JP2001331839A - Method and device for discriminating paper money - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately perform pattern recognition, discrimination and elimination of counterfeited/damaged paper money by improving the discrimination rate of the pattern recognition of paper money that is partially subjected to specific counterfeit/alteration processing and normal paper money whose damage is in progress. SOLUTION: This paper money discriminating method in which the characteristic quantity of paper money to be discriminated is inputted, a reference pattern is learned and subsequently pattern recognition of unknown paper money is performed according to the reference pattern comprises the steps of: sensing the unknown paper money and extracting each characteristic quantity of every first small area, collecting the characteristic quantity in second small areas larger than the first small area; comparing the reference pattern that is provided corresponding to the second small area and can be selected as a denomination/direction candidate of paper money and the characteristic quantity of the second small area to perform pattern recognition; consequently decides a candidate to which the unknown paper money corresponds; excluding an uncorresponding candidate; repeatedly performing such processing over the entire area while changing second small areas; storing a finally remaining candidate; simultaneously comparing the characteristic quantity of the entire unknown paper money and all reference patterns by changing the above mentioned conditions to perform pattern recognition; storing the denomination and direction of the paper money obtained as a candidate; and outputting a common candidate as the denomination and direction of the unknown paper money when the common candidate exists in such candidates.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for recognizing forged or altered banknotes which have been partially treated abnormally or which have been improved in their ability to remove banknotes which have been deformed or altered due to damage or the like. About.

[0002]

2. Description of the Related Art The automation of money transfer and settlement operations in finance, distribution, transportation, etc. is remarkable, and the spread of ATMs, distribution, vending machines, etc. in various financial institutions is well known. is there. In addition, payment devices behind these customer service devices, which totalize and settle the received money, have been almost automated. The functions and performance of these automatic devices have been remarkably improved, from automatic dispensers that dispensed the required amount of money loaded by hand to automatic deposit machines equipped with identification devices that identify the denomination and authenticity of money. Automatic depositing and dispensing machines have been developed.

[0003] After that, with the development of a fitness discrimination technology for discriminating whether the currency is dirty or damaged, the deposited money is separated into genuine coins that can be recirculated and damaged coins to be collected.
A circulation type automatic teller machine which circulates only genuine coins as money for withdrawal to a bank can be developed.

[0004] Further, in the above-mentioned automatic equipment, the cash management function in the equipment has been strengthened, and the automatic equipment has been operated as a substitute safe for storing cash. For the development and spread of these automation devices, high reliability is required for various constituent devices and technologies. Furthermore, an infinite guarantee is required for a money receiving / dispensing technology, a separation transporting and storing technology, and in particular, a money discriminating technology for discriminating the denomination and authenticity of money, and no difference is allowed.

[0005] In particular, regarding banknotes, moisture absorption, drying, dirt,
In addition to creases, wrinkles, holes, tears, graffiti, etc., the paper quality and printing condition vary due to paper fatigue and the like. Furthermore, in recent years, counterfeit color-copy tickets, banknotes improperly stamped, perforated banknotes such as pinholes, counterfeit banknotes partially substituting for normal banknotes, and the like have been actively used illegally. Development of very advanced banknote pattern recognition technology is required, such as the use of the same banknote as a normal banknote to accurately identify forged banknotes as fraudulent banknotes. Then, in any state of the bill, the genuine bill is 100
There is a need to be able to 100% reliably detect and eliminate even the most sophisticated counterfeit and counterfeit bills, which are accurately counted.

As for the currency identification technology, there is a conventional pattern matching method of extracting various characteristic amounts from sample pattern data, thereafter performing arithmetic processing to obtain a reference pattern, and performing recognition by comparing the reference pattern with input data. . Also, as a neural network of a model imitating a human brain nervous system, a hierarchical neural network based on learning by a BP (Back-Propagation) method has attracted attention. L as another neural network
There is a competitive neural network that performs learning by the VQ (Learning Vector Quantization) method. A reference pattern is created by learning by the LVQ method. I can do it.

However, when performing complex pattern recognition by pattern matching, simply calculating the average value of the sample data simply means that a fraudulent bill that has been counterfeited using the same bill, such as a normal bill, is partially used. It is very difficult to accurately identify a banknote as a fraudulent banknote, and a lot of experience and time are required to create a reference pattern. In addition, the hierarchical neural network based on the BP method requires a large amount of computation for pattern recognition (because the three-layer structure has a large number of neurons and uses a sigmoid function).
It takes time to recognize. The operation time can be solved by implementing the neuro operation processing as hardware, but the hardware (use of a DSP or a neuro-dedicated CPU board, etc.) is very expensive, and may cause a local solution in learning,
There are problems such as difficulty in finding the physical causal relationship between input and output due to the complicated structure.

On the other hand, in a competitive neural network, a reference pattern in pattern matching is learned as a coupling coefficient, and can be easily created even when a plurality of reference patterns are required, so that a complicated pattern recognition process can be performed. Further, the competitive neural network has a simple structure, and performs a distance calculation between an input and a coupling coefficient and determines an output value using a threshold at the time of identification, which is more complicated than the hierarchical neural network. It is also possible to easily learn the input of a simple pattern or a multidimensional feature amount, and to perform the calculation at the time of identification at a high speed.

[0009] However, in the conventional competitive neural network, the input feature value data is fired by the nearest neuron, so that papers / banknotes other than paper / banknotes to be identified, such as counterfeit bills, are generated. When the input is made, there is a problem that one of the neurons is fired and erroneously recognized even though the distance between the input data and the coupling coefficient is long.

By the way, a currency discriminator may receive a non-target medium such as currency from another country or a counterfeit bill depending on the purpose. However, in the competitive neural network, the input feature data and each competitive layer neuron (hereinafter, referred to as
(Also simply called a neuron)
Since the nearest neuron fires and the input data is identified as the category to which the firing neuron belongs,
Even when the distance between the input data and the coupling coefficient is considerably large, the competitive layer neuron having the coupling coefficient closest to the input data will fire, and therefore will fire even on non-target media such as counterfeit banknotes and cause an error. It can cause recognition.
In order to prevent such erroneous recognition, a method of estimating an output value distribution of a competitive neural network to secure a range of self-recognition and setting a threshold value for an output value to each neuron is disclosed in Japanese Patent Application No. 11-186135. Has been described.
In a competitive neural network, a complex pattern can be accurately recognized by increasing the number of neurons, but the method is not clear. In the above learning,
The number of neurons is automatically added and deleted based on the results of data recognition and reliability evaluation used. Here, the learning means updating the coupling coefficient of each competitive layer neuron.

The above algorithm of the LVQ method includes an LVQ1 for learning one neuron whose distance is the shortest.
Other than the above, there are LVQ2 and LVQ3 for learning the two neurons having the smallest distance, OLVQ1 for which the learning rate in LVQ1 is optimized, and the like.
(Optimized Learning Vector
r Quantization Method 1) is used. Since the LVQ method uses only the Euclidean distance calculation, unlike the hierarchical neural network using the sigmoid function, the amount of calculation is small and the scale is small. Further, since the coupling coefficient is a quantization vector of the input feature amount data, there is a feature that a physical causal relationship between the input and the output can be easily found.

Meanwhile, the competitive neural network has a two-layer structure as shown in FIG. 1, in which a first input layer receives an input pattern, and an input to a second competitive layer is passed to the input pattern to input the input pattern. Classification of The neurons in the input layer are completely connected to the neurons in the competitive layer, and each connection is represented by a connection coefficient. Here, the coupling coefficient between the neuron j in the input layer and the neuron i in the competition layer is represented by w
ij, where N is the number of neurons in the input layer and M is the number of neurons in the competition layer. In the neuron of the competitive layer, the input data x = (x1, x2,..., XN) and the coupling coefficient wi = (w1, w2,.
The Euclidean distance di with respect to N) is calculated according to the following equation (1).

[0013]

(Equation 1) The neuron c whose distance di defined by the above equation 1 is minimized is selected, and is called a winner neuron.

[0014]

(Equation 2) The output yc of the winner neuron is yc = 1, and the outputs yi (i ≠ c) of the other neurons are yi = 0.

In this method, the neuron of the competitive layer having the coupling coefficient closest to the input pattern vector is fired, so that the input pattern to the competitive neural network can be identified.

[0016]

That is, the entire area of a target banknote is divided into predetermined small areas, and a feature amount corresponding to each small area is extracted. In the case of identifying / determining the direction, as shown in FIG. 21 (A), a new unknown banknote X is input to the reference patterns A, B, etc., which are created by processing a predetermined number of normal banknotes in advance. When the extracted feature amount is subjected to pattern matching processing, the unknown banknote X can be normally identified as the denomination and direction of the reference pattern B from the magnitude of the comparison operation value with each reference pattern and the determination of the inside / outside of the allowable range. .

However, like the unknown input banknote Y shown in FIG. 21B, similar foreign country banknotes, counterfeit banknotes, falsified banknotes, and partially peculiarly falsified banknotes using normal banknotes are used. Is input, if the pattern is simply recognized using the above-described pattern matching technique or only the feature amount extracted from the entire area, the unknown input bill Y
Is partially unique, and most of the feature amounts are similar to the reference pattern B of the forgery source, so that the matching result and the pattern recognition result fall within the allowable range of the reference pattern B, This resulted in misidentification. In addition, for the above-mentioned partially unique counterfeit banknotes, it is possible to exclude counterfeit banknotes by extracting only specific counterfeit areas and comparing / collating them with a normal reference pattern. Since it is difficult to predict in advance which area has been forged / falsified, it has been extremely difficult to correctly identify such a specially falsified banknote.

