Counterfeit Detection and Value Recognition of Euro Banknotes

Sebastiano Battiato, Giovanni Maria Farinella, Arcangelo Bruna and Giuseppe Claudio Guarnera

Image Processing Laboratory, University of Catania, Catania, Italy

Keywords:

Banknote Recognition, Counterfeit Detection, Image Forgery.

Abstract:

This paper describes both hardware and software components to detect counterfeits of Euro banknotes. The

system is also able to recognize the banknote values. The proposed method makes use of images acquired

with near infrared camera and works without mechanical parts. This makes the overall system low-cost. The

effectiveness of the proposed solution has been properly tested on a dataset composed by genuine and fake

Euro banknotes provided by Italy’s central bank.

1 INTRODUCTION

The detection of counterfeit banknotes is one of the

most important task in billing machines. Unfortu-

nately these systems are very expensive and are used

only for ATM machines (i.e., fully unsupervised by

humans) where a high degree of reliability is re-

quired. In the last years, cheaper systems for valida-

tion and classiﬁcation of banknotes have been com-

mercialized. They are usually based on motor actua-

tors aimed to let the banknote passing through light

emitters and sensors in a dark area. On the other

hand, the motors are expensive and the mechanical

parts may frazzle in few years.

In this paper we present both hardware and soft-

ware components useful to detect counterfeit of Euro

banknotes and to recognize their currency values.

This is obtained by exploiting a low cost system with-

out motor actuators or other moving parts. A proto-

type of the proposed system has been built and tested

on a dataset composed by genuine and fake Euro ban-

knotes provided by the Bank of Italy. The system is

composed mainly by an infrared camera, a micropro-

cessor and a control ﬂow that implements a robust de-

cision system. A glass is placed in the focal plane of

the camera in order to acquire a sharp image. The ac-

quired image is then processed through the designed

algorithms. The user should just lean a banknote on

the glass and the system provides the information on

the validity and the related value.

The remainder of the paper is organized as fol-

lows: Section 2 reviews the state of the art algorithms

in the ﬁeld. Section 3 summarizes the proposed al-

gorithms for counterfeit detection and currency value

recognition of Euro banknotes. In Section 4 the de-

signed hardware prototype is described, whereas ex-

perimental results are presented in Section 5. Finally,

conclusions are given in Section 6.

2 STATE OF THE ART

It is well known that, in order to avoid forgeries, the

security systems of money are typically encoded in

the banknotes (in several ways in the different curren-

cies). This fact induced researchers to develop differ-

ent counterfeit detection and value recognition algo-

rithms taking into account the different currencies.

In (Hinwood et al., 2006) authors make use of

light transmittance and pattern recognition techniques

to recognize the value of banknotes. The pro-

posed solution requires the banknote to pass through

light emitters (LED) and receivers (photo transistors)

placed in opposite side. Hence, the system requires

a motor or a moving part. Moreover, the method can

be simply deceivedby counterfeit banknotes(not han-

dled by the authors).

In the approach proposed in (He et al., 2004) the

banknotes are segmented into different regions and

a classiﬁer for each segmented region is employed.

A consensus on the classiﬁcation results obtained on

the different segmented regions provide the ﬁnal de-

cision. Speciﬁcally, genetic algorithms were used for

classiﬁcation of both, the segmented regions and fu-

sion blocks.

A neural network and genetic algorithms has been

exploited in (Takeda et al., 1999; Takeda et al.,

2003) to address the problem of banknote recognition,

Battiato S., Farinella G., Bruna A. and Guarnera G..

Counterfeit Detection and Value Recognition of Euro Banknotes.

DOI: 10.5220/0004205100630066

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 63-66

ISBN: 978-989-8565-48-8

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

whereas in (Khashman and Sekeroglu, 2005) the au-

thors proposed an Intelligent Banknote Identiﬁcation

System (IBIS) based on neural networks technique.

The system is designed for Turkish Lira and Cyprus

Pounds identiﬁcation.

Despite different approaches have been presented

in literature, most of them do not take into account

the Euro banknotes or the validation task. Moreover,

some approaches have been tested only on simulated

data. In the Euro banknotes there are several features

which make challenging the forgery: the sheet, the

watermarks, the particular inks with different behav-

ior in visible, infrared and ultraviolet lights, etc. All

the techniques above could be adapted to other cur-

rencies (e.g., Euro) for the recognition task, but their

adaptation to counterfeit detection is not straightfor-

ward because, under visible light, the counterfeit ban-

knotes usually are identical to the valid ones. Hence,

a system designed for a speciﬁc currency (e.g., Euro)

cannot be employed as it is for other currencies (e.g.,

Turkish Lira); each currency has its own speciﬁc fea-

tures to be exploited for counterfeit detection and typ-

ically those features strongly differ from a currency to

another.

Different than other approaches, in the proposed

system the infrared (IR) technology has been ex-

ploited since it can provide good and robust features

for counterfeit detection of Euro banknotes.

3 SOFTWARE COMPONENTS

The proposed software consists of three main blocks:

Calibration, Training and Use Module.

3.1 Calibration

Since in our setup the LED illumination is not spa-

tially uniform, and taking into account its changes

from prototype to prototype, a calibration phase is

required to remove illuminations variability. Speciﬁ-

cally, in this phase a brightness map is computed and

a compressed version of it is stored inside the ﬂash

memory. The brightness map is obtained by acquir-

ing several images of a white sheet of paper in differ-

ent environments (i.e., lighting conditions), and then

considering the average of all the acquired frames.

This process allows also to reduce noise (usually zero

mean Gaussian noise). The map is then used during

the Training and Use Modules to normalize the input

banknote images.

The Training block is used to learn optimal fea-

tures of the banknotes (i.e., patches), which will be

used to determine the validity and the value of each

Figure 1: A banknote of 50 Euro as acquired by the IR cam-

era of the system, and the set of regions used to test its gen-

uineness. White: Overall ROI. Blue: the threshold to bina-

rize the image is calculated in this area. Green: this area

must be dark under infrared light. Red: when seen under

infrared light, this area must be bright and without notice-

able patterns. Magenta: additional area used to check the

genuineness of some denominations.

banknote. The parameters are learned using a large

data set of both genuine and counterfeit banknotes,

which come in several face values.

In our tests we have used both genuine and real

counterfeit banknotes collected, only for testing pur-

poses, at the Bank of Italy. A training dataset of 1000

images has been acquired taking into account typical

contexts of use of the apparatus, in which there are

very different lighting conditions(e.g., neon, sunlight,

incandescent and ﬂuorescent lamps, etc.) and a high

degree of misalignment with respect to the prosce-

nium (the system must be robust to slightly translated

and or rotated banknotes).

Genuine Euro banknotes are made such that only

speciﬁc visual features are visible under infrared

lighting. In particular banknotes must show darker ar-

eas in different zones depending on their value. Those

areas always show characteristic patterns. However,

there are overlap among all banknote values, for both

dark and bright areas, which can be used as validity

check.

The aim of learning stage was hence to ﬁnd the

best area to check banknote validity by using a train-

ing dataset composed by genuine and fake Euro ban-

knotes.

Given the set of all genuine banknotes G =

, G

, . . . , G

}, the training block is devoted to

search for the largest common dark area F

, and the

largest common bright area T

of genuine banknotes.

3.2 Training

Let F

g,i

= IRSignal(G

) be the infrared highlighted

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

dark area of each G

, then F

g,i

Let T

g,i

= G

\ F

g,i

be the unresponsive bright

area of each G

, then T

g,i

The regions T

and F

need to be reﬁned in order

to be robust to counterfeit banknotes, which in some

case might show a slightly similar infrared response.

In a similar way, given the set C = {C

, C

, . . . , C

}

of all counterfeit banknotes in our data set, we deﬁne

c, j

= IRSignal(C

) and T

c, j

= C

\ F

c, j

Let C

∗

be the set of all C

such that

c, j

6= ∅) ∧ (T

c, j

6= ∅). We can

ﬁnally deﬁne T and F considering C

∗

as follows:

T = T

c, j

and F = F

c, j

For each banknote B

in G

C, for each possible

threshold s ∈ [0, 255], we computed the percentage

percB

k,s

of pixels above s in region T

and the per-

centage percD

k,s

of pixels below s in regions F

. Us-

ing those values we computed optimum percB

∗

and

percD

∗

thresholds to separate genuine and counterfeit

classes.

