JP3332952B2 - Coin identification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for covering a pattern using a ring-shaped mask, which is invariable to the rotation of a pattern image, reduces the pattern image data of coins, performs a separation operation efficiently, and learns by a neural network. The present invention relates to a coin identification device in which the weights are realized by computer software.

[0002]

2. Description of the Related Art Conventionally, when discriminating the authenticity, denomination, etc. of a coin, the coin is regulated so as to pass through a predetermined passage or a conveying path under certain conditions, and a predetermined range of the passing coin is optically or magnetically. To be read. Then, coins are identified by comparison and collation with reference data of a predetermined range for each coin registered in advance. For this reason, it is difficult to perform high-performance identification unless the coin transport conditions at the time of registration of the reference data and the coin transport conditions at the time of identification are set to the same state. As described above, in the conventional coin identification method, it is necessary to regulate the transport state of coins at the time of registration and at the time of identification, and there is a disadvantage that the transport system becomes complicated.

On the other hand, Rumelhart (Rumelhar)
t) et al. (Error Back P
research on neural networks has been very active since the report of the BP method (hereinafter abbreviated as BP method). The BP method is a learning method for a multilayer feedforward network, and is known as Rosenblatt (Rose Brat).
nblatt) perceptron learning and ADAL
INE (Adaptive Linear Neuro)
This is a general form of the Widow-Hoff rule (delta rule) for n). In recent years, improvements in the BP method and applied research using it have been particularly active. The most applicable area of the neural network is pattern recognition. This is because the pattern separation capability of the multilayer network is excellent, and it is difficult to recognize the deformed pattern with the conventional recognition technology. Reid et al. Used a higher-order neural network for the pre-processing unit, and constructed a system invariant to displacement, rotation, and scale conversion (enlargement / reduction) (Neural Networks, Vo).
l. 1, P. 689-692 (1989)), various rotations and the like are all given to various displacements, and explosions occur in combination of the units, which is not practical. In addition to these, many invariant pattern recognition systems using neural networks have been proposed, but simultaneously realize the properties that are invariant to displacement, rotation, scale conversion, and shading,
No practical system has been proposed.

[0004]

Another approach to pattern recognition that is invariant to displacement and rotation is to mathematically process an input pattern so as to extract features that are not affected by displacement, rotation, and the like. And inputting it to an identification unit (for example, a BP type neural network). This approach involves pre-processing Hough
Although there are methods using h-transformation, moment invariance, Fourier-Mellin transformation, and the like, the input image must be binary-coded in advance. Therefore, in order to treat not a binary image but an image of several bits of gray level as a pattern input,
When recognizing coins that require image processing such as boundary extraction as preprocessing of these mathematical transformations,
If the size of the coins to be targeted is different, it can be easily identified. However, coins of almost the same size must be identified by the features (patterns) of the coin surface. An image input by an image sensor is often rotated at a random angle from a standard image. In order to recognize the rotated image, a rotation-invariant pattern recognition structure is required, but no effective method has been proposed so far for this kind of problem.

The present applicant has proposed a money identifying method and apparatus (Japanese Patent Application No. 3-78513) for identifying money by a neural network having a learning function. Since the slab value based on the scrambled image is created in the pre-processing unit assuming the above deformation, a large amount of calculation is required in the pre-processing unit. Further, in the conventional neural network construction method, information obtained from the sensor of the identification device is directly input to the neural network. For this reason, in the neural network, the weight of information transmission must be given to all of the obtained input information, and the scale thereof has also been increased. For example, when 32 samples are taken with four sensors, the sensor information has 128 pieces. In the conventional method, the 128 pieces of information are directly input to the neural network, and even if the network structure is three layers, The scale was 128 × 64 + 64 × 12 = 8960, and the amount of calculation was an obstacle to practical use. Also, if a mounting machine equipped with a neural network is shipped and installed as a coin identification device, it becomes very expensive. After learning to the extent that a predetermined function can be performed as a coin discriminating apparatus, in reality, a fixed function without a learning function can sufficiently cope with it.

The present invention has been made under the circumstances described above, and an object of the present invention is to reduce the pattern image data of a coin optically measured by an image sensor by using a plurality of ring-shaped masks. Another object of the present invention is to provide a coin discriminating apparatus for discriminating coins by executing, with computer software, weights obtained by learning coin pattern recognition by a neural network in advance.

[0007]

The present invention relates to an effect discriminating apparatus, and an object of the present invention is to provide a learning function.
Hardness by pattern recognition using neural network
In a coin recognition device, an image sensor that optically measures a pattern image of a coin to be recognized , and a pattern image data obtained by the image sensor are divided into a number of small sections, and a binary image is provided for each small section. However, the preprocessing unit for converting the plurality of images representative values by using a plurality of ring-shaped mask covering the different parts in a ring shape from the binarized image data, the image representative value previously provided two
Obtained by learning on neural networks
Each weighting factor of the separation function is added to the pattern image of each coin.
With stored correspondingly, a separation calculation section for sequentially executing the separation operation of the image representative value based on the stored contents the each pattern images of the coins, the pattern of the coin which the separation calculation portion outputs A determination unit that determines and outputs a determination pattern having the highest matching degree among separation operation values corresponding to an image as a pattern image of the coin.

[0008]

According to the present invention, pattern recognition of coins and the like is performed by a neural network having a learning function, and pattern image data optically measured by an image sensor is reduced using a plurality of ring-shaped masks to form a plurality of images. Get a representative value. The ring-shaped mask has a large number of small sections, and the ring-shaped region of the small section is partially covered.
The image representative value reduced by such a ring-shaped mask is invariant to the rotation of the pattern image, and is input to and stored in the separation operation unit. In the separation operation unit, each weight coefficient of the separation function is learned and stored in advance by a neural network, and the determination unit determines the most among the separation operation values corresponding to the coin pattern image output by the separation operation unit. A determination pattern having a high degree of collation is determined as a pattern image of the coin. Accordingly, the configuration of the neural network at the time of learning becomes small, and the control device of the entire coin identification device becomes small. In addition, since the image data is reduced using the ring-shaped mask, there is an advantage that the recognition performance is not deteriorated by the fluctuation of the coin rotating direction.

[0009]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an example of the configuration of a coin identification device using a ring-shaped mask.
Is measured by an image sensor 2 composed of a CCD or the like, and subjected to necessary image processing such as normalization and noise removal to obtain an input image 3. The input image 3 is subjected to coordinate conversion from the xy coordinate system to the r-θ coordinate system and input to the ring-shaped masks 41 to 4n in the preprocessing unit 4, and the sum value from each of the ring-shaped masks 41 to 4n is calculated. Thus, image representative values SB1 to SBn are obtained. The image representative values SB1 to SBn are input to the separation calculation unit 5 of the neural network, and the separation calculation values SP1 to SPn are calculated for each category to be identified by weights optimally adjusted in advance for the pattern classification of coins for determination. . Each of the separated operation values SP1 to SPn is input to the determination unit 6, and a pattern having the maximum value (or the minimum value) among the separated operation values SP1 to SPn is output as a pattern image of the coin 1 as the target. Masks 41 to 4n in preprocessing unit 4
Is a ring-shaped mask in which different portions of a plurality of ring-shaped regions are covered.

First, a pre-processing unit 4 using a ring-shaped mask
Will be described. Ring masks 41 to 4 for coin identification
The reason for using n is as follows. As shown in FIG. 2, in a binary image of “0” and “1” on an 8 × 8 matrix, a slab value that is the sum of the number of pixels (in the example of FIG. 2A, “14” is obtained as a value characterizing the character “E” in FIG. 2A, and “12” is obtained as a value characterizing the character “H” in FIG. 2B. Is obtained. Therefore, “E” and “H” can be separated by using the slab value.
However, there are cases where slab values are equal even in images having different patterns. For example, in FIG. 3A, the slab value of the character "F" is "10", but the scrub value of the character "K" in FIG. 3B is also "10", and the character cannot be separated. For such a problem, after converting the input image from the xy coordinate system to the r-θ coordinate system, a specific ring shape corresponding to the pixels of the input image as shown in FIG. The problem can be solved by introducing a ring-shaped mask in which a part of the region is covered. When the image of FIG. 3A is covered with the mask of FIG. 4C, the image shown in FIG. 4A is obtained, and the slab value in this case is “4”. On the other hand, FIG.
When the image of FIG. 3B is covered with the mask of FIG.
, And the slab value is “3”. By introducing the ring-shaped mask in this manner, “F” and “K” can be separated. Furthermore, the slab value is invariant to rotation. First, in order to separate various images, the position of a portion covered with a ring-shaped mask as shown in FIG. 4C is selected in advance. In this case, the probability of generating slab values that can separate various images with only one such mask is extremely small. However, different slab value sequences can be obtained for the same image by using different ring masks as described above. Either of the slab value sequences often differs between images, and it is presumed that the use of various ring-shaped masks makes it possible to stochastically increase the separation capability between images. The use of the plurality of different ring-shaped masks has the following physical meaning. That is, when observing a three-dimensional object from another direction while changing the viewpoint, different information can be obtained even for the same object. Similarly, using various ring-shaped masks means observing an image from a different viewpoint in a two-dimensional plane. As described above, different information can be generated even from the same image.

In this case, in the pre-processing, the input image is covered with various different ring-shaped masks, and the sum of the uncovered pixels becomes slab values SB1 to SBn, respectively, and corresponds one-to-one with the input terminals of the input terminal group. .

FIG. 5 is a diagram for explaining the effect of a ring-shaped mask such as the masks 41 to 4n.
Indicates that the slab value is "6" when the input image is not rotated. Then, even if the input image is rotated as shown in FIG. 5B, the slab value remains at “6”, and a slab value that is invariant to the rotation of the input image can be obtained.

Next, the separation operation section 5 will be described with reference to FIGS. The separation operation unit 5 is roughly divided into the input terminal group 1
0, output terminal group 11, input terminal group 10, and output terminal group 1
It consists of a separation function indicating the relationship with 1. The input terminal group 10 is provided with input terminals so as to correspond one-to-one with each type of mask from the pre-processing unit 4, and has a slab value pre-processed by each mask type (pixels after mask processing). (Sum of numbers) is input to the corresponding input terminal. The output terminals 11 are provided with output terminals so as to correspond one-to-one to categories to be identified. For example, when a 500-yen coin, a 100-yen coin, a 50-yen coin, and a 10-yen coin are to be identified, the terminals correspond one-to-one from the terminal above the output terminal group, and the corresponding output terminal is 1 The closest value (maximum value) is output, and the other output terminals output values close to zero. If the coin is identified as a 500 yen coin, the output terminal group 1
The highest output terminal of 1 will output the maximum value.

The separation function indicating the relationship between the input terminal group 10 and the output terminal group 11 is formed through an intermediate separation function (y ₁ , y ₂ ). Specifically, for example, three input terminals of the input terminal group 10 and three output terminals 11
FIG. 6 shows four output terminals and four output terminals. The separation function corresponding to FIG. 6 first uses information (SB ₁ , SB ₂ , SB ₃ ) input to the input terminals of the input terminal group 10 as variables. As an intermediate separation function,

Y ₁ = f (SB ₁ × w ₁ + SB ₂ × w ₃ + SB ₃ × w ₅ ) y ₂ = f (SB ₁ × w ₂ + SB ₂ × w ₄ + SB ₃ × w ₆ ) holds. Also, between the intermediate separation function and the output terminal of the output terminal group,

[Number 2] _{SP 1 = f (y 1 ×} w 1 '+ y 2 × w 5') SP 2 = f (y 1 × w 2 '+ y 2 × w 6') SP 3 = f (y 1 × w 3 '+ Y ₂ × w ₇ ') SP ₄ = f (y ₁ × w ₄ '+ y ₂ × w ₈ '). Here, the following equation 3 is obtained by substituting equation 1 for equation 2.

SP ₁ = f ｛f (SB ₁ × w ₁ + SB ₂ × w ₃ + SB ₃ × w ₅ ) × w ₁ ′ + f (SB ₁ × w ₂ + SB ₂ × w ₄ + SB ₃ × w ₆ ) × w ₅ ′｝ SP ₂ = f ｛f (SB ₁ × w ₁ + SB ₂ × w ₃ + SB ₃ × w ₅ ) × w ₂ ′ + f (SB ₁ × w ₂ + SB ₂ × w ₄ + SB ₃ × w ₆ ) × w ₆ ′｝ SP ₃ = f ｛f (SB ₁ × w ₁ + SB ₂ × w ₃ + SB ₃ × w ₅ ) × w ₃ ′ + f (SB ₁ × w ₂ + SB ₂ × w ₄ + SB ₃ × w ₆ ) × w ₇ ′｝ SP ₄ = f ｛f (SB ₁ × w ₁ + SB ₂ × w ₃ + SB ₃ × w ₅ ) × w ₄ ′ + f (SB ₁ × w ₂ + SB ₂ × w ₄ + SB ₃ × w ₆ ) × w ₈ ′｝ Note that the function f is

