FAKE CURRENCY DETECTOR

Aprameya Dash- 191IT209 Suyash Chintawar- 191IT109

Information Technology Information Technology
National Institute of Technology Karnataka National Institute of Technology Karnataka
Surathkal, India 575025 Surathkal, India 575025
Email: aprameyadash.191it209@nitk.edu.in Email: suyash.191it109@nitk.edu.in

Maheshwari Mihir Premjibhai- 191IT129 Kunal Singh Lohiya- 191IT128

Information Technology Information Technology
National Institute of Technology Karnataka National Institute of Technology Karnataka
Surathkal, India 575025 Surathkal, India 575025
Email: mihirpm.191it129@nitk.edu.in Email: kunalsingh.191it128@nitk.edu.in

Abstract—Counterfeiting of currency has become a real threat system provides a hassle-free way to authenticate currency
to the livelihood of people as well as the economy of our country. notes quickly and accurately. This automated system can
Though fake currency detectors are available, they are restricted
replace the existing manual methods and can be used by
to banks and corporate offices leaving common people and small
businesses vulnerable. So, in this project, we will investigate the anyone easily to detect fake currency.
various security features of Indian currency and then, prepare
a software-based system to detect and invalidate fake Indian
currency by using advanced image processing and computer A. Commonly Used Security Features to Detect Fake Notes
vision techniques. This currency authentication system is de-
signed completely using Python language in Jupyter Notebook 1) Bleed lines: There are angular bleed lines on 500 and
environment. 2000 note on left and right corner of note in raised print.In
Keywords—Fake currency, counterfeit detection, image pro- 500Rs. note there are 5 bleed lines and In 2000Rs. note, 7
cessing, feature extraction, Bruteforce matcher, ORB detector
bleed lines.
I. INTRODUCTION 2) Security Thread: A colour-shifting security thread with
the inscription Bharat (in Hindi).RBI and 2000 (500 for 500
Currency duplication or production of counterfeit currency
note). The colour changes from green to blue when it is tilted.
notes illegally by imitating the actual manufacturing process
3) Latent Image : Latent images of number 2000 / 500 can
is a huge problem that every country is facing. Fake currency
be seen when note is held at 45 degrees angle.
can reduce the value of real money and cause inflation due to
an unauthorized and unnatural increase in the money supply. 4) Water mark: Watermark of Mahatma Gandhi and elec-
Manual authentication of currency notes is a solution but it trotype of numeral 2000/500.
is a very time-consuming, inaccurate, and difficult process. 5) Denominational Numeral: See-through register, with
Automatic testing of currency notes is, therefore, necessary denominational numeral 2000, can be seen when you hold
for handling large volumes of currency notes and then, getting the note against the light.
accurate results in a very short time span. In this project, we 6) Portrait of Mahatma Gandhi: Portrait of Mahatma
propose a fake currency note detection system using various Gandhi wiht RBI written on his spectacle which can be read
image processing techniques and algorithms. using a magnifying glass
7) Number panel: Numerals growing from small to big
The proposed system is designed to validate Indian size, is printed on the top left side and bottom right side.
currency notes of denomination 500 and 2000 rupees. The 8) Denominational numeral: On left side of Mahatma
system consists of three main algorithms and checks the Gandhi there is a 500/2000 in Devnagari script.
authenticity of various features in a currency note. The 9) Ashoka Pillar: There is Ashoka Pillar on right bottom
first algorithm consists of several steps including image side.
acquisition, pre-processing, greyscale conversion, feature 10) Guarantee and promise clause: The guarantee and
extraction, image segmentation, comparisons of images and promise clause of RBI is present in devnagri as well as in
output, and uses advanced image processing methods such english in the top left and top right corner of the currency
as ORB and SSIM. The second algorithm authenticates the notes respectively.
bleed lines of the currency notes whereas the third algorithm 11) RBI seal: The seal of RBI is present just below the
authenticates the number panel of the currency notes. Finally, Governor’s signature. This seal as well as the guarantee
the processed output is displayed for each currency note. This clause, etc are in intaglio printing.
 12) Denominational value in words: The denominational note detection unit with image processing algorithms. The
value of the currency note is written in devnagiri script in the experimental results indicate that the results achieved are
top central region of the note. nearly accurate.
II. LITERATURE SURVEY
[5] In this paper the recognition and verification of the
[1] In this paper an automatic system is designed for paper currency is explained with the use of image processing
identification of Indian currency note is fake or original. The techniques. The proposed approach consists of multiple
automatic system is very useful in the banking system and other element transactions like Image Acquisition, Feature
fields also. In India increase in the fake currency notes of 100, 500, extraction and comparison, Texture features, and Voice
and 1000 rupees. As there are increases in technology, like output.This system is divided into two parts. The first part
scanning, colour printing, and duplicating, there is an increase in the is to identify the currency denomination through image
counterfeit problem. In this model first the image acquisition is processing. The second part is the oral output to notify the
done and applies to pre-process to the image. In pre-processing visually impaired person about the denomination of the note
crop, smooth and adjust then convert the image into grey colour that he/she is currently having. The desired results will be in
after conversion applies the image segmentation then extract text and voice output of the currency recognized and verified.
features and reduce, finally comparing image.
III. PROBLEM STATEMENT
[2] In this paper, An automatic recognition of fake Indian To test the authenticity of Indian currency notes by prepar-
paper currency note using MATLAB is implemented and feature ing a system which takes the image of currency bill as input
extraction with HSV color space and other applications of image and gives the final result by applying various image
processing. The proposed architecture is as follows Image processing and computer vision techniques and algorithms.
Acquisition, Gray-Scale Conversion, Edge Detection, Image
Segmentation, Characteristic Extraction, Comparison, Output. In the A. Objectives
project setup, a note is placed in front of the camera to check • The main objective of the project is to identify the fake
whether it is fake or genuine. The clicked pictures of notes are Indian currency notes through an automated system by
analyzed by MATLAB program installed on the computer. This using Image processing and computer vision techniques.
algorithm works for Indian denomination 100, 500, 1000. If the note • The system should have high accuracy.
is genuine, the respective message is appeared on the screen and • The system should be able to give the final results in a
vice-versa. short time.
• The system should have an User-friendly interface, to
[3] In this paper, a hybrid fake currency detection model make it convenient to use and understand.
was proposed and implemented on MATLAB. The model was
designed to detect fake Bangladeshi notes. In the proposed model, IV. METHODOLOGY
three image processing algorithms were chosen namely Optical A. Preparation of Dataset
Character Recognition(OCR), Hough Transformation and Face
• The first step is the preparation of a dataset containing
recognition(MSD) to get enhanced results. Then, the results of the
images of different currency notes (both fake and real)
proposed model were compared with each model used. The
and images of different features of each of the currency
algorithm followed six basic steps: data collection, pre-processing
notes
of the collected data, edge detection, feature extraction,
• The dataset will contain the following repositories:
identification and output results. In addition, although the
proposed model took slightly longer to process, it displayed 93.33% – Sub- dataset for Rs. 500 currency notes
accuracy which was significantly higher compared to the individual 1) Images of real notes
algorithms applied. 2) Images of fake notes
3) Multiple images of each security feature (tem-
[4] In this paper two ways are used: one by using analysis plate)
through hyper spectral imaging and the second one is the Extraction – Sub- dataset of Rs. 2000 currency notes (Similar
of different features in fake and real currency notes and through structure)
comparing with each other we can able to differentiate the fake note • The various security features that we are considering are:
from the real note. The different colour lights used for hyper (for Rs. 500 currency notes- Total 10 features)
spectral imaging are Ultraviolet (UV) light, Normal LED Bulb, Red – Rs. 500 in Devanagari and English script (2 features)
LED light, Green LED light and Blue LED light with different – Ashoka pillar Emblem (1 feature)
wavelengths ranging from 360 nm to 800nm respectively. This all
– RBI symbols in Hindi and English (2 features)
modules are implemented in MATLAB. They have implemented
– 500 rupees written in hindi (1 feature)
a fake
– RBI logo (1 feature)
– Bleed Lines on Left and right side (2 features)
– Number Panel (1 feature)
 TABLE I: Summary of Literature Survey
Authors Methodology Merits Limitations

Sonali R. Darade [1] Feature extraction and im- Detection of note is good External camera is used
age processing Cost is low
Binod Prasad Yadav, P.H Feature extraction with effective and efficient im- Whole setup required
Patil [2] HSV image processing age processing
Adiba Zarin ,Jia Uddin [3] Optical Character recogni- 93.33 % accuracy Hard method
tion (OCR)
Shripad veling [4] Hyperspectral Imaging Two ways to get result Cost is high ,complicated
Dr. P. Mangayarkarasi, Image processing, Image cost and time efficient If note is dirty and torn
Akhilendu, Anakha A S Acquisition, Feature ex- than it will give wrong an-
[5], Meghashree K, Faris traction swer
AB

B. Image Acquisition
Next, the image of the test-currency note is taken as input
and fed it into the system. The image should be taken from
a digital camera or preferably, using a scanner. The image
should have a proper resolution, proper brightness and should
not be hazy or unclear. Blurred images and images with less
detail may adversely affect the performance of the system.

C. Pre-processing Fig. 1: ORB Feature detection and Matching

Next, the pre-processing of the input image is done. In
this step, first the image is resized to a fixed size. A fixed 2) Feature Extraction: Now, using ORB location of each
size of image makes a lot of computations simpler. Next template has been detected in the input image within the high-
up, image smoothening is performed by using Gaussian lighted area. The highlighted area is then cropped by slicing
Blurring method. Gaussian blurring removes a lot of noise the 3D pixel matrix of the image. Next, we apply Gray scaling
present in the image and increases the efficiency of the and Gaussian blur to further smoothen the image and now our
system. feature is ready to be compared with the corresponding feature
in our trained model.

D. Gray- scale conversion

Gray scale conversion is mainly used because an RGB im-
age has 3 channels whereas a gray image has only one
channel. This makes the computation and processing on
images much more easier in the case of gray scaled images.

E. Algorithm- 1 : For feature 1- 7

1) Feature detection and matching using ORB: After com- Fig. 2: Features in 500 | currency bill
pleting the necessary processing of the image, feature detec- tion
and matching is done using ORB. Our dataset already contains the
images of different security features present in a currency note
(total 10). Further, we have multiple images of varying brightness
and resolutions corresponding to each security feature (6 templates
for each feature). Using the ORB algorithm,each security feature is
detected in the test image. To make the searching of the security
feature (template image) easier and more accurate, a search area
will be defined on the test currency image where that template is
most likely to be present. Then, ORB will be used to detect the
template in the test image and the result will be highlighted properly Fig. 3: Features in 2000 | currency bill
with a marker. This process will be applied for every image of
each security feature present in the data-set and every time the 3) Feature comparison using SSIM : From the previous
detected part of the test image will be highlighted properly using step, the part of the test currency image which matches with
proper markers. each of the templates will be generated. In this method, the
 Image Acquisition

Pre-processing
Algorithm-1,Features: 1-7 Gray-scale conversion Algorithm-3,Feature: 10

Feature Detection and Matching using ORB Algorithm-2,Features: 8-9

Feature Extraction
Thresholding
Feature Extraction Feature Extraction

Image Contour Detection

Thresholding

Feature
Comparison
Finding
Calculation of number of black regions (bleed lines) Bounding

Assigning SSIM scores to each Feature

Eliminating
erroneous
Rectangles

Counting remaining
rectangles (Number of
characters)
Fig. 4: Flow Diagram
Display Output

Result Analysis
 original template will be compared with the extracted feature characters). This algorithm performs various operations and
and then a score will be given for the similarity between the finally counts the number of characters present in the number
two images using SSIM. panel.
(2µxµy + c1)(2σxy + c2)
SSIM(x, y) = (µ2 + µ2 + c1)(σ2 + σ2 + c2)
x y x y

The Structural Similarity Index (SSIM) is a scoring

system that quantifies the image quality degradation that is
caused by processing such as data compression or by losses
in data transmission. Basically it looks for similarity between
Fig. 5: Number Panel Detection
two images. It is a part of skimage library and uses the
1) Image Thresholding (with multiple values): The 1st step
above mentioned formula to calculate similarity. It returns
in this algorithm is again thresholding with suitable value so
a value between -1 and 1. Closer the SSIM to 1, higher
that only the black characters remain in the number panel on
is the similarity. So, for every security feature, the SSIM
a white background and become easy to detect. But in this
value between each image of that security feature and the
algorithm, thresholding is done using multiple values, i.e. first
corresponding extracted feature from the test image will be
the image is thresholded at the initial value (90), then all other
calculated. Then, the mean SSIM for each security feature is
steps mentioned below are done and number of characters
calculated and stored.
are calculated. After that, the threshold value is increased
F. Algorithm 2: For feature 8 and 9 by 5 every time and the process of calculation of number
Every currency note contains bleed lines near the left and of characters is repeated till either we reach the final value
right edges. There are 5 lines in case of 500 currency note (150, in our case) or we find enough proof that 9 characters
and 7 lines in case of Rs. 2000 currency near each of the two are present in the number panel.
sides. This algorithm is being used to count and verify the 2) Contour Detection: In the 2nd step, contour detection
number of bleed lines present in the left and right sides of a of the thresholded image of number panel is done.
currency note. (feature 8 and 9) 3) Finding Bounding Rectangles: In the 3rd step, the
bounding rectangle for each contour is found. The details of
1) Feature Extraction: In the first step, the region in which each rectangle is put inside a list.
the bleed lines are present are extracted by cropping the image. So, a 4) Eliminating erroneous rectangles: The list of rectangles
part near the left and right edges of the input currency note image is computed in the previous step may contain a number of
carefully extracted. erroneous and unnecessary rectangles due to noise present in
2) Image Thresholding: In the 2nd step, the image is the image. These erroneous rectangles need to be eliminated.
thresholded using a suitable value. This ensures that only the black So, in this step, all rectangles whose area is either too big
bleed lines remain on a white background and makes further or too small are eliminated. Then, the rectangles which are
processing quite easy. bound by a bigger rectangle are also eliminated. Finally, those
3) Calculation of number of bleed lines: The 3rd step rectangles which are positioned completely too high in the
involves calculation of number of bleed lines. In this step, first we number panel are also eliminated.
iterate over each column of the thresholded image. Then we iterate 5) Calculation of number of characters: The rectangles re-
over each pixel of each column. Then, we calculate the number of maining after the previous elimination step are those
black regions in each column by increasing a counter whenever rectangles which bound only each character of the number
current pixel of the column is white and the immediate next pixel is panel. The number of rectangles still remaining is calculated
black. Similarly we, count number of black regions for each and this gives us the number of characters detected in that
particular thresholded image.
column, but, if the number of black regions is too large (>= 10),
The above process is repeated for multiple threshold values
then that column is erroneous and it is discarded. Finally, the
(starting form 90 or 95 and increasing by 5 in each iteration).
average count of black regions is calculated by considering the non-
The algorithm stops if either it detects 9 characters in three
erroneous columns only and the result is displayed as the number of
consecutive iterations or if the threshold value reaches the
bleed lines. This count should be approximately 5 for Rs 500
maximum value (150 in our case).
currency notes and 7 for Rs 2000 currency notes.
H. Displaying Output
G. Algorithm 3: For feature 10
Finally, the result of all algorithms is displayed to the user.
Every currency note contains a number panel in the bottom The extracted image of each feature and the various important
right part where the serial number of the currency note is data collected for each feature is displayed properly in a GUI
displayed. The number of characters present in the number window. Further, the status (Pass/ Fail) of each feature is
panel should be equal to 9 (neglecting the space between the displayed along with the details. Finally the total number of
features that have passed successfully for the input image of
Fig. 6: Intially no image is dislayed
and user is asked to insert image

Fig. 7: Browsing image

Fig. 8: Input image of currency note

Fig. 10: GUI showing final result(Real note)

Fig. 9: Image sent for processing...
 • Algorithm 1 (Feature 1-7): This algorithm finally col-
lects the average SSIM score and the max. SSIM score
for each feature. A feature passes the test and is real if
the average SSIM score is greater than a minimum value
(this value has to be decided after proper testing). A
feature also passes the test if the max. SSIM score is too
high (probably greater than 0.8).
• Algorithm 2 (Feature 8-9): This algorithm finally
returns the average number of bleed lines present in the
left and right sides of a currency note. Each feature
passes the test if the average number of bleed lines is
closer to 5, in case of Rs 500 currency note, and 7, in
case of 2000 currency note.
• Algorithm 3 (Feature 10): This algorithm finally returns
the number of characters present in the number panel
of the currency note. This feature passes the test if the
number of characters detected is equal to 9 (for at least
one threshold value).
Fig 6 to Fig 11 are snapshots of the GUI used in our
implementation.
A. Performance Analysis
The performance analysis of the proposed system was
carried out using various images of currency notes. As we
had already created a dataset for both fake and real currency
notes of denominations of 500 and 2000, all the notes were
tested and then the accuracy was calculated. For the sake of
calculating the accuracy, it was assumed that if the currency
note passed at least 9 features out of 10 then the note is real
otherwise it is fake. Testing of both real and fake notes was
done separately.

