Real Time Fake Note Detection using Deep

Convolutional Neural Network

2022 Second International Conference on Computer Science, Engineering and Applications (ICCSEA) | 978-1-6654-5834-4/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICCSEA54677.2022.9936084

Mangesh Ghonge Tejas Kachare Manisha Sinha

Associate Professor Department of Information Technology Department of Electronics and
Department of Computer Vishwakarma Institute of Technology Communication Engineering
Engineering Pune, India University of Engineering and
Sandip Institute of Technology and [email protected] Management
Research Center Jaipur, India
Nashik, India Siddharth Nigade [email protected]
[email protected] Department of Computer Engineering
Pune Institute of Computer Technology Sandeep Shinde
Siddharth Kakade Pune, India Department of Computer Engineering
Department of Computer Engineering [email protected] Vishwakarma Institute of Technology
Vishwakarma Institute of Technology Pune, India
Pune, India [email protected]
[email protected]

Abstract— Nowadays printing fake notes has become very II. LITERATURE REVIEW
easy due to the availability of high-tech printing machines and
advancements in color printing technology. Notes can be A lot of research is already going on regarding counterfeit
printed with maximum accuracy without any slip mistake so currency note detection. The above study discusses some
detecting fake notes in today’s world is almost unattainable. interesting research that was already done to solve this
Hence the fake notes in the market have reduced the value of problem.
the original currency. This counterfeit negatively affects
nations' wealth. So, today it is necessary to detect fake A. Fake notes detect pen
currency. All the currently available methods are based on A fake note detector pen is a simple device that is used to
image processing techniques. Finding fake currencies with detect whether the note is fake or real. There is ink that
these methods is less efficient, unreliable, and time-consuming. contains a tincture of iodine. If the note is fake, it will turn
The proposed work is based on the deep convolutional neural
black else on the other hand if the note is real, it will turn
network used to detect counterfeit currency through the mobile
application. The self-generated dataset is used to train the yellow [4]. But this technique is usually very inaccurate in
model. Further, it was tested using real-time images captured terms of performance. Hence it is not much useto detect fake
through the smartphone camera. The training and validation notes effectively.
accuracy is 96.66%, while the testing accuracy is 86.65%.
B. Image Processing
Keywords— Convolutional Neural Network, Computer vision, The most popular approach to solve this problem was
Image processing, Denomination image processing [5]. Many techniques in image processing
can help classify the given currency as fake or real. The first
I. INTRODUCTION (HEADING 1) step in this approach is to capture the image of a currency
In the last eight years, more than 3.70 lakh cases of note with the help of a camera. There are various security
counterfeit cash have been discovered in India. The features available in the note which can help to detect real
replacement of paper currency instead of coins has increased notes and fake notes. Hence the next step is to extract
this kind of activity. Although the Indian Government has different features from the given image. Both the easily
banned old 500 and 1000 rupees notes to tackle this problem observable features and the invisible features of the notes are
of fake currency, after demonetization, this crime of extracted. The above steps are then performed over the image
duplication of currency is going on. Indian Banks have to perform this task [5, 6, 4].
noticed that there is a high increment of 480% doubtful
transactions after the demonetization [1]. The authority of
printing new notes is only given to RBI. There are several • Image Acquisition
security features incorporated in the notes by RBI so that the • Image pre-processing
fake notes and real notes can be recognized by the normal
public [2]. But still, these fake notes are identical to the real • Edge Detection
notes that a normal person cannot easily detect with naked
• Feature extraction
eyes. Most people are unaware of these security features of
Indian currency notes which are stated by RBI. A • Image Segmentation
smartphone application will be employed in the proposed
system to determine whether the currency is false or real. The image acquisition device is used to capturethe image
Deep Learning models can effectively perform binary that will highlight all higher as well as lower-level features in
classification on two images [3]. The two classes in which the image.
the model will classify the image are real and fake. The C. VGG16 model
CNN architecture would detect the fake and real notes
without manually extracting features from them. The model The drawbacks that were faced by image processing
is trained using the self-generated dataset which will then techniques have been solved by this approach. The artificial
help to predict counterfeit notes.

