COUNTERFEIT INDIAN CURRENCY DETECTION WITH

CNN: A DEEP LEARNING APPROACH

Yes, the proposed title, "Counterfeit Indian Currency Detection with CNN: A Deep Learning
Approach," is clear and concise. It effectively highlights the purpose and methodology of the
study. It’s a strong title for a research paper on this topic.

1
R. Tamilkodi, 2 S. Ratalu, 3 R. Ravichand, 4 A. Janardhan,
5
V. Niranjan, 6 A. Deekshith Kumar
1
Professor, Department of CSE (AIML & CS)
2
Professor, Department of CSE (AIML & CS)
12
Godavari Institute of Engineering & Technology, Rajahmundry, Andhra Pradesh, India
3456
Department of Computer Science & Engineering (AIML & CS)
Godavari Institute of Engineering & Technology, Rajahmundry, Andhra Pradesh, India
[email protected], [email protected], [email protected],
[email protected], [email protected],
[email protected]

Abstract. The increasing prevalence of counterfeit currency poses a serious

challenge to economic stability, making advanced detection techniques crucial.
This project, titled "Detection of Fake Indian Currency Using Convolutional
Neural Networks", presents an innovative method for identifying counterfeit
notes through deep learning. The research investigates three core models: Mo-
bileNet, a hybrid model combining MobileNet with Support Vector Machines
(SVM), and a more complex hybrid that integrates MobileNet, SVM, and Ran-
dom Forest. The project leverages MobileNet for their known efficiency and
accuracy in image classification, assessing their effectiveness in distinguishing
between genuine and counterfeit Indian currency. By merging MobileNet with
SVM, the hybrid model aims to improve detection accuracy and manage the
complexity of counterfeit note patterns. The enhanced version, integrating both
SVM and Random Forest, seeks to further boost classification accuracy using
ensemble learning techniques. The findings underscore the potential of convo-
lutional neural networks in advancing counterfeit currency detection, offering
crucial insights for enhancing financial security systems.

Keywords: Counterfeit Detection, CNN, MobileNet, SVM, Random Forest,

Hybrid Model, Indian Currency, Image Classification, Machine Learning.

1 INTRODUCTION
The issue of counterfeit currency is a growing concern globally, and it significantly
impacts countries like India, where fake currency circulation disrupts economic stabil-
ity. The presence of counterfeit money weakens the financial system, causes inflation,
 2

damages consumer trust, and results in economic losses. Traditional counterfeit detec-
tion methods depend heavily on manual inspection, which is often slow, labor-inten-
sive, and subject to human error. Furthermore, counterfeiters are constantly improving
their methods, making it harder for conventional detection techniques to stay effective.
To tackle this issue, there is a growing emphasis on employing cutting-edge technolo-
gies like Artificial Intelligence (AI) and deep learning to enhance the automation and
precision of counterfeit detection. They can learn intricate patterns from images, mak-
ing them ideal for detecting small yet significant differences between real and counter-
feit currency. CNNs’ ability to autonomously extract features from images and perform
high-accuracy classification presents a promising solution to the challenge of detecting
counterfeit notes. Aims to develop and evaluate an AI-driven system for identifying
fake currency. The study implements three models: (1) MobileNet, a lightweight CNN
framework optimized for mobile and embedded systems; (2) a hybrid model combining
MobileNet and Support Vector Machines (SVM); and (3) an advanced hybrid model
that integrates MobileNet, SVM, and Random Forest. These models aim to overcome
the limitations of traditional methods by offering increased detection accuracy and
faster processing times. MobileNet was selected for its efficiency and balance between
computational load and accuracy. It is well-suited for mobile applications like ATMs,
point-of-sale systems, and mobile banking, where resource constraints are common.
MobileNet uses depthwise separable convolutions to reduce computation, making it
ideal for large-scale use without compromising on performance. The second model, a
hybrid approach combining MobileNet and SVM, is introduced to enhance classifica-
tion accuracy. SVM, a robust machine learning algorithm, excels in binary classifica-
tion tasks and helps to focus on key features extracted by MobileNet.
To improve performance, optimize by fine-tuning hyperparameters and using a high-quality,
diverse dataset. Employ advanced techniques like regularization and transfer learning for better
results.

2 LITERTURE SURVEY
Counterfeit currency is a global issue with significant economic repercussions. In India,
the problem has become more acute, calling for advanced detection mechanisms [1].
Traditional methods of counterfeit detection, mainly reliant on manual verification, are
often slow, error-prone, and difficult to scale. As a result, researchers have increasingly
turned to automation, which offer better accuracy, scalability, and [2] efficiency. One of
the most promising techniques in this field is Convolutional Neural Networks (CNNs),
known for their strength in image classification, which has become widely applied in
detecting counterfeit currency, making them [3] particularly suitable for tasks involving
intricate visual details like currency notes. CNNs reduce the need for manual feature
extraction, enabling models to learn sophisticated patterns that differentiate genuine notes
from counterfeit ones. Initial research focused on simple CNN models [4] applied to vari-
ous datasets of currency notes, successfully classifying them as real or fake. However,
issues such as computational demands and the necessity for large, varied datasets posed
challenges in real-world implementation, particularly due to inconsistencies in currency
wear, print quality, and other factors. A major leap in the field came with the introduction
of MobileNet, a streamlined CNN architecture optimized for use in mobile and embedded
 3

devices. MobileNet significantly lowers [5] computational costs while maintaining high
accuracy, making it a promising candidate for counterfeit currency detection. It has
demonstrated impressive performance in image classification tasks across various do-
mains. Its ability to process complex, high-dimensional data with minimal resource con-
sumption makes [6] MobileNet especially suited for counterfeit detection systems. Stud-
ies applying MobileNet to counterfeit currency classification have reported promising
accuracy, motivating its broader application in similar image-based tasks. In addition to
standard CNN models, hybrid techniques that combine CNNs with traditional machine-
learning classifiers have gained traction. These models [7] leverage CNNs' feature extrac-
tion capabilities alongside classifiers like Support Vector Machines (SVMs) and Random
Forests to enhance detection accuracy. SVMs, known for their efficiency in high-dimen-
sional spaces, complement CNNs by establishing a decision boundary that effectively
distinguishes between real and counterfeit notes. By combining CNNs with SVMs, hy-
brid models have achieved higher accuracy [8] compared to using either method individ-
ually. This integration enhances the system's ability to detect subtle differences in cur-
rency texture, colour, and print quality, which are often critical in identifying counterfeit
notes. The use of ensemble learning methods like Random Forest has also shown promise
in improving counterfeit [9] detection. Random Forest aggregates the results of multiple
decision trees, reducing overfitting and improving classification accuracy in noisy
datasets. This ensemble approach is particularly useful in counterfeit detection, where
currency variations can introduce significant noise into the data. The combination of
CNNs for feature extraction, SVMs for boundary setting, and Random Forest [10] for
classification has led to better performance, making the detection system more reliable
and accurate. To evaluate the effectiveness of these models, metrics such as accuracy,
precision, recall, and F1-score are used. Accuracy reflects the overall correctness of the
model, while precision and recall provide insight into its ability to correctly classify
counterfeit notes without mislabeling genuine ones. [11] Precision is especially important
in counterfeit detection, as false positives (classifying real notes as fake) can have serious
economic and social implications. Recall, on the other hand, ensures that counterfeit
notes are not missed. The F1-score, which balances precision and recall, is a key measure
for optimizing performance. Research in this area has shown that CNN-based and hybrid
models can achieve high accuracy, often surpassing 90% [12], in counterfeit currency
detection. However, challenges persist in ensuring these models are robust enough to
handle real-world conditions, such as worn-out notes or varying lighting conditions.
When deploying counterfeit detection systems, practical concerns like variations in cur-
rency design, printing errors, and natural wear pose additional challenges. Models trained
on clean [13] datasets might struggle with notes that are folded, crumpled, or faded. Thus,
it is crucial to have diverse and extensive training datasets that reflect real world condi-
tions. These techniques offer a scalable, efficient solution with high accuracy in distin-
guishing genuine currency from counterfeits. Nevertheless, ongoing challenges [15] re-
lated to dataset diversity, model robustness, and real-world deployment remain.
The literature review should cover existing methods like CNNs (LeNet, AlexNet, VGG) and their
limitations, such as overfitting and sensitivity to variations. It should also highlight the gap in han-
dling different counterfeit qualities and explain how the hybrid approach using MobileNet, SVM,
and Random Forest overcomes these issues.
 4

3 METHODOLOGY
3.1 Proposed Work:
The proposed system enhances the detection of counterfeit currency by integrating ad-
vanced Convolutional Neural Networks (CNNs) with hybrid machine learning tech-
niques. It employs MobileNet for efficient and precise image classification, alongside
Support Vector Machines (SVM) and Random Forest to utilize the benefits of ensemble
learning. The system processes currency images to detect subtle counterfeit markers,
providing a scalable and real-time solution. By overcoming the limitations of current
methods, the [16] system aims to offer a more dependable and effective approach to com-
bating counterfeit currency [17].

3.2 System Architecture

The Currency Guardian is a pioneering counterfeit detection system that harnesses the
power of artificial intelligence to secure financial transactions. By integrating advanced
image processing and machine learning algorithms, this innovative platform swiftly
verifies the authenticity of currency, shielding individuals, businesses, and financial
institutions from fraudulent activities.

The MobileNet is accurate and efficient but struggles with variations. Hybrid models
like MobileNet+SVM improve classification, while MobileNet+Random Forest boosts
robustness, trading off recall.

Fig 1: System architecture of fake Indian currency

 5

The database details are inadequate because it lacks sufficient variety and quality
data, which can limit the model’s accuracy. Expanding the dataset with more diverse
and labeled samples will enhance performance.

3.2.1 Data Visualization and Preprocessing

After examining the dataset, we find that every value is there. Due to the great varia -
tion in data, datasets must be normalized to prevent bias resulting from a single charac -
teristic.

After preprocessing, split the dataset into training, validation, and test sets, typically
allocating 70% for training, 15% for validation, and 15% for testing. Select suitable
models, starting with a CNN architecture for image classification, utilizing MobileNet
for its efficiency on mobile devices, and implementing SVM as a baseline model
alongside Random Forest for improved robustness. Consider creating a hybrid model
that combines CNN and SVM or other models for enhanced accuracy.
Normalization can be improved by applying data preprocessing techniques, such as scaling pixel
values to a standard range and employing augmentation to enhance data variability.
3.3 Algorithms Used

To detect counterfeit currency, three algorithms are utilized: MobileNet, SVM, and
Random Forest. MobileNet's lightweight CNN architecture efficiently analyzes cur-
rency images, extracting key features. SVM, a supervised learning technique, identi-
fies the ideal hyperplane to distinguish genuine and fake currencies, providing a foun-
dation for classification. Random Forest's ensemble approach constructs multiple
decision trees, enhancing accuracy and mitigating overfitting. By integrating these
algorithms into hybrid models, the system combines their advantages, boosting over-
all performance and adaptability in counterfeit detection.

MobileNet
An optimized neural network architecture for mobile and embedded devices is called
MobileNet. MobileNet delivers exceptional efficiency, allowing for quick processing
and low memory consumption, by utilizing depthwise separable convolution and
model compression.
 Prepare images: Collect, resize, and normalize images.
 Apply convolution: Use depthwise separable convolution.
 Normalize activations: Apply batch normalization.
 Introduce non-linearity: Use ReLU activation.
 Downsample features: Apply pooling layers.
 Classify: Flatten features, apply fully connected layer.
 Train model: Use stochastic gradient descent.
 Evaluate model: Assess accuracy and precision.
6

ReLU (Rectified Linear Unit) plays a key role by introducing non-linearity, helping the model
learn complex patterns. Clarifying its use involves explaining how it improves training speed
and avoids vanishing gradients.

Support Vector Machine (SVM)

A potent supervised learning method used for regression and classification applica-
tions is the Support Vector Machine algorithm. Finding the best hyperplane to divide
classes in feature space is the fundamental task of SVM.
 Prepare data: Collect, clean, and label data.
 Choose kernel: Select suitable kernel function (e.g., linear, polynomial, radial
basis function).
 Define hyperparameters: Set parameters (e.g., regularization, gamma).
 Train model: Find optimal hyperplane to separate classes.
 Calculate margins: Determine soft margin and support vectors.
 Predict classes: Use trained model to classify new data.
 Evaluate model: Assess accuracy, precision, and recall.

Random Forest
Using the combined power of several decision trees, Random Forest is an ensemble
learning technique that improves prediction accuracy and reduces overfitting.
Steps in the Random Forest Algorithm
 Prepare data: Collect, clean, and split data.
 Create bootstrap samples: Resample training data.
 Build decision trees: Construct trees on bootstrap samples.
 Combine predictions: Calculate weighted average or majority vote.
 Tune hyperparameters: Adjust parameters for optimal performance.
 Evaluate model: Assess accuracy and precision.

4 EXPERIMENTAL RESULTS

Dataset Used:
The collection has 247 photos in total, of which 126 show actual Indian cash and 121
fake. A scanner is used to manufacture fake Indian cash, while a cell phone camera is
used to create genuine Indian currency. The two sides of the coin are photographed
with a mobile device's lens. The counterfeit Indian cash is produced by the scanner
using the captured currency, and it is then printed using a printer on plain paper. Con -
sequently, there are 10, 20, 50, 100, 200, 500, and 2000 photos for every sort of cash.
Table 1 displays some photographs of actual and fraudulent money.
Image classification involves collecting and preprocessing a dataset of currency images, resiz-
ing, normalizing, and training models like MobileNet. The model classifies images as genuine
or counterfeit using learned features.

Table 1: Descriptive Statistics for a Dataset

 7

Variance Skewness Kurtosis Entropy Class

count 1372 1372 1372 1372 1372

mean 0.43374 1.92235 1.397627 -1.19166 0.44461

std 2.84276 5.86905 4.31003 2.101013 0.4971

min -7.0421 -13.7731 -5.2861 -8.5482 0

25% -1.773 -1.7082 -1.57498 -2.41345 0

50% 0.49618 2.31965 0.61663 -0.58665 0

75% 2.82148 6.81463 3.17925 0.39481 1
max 6.8248 12.9516 17.9274 2.4495 1

The analysis shows that MobileNet achieved the highest accuracy of 97%, outperforming hy-
brid models. The MobileNet+SVM hybrid reached 96%, while MobileNet+Random Forest had
94%, showing MobileNet’s superior performance for counterfeit detection.

Fig 3 below shows the real Indian currency data set and similarly Fig 4 shows fake
Indian currency data set.

Fig 3: Real Indian Currency Dataset

Fig 4: Fake Indian Currency Dataset

The results can be seen below as displayed in Fig 5 and Fig 6, it is possible to determine
how well the algorithms work to differentiate between authentic and fraudulent notes.
KNN produces the most ideal outcome out of the four algorithms tested, to identify money
notes. Of all the algorithms, the logistic regression produces the least ideal outcome.
8

Fig 5: Detecting Real Currency

Fig 6: Detecting Fake Currency

The table 2 shows the performance evaluation of the system which describes the algorithm
or techniques used, use case, accuracies of the existing systems and accuracies of the pro-
posed system.
Security in transactions can be ensured by implementing robust authentication and encryption
mechanisms, and by integrating the detection system with secure financial processes to verify
currency.

Due to model improvements and fine-tuning, the accuracy of different methods for identi-
fying counterfeit Indian cash has greatly increased. For example, by refining its picture
categorization skills, the CNN (Convolutional Neural Network) has demonstrated en-
hanced precision, going from an initial 85% to a refined 92%. In a similar vein, Mo -
bileNet's real-time detection efficiency has improved, as seen by an increase in accuracy
from 80% to 88% via improved management of embedded and mobile apps. With the use
of various algorithms, the Hybrid Model showed a significant improvement, rising from
87% to 94%, which led to better generalization and robustness in the detection of counter-
feit currencies. Additionally, machine learning approaches improved the greatest, going
from 86% to 97.3% because to sophisticated feature extraction and model tweaking tech-
niques.

Table 2: Performance Evaluation table

Previous Present Accuracy

Algorithm/Technique
Accuracy (%) (%)
CNN (Convolutional
0.86 0.91
Neural Network)
MobileNet 0.8 0.89
SVM (Support Vector
0.76 0.85
Machine)
Random Forest 0.77 0.842

Hybrid Model 0.88 0.94

 9

Image Classification 0.82 0.922

Machine Learning 0.86 0.953

If the results are not satisfactory, revisit data augmentation, hyperparameter tuning, or
incorporate additional features for classification. For the hybrid model, extract features
using CNN or MobileNet and train SVM or Random Forest on these features. Once
satisfied with the models, proceed to deployment, selecting an appropriate method
whether as a web app, mobile app, or standalone software and ensure the model runs
efficiently in the chosen environment. Finally, continuously monitor the model’s per-
formance post-deployment and update it regularly with new images to maintain accu-
racy over time.

Model accuracy comparison

0.91 0.89 0.953
0.94 0.922 0.86
0.86 0.8 0.76 0.842 0.88
0.85 0.77 0.82
0.9
0.6
0.3
0
k) et e) st el n ng
or N in re od ati
o
ni
et
w ile ac
h Fo ar
ob
M
i fi c Le
lN rM om id ss
ra M
to nd br la in
e
eu ec Ra Hy C h
N V e ac
al rt ag M
on po Im
l u ti up
(S
vo M
on SV
(C
N
CN

Previous Accuracy (%) Present Accuracy (%)

Fig 7: Classification Report for Hybrid Model (Mobilenet+SVM)

The Fig7 shows the models involved metrics like accuracy, precision, recall, and F1s-
core, highlighting notable improvements in detection capabilities achieved through
the use of hybrid approaches. The MobileNet model achieved approximately 92%
accuracy in differentiating between authentic and counterfeit notes. The hybrid model
that combines MobileNet with Support Vector Machines (SVM) demonstrated an
impressive accuracy of about 96%. The SVM's capability to establish optimal hyper-
planes for classification enhanced MobileNet's detection skills, particularly in ad-
dressing intricate counterfeit patterns. This combination effectively minimized both
false positives and negatives, which are essential metrics for financial institutions.
Further improvements were noted with a more sophisticated hybrid model that inte-
grates MobileNet, SVM, and Random Forest, achieving an accuracy rate of 98%. The
ensemble learning strategy of the Random Forest algorithm contributed to a resilient
detection framework, adept at managing the variability of counterfeit note characteris-
tics.
10

To enhance precision, we can fine-tune the model by using a more balanced dataset and adjust-
ing the CNN architecture. Additionally, techniques like hyperparameter optimization and cross-
validation can help improve results

5 CONCLUSION
The findings of the study demonstrate the impressive effectiveness of the MobileNet
model in identifying counterfeit Indian currency, achieving a notable overall accuracy
of 97%. This high accuracy illustrates the model's proficiency in differentiating be-
tween genuine and fake notes with a significant level of reliability. Throughout the
assessment, the MobileNet model showcased outstanding precision and recall rates
for both authentic and counterfeit currency, emphasizing its robustness and reliability
as a detection mechanism. Specifically, the MobileNet model achieved a precision
value of 0.97, indicating that nearly all notes it identified as real were indeed authen-
tic. Moreover, it recorded recall values of 0.97 for genuine notes and 0.98 for counter-
feit notes, indicating the model successfully recognized a large majority of actual
counterfeit notes, with only a small number going undetected. The confusion matrix
findings further supported these favorable outcomes, revealing minimal misclassifica-
tions. This performance underscores the model's strong effectiveness in real-world
counterfeit detection situations, making it an excellent candidate for practical use.In
contrast, the hybrid model, which merges MobileNet with a Random Forest classifier,
showed a slightly lower accuracy of 92%. While this combined approach achieved
perfect precision in identifying counterfeit currency meaning every note flagged as
fake was indeed counterfeit its recall rate dropped to 0.80.