• For testing of real notes, 9 real notes for Rs.2000 and

10 real notes for Rs. 500 were considered, out of which
15 of the total 19 notes gave the correct desired results.
Accuracy: 79%
• Similarly, for testing fake notes, 6 fake notes were taken
into consideration for each denomination (12 notes in
total), for which 10 of the 12 notes gave the correct
required output. Accuracy: 83%
The accuracy for both real and fake currency notes was
calculated separately and the result is shown in the a bar graph
(Fig 13).
Fig. 11: GUI showing final result(Fake note) B. Time Analysis
The proposed system has been implemented using Python
programming language in Jupyter Notebook environment.
currency note is displayed and based upon that it is decided
Along with the final results, a lot of images and other analysis
whether the note is fake or not. The entire GUI is programmed
related data is being printed by our system. If all the data
in python itself using tkinter library.
(consisting of 100+ images regarding the examination of each
V. RESULTS AND ANALYSIS feature) is printed, then the system takes about 35 seconds to
The proposed system authenticates the input image of cur- process, print all the data and give the final results. If only the
rency note through image processing. The input image passes final results are displayed after processing of input image, the
through various algorithms in which the image is processed system takes only 5 seconds for each input image.
and each extracted feature is thoroughly examined. The results So, practically the model takes about 5 seconds to give the
are calculated in the following manner: results of each input image.
 Fig. 12: Gantt Chart

VI. CONCLUSION
In this paper, a fake currency detection model has been
proposed for authentication of Indian currency notes of de-
nomination 500 and 2000 and implemented using OpenCV
image processing library in Python3. In this model, 10
features of the input currency note are considered and then
analyzed using 3 different algorithms. The input image is
taken through a GUI which allows the user to browse the
image in his/ her system. Then the results of the implemented
model are computed and the analysis of each feature is
displayed in detail through a graphical user interface (GUI)
created using Tkinter GUI library.
The model takes less time (about 5 sec- when only final
results are shown leaving unnecessary details) for processing
an input image. The results are also quite decent giving almost
79% accuracy in detecting genuine currency and 83%
accuracy in detecting counterfeit currency.

ACKNOWLEDGMENT
We would like to express our deep sense of gratitude to Ms
Thanmayee S for sharing her expertise as a guide every
week. With deep gratitude, we would like to express our
thanks to Professor Ram Mohana Reddy Guddeti, for his
kind co- operation, valuable guidance. Further, we would also
Fig. 13: Performance Analysis Graph
like to express our sincere thanks to all the faculties of the
department
for their valuable guidance during the development of this
project.
INDIVIDUAL CONTRIBUTION
The Gantt chart of this project has been displayed in Fig.
12. The table mentioned below displays the individual
contribution (Fig14)

Fig. 14: Individual Contribution

IMPLEMENTED/BASE PAPER
The initial idea and inspiration for our proposed system has
been taken from the paper mentioned below. However, the
methodology of this paper is quite different from our proposed
system.
Paper name: Fake Indian Currency Note Recognition
Authors: Mangayarkarasi, P and Akhilendu and A S, Anakha
and K, Meghashree and A B, and Faris
Published year: 2020
REFERENCES
[1] S. R. Darade and G. Gidveer, “Automatic recognition of fake indian
currency note,” in 2016 international conference on Electrical Power and
Energy Systems (ICEPES). IEEE, 2016, pp. 290–294.
[2] B. P. Yadav, C. Patil, R. Karhe, and P. Patil, “An automatic recognition
of fake indian paper currency note using matlab,” Int. J. Eng. Sci. Innov.
Technol, vol. 3, pp. 560–566, 2014.
[3] A. Zarin and J. Uddin, “A hybrid fake banknote detection model using
ocr, face recognition and hough features,” in 2019 Cybersecurity and
Cyberforensics Conference (CCC). IEEE, 2019, pp. 91–95.
[4] M. S. Veling, M. J. P. Sawal, M. S. A. Bandekar, M. T. C. Patil, and
M. A. L. Sawant, “Fake indian currency recognition system by using
matlab.”
[5] F. A. B, P. Mangayarkarasi, Akhilendu, A. A. S, and M. K, “Fake
indian currency note recognition,” vol. 7, pp. 4766–4770, 2020. [Online].
Available: https://www.irjet.net/archives/V7/i5/IRJET-V7I5915.pdf