currency prediction

7/30/2019 Currency Prediction

1/5

IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 6, NO. 1, JANUARY 1995 13

High Speed Paper CurrencyRecognition by Neural NetworksFumiaki Takeda and Sigeru Omatu

Absfracf-ln this paper a new technique is proposed to improvethe recognition ability and the transaction speed to classify theJapanese and U.S. paper currency.Two types of data sets, timeseries data and Fourier power spectra, are used in this study. Inboth cases, they are directly used as inputs to the neural network.Still more we also refer a new evaluation method of recognitionability.Meanwhile, a technique is proposed to reduce the input scale ofthe neural network without preventing the growth of recognition.This technique uses only a subset of the original data set whichis obtained using random masks. The recognition ability of usinglarge data set and a reduced data set are discussed. In addition tothat the resultsof using a reduced data set of the Fourier powerspectra and the time series data are compared.I. INTRODUCTION

N conventional paper currency recognition machines, weI ave developed the recognition algorithm according to thetransaction speed and difference of various specifications. Forexample, we have produced the paper currency recognitionmachines that can transact from 1 to 10 pieces/s [I] , [2].However, development of the algorithm has been based onthe method of trial and error. Namely, engineering designersusually sample the characteristic parameters of conveyed papercurrency and examine whether the paper currency can berecognized or not using the sampled characteristic parametersby evaluating large quantities of paper currency. But manyresearchers have reported that neural networks, (NN's), aresuitable for the pattem recognition because of the ability ofself-organization, parallel processing, and generalization [3].[4]. Especially, we can reduce work time to find characteristicparameters of the paper currency with our experience andknow-how owing to the ability of self-organization. We canhope robustness for defect of data or noise of conveyed papercurrency because of the ability of generalization. Still more,we can transact the paper currency recognition with multi-taskor multi-CPU owing to the parallel processing. In this paper,we propose a method for Japanese and U.S. paper currencyrecognition using NN's. Then we show the effectiveness andpossibility of the present algorithm on developing periodand its recognition ability compared with a conventionalmanual method by discriminative inequalities that engineeringdesigners are used to apply [I] , [2], [ 5 ] . Meanwhile, on

Manuscnpt received October 22, 1992; revised December 19, 1993.F. Takeda is with the Development Center GLORY LTD.,Himejl, 670,Japan.S. Omatu is with the Depart ment of Information Science and IntelligentSystem, Faculty of Engineering, University of Tokushima, Tokushima, 770,Japan.IEEE Log Number 9400110.

the ordinary pattem recognition using NN's, the ability ofrecognition has been evaluated by only maximum value ofthe output unit [6], [7]. In other words, maximum valuemainly has been adopted but its distribution of other outputunit values has not been considered. Here, we pay attentionto not only maximum value but also the distribution of theoutput unit values. We newly introduce a measure of reliabilityas an indicator for the error recognition [ l l , 121, 151, [61.We show that the ability of recognition can be evaluated indetail by introducing this measure of reliability. On the otherhand, when we try to implement the above N N to the usualcommercial products, the NN scale is a serious problem. Tosolve this, the various methods using FFThave been proposed[2], [8]. But owing to complex pre-processing, they are notpreferable for implementation of the commercial products.In this paper, we propose a new method which condensesinput pixels by a simple pre-processing using random masks.The inputs to the NN are not paper currency data or its Fourierpower spectra but sums of these data which have been passedthrough the various random masks. It is shown that we canreduce NN scale using random masks without preventing thegrowth of the recognition ability.

11. CONVENTIONALECOGNITION METHODConventionally, engineering designers are used to apply

discriminative inequalities to recognize paper currency [11,[2], [5]. In this section, we describe this conventional methodand its problems. These discriminative inequalities were de-termined manually. Namely, to recognize the paper currency,discriminative points are selected from each sensor. The char-acteristics of each paper currency from a set of discriminativepoints are extracted by the trial and error. The discriminativeinequalities that contain the representative values of the pa-per currency are then determined. Using this method, papercurrency recognition is executed with these discriminativeinequalities. Fig. 1 shows a rough sketch of the conventionalmethod to determine the discriminative inequalities. It is con-structed by search of discriminative points and determinationof thresholds. The former block explains the searching pro-cedure of discriminative points corresponding to pixels whichinclude sufficient information to discriminate the purposivepaper currency from the rejected ones. A and B valuesenclosed by the rectangulars denote the total sums of themean squared sensor values on the hatched areas A andB, respectively. The aim of the searching procedure is tofind preferable A and B such that the distance ( A value-Bvalue) between the purposive paper currency and rejected

1045-9227/95$04.00 0 1995 IEEE


2/5

14 IEEE TRANSACTIONS ON NEURAL NETWORKS. VOL. 6, NO. 1, JANUARY 1995

SEARCH OF DlSCRlMlNATlA VILUE=180

DETERMINATION 4OF THRESHOLDS.FJ ki~[L:I~UVF -b. -VALUE. .- ~

Fig. 1, Determination of discriminative inequalities.

I NPUT L A Y E R H I D D E N L A Y E R OUTPUT L A Y E RrmPAPER CURPENCY

U l rFig. 2. Configuration of the proposed method with NN .

110,000-head110.000-ta i IY 5 , O O O - h e a dY5,OPO-tai lYI, O O O - h e a dV I , 000-t I

P A P E R C U R R E N C Ya m p l e d p i x e l s/'

c- ONVEYED DIRECTION' s a m p l e n u m b e r 32Fig. 3. Paper currency data and its sampling method.

paper cu r r en cy (spec imen ) paper cu r r en cy (spec imen )

head u p l i g h t t a i l i n v e r s ep a p er c u r r e n c y ( s pe c i m e n l paper Cu r ren cY I sP ec imen )

h ea d i n v e r s e t a i l u p1 ightFig. 4. Conveyed direction of paper currency.

one becomes as large as possible. The latter block is todetermine the thresholds of discriminative inequalities. Using hidden unit number is 64 which have been decided thoughvarious experiments considering the over-.tting problem. Out-this we the purposive Paper currencyfrom rejected One as shown in Fig. is put unit number is 12 which correspondto recognition patterns.Namely, we recognize three kinds of Japanese paper currencyf theassumed 125 in this case. The discriminative inequalities aregiven by such as y lo ooo, y5000, and ylooo and four conveyeddirections such as head upright, head inverse, tail inverse, and

tail upright as shown in Fig. 4where the slope line in papercurrency denotes specimen. We adopt the back-propagationmethod [9] for learning and is given by

f z =fz(a1,a2,. ..,am, bl, b z , . ..,6i =1,2; . . ,L,j = 1 , 2 , . . . , m , k =1 , 2 . . . , n

2 e ; , , a j E A , b k E L?(1) AWj,i(t)=- - ~ d j ~ ;aAWj,;(t- 1)+pAWj,;(t- )

dk = (ok - k)f' (netk), etkwhere L is the number of discriminative inequalities, m andn are the numbers of discriminative points, A and B are aset of discriminative points, Bi is the threshold, and aj and b kare discriminative points. Here, m and n depend on i. In theconventional method, engineering designers have determinedthe number ( L ) and parameters ( a j , k , &) in discriminativerejected one. However, it is difficult to give a systematic

= ~ k , j ( t ) o j for output layer3

inequalities between the purposive paper currency and the netj = W, , ( t ) o i for hidden layer (2)2

procedure to work of determining discriminative inequalities.This is due to the reason that the conventional method dependson the experience and know-how of engineering designers.So it has taken more than 6 months to obtain the finaldiscriminative inequalities by the conventional method.

where Wj,i(t) s weight from unit i to j , AWj,i(t) s thechange of weight AWj,;(t), d is the generalized error, o isthe output value, t is the sample, E is the positive learningcoefficient, a is the proportional coefficient of inertia tcrm,p is the proportional coefficient of oscillation term. Here,term has the role of escaping from a local minimum [9]. Weregard that learning is converged when the sum of squarederror between output unit values and desired values for eachpattern becomes less than a threshold value of 0.001. Thedata is obtained by a high speed conveyed paper currencyrecognition machine [ l] , [2], [5] as shown in Fig. 5 and theyare sampled from real products. This figure shows each sensor

111. RECOGNITIONM E T H O D BY ORDINARY NN'SThe proposed method using the N"s for paper currency

recognition is three-layer configuration as shown in Fig. 2.We input paper currency data consisting of 32 samples x 4sensors as shown in Fig. 3or the Fourier power spectra of 17amplitudes of FFT x 4 sensors to the NN [2], [ 5 ] , [8]. The