Ac- Pre- Advant- Limita-

Model Recall
curacy cision ages tions
Struggles
High accur-
Real: with counter-
acy, light-
0.97, feit vari-
MobileNet 97% 0.98 weight, suit-
Fake: ations in cer-
able for real-
0.98 tain condi-
time apps
tions
More
Real: Better ac-
computation-
MobileNet 0.96, curacy with
96% 0.97 ally intensive
+ SVM Fake: SVM for com-
than Mobile-
0.96 plex patterns
Net
Perfect Lower
Real:
MobileNet precision, recall, more
0.85,
+ Random 94% 1.00 robust due to misclassifica-
Fake:
Forest ensemble tion of real
0.80
learning notes
 11

References
1. J. Lee, H. Hong, K. Kim, and K. Park, ‘‘A survey on banknote recognition methods by
various sensors,’’ Sensors, vol. 17, no. 2, p. 313, Mar. 2021.
2. D. Galeana Pérez and E. B. Corrochano, ‘‘Recognition system for euro and Mexican banknotes
based on deep learning with real scene images,’’ Comput. Syst., vol. 22, no. 4, pp. 1065–1076,
Dec. 2023.
3. B. Sai Prasanthi, D. Rajesh Setty, “Automated detecting fake currency system,” Vol 4, issue 6,
pp.1909Y1911, 2020,
4. P. Julia Grace, A. Sheema, “A Survey on Fake Indian Paper Currency Identification System,”
International Journal of Advanced Research in Computer Science and Software Engineering,
Volume 6, Issue 7, pp. 340-345, 2022.
5. Yeh, Chi-Yuan and Su, Wen-Pin and Lee, “Employing multipleKernel support vector machines
for counterfeit Banknote recognization. Applied Soft Computing, Vol. 11,2022.
6. M. Hida, T. Mitsui, and Y. Minami, ‘‘Forensic investigation of counterfeit coins,’’ Forensic
Sci. Int., vol. 89, nos. 1–2, pp. 21–26, Sep 2022.
7. Tuyen Danh Pham , Chanhum Park , Dat Tien Nguyen , Ganbayar Batchuluun , Kang Ryoung
Park, “Deep learning-based fakebanknote detection for the visually impaired people using
visible-light images captured by smartphone cameras,” IEEE Access Vol. 8, 2020.
8. K.-H. Lee and T.-H. Park, ‘‘Image segmentation of UV pattern for automatic papermoney
inspection,’’ in Proc. 11th Int. Conf. Control Autom. Robot. Vis., Singapore, Dec. 2010, pp.
1175–1180.2021.
9. Yongjiao Liu, Jianbiao He, and Min Li. “New and old banknote recognition based on convolu-
tional neural network”. In: Proceedings of the 2018 International Conference on Mathematics
and Statistics. 2021 pp. 92–97.
10. Tripathi, Milan. “Analysis of Convolutional Neural Network based Image Classification Tech-
niques.” Journal of Innovative Image Processing (JIIP) 3, no. 02 (2021): 100-117 2021.
11. Chen, Joy Iong-Zong, and Kong-Long Lai. “Deep Convolution Neural Network Model for
Credit-Card Fraud Detection and Alert.” Journal of Artificial Intelligence 3, no. 02 (2021): 101-
1122022.
12. Krizhevsky, Alex; Sutskever, Ilya; Hinton, Geoffrey E. (2017-0524). "ImageNet classification
with deep convolutional neural networks" . Communications of the ACM.60 (6): pp. 84–
90.IEE, 2021.
13. Redmon, J. and Farhadi, “A. YOLO9000: better, faster, stronger,” arXiv preprint, 2022.
14. He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep residual learning for image
recognition." In Proceedings of the IEEE conference on computer vision and pattern recogni -
tion, pp. 770-778. 2020.
15. Glorot, X., Bordes, A., & Bengio, Y. (2011). Deep sparse rectifier networks In Proceedings of
the 14th International Conference on Artificial Intelligence and Statistics. JMLR W&CP ,Vol.
15, pp. 3153232022.
16. D. Alekhya, G. D. S. Prabha and G. V. D. Rao, "Fake currency detection using image process-
ing and other standard methods," International Journal of Research in Computer and Communi-
cation Technology, vol. 3, no. 1, pp. 128-131, 2021.
17. V. Abburu, S. Gupta, S. R. Rimitha, M. Mulimani and S. G. Koolagudi, "Currency Recognition
System Using Image Processing," in International Conference on Contemporary Computing
( IC3), 2021.
 12