To classify the banknote value, during train-

ing phase we learn the most discriminant patterns

for each possible banknote value D

∈ D =

{5, 10, 20, 50, 100, 200, 500}, such that the intra-class

distance is minimized whereas the inter-class distance

is maximized. This process is very tricky for some

banknote values which share similar patterns (e.g., 5

Euro and 20 Euro). The ﬁnal output is a set of patches,

with the corresponding position, related to the differ-

ent banknote values. To take into account slightly

translated and/or rotated images, the search area of

each patch has the same shape and center coordinates

of the patch itself, but it is wider: approximately the

search area is 2.5 times the corresponding patch area.

Once the training phase is performed, the selected

banknote features (i.e., the learned patches, locations

and thresholds) are stored in the ﬂash memory and

used to infer validity and value of input banknote im-

ages.

3.3 Banknote Authentication

Once the image has been corrected for the non-

uniform led illumination (see Section 3.1), a proper

threshold needs to be found according to the actual

input data. In our experiments we have noticed that

the average gray value (indicated as MeanRef) of the

blue region shown in Fig. 1 can be robustly used for

this purpose.

To check the genuineness, the percentage of

pixels inside the green region with a gray value

below MeanRef (i.e., percLT) is computed, to-

gether with the percentage of pixels inside the

magenta and red regions with a grey value

above MeanRef (i.e., percGT). If ((percLT >

percD

∗

)AND(percGT > percB

∗

)) the banknote is

classiﬁed as genuine.

To identify the face value, the system makes use

of patches learned during the training stage (i.e., tem-

plates related to the different banknotes values) and

the corresponding search areas coordinates. Search

areas are wider than the patches in order to be ro-

bust to small misalignments. Each template patch is

placed at the center of its search area and a correla-

tion measure between the pattern itself and the corre-

sponding pixels on the search area is used for com-

parison. The procedure then search if a translation

around the neighborhood of the current position could

increase the correlation. This is performed by moving

the pattern position. This step is then repeated until

the correlation reaches a local maximum. The pat-

tern with the highest correlation determines the ﬁnal

value to be assigned to the input banknote. Informa-

tion about genuineness and face value are then sent

to the display, in order to inform the user about the

results of the banknote analysis. The overall compu-

tational speed is of about 1 second for validity and 2

seconds for value recognition using the architecture

described in the Section 4.

4 HW PROTOTYPE

The hardware prototype has been designed to demon-

strate the effectiveness of the proposed framework.

The Infrared LEDs (IR leds block) illuminate the

scene. It is composed by 6 LEDs placed around the

proscenium (the area where the banknote is placed).

The illumination is not uniform in the real system.

Hence, a calibration is needed to take into account

the non uniformity and the performances decay dur-

ing the system life (see Section 3.1). An optical ﬁl-

ter is inserted to avoid the external light source to

inﬂuence the image acquisition system. It is placed

on top of the infrared camera (IR camera block) that

acquires the image. The camera is a common low

cost camera with CCTV output. Since the output

is analog, an analog to digital converter (A/D con-

verter block) is used to obtain a digital standardized

format (i.e., CCIR − 656). The digital image is ac-

quired by the microprocessor. In the prototype we

have used an ATMEL AT91SAM9XE256, containing

a ARM926EJ− S

processor, 200MHz with 256KB

internal high-speedﬂash memoryto store the program

instructions. It contains also an Image Sensor Inter-

face (ISI) port able to capture video sequences com-

pliant with the standard ITU-R BT 601/656. The soft-

ware contains, beside the control logic for the entire

CounterfeitDetectionandValueRecognitionofEuroBanknotes

Table 1: Genuine/Counterfeit classiﬁcation.

True Positive (counterfeit banknote correctly classiﬁed) False Positive (genuine banknote incorrectly classiﬁed)

100% 4.3%

False Negative (counterfeit banknote incorrectly classiﬁed) True Negative (genuine banknotes correctly classiﬁed)

0% 95.7%

Table 2: Banknote value classiﬁcation.

5E 10E 20E 50E 100E 200E 500E Counterfeit

5E 88,00% 0,00% 0,00% 0% 0,00% 0,00% 0,00% 12,00%

10E 1,00% 91,00% 0,00% 0,00% 0,00% 0,00% 0,00% 8%

20E 0,00% 0,00% 98,00% 0,00% 0,00% 0,00% 0,00% 2%

50E 0,00% 0,00% 0,00% 99,00% 0,00% 0,00% 0,00% 1,00%

100E 0,00% 0,00% 0,00% 0,00% 93,00% 0,00% 0,00% 7,00%

200E 0,00% 0,00% 0,00% 0,00% 0,00% 98,00% 1,00% 1,00%

500E 0,00% 0,00% 0,00% 0,00% 0,00% 0,00% 95,00% 5,00%

subsystems (e.g., IR led, IR camera settings, Display,

etc.), the related algorithms (for both validation and

classiﬁcation) described in Sections 3. The prototype

has been also equipped with external SRAM mem-

ory, since the microprocessor contains only 32KB of

internal SRAM.

5 EXPERIMENTAL RESULTS

To evaluate performances of the proposed technique

we used our prototype to acquire a test set of 1750

banknote images, with the same criteria used for the

training dataset (i.e., both counterfeit and genuine

banknotes have been acquired under several environ-

ment lighting conditions, with different illuminants

and brightness). In both cases, training and experi-

mental phases, the banknotes samples have been pro-

vided by the Bank of Italy.

Acquired counterfeit banknotes include also spec-

imens carefully calibrated to mislead digital counter-

feit detectors. To deal with special cases (i.e., fake

banknotes provided by the bank), additional proce-

dures have been included. The overall processing

time is very close to the base algorithm, since it in-

cludes a few average computations on very small ar-

eas. Table 1 reports the results of the validity assess-

ment. Table 2 shows the results of the banknotesvalue

classiﬁcation.

6 CONCLUSIONS

In this paper we have proposed an effective system

to detect counterfeit of Euro banknotes composed by

both hardware and software modules. Conversely to

the state of the art algorithms, the proposed solution

makes use of infrared imaging and low-cost hard-

ware. The proposed system allows recognizing not

only the value, but also forgeries. The described al-

gorithms are robust to changes in environment light-

ing, in terms of illuminant type and incident inten-

sity. Thanks to a training phase it is also robust to

non-uniformity of the infrared light. The experiments

performed on genuine and fake banknotes provided

by the Bank of Italy show good performances in both

validity and value recognition.

ACKNOWLEDGEMENTS

The authors would like to thank Imperial Emporium

Srl and the Bank of Italy for supporting this research

activity.

REFERENCES

He, C., Girolami, M., and Ross, G. (2004). Employing op-

timized combinations of one-class classiﬁers for au-

tomated currency validation. Pattern Recognition,

37(6).

Hinwood, A., Preston, P., Suaning, G., and Lovell, N.

(2006). Bank note recognition for the vision im-

paired. Australian Physical and Engineeing Sciences

in Medicine, 29(2).

Khashman, A. and Sekeroglu, B. (2005). Multi-banknote

identiﬁcation using a single neural network. In Blanc-

Talon, J., Philips, W., Popescu, D., and Scheunders,

P., editors, Advanced Concepts for Intelligent Vision

Systems, volume 3708 of Lecture Notes in Computer

Science. Springer Berlin Heidelberg.

Takeda, F., Nishikage, T., and Omatu, S. (1999). Banknote

recognition by means of optimized masks, neural net-

works and genetic algorithms. Engineering Applica-

tions of Artiﬁcial Intelligence, 12(2).

Takeda, F., Sakoobunthu, L., and Satou, H. (2003). Thai

banknote recognition using neural network and con-

tinues learning by dsp unit. In Palade, V., Howlett, R.,

and Jain, L., editors, Knowledge-Based Intelligent In-

formation and Engineering Systems, volume 2773 of

Lecture Notes in Computer Science. Springer Berlin -

Heidelberg.

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications