An Indian Currency Recognition Model for

Assisting Visually Impaired Individuals

2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE) | 978-1-6654-9491-5/22/$31.00 ©2022 IEEE | DOI: 10.1109/RASSE54974.2022.9989624

Madhav Pasumarthy Rutvi Padhy Raghuveer Yadav Ganesh Subramaniam Madhav Rao
IIIT Bangalore IIIT Bangalore Niyodaya Foundation Niyodaya Foundation IIIT Bangalore
[email protected] [email protected] [email protected] [email protected] [email protected]

Abstract—Visually impaired persons find it extremely difficult consider digital payments, especially in India. The local
to perform cash transactions in outdoor environments. For transportation systems generally thrive on cash payments
assisting the visually challenged individuals, a YOLOv5 based and hence to some extent currency exchange still persists.
deep neural network was designed to detect image based cur-
rency denominations. Thereby aid in completing the authentic Besides, the digital payments in the form of wallet or
transaction. The robust model was trained for images with UPI interface still does not have a user-friendly design for
currency notes in different backgrounds, multiple sides of the VIPs [9], at least in the case of Indian currency notes.
currency notes presented, notes around cluttered objects, notes Additionally there exists some concerns on security and
near reflective surfaces, and blurred images of the currency lapses in preserving privacy during digital transactions, which
notes. An annotated and augmented dataset of around 10,000
original images was created for developing the model. A makes the visually impaired section of the society highly
pre-processing step to rescale all the images to 224 × 224 vulnerable to this problem [10]. Reserve bank of India (RBI)
was applied to standardize the input to the neural network, managed Mobile Aided Note Identifier (MANI) does serve
and generalize the model for different platforms including the purpose to an extent [11], however, the processing and
single board computer and smartphones. The trained model currency estimation is performed in the cloud, which binds
showcased an average denomination recognition accuracy of
92.71% for an altogether different dataset. The trained model the visually impaired users to be always connected to the
was deployed on Raspberry-Pi and Smartphone independently, internet. This especially makes the VIPs handicap in areas
and the outcome to detect the currency denomination from the where connectivity is a problem. Few image processing tech-
image was successfully demonstrated. The model showcased niques to recognize currency and detect denominations are
adequate performance on different platforms, leading to the attempted in the past [12]. However, recently deep learning
exploration of several other assistive applications based on the
currency recognition model to improve the standard of living (DL) method has emerged as a popular technique especially
for visually challenged individuals. for object detection and recognition applications [13]. Few
Keywords—Currency detection, Image processing, CNN, of the staged datasets for currency detection were aggregated
YOLO network. and a neural network model was trained as a single shot
multi-box detector (SSD) for only detecting 3 denomina-
I. I NTRODUCTION tions [14]. CNN was attempted to extract features, to classify
folded notes for a single denomination in [15]. A portable
The ever dependence on a care-taker by the visually
device named ”Note Teller 2” invented by BRYTECH, Ot-
impaired person (VIP) does not last long, hence a large
tawa, Canada, recognizes currency denomination for about
group of assistive devices and software apps are developed
80 % for unfolded notes, but performs poorly for wrinkled
to cater to the demands of this section of society [1]–[8]. A
and worn-out notes [16], [17]. A defined speeded-up robust
VIP faces several challenges in their daily routine and often
features (SURF) was extracted on the banknote image to
need to rely on others for help. A few everyday challeng-
recognize US currency denominations, however, the printed
ing situations include: a) determining the denomination of
face specific to the currency denomination was extracted as
currency notes that are exchanged, b) indoor and outdoor
a feature to classify different notes [17]. Indian currency has
navigation and determining stationary and moving objects
same printed face for all denominations, and hence SURF
and obstacles in the path around them, c) recognizing objects
method will not be effective. An Ethiopian banknote currency
and people in the path, and d) reading printed content in
recognition system based on support vector machine (SVM)
books, menu cards, newspapers, packaged items, medicine
was discussed in [18], however, the currency was detected
strips, and others. Dealing with hard currency notes is one
only for the front portion of the currency note, hence the
of the biggest challenges among the lot. The VIP may have
model may not necessarily fit in real-life scenarios. A low
to rely on someone around to ensure correct currency notes
quality banknote recognition system is stated in [19], and
while doing transactions. The existing mobile-based solutions
a sequential DNN model with data augmentation technique
have challenging issues with orientation of the note, cluttered
to improve accuracy is described in [20]. Both these models
background, usage in low light, speed of inference, accuracy,
perform well for staged dataset and not on real-life scenarios.
and ever dependence on internet connectivity and access to
The primary contribution of the paper is in the development
the cloud. The paper documents the work done to develop a
of image based Indian currency recognition system using
new deep neural network model to accurately recognize the
deep neural network (DNN) model. The DNN model was
currency note in all possible real-life situations.
trained to incorporate real-life currency images acquired at
Although digital payments are now accepted widely, how-
different background, with poor lighting conditions, near
ever, not all impromptu roadside and window shopping

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 23,2025 at 02:57:41 UTC from IEEE Xplore. Restrictions apply.
 (a) (b)

(c)

Fig. 1. Few images of unfolded-dataset showing annotated (a) currency notes with multiple objects, (b) currency notes in poorly lit surroundings, and
(c) currency notes near reflective surfaces.

reflective surfaces, and different sides of the currency notes. the cash transactions in the form of folded notes are also
The model also accommodated the front and back form of possible, hence its imperative for the model to recognize
the currency notes to not miss out on any form of valid folded notes as well. The folded notes dataset is referred
transaction. As a proof of concept, the trained model was to as folded-dataset. The dataset was divided into training,
deployed on a Raspberry-Pi 4 Model B single board computer validation, and testing set, in the ratio of 70%, 20%, and
and on android smartphone to validate its real-time effect. 10% respectively, with images for training set being picked
Additionally, a vast dataset of different Indian currency notes randomly to avoid any bias in the model learning. The folded-
was created and was made freely available at [21] for further dataset was manually annotated to 22 different classes as per
usage to the research and developers community. the type of the currency note, denomination value, and its
orientation, as shown in the table I, for six of the classes from
II. A PPROACH AND METHODOLOGY the 22 defined. Few samples of the annotated real-life images
The YOLOv5 network is an updated module derived involving currency notes in real-life settings, used for training
from the same family of YOLO networks [22], [23], and are shown in the Figure 1. The dataset was annotated to 22
is considered as one of the best predefined models for different classes by drawing bounding boxes on the individual
object detection [24]. To develop an optimal model, a large captured images using Roboflow software. The dataset was
dataset of Indian currency with the following denominations: augmented by manipulating saturation, color, rotation, alter-
10, 20, 50, 100, 200, 500, and 2000, was built and made ing background, making translation to the original images,
available as open source in the Github [25]. This dataset is and randomly flipping image pixels from left to right, to
a culmination of originally acquired images by the authors, create mirror images of the original. All these techniques
and images taken from Kaggle dataset [26], and another open were performed to scientifically enlarge the dataset on the
source dataset available in the Github [21]. A smartphone- run during the training process.
based camera with 1280 × 720 pixel resolution was used to The YOLOv5 network was trained in Google Colab
acquire images for different denominations of the currency. system [27] for 224 × 224 image pixel resolution units
Different illuminations, cluttered background, currency notes over 3000 epochs, and the model, consisting of 283 layers,
near reflective surfaces, and other real-life scenarios were was obtained. 1500 epochs showcased minimum validation
considered for developing a large dataset of unfolded notes loss over training and hence the same was selected for
of size 2000. The dataset of each denomination with front characterizing the results. Pre-processing, to rescale all the
and back view of the currency note was acquired multiple input images to 224 × 224, was performed to standardize
times to develop a robust model. This dataset is referred to as the model for different platforms namely - Google Colab,
unfolded-dataset. Additional 2000 images were captured with embedded systems, and mobile devices. The trained model
similar real-life settings, but for folded notes, considering was further deployed on Raspberry-Pi embedded platform to

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 23,2025 at 02:57:41 UTC from IEEE Xplore. Restrictions apply.
 TABLE I for detecting not only the denominations but also the type,
S IX CLASSES PICKED FROM 22 ANNOTATED CLASSES TO DEMONSTRATE and sides of the notes. The folded-dataset which was not
THE ANNOTATIONS PERFORMED , WHICH WERE BASED ON
DENOMINATIONS , TYPE OF NOTES , AND ITS ORIENTATION . utilized to build the model was applied on the model to
characterize the classification accuracy. The folded-dataset
Denominations Type Orientation Labels involved folded currency notes in similar settings as that of
100 New Back New100Back
100 New Front New100Front unfolded-dataset such as currency in a cluttered background,
10 New Back New10Back multiple notes, poorly-lit surroundings, blurred image, and
10 New Front New10Front notes near reflective surfaces. The dataset had 1970 orig-
50 Old Back Old50Back
50 Old Front Old50Front inally acquired images, and similar augmentation scheme
was applied to enlarge the dataset to 10,000 size. Few of
the detected folded notes with denominations are shown in
showcase the ability to detect the currency denomination on the Figure 2. Confusion matrix of 22-class currency note
edge platform. Raspberry-Pi was interfaced with a Logitech denomination identification was generated from the trained
HD webcam with 720p resolution. TensorFlow Lite (TFL) model for folded notes. Except four classes, all other 18
framework was considered best suited for running deep classes reported more than 80 % accuracy, which is consid-
learning models on mobile devices. The light weight design ered extremely robust considering this dataset was not used
offers extremely low latency, and low power consumption, for training the model. On an average, 85.95% of accuracy
which preserves the charge for running different applications was achieved for currency detection of images with folded
on the mobile devices. The conversion process from Colab notes for 22 classes. Further post-processing to merge same
interface which uses Pytorch implementation to TFL version denominations of different sides and types was performed.
involves three steps of sequential transformations through The merging of 4 classes of similar denomination to one
ONNX, Tensorflow, and TFL. The converted model was class, minimized the 22-class to 7-class inferences. The old
further packaged to a mobile application and deployed on note, new note, front view, and back view of the detected
a smartphone to validate real-life usage. The developed app class was fused to a single class of denomination value. The
is made available in the play-store [28] for further usage. confusion matrix of 7-class of denomination post processing,
is shown in the Figure 3. On an average, the denomination
III. R ESULTS AND A NALYSIS estimated for folded currency note in real-life images were
The model that was trained on unfolded-dataset, and highly accurate to 92.71%. Considering the estimation was
was characterized for validation subset, yielded detection performed on entirely new dataset, the characterized accuracy
accuracy reported in the Table II. An average MAP of 50% of the model for estimating 7 denomination classes attained,
showing 98.74% for the validation dataset is considered is considered very robust. An additional dataset external-
extremely high for the model developed. In the real-life dataset [29] developed by Niyodaya foundation was also
applied on the trained model to showcase the relevance of the
model on different images. The external-dataset with original
TABLE II
C LASSIFICATION RESULTS CHARACTERIZED THROUGH PRECISION , images of 1828 size was augmented to enlarge the size to
RECALL , AND MEAN - AVERAGE - PRECISION (MAP) OF 50% SCORES FOR 9140 images. The model when applied on external-dataset
22 CLASSES AS ACQUIRED FROM THE VALIDATION SUBSET. showcased a very high precision of 92.2%, and average
Class Precision (%) Recall (%) MAP of 50%(%) recall of 91.8%, which is considered extremely high for
New100Back 99.7 100 99.5 an altogether different dataset, which was not used while
New100Front 100 84.3 99.2 training.
New10Back 96.2 100 99.5
New10Front 100 94.7 98.2
New2000Back 100 95.2 99.5
New2000Front 94.1 100 99.5
New200Back 98.8 100 99.5
New200Front 99.3 100 99.5
New20Back 99 100 99.5
New20Front 99 93.8 93.7
New500Back 100 97.2 99.5
New500Front 99.3 100 99.5
New50Back 98.8 100 99.5
New50Front 100 100 99.5
Old100Back 98.6 100 99.5
Old100Front 100 96.4 99.5
Old10Back 96.7 100 99.5
Old10Front 98.8 100 99.5
Old20Back 98.6 100 99.5
Old20Front 99 100 99.5 Fig. 2. Detection of folded currency notes in different real-life scenarios.
Old50Back 99.2 94.4 94.7
Old50Front 94.2 100 95.1 The model was expected to detect multiple currency
notes to estimate the total currency exchanged in a real-
scenario, not all currency exchanges are performed with life scenario. As shown in the Figure 4 (a), multiple notes
unfolded notes, hence folded notes were also tested with the were successfully detected by the model. These images are
developed model. The model showcased a high classification the ones taken from the 10% kept aside as testing subset
accuracy in terms of precision, recall, and MAP of 50% of unfolded-dataset. Additionally, currency notes were also

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 23,2025 at 02:57:41 UTC from IEEE Xplore. Restrictions apply.
 successful detection of the currency denomination as shown
in the Figure 6 (b). Multiple currency notes, folded currency
note, currency note along with different object, and blurred
image were successfully detected by the TFL model running
on the smartphone. The smartphone application additionally
renders a voice feedback of currency notes detected in real
time, assisting visually impaired individuals to estimate the
denomination on the run. Latency of less than 1 second was
characterized for the currency detection TFL model running
on smartphone.

Fig. 3. Confusion matrix generated for the folded-dataset from the trained
model configured to classify 22-classes, and further post processed to 7
denomination class.

successfully detected in a poorly lit settings as shown in

the Figure 4 (b). In real-life, not all images are captured
properly without any blurred effect. Hence the model was
tested on different blurred images comprising of currency
notes. Blurred images showing currency note on a plain
background, and on cluttered background were successfully
detected by the model as shown in the Figure 5 (a,b). (a)

(a) (b)

Fig. 4. Successful detection of (a) multiple currency notes, and (b) currency
notes in poorly lit surroundings.

(b)

Fig. 6. Currency notes detected (a) by the model when deployed on

Raspberry-Pi device, and (b) by the TFL model when ported to the
smartphone.

IV. C ONCLUSION
(a) (b) A robust DNN model was deduced from a predefined
YoloV5 network to detect denominations of Indian currency
Fig. 5. Detection of blurred currency notes (a) in plain background, and
(b) under cluttered environment with multiple objects. notes, from the real-life captured images. The model trained
on unfolded currency dataset, showcased an average detection
accuracy of 92.71% in detecting the denomination value,
The model of size 14.4 MB when deployed on Raspberry- when applied on folded currency dataset, which is considered
Pi device was able to successfully detect the currency notes highly reliable for real-life applications. This is the first time
as shown in the Figure 6 (a). Additionally the model was able a robust and reliable currency detection and denomination
to detect the currency notes in different oblique orientations estimation model was attempted for Indian currency. The
as shown in the Figure. The Raspberry-Pi prototype system trained model when deployed on Raspberry-Pi that was
conceded a latency of less than 1 second to detect the interfaced with a camera, showcased successful detection of
currency denomination from the image. Also, the minimum, currency notes. Similar detection results were achieved for
optimal, and maximum distance to detect the currency notes, TFL model when ported to the smartphone.
were found to be in the range of 3.5-4 inches, 10-25 inches,
and 30-35 inches, respectively; considering 25 inches as the R EFERENCES
length of an average human arm. The Tensor Flow Lite [1] D. Dakopoulos and N. G. Bourbakis, “Wearable obstacle avoidance
(TFL) modified model when ported to smartphone showcased electronic travel aids for blind: A survey,” IEEE Transactions on

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 23,2025 at 02:57:41 UTC from IEEE Xplore. Restrictions apply.
 Systems, Man, and Cybernetics, Part C (Applications and Reviews), [14] Q. Zhang and W. Q. Yan, “Currency detection and recognition based
vol. 40, no. 1, pp. 25–35, 2010. on deep learning,” in 2018 15th IEEE International Conference on
[2] R. Manduchi, J. Coughlan, and V. Ivanchenko, “Search strategies of Advanced Video and Signal Based Surveillance (AVSS), 2018, pp. 1–
visually impaired persons using a camera phone wayfinding system,” 6.
in Computers Helping People with Special Needs, K. Miesenberger, [15] M. JIAO, J. HE, and B. ZHANG, “Folding paper currency recognition
J. Klaus, W. Zagler, and A. Karshmer, Eds. Berlin, Heidelberg: and research based on convolution neural network,” in 2018 Interna-
Springer Berlin Heidelberg, 2008, pp. 1135–1140. tional Conference on Advances in Computing, Communications and
[3] H. Shen and J. Coughlan, “Grouping using factor graphs: An approach Informatics (ICACCI), 2018, pp. 18–23.
for finding text with a camera phone,” vol. 4538, 06 2007, pp. 394– [16] “Note Teller v2,” https://www.eastersealstech.com/2011/06/15/note-
403. teller-2-offers-a-way-for-visually-impaired-to-read-paper-
[4] Y. Tian, C. Yi, and A. Arditi, “Improving computer vision-based money/brytech-note-teller-2/, accessed: June, 2021.
indoor wayfinding for blind persons with context information,” in [17] F. M. Hasanuzzaman, X. Yang, and Y. Tian, “Robust and effective
Computers Helping People with Special Needs, K. Miesenberger, component-based banknote recognition for the blind,” IEEE Transac-
J. Klaus, W. Zagler, and A. Karshmer, Eds. Berlin, Heidelberg: tions on Systems, Man, and Cybernetics, Part C (Applications and
Springer Berlin Heidelberg, 2010, pp. 255–262. Reviews), vol. 42, no. 6, pp. 1021–1030, 2012.
[5] E. Tekin, J. M. Coughlan, and H. Shen, “Real-time detection and [18] E. Ayalew Tessfaw, B. Ramani, and T. Kebede Bahiru, “Ethiopian
reading of led/lcd displays for visually impaired persons,” in 2011 banknote recognition and fake detection using support vector machine,”
IEEE Workshop on Applications of Computer Vision (WACV), 2011, in 2018 Second International Conference on Inventive Communication
pp. 491–496. and Computational Technologies (ICICCT), 2018, pp. 1354–1359.
[6] S. Shoval, J. Borenstein, and Y. Koren, “Auditory guidance with the [19] Y. Jin, L. Song, X. Tang, and M. Du, “A hierarchical approach for
navbelt-a computerized travel aid for the blind,” IEEE Transactions on banknote image processing using homogeneity and ffd model,” IEEE
Systems, Man, and Cybernetics, Part C (Applications and Reviews), Signal Processing Letters, vol. 15, pp. 425–428, 2008.
vol. 28, no. 3, pp. 459–467, 1998. [20] V.-D. Hoang and H.-T. Vo, “Hybrid discriminative models for banknote
[7] A. J. Huete, J. G. Victores, S. Martinez, A. Gimenez, and C. Balaguer, recognition and anti-counterfeit,” in 2018 5th NAFOSTED Conference
“Personal autonomy rehabilitation in home environments by a portable on Information and Computer Science (NICS), 2018, pp. 394–399.
assistive robot,” IEEE Transactions on Systems, Man, and Cybernetics, [21] “Indian Currency Classifier,” https://github.com/piyushdavda007/indi
Part C (Applications and Reviews), vol. 42, no. 4, pp. 561–570, 2012. an-currency-classification, accessed: March, 2021.
[8] Y. Tian, X. Yang, and A. Arditi, “Computer vision-based door detection [22] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look
for accessibility of unfamiliar environments to blind persons,” in once: Unified, real-time object detection,” in 2016 IEEE Conference on
Computers Helping People with Special Needs, K. Miesenberger, Computer Vision and Pattern Recognition (CVPR), 2016, pp. 779–788.
J. Klaus, W. Zagler, and A. Karshmer, Eds. Berlin, Heidelberg: [23] J. Redmon and A. Farhadi, “Yolo9000: Better, faster, stronger,” in
Springer Berlin Heidelberg, 2010, pp. 263–270. 2017 IEEE Conference on Computer Vision and Pattern Recognition
[9] A. Das and P. Das, “Digital payments in india: The road ahead,” IEEE (CVPR), 2017, pp. 6517–6525.
Potentials, vol. 36, no. 6, pp. 14–19, 2017. [24] L. Aziz, M. S. B. Haji Salam, U. U. Sheikh, and S. Ayub, “Exploring
[10] B. Rajendran, A. K. Pandey, and B. S. Bindhumadhava, “Secure deep learning-based architecture, strategies, applications and current
and privacy preserving digital payment,” in 2017 IEEE Smart- trends in generic object detection: A comprehensive review,” IEEE
World, Ubiquitous Intelligence Computing, Advanced Trusted Com- Access, vol. 8, pp. 170 461–170 495, 2020.
puted, Scalable Computing Communications, Cloud Big Data Com- [25] “Currency Detection,” https://github.com/madhav2391/Currency-
puting, Internet of People and Smart City Innovation (Smart- detection, accessed: October, 2021.
World/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), 2017, pp. 1–5. [26] “Indian Currency Note images dataset 2020,”
[11] “Mobile Aided Note Identifier,” https://www.rbi.org.in/scri https://www.kaggle.com/vishalmane109/indian-currency-note-images-
pts/MANI.aspx, note = Accessed: June, 2021. dataset-2020, accessed: January, 2021.
[12] G. V. Rajan, D. M. Panicker, N. E. Chacko, J. Mohan, and V. Kavitha, [27] T. Carneiro, R. V. Medeiros Da NóBrega, T. Nepomuceno, G.-B.
“An extensive study on currency recognition system using image Bian, V. H. C. De Albuquerque, and P. P. R. Filho, “Performance
processing,” in 2018 Conference on Emerging Devices and Smart analysis of google colaboratory as a tool for accelerating deep learning
Systems (ICEDSS), 2018, pp. 228–230. applications,” IEEE Access, vol. 6, pp. 61 677–61 685, 2018.
[13] R. C. Joshi, S. Yadav, and M. K. Dutta, “Yolo-v3 based currency [28] “NETRA App,” https://play.google.com/store/apps/details?
detection and recognition system for visually impaired persons,” in id=com.nethra.android, accessed: January, 2022.
2020 International Conference on Contemporary Computing and Ap- [29] “Indian Currency Dataset,” https://drive.google.com/file/d/1kjLVUenGu
plications (IC3A), 2020, pp. 280–285. N 6uzwj5ibd8imVXVRW2EpaU/view, accessed: October, 2021.

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 23,2025 at 02:57:41 UTC from IEEE Xplore. Restrictions apply.