to download project abstract


The currency has a great meaning in everyday life. Thus currency recognition has
gained a great interest for many researchers. The researchers have suggested diverse
approaches to improve currency recognition Based on strong literature survey, image
processing can be considered as the most widespread and effective technique of
currency recognition. This paper introduces some close related works of paper-
currency recognition. This paper has explained a variety of different currency
recognition systems. The applications have used the power of computing to
differentiate between different types of currencies with the appropriate layer n
Choosing the proper feature would improve overall system performance. The main
goal of this work is to compare previous papers and literatures through reviews these
literatures and identify the advantages and disadvantage for each method in these
literatures. The results were summarized in a comparison table that presented
different ways of reviewing the technology used in image processing to distinguish
currency papers.

Chapter 1

Despite a decrease in the use of currency due to the recent growth in the use of
electronic transactions, cash transactions remain very important in the global market.
Banknotes are used to carry out financial activities. To continue with smooth cash
transactions, entry of forged banknotes in circulation should be preserved. There has
been a drastic increase in the rate of fake notes in the market. Fake money is an
imitation of the genuine notes and is created illegally for various motives. These fake
notes are created in all denominations which brings the financial market of the
country to a low level. The various advancements in the field of scanners and copy
machines have led the miscreants to create copies of banknotes. It is difficult for
human-eye to recognize a fake note because they are created with great accuracy to
look alike a genuine note. Security aspects of banknotes have to be considered and
security features are to be introduced to mitigate fake currency. Hence, there is a dire
need in banks and ATM machines to implement a system that classifies a note as
genuine or fake. In the recent years, Soft computing techniques have been widely
used to solve problems that are difficult to solve using conventional mathematical
methods. Supervised learning techniques are widely used in classification problems.
This paper evaluates supervised machine learning algorithms to classify genuine and
fake notes, and compares algorithms on the basis of accuracy, sensitivity, and
specificity. Consider someone wants to deposit money in the bank. The notes that are
to be deposited are given to a human being to check for their authenticity. As the fake
notes are prepared with precision, it is difficult to differentiate them from genuine
ones. A recognition system must be installed to detect legitimacy of the note. The
system should extract the features of the note using image processing techniques.
These features will be given as input to the machine learning algorithm which will
predict if the note is true or fake. Supervised machine learning techniques such as
BPN and SVM were implemented. The dataset used to train these algorithms was
collected by extracting features from banknote images. The dataset also classifies all
the samples into a particular class i.e. genuine or forged. A comparative study of these
techniques with respect to their accuracy, sensitivity, specificity and precision rate is

Currently, the use of paper money remains one of the main options for the exchange
of products and services. However, one of the remaining problems is the detection of
counterfeit banknotes, which increasingly resemble originals, making it diffificult for
someone who is not an expert in the fifield to detect them. On the other hand, there
are machines for detecting counterfeit banknote; however, these are often expensive,
so the identifification and retention of counterfeits ends up falling on fifinancial and
government entities, with minimal community involvement.

In order to solve this problem and to present alternative solutions, in the state-of-
theart, there are proposals based on classical computer vision techniques. For example,
from histogram equalization, nearest neighbor interpolation, genetic algorithms and
fuzzy systems . However, the main problem of this type of methods is its low capacity
of generalization for new examples as well as its locuracy. Another group corresponds
to those methods based on deep learning w ac(DL) using convolutional neural
networks (CNNs), which have outperformed to the classic machine learning

and humans too in classifification tasks. Considering the current importance of the
CNNs in the fifield of computer vision, there are some proposals in the area of
banknote recognition and counterfeit detection. For example, transfer learning (TL)
with Histograms of Oriented Gradients for Euro banknotes , a YOLO net for Mexican

banknotes or custom CNN architectures for dollar, Jordanian dinar and Won
Koreano banknotes have been proposed. However, one of the main disadvantages

of proposals using CNNs that focus on fake banknote recognition is that there is no
clarity about which design strategy is more appropriate, either custom or by transfer
learning. When using transfer learning-based networks, there are many types of
patterns that the network has learned, but they are not specifific to the current task. On
the other hand, custom networks are trained with a much smaller dataset than the pre-
trained networks, but they specififically learn the patterns of this type of
classifification task. Another shortcoming found in the literature is that the impact of
the freezing point (FP) of the pre-trained network on the performance of the
classififier has not been analysed. According to the above, the main contributions of
this research are as follows:

• A methodology to identify the best freezing point in models by transfer learning, for
three different types of architectures: sequential, residual and Inception is proposed.

• A custom model inspired in AlexNet that has faster inference times in an
embedded system than models by transfer learning is proposed.

• A comparative study for the fake banknote recognition task between a custom
model and several models obtained by transfer learning, in terms of accuracy and
inference times is given. The rest of the paper is organized as follows. Section
presents the background of Convolutional Neural Networks and transfer learning.
Section shows the proposed system of image acquisition and the used dataset. Section
explains the proposed methodology for selecting the freezing point in the design by
transfer learning. Section presents the proposed custom model. Section shows the
results of the research. Finally, Section summarizes the work.

 Fake Currency has always been an issue which has created a lot of problems in the

 The increasing technological advancements have made the possibility for creating
more counterfeit currency which are circulated in the market which reduces the
overall economy of the country.

 There are machines present at banks and other commercial areas to check the
authenticity of the currencies.

 But a common man does not have access to such systems and hence a need for a
software to detect fake currency arises, which can be used by common people.

 This proposed system uses Image Processing to detect whether the currency is
genuine or counterfeit.

The system is designed completely using Python programming language.

 It consists of the steps such as grayscale conversion, edge detection, segmentation, etc. which
are performed using suitable methods.

Leave a Comment


No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *