International Journal of Engineering Applied Sciences and Technology, 2020

Vol. 5, Issue 1, ISSN No. 2455-2143, Pages 539-542

Published Online May 2020 in IJEAST (http://www.ijeast.com)

CURRENCY RECOGNITION SYSTEM USING

IMAGE PROCESSING
Shivam Singh, Anshuman Tiwari, Sameer Shukla, Sankalp Pateriya
Department of Computer Science and Engineering
Student, IMS Engineering College
Ghaziabad, India

Abstract— In this paper, we invented a system for top of Tensor Flow that makes it easy to construct, train and
currency identification using a technique called image deploy object detection models.
processing. The suggested approach can be used for
identifying the nation (country) or origin and the cost or II. BACKGROUND MOTIVATION
the denomination of a given note as well. We have taken As there are more than 150 different currencies available
simplest paper currencies into consideration. This system around the world so for the currency recognition the demand
works by first pointing the nation of origin using assured of an automated system has been exponentially growing. The
existing region of interest, and then extricating the cost need for constructing a system for processing notes with
value using attributes like color, text, or size of the negligible human interference for various different uses had
currency, based on how much the notes in the same nation significance for the development of system which helps in
varies. We have examined some of the most popular detecting or identifying the paper notes. However this work
currencies, along with their cost value or denominations. becomes very critical due to the differing characteristics in
Our system is able to accurately and expeditiously classify each notes and the security measures involved in different
sample notes. currencies. Various methods were suggested like pattern
matching to construct a system that uses a method for the
Keywords— Machine Learning, Currency detection, recognition and identification of currencies. Not even a single
Feature extraction, Neural Network. technique has been proven to be sufficient enough for genuine
improvement therefor making the issue an interesting area of
I. INTRODUCTION research. In the early 90’s proposal of an approach was made
As per the survey conducted, there exist more than 150 to identify the paper currency using image processing
different currencies in the world presently. These currencies technique. It has been supposed that the notes taken were in
varies substantially in characteristics such as color, size and good condition and the images are obtained as needed. It's
shape. Unlike the ancient times, the business and exchange critical to make reference that the system suggested need the
among various nations has expanded in all varieties of levels. input images which are taken in a predefined dimensions. The
The need for acquiring knowledge about approximately all of methodology discussed performs pre-processing steps in a
the currencies by means of the banks has been extraordinarily sequence on input pictures and several features are then
critical. However for any human being it is almost impossible selected as hue, saturation and value criterion so that
to recognize or memorize all the types of currencies regulating calculation of Euclidean distance using these parameters can
throughout the world. Thus it becomes necessity to have an be performed and this will be matched with the predefined
efficient automatic system that helps in recognizing paper parameters, set as standards. Though an algorithm was
currency which is pivotal for the future. In this paper, we proposed in this system for the multiple currencies, but as the
suggest an automated system for currency recognition using various notes across nations have similar features so this
Image processing techniques. Our system works for the method is not an effective method to identify all the
currencies used most commonly. It use the sample image currencies.
dataset to implement object recognition. We implement a III. PROPOSED MODEL
model using tensor flow in python with a sample dataset of
various classes to retrieve various elements in labelled images. Our proposed system needs to be taught how an object or
The advancement we have found is the precision of finding currency looks like in order to teach it to recognize that
the objects in images. It present Convolution neural networks currency or object. We must teach the system how an rs10,
for object recognition and classification in this paper. In rs20, rs100 etc. notes looks like before it can identify it. The
conviction with other software, tools and APIs such as Tensor precision of the system can be extended with the increase in
Flow Object Detection API, is an open source schema built on number of labeled pictures the system takes to train the
classifier. This type of learning is known as Supervised
Learning.

539
 International Journal of Engineering Applied Sciences and Technology, 2020
Vol. 5, Issue 1, ISSN No. 2455-2143, Pages 539-542
Published Online May 2020 in IJEAST (http://www.ijeast.com)

Prior to starting with the preparation of system's model

architecture, collecting proper data to train the model is an
important step, in this case data refers to collection of several
images of different categories which the system must
recognize. A computer can only perform numerical
computations but cannot interpret images as humans do. So in
order to make the system understand the input, we need to
convert the pictures into numbers or bytes.
Our object detection model has the following steps:

A. Model Generation:-

a. Dataset:-
The dataset include 7 classes which do not overlap with
200 images in every class. The images are perfectly
labelled and are without any noise.

b. Splitting our dataset:-

It consumes very long period of time to calculate the
gradient of model using a vast dataset as a whole.
Therefore, during each iteration, a less number of images Fig. Faster R-CNN
will be used. Then the dataset will be divided into 80% as
image training set and 20% as image test set. It is made out of 3 sections:

c. Pre-processing:- 1. Convolution layers:-

We will perform the combination to the images randomly This layer is used in training the filters to extricate the
as mentioned. Cropping the images, horizontally flipping images, for example let us assume that these filters are
the images, resizing the images and convert the above going to instructed in order to extract the suitable
dataset into required xml file. characteristics of a man face, then the filters that only
existing in the human face are colors and shape of
d. Build a model:- human face.
Convolutional Neural Networks (CNNS) are among the Convolution network are commonly made out of
most popular techniques which used to enhance accuracy convolution layers, pooling layer and a last segment
for images classification. It is a type of neural network that which is the completely associated or another expand
works similarly as a regular neural network, which would things that will be utilized for a proper errand like
be the backbone of our model. characterization or identification.
We perform computation of convolution by adjusting
B. Training the model:- the filters along our pictures given as input, the output
is the matrix of two dimensions known as feature map.
A model is trained on the training dataset and tested on the In the feature map the quantity of features will be
validation set, and the model can be then improved over decreased by eliminating the pixels having less
time over the validation set. The model can be applied to intensity values, which is comprised in pooling. And
our actual test dataset, once we are satisfied with the the last step in Faster R-CNN is to arrange the features
results. Thus, on our validation set, we can see whether the that do not belong to our case by utilizing the fully
model is over-fitting or under-fitting, which can help us to associated layers.
fit the model better and get better accuracy.
There are several prior models present in the tensor flow 2. Region Proposal Network (RPN):-
which we can use to train our dataset. We are using the RPN is a type of neural network based upon the final
faster R-CNN model to train our dataset. feature map of convolution layers and it also guess that
Faster R-CNN is an object detection architecture which whether an object present or not and those object’s
uses CNN like YOLO (You Look Only Once) and SSD bounding boxes.
(Single Shot Detector).
3. Classes and Bounding Boxes prediction:-

540
 International Journal of Engineering Applied Sciences and Technology, 2020
Vol. 5, Issue 1, ISSN No. 2455-2143, Pages 539-542
Published Online May 2020 in IJEAST (http://www.ijeast.com)

Now the range suggested by RPN will be given as an

input to another completely associated neural network
and speculate bounding boxes as well as object class.

Fig. Processed image

Overall, it have been found that the accuracy of the system

tends to be 93% i.e., the system is capable of identifying the
cost of the notes as well as the country of origin of most of the
notes accurately and correctly. This is a much better result.
Fig . Model training

C. Testing the model:- V. CONCLUSION

A system have been developed for the accurate identification
As we have trained our neural network on the custom of country as well as the denomination of the note. The dataset
dataset, now the model is ready to predict the output. upon which our system is trained, comprises of some of the
most common currencies, but this dataset can easily be
IV. RESULT increased by using the same approach we have previously
To recognize the nation of currency, the banknotes have been described. As observed that our system is capable to identify
segmented into the groups on the basis of blank region’s the denomination and country of the currency in approx. 5-6
presence and absence. Then, the banknotes will be classified seconds, which is considered as an improvement. As there are
by performing the template matching from the already stored more than 150 currencies present in the world, but our system
templates that identifies the banknotes from the countries in is concerned about few of the most commonly used
the each group. For demonstration purpose, the Indian 500 currencies, but the number of different currencies to be
rupee note is chosen. It is shown in below figures, before and recognized can be increased. Our system is not much efficient
after preprocessing steps. to recognize the damaged currencies which can be worked on
in future.
VI. REFERENCE

[1] Vedasamhitha Abburu “Currency Recognition System

Prof. Sagar S.Rajebhosale”, 2017, Tenth International
Conference on Contemporary Computing (IC3) .
[2] Prof. Sagar S.Rajebhosale, March 2017, “Currency
Recognition System Using Image Processing”,
International Research Journal of Engineering and
Technology (IRJET), Vol.1, Issue 03.
[3] Dr. Ritu Vijay, Vipin Kumar Jain, January 2013, Indian
Fig. Original note Currency Denomination Identification Using Image
Processing Technique IJCSIT, Vol.4, issue 1,pp.126-128.
Once the image is done with preprocessing to recognize the [4] Mahesh Pawar, Ketan Desai, “Object detection
country of origin of note, then we perform the classification on andCurrency Recognition using CNN”, Cikitusi Journal
the templates of all the nations within its group. for Multidisciplinary Research.
[5] Rubeena Mirza,Vinti Nanda,February 2012, Paper
Currency verification System Based On characteristic

541
 International Journal of Engineering Applied Sciences and Technology, 2020
Vol. 5, Issue 1, ISSN No. 2455-2143, Pages 539-542
Published Online May 2020 in IJEAST (http://www.ijeast.com)

Extraction Using Image Processing, IJEAT, Vol.1,Issue

03, pp.68-71.
[6] Padam Kumar, Harish Agarwal, “ Indian Currency Note
Denomination Recognition in Color Image, Int. Journal
on Advanced Computer Eng. AndCommunication Tech.
Vol.1.
[7] JAmudhavel, Raswitha Bandi, Object Recognition Using
Keras with Backend Tensor Flow.
[8] W. Li, Y. Li and K. Luo, 2008 "Application of image
processing technology in paper currency classification
system", Pattern Recognition 2008. CCPR ‘08. Chinese
Conference on, pp. 1-5.
[9] F.Takeda , T. Nishikage, 2000 ,”Multiple kinds of paper
currency recognition using neural network and application
of Euro currency”, IEEE International Joint Conference
on Neural Networks, Como, pp. 143-147.
[10] Ms.Trupti Pathrabe and Dr. N.G.Bawane,2010, Paper
Currency Recognition System Using Characteristics
Extraction and Negativity Correlated NN Ensemble, Int.
Journal of Latest Trends in Computing.
[11] Binod Prasad Yadav, 2014 “Indian Currency Recognition
and VerificationSystem Using Image Processing”. In:
International Journal of Advanced Research in Computer
Science and Software Engineering 4.12.
[12] Muhammad Sarfraz, 2015 “An Intelligent Paper Currency
Recognition System”. In: Procedia Computer Science 65,
pp. 538–545.

542