JPH03290786A - Pattern recognizing device for coin or the like - Google Patents

Abstract

PURPOSE:To enable pattern recognition in a short period of time even in the case of an object such as a coin or the like is free in the rotational direction by setting a concentrical ring in respect to the circular object, fetching the data of an input picture along the ring, exerting Fourier transform upon the data and comparing the transformed data with reference pattern data. CONSTITUTION:The picture signal of the coin or the like is inputted from a picture input part 1, converted to a multilevel digital picture signal by an A/D conversion part 2 and stored into a picture memory part 4. Next, in a ring forming part 5, the central position of the picture signal is calculated and further, the plural rings are formed concentrically to the calculated central point. In a ring data fetch part 6, the picture data is fetched from the picture memory part 4 along these plural rings and in a Fourier transform calculation part 7, the Fourier transform is exerted upon the data. Then, the frequency strength of the picture data is calculated. Afterwards, the pattern recognition is executed by comparing the picture data with the stored reference pattern upon which the Fourier transform is exerted in advance. Thus, the pattern recognition can be executed at high speed even in the case of the object such as the coin or the like free in the rotational direction.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an apparatus capable of recognizing and sorting patterns on the surface of circular objects such as coins.

(Prior Art) Conventionally known pattern recognition devices input an image of a subject and then process it to recognize a pattern, and there are the following types depending on the image processing method.

■A method in which the input image data is binarized, the area, center of gravity, perimeter, and other geometrical precision quantities are determined, and these are compared with reference values.

■A method of recognizing input cover patterns by pattern matching on the binary image or multilevel side. For example, a method in which an input character pattern of an object is superimposed on a standard character pattern prepared in advance, and the input character pattern is divided into 21 meters depending on the degree of similarity.

(2) A method of performing density projection on the X-Y coordinates or r-θ coordinates of the input image and comparing the characteristics of the intensity distribution.

For example, FIG. 9 shows an example of recognizing a character pattern by performing density projection on an X-Y/I mark, in which density projection 21 of an input character pattern 20 onto the X coordinate axis and density projection onto the Y coordinate axis are performed. 22, and the pattern is recognized using the waveform of the intensity distribution as a feature.

The principle of the method of recognizing a pattern by determining the density projection on the r-θ coordinate is the same, and the only difference is that the density projection in the rotational direction is determined.

(Problem to be Solved by the Invention) Regardless of the method used by the conventional pattern recognition device described above,
The positional relationship in the rotational direction is not constant, and it cannot be said to be rational for recognizing objects that are free in the rotational direction, such as coins. In other words, when attempting to recognize coins, etc. using the method described in (■) above,
After binarization, a process called labeling processing or area division is performed to divide the area for each character, for example when multiple characters are input at - degrees, and the center coordinates, area, and outer area of each area are divided. In order to obtain geometrical features such as diameter and circumference, a considerably troublesome process is performed, and there is a problem in that pattern recognition takes a long time.

When attempting to recognize a coin, etc. using the method (2) above, it is necessary to normalize the position of the coin, etc., including the direction, and there is a problem in that it takes time to determine the direction.

According to the method of projecting the density onto the r-θ coordinate of the method (2) above, it is possible to recognize coins, etc., but according to this method, the center of the coin, etc. is determined from the input image data,
Next, it is necessary to perform density projection and find the correlation with a reference coin or the like for one rotation, which poses a problem in that pattern recognition takes a long time.

In summary, conventional pattern recognition devices have a problem in that image processing for recognizing patterns on coins and the like takes a long time.

The present invention has been made to solve the problems of the prior art, and allows pattern recognition of objects such as coins that are free in the rotational direction by high-speed processing.
It is an object of the present invention to provide a pattern recognition device for coins, etc., which can maintain a high recognition rate.

(Means for Solving the Problems) The present invention improves the method of projecting the density onto the r-θ coordinates, and includes an image input section for inputting a pattern of a coin or the like;
An image memory unit that processes and stores the input pattern of coins, etc. into a multivalued image, and a ring formation unit that determines the center of the pattern of coins, etc. stored in the image memory unit and forms a plurality of rings concentric with this center. a ring data capture unit that captures data from the image memory unit along the plurality of formed rings and stores this data;
a Fourier transform calculating section that performs Fourier transform on the data along the plurality of rings stored in the ring data import section; The present invention is characterized by comprising a comparison section that compares the pattern data with reference pattern data.

(Function) The image data stored in the ring data import unit along a plurality of concentric rings changes in brightness according to the pattern of the portion where the rings are located. When this image data is Fourier-transformed by a Fourier-transform calculation section, the frequency intensity of the image data is determined. Since this frequency intensity varies depending on the pattern of the target object, such as a coin, pattern recognition can be performed by comparing it with a reference pattern that has been Fourier-transformed and stored in advance.

Furthermore, the frequency intensity is independent of the rotational position of the object such as a coin.

(Example) Hereinafter, an example of the pattern recognition apparatus for coins, etc. according to the present invention will be described with reference to FIGS. 1 to 8.

In FIG. 1, an image input section 1 is composed of a television camera or the like, and receives an image signal obtained by imaging a pattern of a coin or the like. The image input unit 1 uses a one-dimensional sensor or a two-dimensional sensor. In any case, it should be possible to process it as a two-dimensional image. Therefore, in the case of a -dimensional sensor, one screen is input by moving it relative to the object. The image signal taken in by the image input section 1 is converted into a multivalued digital image signal by the A/D conversion section 2 and input to the image processing section 3.

The image processing section 3 includes an image memory section 4, a ring forming section 5, a ring data import section 6, a fast Fourier transform calculation section (hereinafter referred to as "rFFT calculation section") 7° reference pattern storage section 8
and a comparison section 9.

The image memory section 4 stores the multivalued digital image signal. The ring forming section 5 finds the center of the pattern of a coin or the like stored in the one-image memory section 4, and forms a plurality of rings concentric with this center.

The ring data importing section 6 includes the ring forming section 5.
Data is taken in from the image memory unit 4 along the ring formed by the ring, and this data is stored.

The FFT calculation section 7 performs fast Fourier transform on the data along the plurality of rings stored in the ring data acquisition section 6.

The reference pattern storage section 8 previously stores reference pattern data such as coins in a Fourier transformed form.

The comparison section 9 compares the data Fourier transformed by the FFT calculation section 7 with the reference pattern data stored in the reference pattern storage section 8, and determines whether or not the data of both are similar. Pattern! LIL, output the result.

Next, details of the pattern recognition operation according to the above embodiment will be explained in the second section.
This will be explained with reference to FIGS. 6 through 6.

First, an image signal of a coin is input from the 1-image input section 1. FIG. 3(a) shows an example of an image 10 of a coin.

The input image signal is converted by the A/D converter 2 into a multivalued digital image signal, for example, 256 gradations, and is stored in the image memory 4.

Next, the ring forming section 5 calculates the image signal to find the center position. Here, for example, a plurality of points are taken on the edge of the image signal, and the least squares method is used to obtain the information as accurately as possible from the plurality of points.

FIG. 3(b) shows edge 1 of the image signal as described above.
The obtained center point ◯ of the image signal, which indicates the state in which the center point ◯ is obtained from 1, is stored in the image memory section 4. Ring forming part 5
Next, we will further form a plurality of rings that are concentric with the found center point ○. FIG. 3(c) shows an example of a plurality of rings 13, 14, 15 obtained as described above.

Next, the ring data import section 6 imports image data from the image memory section 4 along the plurality of rings, and stores the image data for each ring. Here, it is desirable to perform smoothing processing for noise removal for each ring. However, the smoothing process does not necessarily have to be performed in this step, and may be performed in the multivalued step described above. However, if the smoothing process is performed by the ring data importing section 6, only the ring-shaped data needs to be smoothed, so there is an advantage that the processing time is shortened. The image data captured along each ring becomes data whose density changes according to the pattern along each ring. The left half of FIG. 3(d) shows various examples of density changes along the rings, and the included frequency components differ depending on the pattern changes along each ring.

Next, the FFT calculation unit 7 performs FFT calculation for each data along each ring. By performing FFT calculation, the frequency intensity of each ring-shaped data is determined.

Frequency intensity refers to the change in concentration during one rotation of the ring.
When expressed as a frequency such as Hz, 2Hz, etc., it refers to the frequency at which the same frequency component is included. The right half of Figure 3(d) shows the data of the ring-shaped concentration change mentioned above.
This shows an example obtained by FT calculation, with the horizontal axis representing frequency and the vertical axis representing amplitude (intensity). As can be seen from the right half of Figure 3(d), when the frequency component of the ring-shaped concentration is relatively small, the FFT amplitude is large in the low frequency range, and when the frequency component of the ring-shaped concentration is relatively large, the FF amplitude is large.
T amplitude is large in high frequency range, in other words +FFT
The amplitude waveform will vary depending on the pattern along the ring.

Then, in the comparison section 9, the FFT amplitude data obtained for each ring and the FFT amplitude data of the reference pattern which has been Fourier transformed and stored in the reference pattern storage section 8 in advance are compared.
The FT amplitude data is compared with the evaluation amount, and it is recognized to which coin, etc., the target object corresponds.

Here, the evaluation amount Sk is as follows: Wij is the weight, Fij is the FFT amplitude of the input image, and R
i j k is represented by the FFT amplitude of the reference coin. The weight Wij weights a portion representing a characteristic pattern of the object as a large value. In the above formula, the evaluation amount S
It is determined that the smaller k is, the more similar the pattern is to the reference pattern. As described above, patterns on coins, etc. can be recognized.

Among the pattern recognition through the above series of image processing, FF
The T calculation will be further explained using the case of a 100 yen coin as an example. As shown in FIG. 4, since the pattern 10 on the back side of the 100 yen coin is a relatively simple pattern, a ring 14 of a predetermined radius is set on it, and the brightness level of the image signal along this ring 14 is Taking the key data, Figure 4 (
As shown in b), the period of change in brightness gradation during one rotation of the ring 14 until its rotation angle θ changes from O to 2π is relatively long, and the frequency component is relatively low.

Therefore, when this gradation data is subjected to FFT processing, as shown in Fig. 4 (
As shown in c), the amplitude (intensity) is large in a low frequency region such as 1Hz or 2Hz, and the amplitude (intensity) is small in a relatively high frequency region.

On the other hand, as shown in Fig. 5, since the pattern IOA on the surface of the ↓OO yen coin is a complex pattern, a ring 14 of a predetermined radius is set on it, and the brightness of the both meeting signal along this ring 14 is set. When the gradation data is taken, as shown in FIG. 5(b), the ring 14 rotates once and the rotation angle θ is O.
The change period of the brightness gradation from 2π to 2π is short, and the 1-frequency acid content is relatively high. Therefore, when this gradation data is subjected to FFT processing, a peak is seen in the amplitude (intensity) at a relatively high frequency of 5 Hz, as shown in No. 5@(C), which is a characteristic feature.

The vibration @ (intensity) for each frequency in Fig. 4 (c) and Fig. 5 (c) is determined by the FFT vibration @ F i j of the input coin or the F i j of the reference coin in the formula for calculating the evaluation amount Sk described above.
FT swing [Represents Rijk, FFT swing of input coin @
By evaluating the difference between F i j and the FFT amplitude Rijk of the reference coin, the type of coin and whether the coin is heads or tails can be recognized.

According to the embodiment described above, a concentric ring is set for a circular object such as a coin, input image data is captured along this ring, and this is further subjected to FFT calculation to generate ring-shaped data. Since the frequency intensity of the object is determined and compared with the reference pattern data to perform pattern recognition, the above frequency intensity finally used for pattern recognition becomes data that is unrelated to the rotational direction of the object, and is free in the rotational direction. Even for objects such as coins, pattern recognition can be performed in a short time with relatively simple processing, and the recognition rate can be kept high. Also, since image data is captured along concentric rings, the value at rotational position O becomes the value at rotational position 2π, which eliminates the effects of data aliasing. There is no need to set the value of O and the value of 2π to match.

In addition, multiple concentric rings are set up and image data is captured along each ring and processed, allowing for more accurate recognition of the pattern characteristics of the target object and improving the recognition rate. can be improved. Therefore, it is desirable to set each ring at a portion where the characteristics of the pattern of the coin or the like can be best captured.

Note that the reflection intensity of the object such as a coin varies depending on whether it is new or used, or depending on the material of the coin, etc., and the image captured along the ring. Differences in the signals may occur, reducing the reliability of pattern recognition. Therefore, reliability is ensured by comparing other data, for example, the sum of brightness for each ring, and normalizing the data according to the results. FIG. 6 shows an example of data normalization, and is an example in which the intensity after FFT calculation is corrected. Curve a shows the case of new coins, and curve S shows the case of circulating coins. The strength of circulating coins at IHz is converted to the strength of new coins at IHz, and the strength of circulating coins at other frequencies is shifted in accordance with this conversion value, and this is used for evaluation.

Even if the width of the ring to be formed is the width of ■ pixels of the image input section, the reliability of pattern recognition can be improved by forming multiple rings as described above. If you do it.

Even with one ring, the reliability of pattern recognition can be improved. FIGS. 7 and 8 show examples thereof.

In FIG. 7, the width W of a ring 17 formed concentrically with respect to an input image 10 such as a coin is set to three pixels. FIG. 7(c) is a developed view of the ring 17 set to have a width of three pixels, and three pixel columns are arranged in the upper row from the first row to the nth row. Therefore, Fig. 7(d)
Find the sum of the intensities of each pixel in each column as shown below, and apply this to FFT
Submit for calculation. Here, as shown in FIG. 8, all a21 a is selected from three pixels forming each pixel column.

When it is assumed that an intensity signal of 1 is obtained, it is preferable to weight each pixel. For example, set the weight of a engineering a3 to 1
If the weight of a2 is 2, then the strength A of one column is A=a1+2 a2+a.

becomes. By performing such processing, the characteristics of the pattern of the target object can be strongly brought out, and even if the center of the ring is shifted, the recognition rate can be increased.

Note that instead of performing FFT calculation on the pixels of the entire formed ring, a predetermined number of pixels may be extracted and calculation may be performed only on the extracted pixels. In this case, the number of data to be extracted and subjected to FFT calculation is 2N(
N is an integer).

If the number of data to be extracted is small, smoothing processing is performed to remove noise. If there is a large amount of data to be extracted, it is not necessarily necessary to perform smoothing processing.

(Effects of the Invention) According to the present invention, a concentric ring is set for a circular object such as a coin, input image data is captured along this ring, and the data is further Fourier transformed to form a ring shape. Since the frequency intensity of the data is calculated and compared with the reference pattern data that has been Fourier transformed in advance for pattern recognition, the frequency intensity used for pattern recognition is ultimately related to the rotation direction of the object. Even if the object is free in the direction of one rotation, such as a coin,
Pattern recognition can be performed in a short time with relatively simple processing, and the recognition rate can be kept high.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the coin pattern recognition device according to the present invention, FIG. 2 is a flowchart showing the operation of the same embodiment, and FIG. 3 is a conceptual diagram showing the signal processing process according to the above embodiment. 4 is an explanatory diagram showing step by step the principle of recognizing the back side of a coin according to the above embodiment, FIG. 5 is an explanatory diagram showing step by step the principle of recognizing the back side of a coin according to the above embodiment, and FIG. A line diagram showing the concept of normalization performed in response to changes in the conditions of the target object. Figure 7 is an explanatory diagram showing step-by-step another example of pattern recognition processing for coins, etc., and Figure 8 shows the same pattern recognition processing as above. FIG. 9 is a cross-sectional view showing the appearance of one row of pixels used in the process, and FIG. 9 is a conceptual diagram for explaining one of the conventional pattern recognition methods. 1 Image input unit, 4 Image memory unit, 5 Ring forming unit, 6 Ring data import unit, 7 Fourier transform calculation unit, 9 Comparison unit, 10 IOA pattern, 13, 14.15.・Ring, ○・Center of pattern. Figure Frequency T)-1z) (0) Figure (0) (b) (C) Figure

Claims (1)

[Scope of Claims] An image input unit for inputting patterns of coins, etc., an image memory unit for processing and storing the input patterns of coins, etc. into multivalued images, and an image input unit for inputting patterns of coins, etc. stored in the image memory unit. a ring forming section that finds a center and forms a plurality of rings concentric with the center; a ring data import section that captures data from the image memory section along the formed plurality of rings and stores this data; a Fourier transform calculation unit that performs a Fourier transform on the data along the plurality of rings stored in the ring data import unit; A pattern recognition device for coins, etc., comprising a comparison section for comparing the pattern data with reference pattern data.