(Equation 4) It is. By executing the above equations 1 to 3 with computer software, the weights w _{1 to} w ₆ and w ₁ ′ to w ₈ ′ are all predetermined, so that combinations SB _{1 to} SB _{3 of the} input image representative values are set. Output SP _{1 to} SP ₄ (for example, 50
The matching degree of 0 yen coin, 100 yen coin, 50 yen coin, 1 yen coin) can be obtained.

The weights W _{1 to} W ₆ and W ₁ ′ to
For the selection of W ₈ ′ in advance, for example, FIG.
Can be performed using neural network learning as shown in

(Equation 5) There is a back propagation method given by Since this learning is already generally known, the description is omitted. Next, the determination unit 6 will be described. The determination unit 6 receives information from each output terminal of the separation operation unit 5, determines the maximum value (may be set to the minimum value) from the information, and corresponds to the output terminal having the maximum value. Determine the category.
For example, if the output terminals of the output terminal group 11 correspond to the category of 500 yen coin, 100 yen coin,... 10 yen coin from the top, if the output of the highest value is output from the highest output terminal, 500 yen Judge as coins. In the above embodiment, the input image is temporarily converted from the x, y coordinate system to the r-θ coordinate system to perform processing. However, this coordinate conversion is omitted and the processing is directly performed in the xy coordinate system. Processing is also possible. In particular, in the case of an input image having a high resolution, there is little need to convert the input image into the r-θ coordinate system, and the ring-shaped processing can be performed with the xy coordinate system. Further, in the above embodiment, the separation operation unit 5 and the discrimination unit 6 are configured by different hardware.
And the same processing can be performed by a software program. Further, in the above embodiment, the intermediate separation function (y ₁ , y ₂ ) is general, but the separation operation between the input terminal group 10 and the output terminal group 11 is performed via the intermediate separation function of several stages. You may do it.

[0016]

As described above, according to the coin discriminating apparatus of the present invention, since the learning is performed in advance using the neural network, the coin can be recognized invariably in the rotation and density of the coin. Further, according to the present invention, since the image representative value (slab) is obtained by using the ring-shaped mask in the preprocessing unit, the configuration of the discrimination function unit is small, and the control device of the entire coin discrimination device is small. Become. Furthermore, since a discriminant function is configured using the weighting coefficients obtained by learning and discriminant calculation is performed by computer software, a relatively inexpensive mounting machine can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration example of a coin identification device using a neural network as a premise of the present invention.

FIG. 2 is a diagram for explaining a slab value used in the present invention.

FIG. 3 is a diagram for explaining a slab value used in the present invention.

FIG. 4 is a diagram for explaining a ring-shaped mask used in the present invention.

FIG. 5 is a diagram for explaining the effect of the ring-shaped mask.

FIG. 6 is a diagram for explaining weights of a neural network according to the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Coin 2 Image sensor 3 Input image 4 Preprocessing part 5 Separation calculation part 6 Judgment part 7 Storage calculation part

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-4-118796 (JP, A) JP-A-1-245386 (JP, A) JP-A-59-62980 (JP, A) (58) Survey Field (Int. Cl. ⁷ , DB name) G07D 5/00

Claims (1)

(57) [Claims] 1. A neural network having a learning function
In a coin recognition device based on pattern recognition using, an image sensor that optically measures a pattern image of a coin to be recognized , and a pattern image data obtained by the image sensor is divided into a number of small sections. binarizing each small compartment, a preprocessing unit for converting the plurality of images representative values by using a plurality of ring-shaped mask covering the different parts in a ring shape from the binarized image data, previously provided the Image representative value is calculated using the neural network
Each weight of the separation function obtained by learning on the work
A separation operation unit that stores the only coefficient in correspondence with the pattern image of each coin, and sequentially executes the separation operation of the image representative value based on the storage content for each pattern image of each coin, A determination unit that determines and outputs, as a pattern image of the coin, a determination pattern having the highest matching degree among the separation operation values corresponding to the pattern image of the coin output by the separation operation unit. Coin identification device.