Further, there are the following problems with a banknote that has been damaged. That is, for an unknown input bill, the entire bill is divided into predetermined small areas, and the feature amount of each small area is sequentially extracted each time the feature amount is extracted.
Input to the pattern recognition device, the denomination of the bill, the method of identifying the direction and, for unknown input bills, divide the whole bill into predetermined small areas, extract the characteristic amount for each small area, Even if the method of collectively inputting into the pattern recognition device and identifying the denomination and the direction of the banknote is used in a simple combination, even if the similar two normal reference patterns M and N are registered, In this case, there is a problem that the identification rate of a damaged normal bill is not improved.

More specifically, when two similar normal reference patterns M and N are registered as shown in FIG. 22A, the input of a normal bill to be identified as the reference pattern N is performed. With respect to the input pattern X which is a pattern and is partially damaged only in the region 5, in the determination for each small region, the reference pattern N is excluded due to the damage in the region 5,
The reference pattern M remains in the identification candidate as being within the allowable range.

Next, the feature values of the entire area are input collectively,
When calculating the similarity, simply extracting the most similar reference pattern within the allowable range only for the reference patterns remaining in the above-mentioned identification candidates, the reference pattern N becomes the small area identification. Since it is excluded, the reference pattern M is determined to be the most similar reference pattern within the allowable range, and the damaged input banknote X is erroneously recognized as the reference pattern M.

The present invention has been made in view of the circumstances described above, and it is an object of the present invention to identify a banknote that can be accurately pattern-recognized, discriminated, and eliminated from a partially unique counterfeit / falsified banknote. It is to provide a method and an apparatus. Further, another object of the present invention is to provide a method and an apparatus for improving the identification rate of pattern recognition of a damaged normal banknote, thereby accurately recognizing a damaged banknote as a pattern, and identifying and rejecting the damaged banknote. There is also.

[0022]

According to a first aspect of the present invention, there is provided a banknote identification method for inputting a feature amount of a banknote to be identified, learning a reference pattern, and then inputting an unknown unknown value. A banknote recognition method for recognizing a pattern of the banknote using the reference pattern, wherein the unknown banknote is sensed and each characteristic of the unknown banknote is extracted for each predetermined first small area (basic unit). A first step of assembling an amount into a second sub-region larger than the first sub-region; and providing and corresponding to each of the second sub-regions (regions 1 to P of 103c). A reference pattern selectable as a candidate for a banknote denomination / direction and the second small area (1
A second step of recognizing a pattern by comparing a feature amount of the first small area (basic unit) constituting the first to third areas 03c and a denomination to which the unknown banknote corresponds as a result of the pattern recognition A third step of determining candidate directions and excluding non-applicable denominations and directions from the candidates; and replacing the second step and the third step by replacing the second small area with the entire area. (From 1 to P of the window pattern 103c), obtaining the result of the second step from among denominations and directions, and finally denominations of banknotes remaining as candidates for the unknown banknote And a fourth step of storing the direction, and simultaneously performing the feature recognition of the first predetermined small area and the reference pattern for the entirety of the unknown banknote to perform pattern recognition and obtain a candidate as a candidate. Remembers banknote denominations and direction candidates 5 and step, the fourth step and the fifth banknote denomination stored in step,
If there is a common candidate for each direction,
A sixth step of outputting the denomination and direction of the common candidate banknote to the outside as the denomination and direction of the unknown banknote.

[0023] Further, in the bill recognizing method according to the second aspect, the pattern recognition may be performed by an LVQ (Learning Vector Quantizon).
ation) method is used.

[0024] Further, a bill discriminating method according to claim 3 is characterized in that:
A banknote identification method in which a feature amount of a banknote to be identified is input and a reference pattern is learned, and then the input unknown banknote is pattern-recognized using the reference pattern. Quantizatio
Using a neural network according to the n) method, a denomination and a direction to be fired in the output layer neuron are set, the bill to be identified is sensed, and the whole bill is set to a predetermined first.
And input all the feature amounts extracted for each of the first small regions to the input layer neurons of the neural network, and learn and store a reference pattern in pattern matching as a coupling coefficient. A first step of causing each of the feature amounts of the banknote to be learned extracted for each first subregion (basic unit) of the banknote to be identified to be a second subregion larger than the first subregion. The feature amount of the second small area is repeatedly input to the input layer neurons of the neural network over the entire area while changing the small area, and the distance from the reference pattern is calculated. A second step of setting a threshold value so that the output layer neuron of the corresponding denomination and direction is fired; and a whole feature amount banknote extracted for each first small area (basic unit). Simultaneously inputting into the neural network (103b), calculating the distance from the reference pattern, and setting a threshold value so that the output layer neuron of the corresponding denomination and direction fires; The entire bill is divided into the first small areas, and the characteristic amounts of the unknown bills extracted for each of the small areas (basic units) are collected in the second small area (103C).
The feature amount of the second small area is repeatedly input to the input layer neuron of the neural network over the entire area while changing the small area, and the distance from the reference pattern is calculated. A fourth pattern for calculating a seed and a direction and performing a first pattern recognition for inactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value;
And simultaneously inputting the feature amounts extracted for each of the first small areas (basic units) of the entire unknown bill to the neural network (103b), and denominations of candidate bills, Comparing the fifth step of performing the second pattern recognition for calculating the direction and the calculation results of the fourth and fifth steps, and when the same candidate is extracted, the denomination of the banknote of the same candidate; And a sixth step of outputting a direction.

According to a fourth aspect of the present invention, there is provided a method for recognizing a bill, wherein a feature amount of the bill to be recognized is inputted and a reference pattern is learned.
Thereafter, a bill identification method for recognizing an inputted unknown bill using the reference pattern, wherein the pattern recognition includes LVQ (Learning Vector Quantization).
The denomination and the direction to be fired in the output layer neuron are set using the neural network by the method, the bill to be identified is sensed, and the whole bill is divided into a predetermined first small area (basic unit). A first step of inputting all the feature amounts extracted for each of the first small areas to an input layer neuron of the neural network, learning and storing a reference pattern in pattern matching as a coupling coefficient, and The feature amounts of the banknote to be identified extracted for each first subregion (basic unit) of the banknote to be collected are collected in a second subregion larger than the first subregion (103C). The feature amount of the second small area is repeatedly input to the input layer neuron of the neural network over the entire area while changing the small area, and the distance from the reference pattern is calculated. A second step of setting a threshold value so that the output layer neuron of the corresponding denomination and direction is fired, and calculating the whole amount of the feature amount banknote extracted for each of the first small areas (basic units) The third step of simultaneously inputting to the neural network, calculating the distance from the reference pattern, setting a threshold value so that the output layer neuron of the corresponding denomination and direction fires, ,
While dividing into the first small area, each characteristic amount of the unknown bill extracted for each of the first small areas (basic unit) is collected in the second small area (103C), The feature amount of the second small region is repeatedly input to the input layer neuron of the network over the entire region while changing the small region, and the distance from the reference pattern is calculated. A fourth step of calculating a direction and performing a first pattern recognition for inactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value; The feature amount extracted for each small area (basic unit) is simultaneously input to the neural network including neurons whose output layer neurons have been deactivated by the first pattern recognition means,
The fifth step of performing the second pattern recognition for calculating the denomination and the direction of the candidate banknotes is compared with the calculation results of the fourth and fifth steps. And a sixth step of outputting the denomination and the direction of the same candidate banknote.

Further, the bill recognition method according to the fifth aspect is characterized in that:
A banknote identification method in which a feature amount of a banknote to be identified is input and a reference pattern is learned, and thereafter, an input unknown banknote is pattern-recognized from the reference pattern. In the pattern recognition, LVQ (Learning Vector Quantization) is used. A denomination and a direction to be fired in an output layer neuron are set using a neural network based on the method, and the input layer neuron of the neural network senses the bill to be identified and performs a predetermined first small area A first step of inputting all the feature amounts extracted in (basic unit) and learning and storing a reference pattern in pattern matching as a coupling coefficient, and for each first small area of the banknote to be learned (basic unit) The feature amounts of the banknotes to be learned that are extracted in the second small area larger than the first small area are collected (103C). The feature quantity of the second small region in the input layer neurons-menu neural network,
A second step of repeatedly inputting over the entire area while replacing the small area, calculating the distance to the reference pattern, and setting a threshold value so that the output layer neuron of the corresponding denomination and direction fires; The entire banknote to be learned is divided into the first small areas, and all the feature amounts extracted for each first small area (basic unit) are simultaneously input to the neural network (103b). A third step of calculating a distance from the reference pattern and setting a threshold value so that an output layer neuron of a corresponding denomination and direction is fired, and extracting the first small area (basic unit) of the unknown bill; The characteristic amounts of the unknown bills thus collected are collected in the second small area (103C), and the input amount neuron of the neural network is replaced with the characteristic amount of the second small area, and the entire area is replaced by the small area. Over Return input, calculate the distance to the reference pattern, calculate the denomination and direction of the candidate banknote, and deactivate the output layer neuron whose distance to the reference pattern is equal to or greater than a predetermined threshold. A fourth step of performing pattern recognition, and dividing the entirety of the unknown banknote into the first small areas, and extracting all feature amounts extracted for each first small area (basic unit). A fifth step of simultaneously inputting to the neural network (103b) and performing a second pattern recognition for calculating the denomination and direction of a candidate banknote; and a first step for each first small area of the unknown banknote (basic unit). ), The feature amounts of the banknotes to be learned are collected in a third sub-region that is larger than the first sub-region and has a feature that can be identified for a candidate denomination. (103a), the new The feature amount of the third small region is input to the input layer neuron of the network, the distance to the reference pattern is calculated, and the output layer neuron closest to the reference pattern is used as the identification result. A sixth step of performing pattern recognition of
A seventh step of comparing the calculation results of the fourth, fifth, and sixth steps and, when the same candidate is extracted, outputting the denomination and direction of the banknote of the same candidate. And

Next, the bill discriminating apparatus according to claim 6 inputs a feature amount of a bill to be discriminated and learns a reference pattern,
Thereafter, the input unknown banknote is pattern-recognized from the reference pattern, and LV is used for the pattern recognition.
A neural network based on the Q (Learning Vector Quantization) method is used, and the input layer neurons of the neural network are selectively input with the obtained feature amounts for each of the divided small regions according to the positional dependency. A bill discriminating device for performing a neuro operation, wherein the bill sensing device performs sensing of the bill and divides the whole bill into a first predetermined small area to extract a feature amount; and a predetermined number of learning bills. For each of the banknotes to be identified, sequentially inputting the characteristic amount of each small area obtained by the characteristic amount extracting means to the input layer neuron of the neural network, and setting the reference pattern to a predetermined value. A first learning unit that stores the characteristic amount of the banknote to be identified extracted for each first small area by the characteristic amount extracting unit. A second learning step in which the reference amounts are gathered in a second small area larger than the area, and the feature amounts for the entire area of the bill are simultaneously input to the input layer neurons of the neural network, and the reference patterns are stored in predetermined neurons, respectively. Means, the feature amounts obtained by the feature amount extraction means from the unknown banknotes are collected in the second small area, and sequentially input to the input layer neurons of the neural network, and the distance from the reference pattern is calculated. A first pattern recognition means for calculating a denomination and a direction of a candidate banknote and deactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value; The output layer neuron is deactivated by the first pattern recognition means by using the feature amount of the entire unknown bill extracted for each first small area by the first pattern recognition means. The Type neural network simultaneously comprising of neurons, and calculates the distance between the reference pattern, the output layer neurons of the distance is within the predetermined threshold value,
Comparing the calculation results of the second pattern recognition means, which can be selected as banknote denominations and direction candidates, and the first and second pattern recognition means, and when the same candidate is extracted,
And a calculation result integrating means for outputting a denomination and a direction of the banknote of the same candidate.

[0028] The bill identifying method according to claim 7 is characterized in that:
A banknote identification method in which a feature amount of a banknote to be identified is input and a reference pattern is learned, and thereafter, an input unknown banknote is pattern-recognized from the reference pattern. In the pattern recognition, LVQ (Learning Vector Quantization) is used. A denomination and a direction to be fired in an output layer neuron are set using a neural network based on the method, and the input layer neuron of the neural network senses the bill to be identified and performs a predetermined first small area A first step of inputting all the feature amounts extracted in (basic unit) and learning and storing a reference pattern in pattern matching as a coupling coefficient, and for each first small area of the banknote to be learned (basic unit) The feature amounts of the banknotes to be learned that are extracted in the second small area larger than the first small area are collected (103C). The feature quantity of the second small region in the input layer neurons-menu neural network,
A second step of repeatedly inputting over the entire area while replacing the small area, calculating the distance to the reference pattern, and setting a threshold value so that the output layer neuron of the corresponding denomination and direction fires; The entire banknote to be learned is divided into the first small areas, and all the feature amounts extracted for each first small area (basic unit) are simultaneously input to the neural network (103b). A third step of calculating the distance from the reference pattern and setting a threshold value so that the output layer neuron of the corresponding denomination and direction fires, and an associative neuron table in which similar candidate neurons are registered in advance A fourth step of providing, and each feature amount of the unknown bill extracted for each first small area (basic unit) of the unknown bill into a second small area larger than the first small area. Collection (103C), the feature amount of the second small area is repeatedly input to the input layer neuron of the neural network over the entire area, replacing the small area, and the distance from the reference pattern is calculated. A fifth step of calculating the denomination and direction of the banknote to be performed, and performing first pattern recognition for deactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value;
The entire unknown banknote is divided into second predetermined small areas, and each small area is collectively processed, and the obtained feature amount is output to the output layer neuron by the first pattern recognition means. In the neural network including activated neurons, all neurons registered as candidate neurons in the associative neuron table are simultaneously input to the activated neural network, and the distance from the reference pattern is determined. The output layer neurons whose distance is within a predetermined threshold are all denominations of banknotes,
The sixth step of recognizing a second pattern that can be selected as a direction candidate is compared with the calculation results of the fifth and sixth steps, and when the same candidate is extracted, the money of the banknote of the same candidate is extracted. And a seventh step of outputting a seed and a direction.

[0029] Further, the bill identification method according to claim 8 is characterized in that:
A banknote identification method in which a feature amount of a banknote to be identified is input and a reference pattern is learned, and thereafter, an input unknown banknote is pattern-recognized from the reference pattern. A denomination and a direction to be fired in an output layer neuron are set using a neural network based on the method, and the input layer neuron of the neural network senses the bill to be identified and performs a predetermined first small area A first step of inputting all the feature amounts extracted in (basic unit) and learning and storing a reference pattern in pattern matching as a coupling coefficient; and a first step for each of the first small areas of the banknote to be learned (basic unit). ), The characteristic amounts of the banknotes to be learned extracted in (2) are aggregated in a second small area larger than the first small area (103C). Serial characteristic quantity of the second small region in the input layer neurons of the neural network, and enter repeatedly over the entire area instead of the small regions,
Calculating a distance from the reference pattern and setting a threshold value so that the output layer neuron of the corresponding denomination and direction is fired;
And the whole of the banknotes to be learned
, And simultaneously input all feature amounts extracted for each first small area (basic unit) to the neural network (103b), calculate the distance from the reference pattern, and calculate A third step of setting a threshold value so that an output layer neuron of a seed and a direction fires, a fourth step of providing an associative neuron table for registering previously similar candidate neurons, and the unknown banknote. The characteristic amounts of the unknown banknotes extracted for each of the first small areas (basic units) are aggregated into a second small area larger than the first small area (103C), and In the input layer neuron, the second
The feature amount of the small area is repeatedly input over the entire area while changing the small area, and the distance from the reference pattern is calculated,
A fifth step of calculating a denomination and a direction of a candidate banknote and performing a first pattern recognition for deactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value; The entire banknote was divided into the second sub-regions, and the respective sub-regions were collectively processed, and the obtained feature amounts were used to deactivate the output layer neurons by the first pattern recognition means. The neural network including neurons, and all the neurons registered in the candidate neurons of the associative neuron table are simultaneously input to the activated neural network, and the distance from the reference pattern is calculated. The output-layer neurons whose distances are within the predetermined threshold value are all in the sixth step of recognizing the second pattern selectable as a candidate for the denomination and direction of the banknote. A region where each feature amount of the banknote to be extracted extracted for each first subregion (basic unit) of the unknown banknote is larger than the first subregion and is a candidate denomination Are collected into a third small area having distinguishable features (103a), and the feature amount of the third small area is input to the input layer neuron of the neural network, and the distance from the reference pattern is calculated. A seventh step of performing third pattern recognition using an output layer neuron closest to the reference pattern as a classification result;
An eighth step of comparing the calculation results of the fifth, sixth, and seventh steps and outputting the denomination and direction of the same candidate banknote when the same candidate is extracted. And

[0030]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The details of the present invention will be described below with reference to the drawings.

FIG. 2 shows a feature of the present invention in which a feature amount such as an image of a banknote or time-series data is input using the above-mentioned competitive neural network and a similarity with a pre-registered reference pattern is calculated. 1 shows an example of a configuration of a pattern recognition unit, in which a determination unit 1 is connected to a competitive layer neuron of a neural network. That is, the feature quantity such as an image or time-series data extracted from an input bill is determined by the output of a reflection / transmission type optical sensor using a predetermined number of different wavelengths or the magnetic sensor (single or plural). The output is obtained by arithmetic processing, and the feature amount of the sensor output is clustered by a competitive neural network.

By the way, conventionally, when extracting a characteristic amount from an input bill, the output from each sensor is subjected to a mosaic processing shown in FIGS. To reduce it, normalization was performed. In other words, as shown in FIG. 5, the outputs of the sensors 2a to 2n are sequentially processed in the transport direction, mosaic processing is performed for each small area divided in the transport direction, and further, over the entire area of the input bill. After the end of the input scanning, normalization is performed to reduce the above-described offset and gain variations with respect to the input sensor data. For this reason, the above-mentioned mosaic processing and the normalization processing for reducing the variation of the offset and gain with respect to the input sensor data cannot be executed until the scanning of the entire area for the input bill is completed. And the pre-treatment time is prolonged.

In the conventional processing for each small area divided in the transport direction, no attempt has been made to set a small area for feature extraction by overlapping each divided small area. For this reason, there has been a problem that a feature extraction process is not stable because a partial feature can be extracted or cannot be extracted well.

Thus, data is classified according to the firing neuron. A data set firing in the competitive layer neuron i is defined as one population Si. For example, when 100 bill data are input, 40 bills fire on neuron 1,
Assuming that 60 neurons fired on the neuron 2, it is classified into a set S1 composed of 40 data and a set S2 composed of 60 data. This is called clustering.

The determination unit 1 finds the minimum value dc from the distance di (1 to M), compares the minimum value dc with the threshold value θc of the neuron c, fires the neuron c if dc ≦ θc, and dc> θc
If so, the neuron c is not fired. However, the threshold θc
Is given assuming that the average value of the minimum distance in the neuron c is mc and the standard deviation is σc with respect to the learning data calculated from the feature amount extracted from the normal banknote, and θc = mc + kc · σc. The competition layer has one or more neurons for one category (denomination, direction: four bill conveyance directions are represented by A to D). For example, in the determination unit having four competitive layer neurons in the D direction of US $ 5 banknote, a threshold value is set for each neuron in the D direction of $ 5.

FIG. 3 shows an example of an overall operation example of the learning process of the neural network according to the present invention. First, a feature extracted from a normal banknote without forgery or the like is used as learning data. Data preprocessing is performed on these data (step S10), initialization of the coupling coefficient Wij is performed (step S20), and LVQ learning is performed (step S30). In the initialization of the coupling coefficient Wij, the average value of the input data corresponding to each neuron is used. For example ＄
The average value of the data of ＄ 1A is set as the initial value for the neuron of 1A. A neural network is constructed by learning the coupling coefficient Wij, a threshold is created for the learning data (step S40), and non-firing neurons are deleted (step S50). Next, the learning data is evaluated (step S60). If the result of the evaluation indicates that the recognition rate and the reliability are low, it is necessary to add a neuron. Therefore, the neuron addition is checked (step S70). It is determined whether or not there is not (step S80), and if there is no addition, the threshold is extended (step S90). Also,
If a neuron is added in step S80, the process returns to step S30. The details of each part of the above flow will be described below.

FIG. 4 shows a processing flow of the data pre-processing (step S10). As shown in FIG. Mosaic processing is performed on the bill data (step S11). In this mosaic processing, 2
The sensor output data of 56 pixels is converted into input data of 32 pixels.

Specifically, the data of 256 pixels is first divided into eight pixels, and the average value of each pixel value in the area is calculated. As a result, each of the calculated average values is set as a pixel value 3
Two pieces of data are obtained as the mosaiced input data. That is, the data collected by the sensors 2a to 2n is, for example, time-series data as shown in FIG. 6A, and when this time-series data is graphed by an average value in units of eight pixels, it is shown in FIG. Become like By this mosaic processing, it is possible to expect a reduction in identification calculation time due to noise removal and a reduction in the number of data, and it is possible to reduce errors in edge extraction. When the mosaic processing is performed on an image (area), the result is as shown in FIGS. In this example, the mosaic processing is performed with an average value of 3 × 4 pixels.

The input image mosaiced as described above is subjected to normalization processing for eliminating offset and gain variations. Xi the mosaiced data
(I = 1,..., N), the conversion table data Xi
(I = 1,..., N) becomes the following equation 3. Here, A is a gain constant, and B is an offset constant.

[0040]

(Equation 3) In Equation 3, x overbar represents the average of data xi, and Sx represents the sum (area) of the absolute value of the difference between the average x overbar and data xi. The first term on the right side represents the variation of the data xi from the average as a percentage of the overall variation.

FIG. 8 shows 32 pieces of mosaiced data (0 to 0).
31) shows the sum (area) of the average x overbar and the absolute value. For this reason, a value that is invariant with respect to the data gain variation and offset variation is obtained. The offset constant B is a constant for adjusting the offset of the converted data. Generally, the constant B is set as the median of the dynamic range.

Hereinafter, the fact that the normalized data is invariant with respect to the gain fluctuation and the offset fluctuation will be described with reference to mathematical expressions. First, change data

Yi = axi + b (i = 1,..., N) a is a gain variation (> 0) and b is an offset variation.

Then, the average is

(Equation 5) And the total sum Sy of the absolute values of the differences between the average y overbar and the data yi is as shown in the following Expression 6.

[0044]

(Equation 6) Here, the normalized data Yi (i = 1,..., N)
Becomes the following Expression 7.

[0045]

(Equation 7) By performing such normalization, fluctuations in gain and offset can be removed.

Using the preprocessed data, learning of the coupling coefficient of each neuron and pattern recognition by a neural network are performed. In order to recognize, identify, and eliminate, in the feature extraction of the input bill, in the transport direction, as shown in FIG.
Feature extraction was performed for each small area.

Specifically, the whole is compressed in the transport direction to 30 pixels, which is the first small area (basic unit), and this is reduced to 10 pixels.
It is divided into a second small area that overlaps in pixel units.

Further, after the input scanning is completed over the entire area of the input bill, a normalization process for reducing the offset and gain variations with respect to the input sensor data is performed. 23, the normalization process cannot be executed (FIG. 23A), and the input data correction process takes time and the pre-processing time becomes long. As shown in FIG.
At the time when the input scanning for the correction 1 area is completed.
A first normalization process for reducing the variation of the offset and the gain with respect to the input sensor data is performed,
Utilizing this result, the pattern recognition process of the front part is executed, and at the time when the scanning of the entire area with respect to the input bill is completed, a second offset for reducing the offset and the gain variation with respect to the input sensor data is performed. Is performed, and a pattern recognition process at the rear portion is executed using the result.

A neural network learning and pattern recognition process using the feature data preprocessed for normalization of the sensor data sequentially input in a time-division manner will be described below.

First, in the learning process of the neural network as shown in FIG. 3, an initial value of the coupling coefficient is set (step S20). At the beginning of learning, one neuron per category is set in the competitive layer, and the initial value of the coupling coefficient of each neuron is the average value of data belonging to the category.

Next, the algorithm of the LVQ method will be described. First, a conceptual diagram of the LVQ method is shown in FIG. As shown in FIG. 20, feature amount data x extracted from an unknown input bill
If (t) and the firing neuron c are in the same category, the coupling coefficient wc (t) of the neuron c is calculated as input data x
(T), and if the input data x (t) and the firing neuron c are in different categories, the coupling coefficient wc (t) of the neuron c is moved away from the input data x (t).
Thus, the coupling coefficient is learned in the LVQ method.

The LVQ method includes LVQ1, LVQ2, LVQ
3. There are algorithms such as OLVQ1.
Shown in

[0053]

(Equation 8)

(Equation 9) Here, t is the number of times of learning, αc (t) represents the learning rate of each neuron, and s (t) = 1 (when correctly recognized) or s (t) = − 1 (when incorrectly recognized). . Number 8
Now we update the coupling coefficient of the neuron that fired during learning,
Equation 9 indicates that the other neurons are not updated.

After the initialization of the coupling coefficient in FIG. 3 (step S20) and the LVQ learning (step S30), a threshold is created (step S40), the details of which are as shown in the flowchart of FIG.

That is, first, the learning data number l (l = 1 to 1)
L) (steps S41 to S41)
In S43), learning data is clustered for each firing neuron (step S42) to obtain M data sets. Then, the neuron number i (i = 1 to M)
In the loop for creating thresholds (steps S44 to S49), the distance distribution of the set of clustered learning data is estimated, and the average value mi, the standard deviation σi, and the maximum value d
max is obtained (step S45), and a judgment threshold θi is created (step S46). Then, it is determined whether the threshold value θi is smaller than the maximum value dmax (step S4).
7) If the value is smaller than the maximum value dmax, a value obtained by adding a predetermined number α to the maximum value dmax is set as the threshold value θi (step S4).
8). This is repeated for M neuron numbers (step S49).

As shown in FIG. 10, the threshold θi as the range of self-recognition of the competitive layer neuron i is set to θi = mi + kiσi from the estimated distribution. Thus, the Euclidean distance d between the input data and the competitive layer neuron i
If i is di ≦ θi, the sample is recognized as belonging to the population Si. Here, ki can be set to any value as the range of self-recognition. In the case of ki = 4.5, the probability that the neuron i fires (recognizes correctly) with respect to the data of the population Si is expressed by a standard normal distribution table. To the power of 1-3.4 × 10−6, and when ki = 6.5, the power is the power of 1-4.0 × 10−11. Depending on the learning data, the threshold value set in this way may be exceeded. However, if such data needs to be recognized, it is sufficient to set it to θio as shown in FIG.

After the threshold value is created as described above, the non-firing neurons are deleted (step S50).
The details will be described with reference to the flowchart of FIG. First, in a loop for deleting a neuron with a neuron number i (i = 1 to M) (steps S51 to S54), it is determined whether or not the neuron i has fired (step S52), and if not, the neuron i is deleted. (Step S53).

Thereafter, the learning data is evaluated (step S60), and the details will be described with reference to the flowchart of FIG.

First, a learning data identification loop based on the learning data number l (l = 1 to L) (steps S61 to S64)
In step S6, identification of learning data is determined (step S6).
2) The learning data is clustered for each firing neuron (step S63). Then, in a reliability evaluation loop based on the neuron number i (i = 1 to M) (steps S65 to S68), the distance distribution of the set of clustered learning data is estimated (step S66).
A reliability evaluation is performed (step S67). This is performed over all neuron numbers.

When the evaluation of the learning data is completed, a neuron addition check is next performed (step S70). Details of the check will be described with reference to the flowchart of FIG.

First, a loop for checking neuron addition by the neuron number i (i = 1 to M) (steps S71 to S71)
In S76), it is determined whether or not there is a misrecognition (step S72). If there is a misrecognition, an average value of the misrecognition data is obtained for each category, and a neuron using the average value as a coupling coefficient is added (step S76). S73). Thereafter, it is determined whether or not the neuron i satisfies the predetermined reliability (step S74), and a copy of the neuron i or a neuron whose data set to be evaluated for reliability fires is added (step S74).
75).

In the reliability evaluation, as shown in FIG. 15, the output value distribution of neuron j (j ≠ i) with respect to data set Si firing on neuron i is considered. Assuming that this output value distribution is a normal distribution ND as shown in FIG. 15, the probability that the normal distribution ND becomes smaller than the threshold value θj is defined as reliability. This represents the area of the black portion BA in FIG.
This means the probability that data to be fired on neuron i will fire on neuron j.

After setting the threshold value, the discrimination judgment is performed on the learning data, and if there is any erroneously recognized data, a neuron for the data is added. The coupling coefficient of the neuron is an average value for each category of the misrecognition data, such as the banknote denomination direction. This makes it possible to improve the recognition rate. In addition, a reliability evaluation is performed on the learning data, and a duplicate of the neuron or the evaluation target neuron (a neuron firing the evaluation target data) is added to a neuron for which a predetermined reliability cannot be obtained. This allows
Since a new neuron is provided for erroneous recognition data determination, the dispersion of the output value distribution of the neuron is reduced, and the reliability is improved.

FIG. 16 shows the addition of a neuron to the misrecognition data, and shows how to add a neuron to the neuron i with the average of the input values of the misrecognized bill data to the input layer as the coupling coefficient. . FIG. 17 shows the addition of a neuron when the default reliability is not satisfied, and a copy of the neuron i or the evaluation target neuron that does not satisfy the default reliability is added.

Next, a method of dividing a neuron when the reliability of the neuron i for different denomination data Y is low will be described. Here, neuron i is denomination data X
Is assumed to have been learned so as to fire when input is made, and the neuron j has been learned to be fired when the denomination data Y is input.

(Division method 1) Low reliability neuron i
Add a duplicate.

(Division Method 2) Output value distribution Di (X) when denomination data X is input to neuron i, and output value distribution Di (D) when denomination data Y is input to neuron i Y), and the standard deviations σi (X) and σi (Y) of the respective distributions are calculated. And
As shown in FIG. 24B, the standard deviations σi (X) and σi (X)
In (Y), when σi (Y) is larger, it is considered that the variation of the denomination data Y is large, so a copy of the neuron j showing a large standard deviation is added.

(Division method 3) Output value distribution Di (X) when denomination data X is input to neuron i, and output value distribution Dj (D) when denomination data Y is input to neuron j Y), and the standard deviations σi (X) and σj (Y) of the respective distributions are calculated. And
As shown in FIGS. 24C and 24D, the standard deviation σi
(X) and σj (Y), when σj (Y) is larger, it is considered that the variation of the denomination data Y is large, so a copy of the neuron j showing a large standard deviation is added.

Check of neurons in FIG. 3 (step S
In step 70), if there is no additional neuron thereafter (step S80), the threshold is extended (step S9).
0), and its operation will be described with reference to the flowchart of FIG.

The denomination data X may have a variation that cannot be predicted by actual data, and may exceed a statistically set threshold. The threshold may be extended to provide such data with generalization ability. Therefore, based on the results of reliability evaluation,
If the range is extended within a range in which the default reliability can be obtained, the recognition rate can be improved while the default reliability is secured. In the specific example shown in FIG. 19, the threshold θi is set to θi = mi +
After setting to 4.5σi, reliability evaluation is performed. As a result, the evaluation target neuron calculates a boundary value that satisfies the predetermined reliability, and resets the boundary value as an extended threshold. At this time, in order to prevent excessive expansion, mi + 6.5σ
i is set as a limit value of the threshold expansion, and if the boundary value is larger than the limit value, the limit value is set as the expanded threshold.

First, in a threshold expansion loop (steps S91 to S96) based on the neuron number i (i = 1 to M), a boundary value θB of a threshold satisfying a predetermined reliability is obtained (step S92), and the threshold expansion limit is set. The value θL is obtained (step S93). Then, the boundary value θB becomes the limit value θL.
It is determined whether or not the threshold value is smaller (step S94), and if smaller, the threshold value θ is extended to the boundary value θB (step S95),
Otherwise, the threshold value θ is extended to the limit value θL (step S97).

Thus, learning of the neural network is completed. Next, a more specific configuration and operation of the learning and identification / judgment processing for the input bill 4 will be described in detail.

First, FIG. 25 is an example of an overall block diagram of the bill validating apparatus 100 of the present invention. An input bill 4 is transferred to a sensor means 2 (including an optical sensor and a magnetic sensor) by a transport means (not shown). , And the output is sequentially sent to the sensor data preprocessing correction unit 10.
1 is input.

Next, the sensor data pre-processing correction section 101 performs the compression processing shown in FIGS. 4 to 7 on the sequentially input sensor data, and further reduces the offset and gain variations with respect to the input data. Normalization as shown in FIG. 8 is performed.

That is, FIG. 4 shows a processing flow of the data pre-processing, and as shown in FIG. The data is subjected to compression processing, and 256
The sensor output data of the pixel is converted into input data composed of 32 pixels.

As a result, 32 data having the calculated average value as a pixel value are obtained as the compressed input data, and the data collected by the sensors 2a to 2n is, for example, as shown in FIG. The time-series data is as shown, and when this time-series data is graphed (moving average graph) with an average value for each of eight pixels, the graph becomes as shown in FIG. By this compression processing, it is possible to expect a reduction in discrimination calculation time accompanying noise removal and a reduction in the number of data, and it is possible to reduce errors in edge extraction.

When the compression processing is performed on an image (area), the result is as shown in FIGS. 7A and 7B. Next, the offset and gain of the compressed input image are eliminated. Perform normalization processing. Sx in FIG.
Average x for 32 mosaiced data (0 to 31)
The overbar and the sum (area) of the absolute values are shown. Using the data thus preprocessed as a feature amount, learning of the coupling coefficient of each neuron and pattern recognition by a neural network are performed.

That is, the characteristic amounts of the bills to be identified are generated by the sensor data preprocessing correction unit 101 from the sensors 2 i (i = a to n), and these characteristic amounts are input to the pattern recognition unit 110. Then, a reference pattern is learned for each denomination direction, and the obtained coupling coefficient of each neuron is stored in the reference pattern storage means 116d as a reference pattern.

The pattern recognition means 110 includes a sensor data pre-processing correction section 101, a data buffer 102, an input window pattern storage section 103, a neural network 112,
It comprises a threshold reference memory 113, a mode-specific determination unit 114, a reference pattern storage unit 116, a reference pattern learning and updating / setting unit 118, and a control unit 119.
4 are integrated by the determination result integration unit 120,
The final processing is performed and output to the outside.
Here, the mode-specific determination unit 114 has a minimum value determination mode 114a, a minimum value determination and threshold value determination mode 114b, and a threshold value determination mode 114c. The input window pattern storage unit 103 stores an input window pattern 103a, an input window pattern 103b, and an input window pattern 103c.

The input window pattern 103c uses a region composed of a plurality of basic unit small regions as windows, and selects one of the P windows to input normalized data to the input layer. The size of the window at this time is the number of neurons in the input layer, for example, ten.

The input window pattern 103b is an input window pattern 103c in which all windows are open. That is, all inputs are simultaneously input to neurons in the input layer. The input window pattern 103a is a window provided so that a feature that can be accurately identified as a specific one denomination from a plurality of candidate denominations is extracted, and is a window of an appropriately provided size. Is fixed, but is usually provided for each output neuron.

In the present invention, in order to accurately recognize, discriminate, and eliminate partially unique forged / falsified banknotes and the like, the feature amount extraction of the input banknote is performed in the transport direction in FIG. As shown in B), the input data is divided into small areas 1 to 5 while overlapping, and a feature amount is extracted for each small area.
The whole is compressed to 30 pixels in the transport direction, and is divided into overlapping small areas in units of 10 pixels. here,
Each of the compressed 30 pixels becomes a basic unit small area (first small area) shown in FIG.

Further, after the input scanning is completed over the entire area of the input bill, the sensor data pre-processing correction unit 101 performs normalization for reducing the offset and gain variations with respect to the input sensor data described above. Until the scanning of the entire area for the input bill is completed, the normalization processing cannot be executed (FIG. 23A), and the correction processing of the input data takes time and the preprocessing time becomes longer. As shown in FIG. 23 (C), the entire area of the input bill is divided into an area for correction 1 and an area for correction 2 while overlapping, and when input scanning for the area for correction 1 is completed, First normalization is performed to reduce the offset and gain variations with respect to the input sensor data described above, and the first reference pattern learning and reference pattern Update processing, and pattern recognition processing using the first reference pattern execution.

More specifically, the structure and operation of the pattern learning means 110 will be described.
Is characterized in that the entire input bill is divided and processed into second predetermined small areas each including an overlapped area. In learning a reference pattern, the pattern recognition unit 110 controls the control unit 119. , The neural network 112 is initialized, the reference pattern learning,
The updating and setting unit 118 is switched to the learning mode, a predetermined number of normal banknotes are input, and sensor outputs input from each banknote are sequentially stored in the sensor data preprocessing correction unit 101, and the entire area of each banknote is When the scanning and input of are completed, the pre-processing and correction processing are executed by the sensor data pre-processing correction unit 101, and are temporarily stored in the data buffer 102. After that, the feature amounts filtered by the input window pattern 103b are collectively stored in the neural network 11.
2 and learning is performed. Then, a coupling coefficient of each neuron is obtained, and this is stored in the storage unit 116d as a reference pattern. At this time, an output when a predetermined number of normal banknotes used for learning are input is obtained, a threshold value of each neuron is determined from this threshold value, and this threshold value is stored in the storage means 116b.

The neural network 11 uses the feature amount filtered by the input window pattern 103.
As shown in FIG. 29A, it is also possible to selectively input a feature amount obtained for each of the divided predetermined small regions to the input layer neurons 2 according to the positional dependence,
For example, small areas 1, 4, 7, 2, and
Each of the feature amounts 5 and 8 is subjected to a neuro operation without being input to the input layer neurons 3, 6 and 9, and the small regions 2, 5, 8,
When the neuro operation is executed without inputting the neurons 1, 4, and 7 to the feature amounts of 3, 6, and 9, it is effective that
It was found by experiment.

Next, the mode-specific determination unit 114 is switched to the minimum value determination and threshold determination mode 114b, and the same predetermined normal banknotes are sequentially input to update the addition and deletion of each neuron of the neural network 112. Perform processing. Thereafter, reference pattern learning, updating, and setting unit 118
Is switched to the learning mode, learning is performed again, and the reference pattern and the threshold are updated.

Finally, the input window pattern 103c is applied to obtain an output value when a predetermined number of normal banknotes used for learning are input, and the threshold value of each neuron is determined from the output value. It is stored in the storage means 116c.

Next, the identification processing will be described.

In the threshold determination mode of the first pattern recognition process of the pattern recognition means 110, the reference pattern learning, update and setting section 118 is switched to the setting mode under the command of the control section 119, and the neural network 112 For each neuron, a coupling coefficient as a reference pattern is read from the reference pattern storage means 116d and set. Further, a threshold value is read from the threshold value storage means 116c,
It is set in the threshold reference memory 113. Then, after switching the mode-specific determination unit 114 to the threshold determination mode 114c, the unknown input banknotes are sequentially scanned and read by the sensor, and the feature amount processed by the sensor data preprocessing correction unit 101 is stored in the data buffer. The input signal is stored in the input window pattern 102, filtered by the input window pattern 103c, input to the neural network 112, and the distance between each neuron and the coupling coefficient is calculated.
Then, by comparing with the threshold value stored in the threshold value reference memory 113 and firing a neuron at a distance lower than the threshold value, the denomination and direction of the unknown input bill are determined and identified.

In the minimum value judgment and threshold judgment modes of the second pattern recognition processing of the pattern recognition means 110, the reference pattern learning, update and setting section 118 is switched to the setting mode under the command of the control section 119, and For each neuron of the neural network 112, a coupling coefficient as a reference pattern is read from the reference pattern storage means 116d and set. Further, the threshold value is read from the threshold value storage means 116 b and set in the threshold value reference memory 113. Thereafter, the mode-specific determination unit 114 is set to the threshold determination mode 114b.
After switching to, the unknown input banknotes are sequentially scanned by the sensor, the feature amount processed by the sensor data preprocessing correction unit 101 is stored in the data buffer 102, and the input window pattern 103b performs a filtering process. The distance is input to the neural network 112 and the distance between each neuron and the coupling coefficient is obtained.
As a result, the denomination and the direction of the unknown input bill are determined and identified by firing the neuron whose distance is lower than the threshold value and whose distance is the smallest.

In the minimum value determination mode of the third pattern recognition process of the pattern recognition means 110, the reference pattern learning, update and setting section 118 is switched to the setting mode under the command of the control section 119, and the neural network 112 is switched to the setting mode.
, The reference pattern storage means 116d
, A coupling coefficient as a reference pattern is read and set. Then, after switching the mode-specific determination unit 114 to the threshold determination mode 114a, unknown input bills are sequentially scanned by the sensor, and the sensor data pre-processing correction unit 101
Are stored in the data buffer 102, filtered by the input window pattern 103a, input to the neural network 112, and the distance between each neuron and the coupling coefficient is determined. In the mode 114a, the denomination and direction of the unknown input bill are determined and identified by firing the neuron whose distance is the minimum. FIG. 26 shows a state in which the above-described first pattern recognition, second pattern recognition, and third pattern recognition are performed in a cascade. At the time of actual bill discrimination, the calculation result of the first pattern recognition process, the calculation result of the second pattern recognition process, and the calculation result of the third pattern recognition process are combined with the determination result integrating means 120.
When the operation results of the respective pattern recognition units 110 are all the same or when the same result is obtained by majority decision, the same operation result is
It is output to the outside as the final result of 0.

The operation of the bill validator 100 will be described with reference to FIG.
7, the learning and updating of each reference pattern are performed by the pattern recognition unit 110.
It is assumed that all the processes have been completed, and a pattern recognition process is performed as follows. In the first embodiment, first, FIGS. 27A, 27B, and 27C are performed based on the first pattern recognition processing of the pattern recognition unit 110. FIG.
In FIG. 27A, the filtering process for the region 1 is performed using the input window pattern 103c, and the threshold value corresponding to the region 1 is stored in the threshold value reference memory 113 from the threshold value storage unit 116c by selecting the threshold value determination mode 114c. Is stored. Then, the feature amount of the area 1 is input to the neural network 112. The distance between the feature input and the coupling coefficient of each neuron is obtained and compared with a threshold stored in the threshold reference memory 113. By firing a neuron whose distance is shorter than the threshold, the denomination of the unknown input banknote, Judge and identify the direction. At this time, only one neuron does not necessarily fire one neuron, and even if one neuron fires, it is a result of some feature amounts, and the denomination and direction are determined. Is not enough. Thus, the one or more neurons fired here are the final denomination,
The threshold reference memory value is set to 0 for other non-firing neurons, which are left as direction candidates, so that the calculation result is not reflected in the result.

Next, in FIG. 27 (B), by moving the window in the input window pattern 103c, a filtering process is performed on the area 2 which is another area.
The threshold value corresponding to the area 2 is stored in the threshold value reference memory 113 from the threshold value storage unit 116c. Here, the value of the threshold reference memory corresponding to the neuron that has not fired in the determination for the area 1 is not updated, and is kept at 0. At this time, the feature amount of the area 1 is input to the neural network 112, the distance between this input and the coupling coefficient of each neuron is calculated, and this is compared with the threshold value stored in the threshold value reference memory 113. By firing the neuron of the distance, the denomination and direction of the unknown input bill are determined,
Identify.

As in the case of the area 1, the 1
One or more neurons are left as candidates for the denomination and direction to be finally determined, and for the other neurons that have not fired, the threshold reference memory is set to 0 so that the calculation result is not reflected. Further, the input window pattern 10
In 3c, the judgment is performed while sequentially moving the window,
The candidates for the directions are narrowed down, and finally, as shown in FIG.
The determination result is stored in the determination integration unit 120 as the determination result of the embodiment.

Next, a second embodiment, FIG. 28 (F), will be implemented. In FIG. 28F, filtering processing is performed on the entire bill using the input window pattern 103b, that is, all the feature amounts of the bill are input, and the threshold determination mode 1 is used.
At 14b, the threshold values corresponding to all the areas are stored in the threshold value storage means 116.
b is stored in the threshold reference memory 113. At this time,
The feature amount of the whole bill is input to the neural network 112, but the one in which 0 is stored in the threshold reference memory in FIG. 27C is not updated. Then, the distance between this input and the coupling coefficient of each neuron is obtained and compared with the threshold value stored in the threshold value reference memory 113. By firing the neuron whose distance is shorter than the threshold value, the denomination of the unknown input bill is calculated. Judge and identify direction. Then, when the determination result matches the value stored in the determination integration unit 120, the determination integration unit 120 holds the stored value, and when they do not match, the determination integration unit 120 determines that the determination result is rejected. Output.

As a modification of the second embodiment, FIG.
(G) can also be performed. That is, the neurons narrowed down in (C) of FIG. 27 are released, and all of them can be used as candidates without inheriting the fact that 0 is stored in the threshold reference memory. Using the input window pattern 103b, a filtering process is performed on the entire banknote, that is, all the feature amounts of the banknote are input, and the threshold determination mode 11
4b, the threshold values corresponding to all the areas are stored in the threshold value storage means 116b.
It is stored in the threshold reference memory 113. At this time, the feature amount of the entire bill is input to the neural network 112, the distance between this input and the coupling coefficient of each neuron is calculated, and the distance is compared with the threshold stored in the threshold reference memory 113. By firing, the denomination and direction of the unknown input bill are determined and identified. Then, when the determination result matches the value stored in the determination integration unit 120, the determination integration unit 120 holds the stored value, and when they do not match, the determination integration unit 120 determines that the determination result is rejected. Output.

Further, when there are different denominations whose coupling coefficients are similar to the denominations indicated by the values stored in the judgment integration unit 120, the third embodiment and FIG. I do. In FIG. 28, the input window pattern 103a is provided for filtering the input so as to extract a feature that can accurately identify the denomination to be implemented. Then, the minimum value determination mode 114a is adopted, and 1 is stored in the threshold reference memory for the neuron corresponding to the denomination to be implemented, and 0 is stored in the other threshold reference memories. At this time, the input window pattern 103a
Is input to the neural network 112, and the distance between the input and the coupling coefficient of each neuron corresponding to the denomination to be implemented is determined.
By firing the smallest neuron in each distance, the execution target denomination is determined and identified as a unique denomination. If the determination result matches the value stored in the determination integration unit 120, it is output as a formal determination result, and if not, the determination integration unit 120 outputs the determination result as a reject.