978-1-6654-5834-4/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:14:47 UTC from IEEE Xplore. Restrictions apply.
 datasets are used to train and test the model and by using the trained with two thousand, five hundred, two hundred, and
Deep Convolutional Neural Network Feature Extraction of fifty Indian rupee note data sets. After learning the feature
the image is done [7]. This model is then incorporated into map, the network is ready to detect bogus cash in real-time.
the mobile application. Through this mobile application, any The suggested method efficiently detects fake currencies in
person can successfully detect whether the note is suspicious the denominations of 2000, 500, 200, and 50 while
just by uploading the image on the application. Below is the consuming less time. The Automatic Fake Currency
stepwise working procedure of the above-mentioned model. Recognition System (AFCRS) is described in [15] as a
• Image classification system for detecting counterfeit paper currency and
The basic goal is to divide the image into two categories. determining if it is genuine. AFCRS, the suggested
There are two types of classes: fake and real. The model is technology, may also be used as a smartphone application to
trained using datasets in such a way that it can label the image assist society to discern between false and genuine money
either in one of the categories which is either fake or real [4]. notes. The suggested system's accuracy can be improved by
• Transfer Learning using original fake notes, whereas the proposed method uses
While training the model on a particular dataset the photosfrom the Churan label of the children's bank.
weights and bias parameters are constantly updated after each
epoch [3]. When the training is complete accurate weights III. PROPOSED APPROACH
and bias values are used with new input to test whether it The approach that is used is as follows.
correctly classifies the given image or not [3]. This is known
as Transfer Learning. The drawback in this approach was that 1. There was no readily available large dataset hence, it was
for a case of 2000 rupees notes it was overfitting. Secondly, necessary to use a self-generate dataset. Considering the
only the phones with high-end processing power could be four categories, the real and fake for both 500 and 2000
able to best fit the model. rupees notes each category has 10,000 images. So total
The research in [8] uses Deep Learning to create a dataset is 40,000 images.
convolution neural network (CNN) model with the goal of 2. The dataset splitting is done in the ratio of 80% for the
detecting counterfeit notes on mobile devices such as training set and 20% for testing purposes. As a result,
smartphones and tablets. A self-generated dataset was used to 40,000 photos were taken into consideration. 32000
train and test the model. The CNN network receives images images are used in the training process, whereas 8000
captured using a smartphone camera. The results are images are used in the testing process.
promising, and they may be improved with more study and 3. As a pre-processing step, a median blur filter is applied
modifications to the Deep CNN model's design. The testing to every image before passing it to the model.
accuracy was 85.6 percent, whereas the training and 4. The CNN architecture will contain 5 Convolutional
validation accuracies were 98.57 and 96.55 percent, layers with one flattened layer and then there will be four
respectively. The suggested system in [9] proposes a method fully connected layers.
for verifying Indian banknotes. The ideas of image 5. Again, the training dataset is also divided into two parts.
processing are used to verify monetary notes. The The first set of 30,000 photos is used for training, while
characteristics of Indian currency notes are extracted in this the second set of 2000 images is used for cross-
article. The characteristics of the note are extracted using validation. By considering the above-designed
MATLAB software. The suggested system provides benefits architecture the model was trained using the dataset
such as ease of use and high-performance speed. The images. On the completion of this task, cross-validation
outcome will indicate whether the cash note is genuine. [10] accuracy is obtained.
used a MATLAB technique to create a false note detecting 6. The same pre-processing technique of median-blur filter
device. This study is based on the same topic of fake note is done on testing images.
detection and aims to provide a solution to the problem of Based on the results, a confusion matrix was created.
counterfeit money. [11] includes a comparison of picture
segmentation or thresholding, feature extraction, IV. DATASET GENERATION
classification, and ultimately selection methods. Also After the demonetization, the new 500 and 2000 notes
included is some analytical work that can be done using were introduced. These two notes contain a total of 17
certain current methods. security features. Out of those 17 security features, 4 are
Image Processing is used in the proposed system in [12] present on the back and 2 are specifically used for vision
to determine if the coin is real or counterfeit. The Python impaired persons.
programming language was used to create the entire system. The dataset is generated by holding the note against the
It entails stages such as grayscale conversion, edge detection, light so that most of the security features will be visible [2].
segmentation, and so on, all of which are carried out using By considering 2 categories of fake and real for 2
appropriate methodologies. The model in [13] is a technique denomination currencies 500 and 2000 total dataset is of 4
that uses the picture of fake cash notes to detect them in categories. The first is 500 real, the second 500 fake, the third
circulation in our nation. The model offers most users the 2000 real, and the fourth 2000 fake.
essential mobility and adaptability, as well as credible
accuracy for detecting fraudulent currencies. To make this
program portable and economical, we used image processing
and cloud storage. [14] examines the currency photos to
identify fraudulent cash. To learn the feature map of the
currencies, the transfer learned convolutional neural network
is

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:14:47 UTC from IEEE Xplore. Restrictions apply.
 The following process of data augmentation is done on the
images:
1) Rotation of image
First, the images are rotated from minus 15 degrees to plus
15 degrees. It will be incremented by one in each gap. The
rotation of the image will be done from the central axis [4].
The black pixels are used to fill the void created by rotation
and the axis of rotation is passed through the center of the
Fig. 1. (a). Real currency: Holding it against the image.
light.

2) Cropping of image
To reduce the unnecessary background all the images in
the dataset were cropped and the background was removed as
much as possible maintaining the size and dimension of every
image to equal [4]. With the data augmentation process,
40,000 images were generated from it. 90 images were
generated from one sing image which was a replication of
Fig. 1. (b). 500 Fake currency.
itself with minor changes. As a result of this approach, 10,000
photos for each category were created.

Fig. 1. (c). 2000 Real currency. Fig. 2. Data augmentation result.

• Security Features
The following were the primary security factors that
were considered while training this model,

1) Mahatma Gandhiji’s portrait is in the form of a watermark.

2) The second feature is the ‘See-through register’.
3) The third feature is Angular bleed.
4) Rectangle/circular-based raised print.
Fig. 1. (d). 2000 Fake currency.

Fig. 1. (a) This picture delineates the watermark of Considering these above-mentioned security features
Mahatma Gandhi in the right white; (b) Delineates clear white encouraging results were obtained through the model.
portion within the right, nonappearance of the security string,
precise breed lines, and raised print in the right corner; (c) V. METHODOLOGY
Delineates the watermark of Mahatma Gandhi in the right The main goal of the proposed system is to detect
white portion; (d) Portrays clear white portion in right. whether the note or currency is fake or not with the help of a
smartphone camera. It is a mobile application that can be used
• Real currency Dataset
by common people to detect whether a currency note is fake
In terms of real currency, there were initially 100 images of
simply by using this application. The actual methodology that
Rs.2000 and Rs. 500 notes. The Mahatma Gandhiji
is used by this system is shown in the form flowchart in the
watermark and the ‘see-through register' are evident in the
figure… At first, when the user opens the application, he/she
above figure. [3]. All the images of notes were taken with the
must either capture the image through the camera or select
help of a digital camera and holding it in front of light
the pre-captured image from the gallery of his/her
exposure. The frame of the image occupies about 80% to 90%
smartphone. The images will be stored on cloud firestorm in
area of the total image.
the firebase database as a backend system. Firebase is a real-
time database hence to obtain the results the images are sent
• Fake currency Dataset:
to this database. The image is then passed to the Deep
Real photos were used to train the model with false cash. Real
Convolutional neural network model and the results which
photographs of Rs.500 and Rs.2000 notes were captured, and
are produced by CNN will be displayed on the screen of the
the security elements of the real notes were disrupted using a
smartphone in the form of a text box in a fraction of seconds.
photoshop tool. So basically, to generate fake currency, the
As Firebase is a real-time database so results of real-time
security features of normal notes were manually removed [3].
images can be obtained easily. First, the images are taken
In this way, 100 fake notes images were generated of Rs.500
from
and Rs.2000 notes.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:14:47 UTC from IEEE Xplore. Restrictions apply.
 the smartphone camera is uploaded to the firebase database
and then that image is fed to CNN. When the results are
computed by the CNN concerning input data the output is
again pushed back to the database. Then these results are
displayed to the user through the database. Before feeding the
image directly to CNN it is first converted in 80 * 80 pixels
dimension and pre-processing are also performed on the
image.
• Convolutional layer:
The first layer of this model is a convolutional layer
which is used to extract the features from the input image. It
uses the images in the form of a matrix and then applies a
filter to them. This operation helps CNN to extract and learn
the features from the input image. While doing this it also
preserves the relationship between the neighboring pixels.
The features are extracted from a small square of input data.
RELU is the activation function that is applied to the output
of the convolution layer.
Fig. 2. Flowchart of the proposed diagram.
• Pooling layer:
Due to many parameters, the pooling layer is
implemented to reduce the number of computations required.
It reduces the number of learning parameters whicheventually
reduces the computational cost involved in convolution
operation. The max-pooling layer and – the average pooling
layer are two examples of pooling layers.

• Flatten layer
The flattened layer receives the feature map matrix,
which is obtained from the feature extraction. In the flatten
layer that feature matrix which is the input is then converted
into a single vector. Then this single vector is fed as an input Fig. 4. CNN Architecture model.
to the dense fully connected layer.
The above figure 4 elaborates on the architecture
• Dense Fully connected layer CNN model. The python libraries that were used to train this
Each node in a completely connected layer is linked to the model were Tensor-flow, OpenCV, and T-flite. As enough
node of the next layer. Most of the computational features are data was available to train the model the system did not
analyzed in this layer. employ the concept of transfer learning. The pre-trained
VI. FUZZY LOGIC exiting models like the VGG16, Alex-Net, and ImageNet,
were not used. Instead, the system architecture is comprised
The threshold value that is considered for a note to be of 5 convolution layers, followed by the flatten layer and
fake or real is a probability between 0 and 1. If the probability then the four fully connected layers. The activation function
of a note being fake is 0.55% then in the proposed system the used to select the output feature is used at each layer. A
concept of fuzzy logic is used. In fuzzy logic, a gamma is set regression model was used to classify the output image into
to 1 and the support range is between 0 4 to 0.6. So, if the two classes which are fake and real. The input image that is
output value comes in a range of 0.4 to 0.6 then the fuzzy fed to the model has the dimension 80 * 80 and has 3
logic will assign it the membership for either being fake or channels. The first convolution is performed with 32 filters
real. It specifies whether a note is real or fake. If the each having 3 channels. The filters are of size 5 * 5. Padding
membership is more than 1 then it is classified as fake and on of 2 is applied at the edge of the input image. The stride of 2
the other hand if the membership value is closer to 0 the note is also applied in vertical as well as horizontal directions
is real. while performing the convolution with every filter. In total as
32 filters are used hence, the output is 32 channels followed
by a max-pooling layer which has a block size of 5 * 5.
Similarly, the same process is applied to the other layers, and
the model is trained. The model's projected outcomes are
uploaded to a real-time database. The application will
immediately retrieve the results and display them on the
user's screen.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:14:47 UTC from IEEE Xplore. Restrictions apply.
 IX. CONCLUSION
Tremendous success has been achieved by deep
learning in image classification tasks. Instead of manually
checking the features of the notes to classify them. The Deep
learning CNN model extracts the features by itself and
performs the classification together within the single
framework. The generated dataset of the fake notes has
helped a lot to simulate the real-world situation of detecting
Fig. 5. Confusion matrix of Performance of the system.
a currency note as fake or real. This application is a
VII. TRAINING AND TESTING THE DEEP CNN smartphone user application so that any common person can
easily be able to detect whether the given note is fake or real.
After that, the model is trained and evaluated on the
The future scope of this project will be using different pre-
dataset using 10,000 photos from each category. The actual
trained models and making changes in their architecture to
500, the fake 500, the real 2000, and the fake 2000. There are
improve the accuracy. Also, the dataset can be increased so
40,000 photos in the.jpg format in the dataset. The overall
that a more accurate result can be obtained from the model.
dataset was divided into training and testing sets. 80% of the
Deep learning has picked up the colossal victory in
images in the dataset were used for training purposes and
picture classification assignments. The proposed design
20% of images were used for testing purposes. Total images
which is based on Profound CNN works as a highlight
were 40,000 among which 32000 images of Rs.500 fake, Rs.
extractor dispensing with the must apply picture handling
500 real, Rs 2000 fake, and Rs. 2000 real. The testing dataset
techniques and physically check the nearness of security
was made of 8000 images. The entire training time was 10
highlights within the note. The produced dataset has
hours to train the model. CNN architecture was the backbone
effectively made a difference conduct test and attempting to
for detecting counterfeit notes. Figure 3 shows the overall
imitate the real-world situation. The application built will be
performance of this system. 8000 times the model is tested to
valuable to any common individual to distinguish a fake note.
calculate the confusion matrix. The method gives the correct
The future scope incorporates attempting modern Profound
answer 6777 times out of 8000 times.
CNN structures to extend the exactness of the show.
VIII. RESULTS Expanding the dataset, so that the demonstration gets
prepared way better and produces way better comes about.
The accuracy of the testing was found to be 85.60
percent. The model's training and validation accuracy were
both about 97.65% and 95.45%, respectively. The recall REFERENCES
accuracy is 84% and the Precision is 87.80%. Pre-processing [1] The Financial Express, https://www.financialexpress.com
is done on the images by shrinking them into a size of 80* 80 [2] Reserve Bank of India, http://rbi.org.in/scripts/icbanknotesecurity.aspx
[3] Simonyan, K. and Zisserman, A. (2015) Very Deep Convolutional Networks
pixels as it was the requirement of this architecture as an for Large-Scale Image Recognition. The 3rd International Conference on
input. The notes in the images occupy 80% to 90% of the Learning Representations (ICLR2015). https://arxiv.org/abs/1409.1556
area. The results can be improved further by applying the [4] K, Manikandakumar & K, Gokul & Chidambaram, Ranganathan &
edge detection technique before passing the cropped image to Srikumar, Ramasundaram & R, Vijayakumar & Chandrasekaran, Naveen &
K, Amirthavalli. (2017). Barcode Currency Notes in Future Necessity and its
the model. Benefits. Journal of Computer Science & Systems Biology. 10.
10.4172/jcsb.1000238.
[5] M. R. Pujar.” Indian Currency Recognition and Verification using Image
Processing”, International Journal of Advance Research, Ideas and
Innovations in Technology vol. 3, pp. 175–180, 2018.
[6] P. P. M Deborah and P. CSoniya PrathapME, “Detection of Fake currency
using Image Processing,” IJISET -International J. Innov.
[7] Azra Yildiz, Ali Abd Almisreb, Šejla Dzakmic, Nooritawati Md Tahir,
Sherzod Turaev, Mohammed A. Saleh, International Journal of Advanced
Trends in Computer Science and Engineering, 9(5), September - October
2020, 8115 – 8122
[8] K. Kamble, A. Bhansali, P. Satalgaonkar and S. Alagundgi, "Counterfeit
Currency Detection using Deep Convolutional Neural Network," 2019 IEEE
Pune Section International Conference (PuneCon), 2019, pp. 1-4, DOI:
10.1109/PuneCon46936.2019.9105683.
[9] Tushar Agasti et al 2017 IOP Conf. Ser.: Mater. Sci. Eng. 263 052047
[10] P. P. Binod Prasad Yadav, C. S. Patil, R. R. Karhe, “An automatic
recognition of fake Indian paper currency note using MATLAB,” Certif. Int.
J. Eng. Sci. Innov. Technol., vol. 9001, no. 4, pp. 2319– 5967, 2008.
[11] Mandankandy, Arun & k.E, Kannammal. (2020). Fake currency detection:
A survey. Gedrag en Organisatie. 33. 622 -638. 10.37896/GOR33.04/405.