3/5

TAKEDA AND OMATU:HIGH SPEED PAPER CURRENCY RECOGNITION 75

s e n s o r 1255 I I

32samp Ies e n s o r 3

I 32sanp IeFig. 5. Sensor data for Y l O O O O head upright.

data for YlO 000 head upright for 10pieces. The data set usedare as follows:

1) Learning data; the kinds of paper currency areYlOOO, Y5000, and Y l O O OO . Each paper currency isnot worn out and have not the defected comer. Numberof each paper currency is 10 pieces.2) Testing data; the number of each paper currency is 100pieces, each of which is wom out and has the defectedcomer.

Henceforth, we represent the sensing data in a time seriesform. First, we adopt pass ratio ES1 [6] as a recognition indexand is given by (see (3) at bottom of page).For comparison of recognition ability, we consider threemethods. Namely, method (i) is the conventional methodusing discriminative inequalities, method (ii) is NN methodwith time series data, and method (iii) is NN method withFourier power spectra. ES1 for each method is 100% inthis experiment. Still more, we need less than one month todetermine the configuration of the NN and their parameters.This shows that the methods (ii) and (iii) are more superiorthan the method (i) on developing period.

I v . SCALE REDUCTION F NNS BY RANDOM MASKSWe show the effectiveness of NNs for paper currency

recognition in Section 111. However, the NN scale is one ofthe important design factors when we apply N s to devel-opment of paper currency recognition machines. Generally,it is difficult to implement this type of NNs to usual papercurrency recognition machines owing to its large NN scale. Tosolve this problem, the various methods which condense inputpixels with or other transformation have been proposed[101, [111. But complex pre-processing is required for currentpaper currency recognition machines. Therefore, they are notpreferable for implementation.

s e n s o r 2255 r - I

1 32sanple255r 9

I s a n p l e 32

. . . .m a s k I

_...I INPUT HIDDEN OUTPUTj LAY ER l A Y E R l A Y E R

?part of generatings l a b values- -. . - -

Fig. 6. Configuration of the scale reduction method by random masks

We propose a method which condenses input pixels withsimple pre-processing. Namely, we adopt a slab-like architec-ture [lo]--[12]. In the present method, we use a slab valuethat is sum of pixels as the input to the NN. However, it maygenerate the same slab values even if the inputs are different.To solve this problem, we cover some parts of the input byusing random masks. Then we apply this method to papercurrency recognition.

Here, we determine the masking pixels in the followingway. We generate random values [--1, 11 using randomfunctions where the total number of random values is equalto input pixels. We cover the pixels which correspond to thenumber of random values whose values are minus values. Theconfiguration of the proposed NN with random masks is shownin Fig. 6.In this figure, some parts of input are covered withvarious masks in pre-processing. The sum of the pixels whichare not masked becomes one slab value, which is used as aninput value for the NN. Namely, both numbers of the maskkinds and slab values are the same. The configuration of theproposed NN is three layers and each layer has 16, 16, and12 units, respectively.

(the number of correctly recognized paper currency)ES1 = x 100.(total number of evaluated paper currency) (3 )


4/5

\ \

ITERATIOlj NUMBERFig. 7. Learning curves for time series data and FFT data.

TABLE ICOMPARISONF THE NN SCALE ETWEENPROPOSED METHOD ND ORDINARY NE

We experiment in the following two cases. One is to usethe time series data as an input to the proposed NN. Anotheris to use the Fourier power spectra to the proposed one. Fig. 7shows the learning curves for the proposed method whereFFf data means Fourier power spectra. Pass ratios ES1 forboth data are 100%. Forcomparison of the NN scale, we adoptthe methods (ii) and (iii) that have been already described inSection 111. In these methods (ii) and (iii), henceforth, we callthe methods (ii) and (iii) the ordinary method. From this figure,the proposed method needs more time to converge than theordinary one. However, we do not need on-line learning inpaper currency recognition machines because of the productsspecification. We define the NN scale by the weight number.Table I shows comparison of the NN scales between theproposed method and the ordinary one. In these experiments,we can reduce the NN scale less than 1/10 compared with theordinary method. Therefore, we find that the proposed methodis effective to paper currency recognition and does not makeworse recognition ability.

IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 6, NO. 1, JANUARY 1995

10000 Fig. 8.masks.ou

Fig. 9.

v. ANOTHER ERIFICATION FOR SCALEREDUCTIONY RANDOM MASKSWe verify an effect of the scale reduction by random masks

using another paper currency recognition machine as shownin Fig. 8. The sensor system in the recognition machine cansample 216 x 30 pixels from paper currency. The sampleddata has 1 byte gray level. Paper currency is conveyed inthe parallel direction of its short part and its conveyed speedis 10 pieces/s. This experimental system can sample papercurrency data by recognition machine on-line. Learning isexecuted off-line. Once we can obtain the suitable weights,this system recognizes paper currency using them on-line andits transaction speed is more than 3 piecesh. The data set usedhere are as follows:

An experimental system using the scale reduction method by random

d i s t r i b u t i o n o fo u t p u t v a l u e sc r o s s p a i n t u n i t 1 u n l t 2

iA_ _ . a - L l1 u l i t v a l u e -

Distribution of NN output values.

1) Learning data; the kinds of paper currency are $1, $5,$10, $20, $50, and $100. Each paper currency is notworn out and does not have the defected comer. Thenumber of each paper currency is 10 pieces.2) Testing data; the number of each paper currency is 2000pieces, each of which is wom out and has the defectedcomer.

When we evaluate the recognition ability using ES1, itbecomes more than 92% for each paper currency kind.

VI. EVALUATIONETHODOF RELIABILITYA . Standard of Reliability Evaluation

We newly introduce a standard of reliability evaluation [11,[2], [ 5 ] , [6] as an index for the recognition ability. Generallyspeaking, the pattem classification by NN is based on thewinner-take-all. However, considering output unit values, wecan regard the difference between maximum value and othervalues of output units as a reliability measure. Furthermore,using several data for the same kind of paper currency,the output unit values are fluctuating and obey stochasticdistribution. Error probability appears when the distributionof purposive mode (unit 1) and that of rejected one (unit 2)crosses each other as shown in Fig. 9.

We define upper probability as a standard of reliabilityevaluation and describe it as ES2 [l], [2], [ 5 ] , [6]. But thedistribution of output unit values is asymmetry in [0, I].However, if we assume that this obeys Gaussian distribution


5/5

TAKEDA AND OMATU: HIGH SPEED PAPER CURRENCY RECOGNITION 77-O. 000 % O 000 f5 . 000 15 , 000 %I.OO 1I.000i i I 1 i J I I 1E R R O R R E C o C N I T I O N P A T T E R N5 0 0 0 % l . O O O 1110000 II.000 110,000 15.000

Fig. 10. Comparison of reliability.

N(p: average value, c: tandard deviation), then ES2 can begiven by

where x p is the output unit value of purposive mode, %Gpis the mean value of z P ,oP is the standard deviation ofpurposive mode, 8 is the x coordinateof cross point betweenthe distribution of purposive mode and that of rejected one.We can regard that the recognition reliability is high whenESP is small.B . Comparison of Reliability

Fig. 10 shows the comparison of reliability using ES2 forJapanese paper currency. We choose the worst ES2 (the largestES2) as the reliability value for each error recognition pattem.From Fig. 10,the reliability of the method (ii) is lower thanthat of the method (i). It is supposed that the discriminativepoints of the method (i) are independent of conveying papercurrency with the experience and know-how. So Es 2 by themethod (i) is not influenced so much by conveying papercurrency [ I ] [2]. But ES2 by the method (iii) is much higherthan that of the methods (i) and (ii). Because we can reducethe influence of fluctuation of conveyed paper currency usingFourier power spectra [ 2 ] .Still more, there is no differenceofrecognition ability among the methods (i), (ii), and (iii) if ES1is used. However, we can find the difference among them if weuse ES2. Thus, ES2 is effective for paper currency recognition.

VII. CONCLUSIONIn this paper, we have applied the NN to paper currency

recognition and showed the effectiveness compared with aconventional manual method. Furthermore, we have proposeda structure reduction method of the NN using random masksand showed its effectivenessfor time series data and its Fourierpower spectra. Finally, we also introduced the new evaluationmethod for reliability and showed its effectiveness usingreal data. We expect that the present method will promotecompactness, high speed transaction, and low cost of papercurrency recop-iition machines.

REFERENCESF. Takeda, S . Omatu, and T. Inoue, An expert system for determiningthe discriminative functions for bill money recognition, Trans. Info.Process. Soc. Japan, vol. 33, no. 7. pp. 980-9 90, 1992, in Japanese.E Takeda, S. Omatu, T. Inoue, S . Onami, and K. Konishi, High speedconveyed bill money recognition with neural network, Trans. Inst.Elect. Eng. Japan. vol. 112-C, no. 4, pp. 101-110, 1992, in Japanese.Sonehara, An application of neural network to image data, Trans.Inst. Syst., Control and Info.Eng., vol. 35, no. 1, pp. 11-18, 1991, inJapanese.S. Amari, The Present and the Future of Neural Network. KyoritsuPress, 1990, in Japanese.F. Takeda, S . Omatu, T. Inoue, and S . Onami, High speed conveyedbill money recognition with neural network, in Proc. IMACSISCINE Int.Symp. Robotics, Mechatronics and Manufact. Syst. Kobe, Japan, 1992,N. Suzuki, Y. Yasuoka, and T. Shimura, Probability and StatisticsSeminar. Kyoritsu Press, 1979, in Japanese.M. Nagao, Panern Information Processing. Corona Press, 1986, inJapanese.F. Takeda, S. Omatu, T. Inoue, and S.Onami, Bill money recognitionusing neural network with as pre-processor, Trans. Inst. Syst .,Control and Info. Eng.. vol. 5 , no. 7, pp. 265-2 73;1992, in Japanese.S . Nagata, M. Sekiguchi, and K. Asakawa, mobile robot control by astructured hierarchical neural network, IEEE Control Syst. Mag . , vol.10, no. 2, pp. 69-76, 1990.M. Fukumi, S. Omatu, F. Takeda, and T. Kosaka, Rotation-invariantneural pattern recognition system with application to coin recognition,IEEE Trans. Neural Net, , vol. 3, pp. 272-279, 1992.B. Widrow, R. G. Winter, and R. A. Baxter, Layered neural nets forpattem recognition, IEEE Trans. Acou st., Speech Sig. P rocess., vol.ASSP-36, pp. 1109-1118, 1988.F. Takeda and S . Omatu, Bank note recognition system using neuralnetwork with random masks, in Proc. World Cong. Neural Net.,

vol. 1, pp. 16-20.

Portland, OR, 1993, vol. 1, pp. 241-244.

Fumiaki Takeda received the B. E. and M. E.degrees in mechanical engineering rom the NayogaInstitute of Technology, Aichi, Japan, in 1982 and1984, respectively. He received the Ph. D. de-gree in the Facult y of Engineenng, University ofTokushima, Japan UI 1994.From 1984 to 1986, he was with Toyota CentralResearch and Development and in 1986 he joinedGLORY Ltd., Himeji, Japan. Since then, he has beenengaged in research and development work on arecognition system for paper currency. His researchinterests include neural networks, expert system, and genetic algorithms asapplied to paper currency recognition.

Sigeru Omatu received the B. E. degree in me-chanical engineering from the University of Ehime,Japan in 1969, and the M. E. and Ph. D. degreesin electronics engineering from the University ofOsaka Prefecture , apan, in 1971 and 1974, respec-tively.From 1974 to 1975 he was a Research Associate,then a Lecturer (1975-1980), an Associate Professor(1980-1988), and finally a Professor at the Univer-sity of Tokushima, Japan. From November 1980 toFebruary 1981 and from June to September 1986, hewas a Visiting Associate on Chemical Engineering at the Califomia Instituteof Technology, Pasadena. From November 1984 to October 1985 he was aVisiting Researcher at the International nstitute for Applied Systems Analysis,Austria. His current interests center on neural networks and distributed

parameter system theory.Dr. Omatu received the Excellent Young Researcher Award from theSociety of Instrument and Control Engineers of Japan in 1972 and a BestPaper Award from the Institute of Electrical Engineers of Japan in 1991.He is an Associate Editor of the International Journal of Modeling andSimulation, the IMA Journal of Mathematical Control and Information, andof the LEEE TRANSACTIONSN NEURAL NETWORKS.e is a co-author ofDistributed Parameter Systems: Theory and Applicarion (Oxford UniversityPress).

currency prediction

Documents