Subsequently, a similar pattern neuron table as shown in FIG. 28E is created in advance for each neuron of the neural network 112, and a plurality of candidate neurons remaining as a result of the first embodiment are obtained. The denomination, direction neurons, and neurons registered in the similar pattern neuron table are activated by setting a threshold value in the threshold reference memory 113, and the second embodiment is executed for these neurons. I do. In this manner, even if a normal banknote that has been deformed or deteriorated due to damage or the like is determined as a non-candidate in the region to which the deformed or deteriorated portion belongs according to the first embodiment, And the input banknote can be determined as a valid denomination and direction. In this manner, partially unique forged / falsified banknotes can be accurately pattern-recognized, identified, and eliminated. In addition, it is possible to improve the identification rate of pattern recognition of a normal banknote that has been damaged, and to accurately recognize a pattern as a damaged banknote, and identify and eliminate the banknote.

[0099]

As described above, according to the present invention, first, a small area portion of sensor data such as an image is input to perform pattern recognition, and then the whole is collectively input to perform pattern recognition. By performing the determination processing with limiting the neurons used for the pattern recognition and determination at the stage, even if the part is a unique counterfeit / falsified banknote, the pattern can be accurately recognized and eliminated as an abnormal banknote. In addition, it is possible to improve the identification rate of pattern recognition of a normal banknote that has been damaged, and to accurately recognize and eliminate a damaged banknote in pattern.

Further, by performing a determination process including a step of feeding back a determination result when image data of a small region over multiple stages by a neural network is input and gradually reducing the number of candidate neurons, By inputting image data of the entire area to a neuron to which a similar candidate that is likely to be erroneous with respect to the initially determined denomination candidate is added, a discrimination result is obtained by inputting image data of the entire region. Can be improved.

[Brief description of the drawings]

FIG. 1 is a diagram showing a basic structure of a competitive neural network used in the present invention.

FIG. 2 is a block diagram illustrating a configuration example of the present invention.

FIG. 3 is a flowchart showing an overall operation example of the present invention.

FIG. 4 is a flowchart illustrating an operation example of data preprocessing.

FIG. 5 is a diagram showing a state of data collection of a bill.

FIG. 6 is a diagram for explaining a mosaic processing (time-series data).

FIG. 7 is a diagram for explaining a mosaic processing (image).

FIG. 8 is a diagram illustrating a normalization process.

FIG. 9 is a flowchart illustrating an example of a process for creating a threshold.

FIG. 10 is a diagram for explaining setting of a threshold.

FIG. 11 is a diagram for explaining adjustment of a threshold.

FIG. 12 is a flowchart illustrating a processing example of neuron deletion.

FIG. 13 is a flowchart illustrating an example of a process of evaluating learning data.

FIG. 14 is a flowchart illustrating an example of a neuron addition check process.

FIG. 15 is a diagram for explaining reliability evaluation of a competitive neural network.

FIG. 16 is a diagram for explaining the addition of neurons to erroneously recognized data.

FIG. 17 is a diagram for explaining addition of a neuron when the default reliability is not satisfied.

FIG. 18 is a flowchart illustrating a processing example of threshold value extension;

FIG. 19 is a diagram for explaining extension setting of a threshold.

FIG. 20 is a diagram for explaining the concept of the LVQ method.

FIG. 21 is a diagram illustrating an example of processing for identifying a partially unique bill.

FIG. 22 is a diagram illustrating a processing example of erroneously identifying a damaged banknote as a similar banknote.

FIG. 23 is a diagram illustrating an example of input data correction processing.

FIG. 24 is a diagram illustrating an example of neuron division.

FIG. 25 is an example of an overall block diagram of the bill validator.

FIG. 26 is a diagram illustrating an example of data processing of the identification device.

FIG. 27 is a diagram illustrating an example of processing of the identification device.

FIG. 28 is a diagram illustrating an example of processing of an associative neuron table.

FIG. 29 is an example showing a relationship between input layer neurons and feature amounts of divided small regions.

[Explanation of symbols]

 1, 114 determination unit 2, 3 sensor 4 banknote 102 input data preprocessing correction unit 110 pattern recognition unit 112 neural network 116 reference pattern storage unit 118 learning, updating, setting unit 120 determination result integration unit

 ────────────────────────────────────────────────── ─── Continued from the front page (72) Inventor Kunihiro 1-3-1 Shimoteno, Himeji-shi, Hyogo Glory Industries Co., Ltd. F-term (reference) 3E041 AA02 AA03 AA04 AA05 BA11 BB02 BB03 BC03 CA01 CA08 CA09 CB03 CB04 5L096 AA06 BA03 BA18 FA37 FA41 GA51 HA08 HA11 JA11 LA05

Claims (8)

