pachinko spheres of other stores based on the majority rule. A pachinko ball marking recognition system comprising means for removing all pachinko balls only when the pachinko balls are removed.   The complete image of the store and several types of missing images are used as learning data for the pachinko ball stamp pattern of the store, and the complete image of several stores and several types of missing images are used as learning data for the pachinko ball stamp pattern of other stores. 3. The pachinko ball marking recognition system according to claim 1 or 2, wherein the defect images of the own store and other stores satisfy 1> σ> 0 while the complete image is σ = 1.   The pachinko ball according to any one of claims 1 to 3, wherein a floor surface of the pachinko ball passage is formed in a spiral shape, and guide portions of the pachinko ball are provided along both sides of the passage. Stamp recognition system.