Fig. 6. Result of status shown to the user.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:14:47 UTC from IEEE Xplore. Restrictions apply.
 [12] Vidhi Roy, Gangey Mishra, Rahul Mannadiar, Sushant Patil [14] M. Laavanya, V. Vijayaraghavan, 2019. Real Time Fake Currency Note
International Journal of Computer Science and Mobile Computing, Detection using Deep Learning. URL: https://www.ijeat.org/wp-
Vol.8 Issue.4, April- 2019, pg. 88-93 content/uploads/papers/v9i1s5/A10071291S52019.pdf
[13] Ankush Singh, Ankur Pandey, Aman Tekriwal, Prashant Mankani, [15] KG Navya, GS Pooja, BN Ram, VY Radha, and P Rajarajeswari,
Prof. Ketaki Bhoyar, 2019, Detection of Fake Currency using Image “Recognition of Fake Currency Note using Convolutional Neural
Processing, INTERNATIONAL JOURNAL OF ENGINEERING Networks”, International Journal of Innovative Technology and Exploring
RESEARCH & TECHNOLOGY (IJERT) Volume 08, Issue 12 Engineering, Vol. 8, No. 5, pp. 58-63, 2019.
(December 2019),

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:14:47 UTC from IEEE Xplore. Restrictions apply.