[Claims] 1. A bill identifying method for inputting a feature amount of a bill to be identified, learning a reference pattern, and then recognizing the inputted unknown bill using the reference pattern. Performing a first step of collecting the characteristic amounts of the unknown banknotes extracted for each predetermined first small area into a second small area larger than the first small area; and A comparison is made between a reference pattern provided corresponding to each of the small areas and selectable as a denomination / direction candidate of the banknote and a feature amount of the first small area constituting the second small area. Pattern recognition 2
And the third step of determining candidates for the denomination and direction to which the unknown bill applies as a result of the pattern recognition, and excluding non-applicable denominations and directions from the candidates, and the second step And the third step are repeatedly executed over the entire area by replacing the second small area,
A fourth step of obtaining the result of the second step from among the directions and storing the denomination of the banknote finally left as a candidate for the unknown banknote, and the direction; A fifth step of simultaneously comparing the feature amount of the first predetermined small area with all of the reference patterns, performing pattern recognition, and storing banknote denomination and direction candidates obtained as candidates; If there is a common candidate for each of the banknote denomination and direction candidates stored in the step and the fifth step, the denomination and direction of this common candidate banknote are And a sixth step of outputting the denomination and direction to the outside. 2. An LVQ (Learning) is used for the pattern recognition.
The method for identifying bills according to claim 1, wherein a neural network is used by a Vector Quantization method. 3. A banknote identification method for inputting a feature amount of a banknote to be identified, learning a reference pattern, and thereafter performing pattern recognition of the input unknown banknote using the reference pattern. (Learning Vector Quantiz
ation) method, the denomination and the direction to be fired in the output layer neuron are set, the bill to be identified is sensed, and the whole bill is
In addition to dividing into predetermined first small regions, all the feature amounts extracted for each of the first small regions are input to the input layer neurons of the neural network, and a reference pattern in pattern matching is learned and stored as a coupling coefficient. A first step of causing each feature amount of the banknote to be learned extracted for each first subregion of the banknote to be identified to be collected in a second subregion larger than the first subregion. The input layer neurons of the neural network
The feature amount of the small area is repeatedly input over the entire area while changing the small area, and the distance from the reference pattern is calculated,
A second step of setting a threshold value so that an output layer neuron of a corresponding denomination and direction is fired, and simultaneously inputting the entire amount of the feature amount banknote extracted for each of the first small areas to the neural network. A third step of calculating a distance from the reference pattern and setting a threshold value so that the output layer neuron of the corresponding denomination and direction is fired; And the feature amounts of the unknown banknotes extracted for each of the first small areas are aggregated in the second small area, and the input layer neurons of the neural network are assigned to the second layer. The feature amount is repeatedly input over the entire area, replacing the small area, calculating the distance to the reference pattern, calculating the denomination and direction of the candidate banknote, and calculating the distance from the reference pattern. Predetermined threshold A fourth step of performing a first pattern recognition for inactivating the upper output layer neuron; and simultaneously extracting all the feature amounts extracted for each of the first small regions of the entire unknown bill into the neural network. A fifth step of inputting and performing a second pattern recognition for calculating the denomination and direction of a candidate banknote;
And a sixth step of outputting the denomination and direction of the banknote of the same candidate when the same candidate is extracted by comparing the calculation results of the above steps. 4. A banknote identification method for inputting a feature amount of a banknote to be identified and learning a reference pattern, and thereafter recognizing the input unknown banknote using the reference pattern. LVQ (Learning Vector Quantiz
ation) method, the denomination and the direction to be fired in the output layer neuron are set, the bill to be identified is sensed, and the whole bill is
The neural network is divided into predetermined first small areas, and all the feature amounts extracted for each of the first small areas are input to an input layer neuron of the neural network, and a reference pattern in pattern matching is learned and stored as a coupling coefficient. A first step of causing each feature amount of the bill to be identified extracted for each first small area of the bill to be identified to be collected in a second small area larger than the first small area. The feature amount of the second small area is repeatedly input to the input layer neuron of the neural network over the entire area while changing the small area, and the distance from the reference pattern is calculated. A second step of setting a threshold such that the output layer neurons of
Simultaneously inputting the entire amount of the bill of the feature amount extracted for each of the first small areas to the neural network, calculating the distance from the reference pattern, and firing the output layer neuron of the corresponding denomination and direction. Setting the threshold value as described above, and dividing the entirety of the unknown banknote into the first small areas, and extracting each feature amount of the unknown banknote extracted for each of the first small areas. Are gathered in the second sub-region, and the feature value of the second sub-region is repeatedly input to the input layer neurons of the neural network over the entire sub-region by changing the sub-region. A fourth step of calculating a distance, calculating a denomination and a direction of a candidate banknote, and performing a first pattern recognition for deactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value. When, Simultaneously inputting the feature amount of the entire unknown banknote extracted for each first small area to the neural network including neurons whose output layer neurons have been deactivated by the first pattern recognition means, The fifth step of performing the second pattern recognition for calculating the denomination and direction of the candidate banknotes is compared with the calculation results of the fourth and fifth steps, and when the same candidate is extracted, Denominations of banknotes of the same candidate,
And a sixth step of outputting a direction. 5. A banknote identification method in which a feature amount of a banknote to be identified is input and a reference pattern is learned, and then the input unknown banknote is pattern-recognized using the reference pattern. LVQ (Learning Vector Quantiz
A denomination and a direction to be fired in an output layer neuron are set by using a neural network based on the method of the present invention. A first step of inputting all the feature amounts extracted for each area and learning and storing a reference pattern in pattern matching as a coupling coefficient, and the learning extracted for each first small area of the banknote to be learned. The feature amounts of the bills to be collected are collected in a second small region larger than the first small region, and the feature amount of the second small region is changed to the input layer neuron of the neural network by replacing the small region. A second input for repeatedly inputting over the entire area, calculating a distance from the reference pattern, and setting a threshold value so that an output layer neuron of a corresponding denomination and direction fires. And the step,
The entire banknote to be learned is divided into the first small areas, and all the feature amounts extracted for each first small area are simultaneously input to the neural network, and a distance from the reference pattern is set. A third step of calculating and setting a threshold value so that the output layer neuron of the corresponding denomination and direction fires; and dividing the entirety of the unknown bill into the first small area, Gathering each characteristic amount of the unknown bill extracted for each small area in the second small area,
The feature amount of the second small area is repeatedly input to the input layer neuron of the neural network over the entire area while changing the small area, and the distance from the reference pattern is calculated. A fourth pattern for calculating a seed and a direction and performing a first pattern recognition for inactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value;
And a second pattern for simultaneously inputting the feature amounts of the entire unknown banknote extracted for each of the first small areas to the neural network and calculating the denomination and direction of the candidate banknotes A fifth step of recognizing, each of the feature amounts of the banknote to be learned extracted for each of the first small regions of the unknown banknote is a region larger than the first small region, and a candidate The denominations are collected in a third sub-region having a distinguishable characteristic, the feature amount of the third sub-region is input to the input layer neuron of the neural network, and the distance from the reference pattern is calculated. And comparing the sixth step of performing third pattern recognition with the output layer neuron closest to the reference pattern as the identification result with the calculation results of the fourth, fifth, and sixth steps, Same candidate is extracted And a seventh step of outputting the denomination and direction of the same candidate banknote when the banknote is identified. 6. A reference pattern is learned by inputting a feature amount of a bill to be identified. Thereafter, the inputted unknown bill is recognized from the reference pattern, and an LVQ (Learning Vector Quantization) is used for the pattern recognition. The input layer neurons of the neural network are selectively input into the input layer neurons of the divided small regions in accordance with the positional dependency, and the banknotes for which the neural operation is performed are used. In the identification device, a banknote to be recognized is processed by sensing a banknote and dividing the entire banknote into a first predetermined region to extract a feature amount, and processing a predetermined number of learning banknotes. Are sequentially input to the input layer neurons of the neural network with the feature amount of each small region obtained by the feature amount extraction means. A first learning unit for storing the reference pattern in a predetermined neuron; and a feature amount of the banknote to be identified extracted for each first small region by the feature amount extracting unit. A second learning step of collecting the reference patterns into predetermined neurons by simultaneously collecting feature amounts for all the banknote regions into the input layer neurons of the neural network and collecting the reference patterns into predetermined neurons; Means,
The feature amounts obtained by the feature amount extraction means from the unknown banknotes are collected in the second small area, sequentially input to the input layer neurons of the neural network, and the distance from the reference pattern is calculated, A first pattern recognition unit that calculates a denomination and a direction of a candidate banknote and that deactivates an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value; The feature amounts of the entire unknown banknote extracted for each of the small regions are simultaneously input to the neural network including neurons whose output layer neurons have been deactivated by the first pattern recognition means, and The output layer neurons in which the distance to the pattern is calculated and the distance is within a predetermined threshold value are all the second patterns that can be selected as banknote denomination and direction candidates. Recognizing means, and calculating result integrating means for comparing calculation results of the first and second pattern recognition means and outputting a denomination and a direction of a banknote of the same candidate when the same candidate is extracted. A bill discriminating device characterized by the following. 7. A banknote identification method in which a feature amount of a banknote to be identified is input and a reference pattern is learned, and then the input unknown banknote is pattern-recognized from the reference pattern. Vector Quantiz
A denomination and a direction to be fired in an output layer neuron are set by using a neural network based on the method of the present invention. A first step of inputting all the feature amounts extracted for each area and learning and storing a reference pattern in pattern matching as a coupling coefficient, and the learning extracted for each first small area of the banknote to be learned. The feature amounts of the bills to be collected are collected in a second small region larger than the first small region, and the feature amount of the second small region is changed to the input layer neuron of the neural network by replacing the small region. A second input for repeatedly inputting over the entire area, calculating a distance from the reference pattern, and setting a threshold value so that an output layer neuron of a corresponding denomination and direction fires. And the step,
The entire banknote to be learned is divided into the first small areas, and all the feature amounts extracted for each first small area are simultaneously input to the neural network, and a distance from the reference pattern is set. A third step of calculating and setting a threshold value so that the output layer neuron of the corresponding denomination and direction is fired; a fourth step of providing an associative neuron table for registering previously similar candidate neurons, respectively; ,
Each feature amount of the unknown bill extracted for each of the first small areas of the unknown bill is set to a second value larger than the first small area.
The feature amount of the second small area is repeatedly input to the input layer neurons of the neural network over the entire area while changing the small area, and the distance from the reference pattern is calculated. A fifth step of calculating a denomination and a direction of a candidate banknote and performing a first pattern recognition for deactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value; Is divided into second predetermined small areas, and each of the small areas is collectively processed, and the obtained feature amounts are used to deactivate the output layer neurons by the first pattern recognition means. The neural network including the activated neurons and the neurons registered in the candidate neurons of the associative neuron table are all activated. Simultaneously, the distance from the reference pattern is calculated, and the output layer neurons whose distance is within a predetermined threshold value are all denominations of banknotes,
A sixth step of recognizing a second pattern that can be selected as a direction candidate is compared with the calculation results of the fifth and sixth steps. If the same candidate is extracted, the money of the bill of the same candidate is extracted. seed,
And a seventh step of outputting a direction. 8. A banknote identification method in which a feature amount of a banknote to be identified is input and a reference pattern is learned, and thereafter, the input unknown banknote is pattern-recognized from the reference pattern. Vector Quantiz
A denomination and a direction to be fired in an output layer neuron are set by using a neural network based on the method of the present invention. A first step of inputting all the feature amounts extracted for each area and learning and storing a reference pattern in pattern matching as a coupling coefficient, and the learning extracted for each first small area of the banknote to be learned. The feature amounts of the bills to be collected are collected in a second small region larger than the first small region, and the feature amount of the second small region is changed to the input layer neuron of the neural network by replacing the small region. A second input for repeatedly inputting over the entire area, calculating a distance from the reference pattern, and setting a threshold value so that an output layer neuron of a corresponding denomination and direction fires. And the step,
The entire banknote to be learned is divided into the first small areas, and all the feature amounts extracted for each of the first small areas are simultaneously input to the neural network, and a distance from the reference pattern is set. And setting a threshold value so that the output layer neuron of the corresponding denomination and direction fires, and a fourth step of providing an associative neuron table for registering previously similar candidate neurons, respectively. When,
Collecting each feature amount of the unknown bill extracted for each of the first small areas of the unknown bill into a second small area larger than the first small area; and inputting neurons of the neural network The feature amount of the second small area is repeatedly input over the entire area while changing the small area, and the distance from the reference pattern is calculated.
Calculating a direction and deactivating an output layer neuron whose distance from the reference pattern is equal to or greater than a predetermined threshold value;
A fifth step of performing pattern recognition of the above, and the entirety of the unknown banknote is divided into the second small areas, and the obtained small areas are collectively processed, and the obtained feature amounts are The neural network including a neuron in which an output layer neuron is inactivated by the first pattern recognition unit, and a neural network in which all neurons registered as candidate neurons in the associative neuron table are activated. At the same time, the distance from the reference pattern is calculated, and all the output layer neurons whose distance is within a predetermined threshold value are recognized as the second pattern that can be selected as a denomination and a direction candidate of the banknote. The steps of: extracting each feature amount of the banknote to be learned extracted for each first small area of the unknown banknote in an area larger than the first small area; And collecting the feature amounts of the third sub-region into an input layer neuron of the neural network, and collecting the reference amount of the reference pattern. A seventh step of calculating a distance between the reference pattern and the output pattern neuron whose distance to the reference pattern is the shortest, and performing a third pattern recognition using the output layer neuron as an identification result; and calculating the fifth, sixth, and seventh steps. Comparing the results and, when the same candidate is extracted, outputting the denomination and the direction of the same candidate banknote in an eighth step.