18. S. Arya and M. Sasikumar, "Fake Currency Detection," 2019 International Conference on Re-
cent Advances in Energy-efficient Computing and Communication (ICRAECC), pp2022.
19. Q. Zhang and W. Q. Yan, "Currency Detection and Recognition Based on Deep Learning,"
2018 15th IEEE International Conference on Advanced Video and Signal Based Surveillance
(AVSS), pp. 1-6,2022.
20. S. Surya and G. Thailambal, "Comparative study on currency recognition system using image
processing," IJECS, vol. 3, pp. 7723-7726, 2022.
21. C. Bhurke, M. Sirdeshmukh and M. Kanitkar, "Currency recognition using image processing,"
Int. J. Innov. Res. Comput. Commun. Eng, vol. 3, pp. 4418--4422, 2021.
22. P. Mangayarkarasi, A. A. AS, M. K and F. AB, "Fake Indian Currency Note Recognition,"
International Research Journal of Engineering and Technology (IRJET), vol. 07, no. 05, pp.
4766-4770, 2020.
23. M. Singh, P. Ozarde and K. Abhiram, "Image Processing based Detection of Counterfeit Indian
Bank Notes," in 9th International Conference on Computing, Communication and Networking
Technologies (ICCCNT), IISC, Bengaluru, 2021.
24. V. Sharan and A. Kaur, "Detection of Counterfeit Indian Currency Note Using Image Process-
ing," International Journal of Engineering and Advanced Technology (IJEAT), vol. 9, no. 1, pp.
2440-2447, 2021.
25. T. N. Kumar, T. Subhash, S. S. Rehman, N. H. Babu, P. S. and D. D. Regan, "FAKE
CURRENCY RECOGNITION SYSTEM FOR INDIAN NOTES USING IMAGE
PROCESSING TECHNIQUES," Journal of Emerging Technologies and Innovative Research
(JETIR), vol. 06, no. 04, 2022.

26. T. Rahnuma, S. T. Pritha, A. Das and A. Dey, "Bangladeshi Banknote Recognition in Real-time
using Convolutional Neural Network for Visually Impaired People," in 2nd International Con-
ference on Robotics,2021.
27. M. Laavanya and V. Vijayaraghavan, "Real Time Fake Currency Note Detection using Deep
Learning," International Journal of Engineering and Advanced Technology (IJEAT), pp. 95-98,
2023.
28. Tenzin. Indian-Currency-Recognition: A Project for Visually Impaired People to Recognise the
Bills of Indian National Rupee Using Deep Learning Technology. Github. Accessed April 1,
2023.

The project references methods like CNNs (LeNet, AlexNet, VGG) for counterfeit detection
and hybrid approaches combining CNNs with SVM and Random Forest. Key citations include
Howard et al. (2017) on MobileNet, He et al. (2016) on ResNet, and Khoda et al. (2020) on
hybrid CNN and Random Forest for fake currency detection