Techno.COM, Vol. 23, No.

1, Februari 2024: 176-186

Integrate Yolov8 Algorithm For Rupiah Denomination

Detection In All-In-One Smart Cane For Visually
Impaired
I Made Surya Kumara1, Gde Putu Rizkynindra Sukma Jati2, Ni Putu Widya Yuniari3
1
Universitas Warmadewa,2BPJS Ketenagakerjaan, 3Universitas Warmadewa
E-mail: [email protected] , [email protected],
3
[email protected]

Abstract
The eyes are crucial tools for human observation and perception, facilitating various tasks in
daily life. Individuals, including those with visual impairments or blindness, engage in currency
transactions, posing challenges in recognizing notes and preventing mishaps with counterfeit
money. Despite government efforts, features like embossing on banknotes have limited
effectiveness due to the circulated currency's disheveled condition. Addressing the visually
impaired community's needs is imperative. An innovative solution, the "all-in-one smart white
cane," integrated with machine learning supports daily activities, enhancing independence for
visually impaired individuals. The YOLOv8 algorithm is employed for the precise detection of
monetary denominations, subsequently recorded through a camera and seamlessly integrated
into a smart cane, resulting in a consolidated device. This device, designed with standout features,
excels in detecting Indonesian Rupiah banknote denominations. Detection performance testing,
incorporating methods like object rotation, utilized a dataset divided into training (70%),
validation (20%), and test (10%) segments. Modifications to contrast and variability rotation are
essential in the context of real-time nomination recognition. These adjustments are implemented
to ensure accurate and swift identification in dynamic, real-world scenarios. Testing results
reveal a 99% average accuracy in recognizing currency note denominations, presenting an
effective solution for the visually impaired community.

Keywords: detection, algorithm, smart-white-cane, visually impaired, machine learning

1. INTRODUCTION

Visual impairments represent a significant challenge faced by more than 285 million
individuals worldwide, with 39 million among them experiencing blindness and 124 million
grappling with low vision, according to data presented by the World Health Organization (WHO)
[1]. The substantial population affected underscores the importance of providing special attention
to the visually impaired community. Visual impairment is not an isolated ophthalmic condition,
but various conditions can cause this visual manifestation [2]. It is crucial to note that despite their
visual impairments, they possess the potential to engage in daily activities just like any other
individual. Therefore, it is imperative to seek adequate solutions to facilitate their livelihoods.
Financial transactions play a pivotal role in supporting the sustenance of every individual,
including the visually impaired community. These transactions encompass purchasing goods,
economic activities, and various financial exchanges. Unfortunately, individuals with visual
impairments often encounter challenges when engaging in buying and selling, such as confusion
in distinguishing currency notes or even the risk of falling victim to counterfeit money.
Bank Indonesia has responded to feedback from the visually impaired community
regarding their challenges with physical currency transactions [3]. One of the measures taken is
the adjustment of the length differences between currency notes, from the previous 2 millimeters
to 5 millimeters. Although this step is positive, many visually impaired individuals still experience
discomfort in distinguishing currency denominations based solely on length. Moreover, currency
notes in circulation are often not in pristine condition, which can affect the visually impaired's
176
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

ability to identify their value. To address this issue, Bank Indonesia has introduced tactile features,
such as embossed relief, on the sides of currency notes to facilitate denomination recognition.
However, in practice, currency in circulation often lacks uniformity and may show signs
of wear and tear. The method still has several weaknesses, namely in terms of memory capacity,
physical condition, and the absence of a determinant of honesty in the process of buying and
selling goods and services [4]. Previous studies have employed the TCS3200 color sensor to
determine the value of banknotes; however, the performance of the system heavily relies on the
physical condition of the currency [5]. Conventional image processing methods are less effective
in detecting the authenticity of money because of the various types of counterfeit money in
circulation [6]. Therefore, the idea of implementing Computer Vision technology, specifically
Convolutional Neural Networks (CNN), to facilitate currency denomination recognition has
emerged, because CNN can apply to image and text classification [7]. The utilization of machine
learning methods for verifying the authenticity of banknotes emerges as an optimal solution,
ensuring accurate recognition of currency irrespective of its condition, whether worn or otherwise.
This approach enhances the robustness of currency identification [8].
The "All-in-one smart cane" is the latest innovation in the form of a smart cane that
incorporates various intelligent features to support the lives of visually impaired individuals. One
of the standout features developed is the ability to automatically recognize Indonesian Rupiah
denominations. This device is designed to be portable and rechargeable, making it highly practical
for carrying anywhere. Furthermore, it utilizes Machine Learning, employing a CNN approach
based on the YOLO V8 algorithm.
A. Related Research
1) A. Hermawan, L. Lianata, Junaedi, and A. R. K. Maranto, "Implementasi Machine
Learning Sebagai Pengenal Nominal Uang Rupiah dengan Metode YOLOv3,". This research
presents an application of currency denomination detection employing the YOLO version 3 (V3)
methodology. The investigation involves the development of a mobile application designed for
currency denomination detection. Despite achieving commendable accuracy, the device under
consideration exhibits limitations, particularly in accessibility for the visually impaired and the
complexity of navigation[3].
2) A. K. E. Lapian, S. R. U. A. Sompie, and Pinrolinvic D. K. Manembu, "You Only
Look Once (YOLO) Implementation For Signature Pattern Classification,". This study focuses
on detecting the authenticity of signatures through the utilization of machine learning techniques.
There exists a necessity for the augmentation of datasets and the incorporation of precise signature
resolutions to enhance the overall efficacy of the system[9].
3) S. Kudalkar, P. Patil, and N. Shirdhone, "Fake Currency Detection Using Image
Processing,". This research endeavors to identify the legitimacy of currency through the
application of image processing techniques. The process involves sequential stages, including
grayscale conversion, edge detection, segment detection, feature extraction, and comparison.
Despite these efforts, the average accuracy in currency recognition remains at 80%[10].
4) A. Bhatia, V. Kedia, A. Shroff, M. Kumar, B. K. Shah, and Aryan, "Fake Currency
Detection with Machine Learning Algorithm and Image Processing,". This study employs deep
learning for the detection of counterfeit currency. With a dataset comprising approximately 2000
samples, the model achieves a remarkable accuracy of 99.9%. However, the authors posit that
further strides are required for real-world implementation, emphasizing the need for substantial
additional datasets[11].
5) A. Antre, O. Kalbhor, P. Jagdale, and G. Dhanne, "Fake Currency Detection Using
Convolution Neural Network,". This investigation leverages genuine and counterfeit Indian
currency for analysis. Preprocessing steps involve the conversion of RGB images to grayscale,
accompanied by Gaussian blur application for clarity enhancement. Despite achieving an
accuracy rate of 97.72%, the study advocates for the incorporation of additional datasets and
further refinement in the methodology[12].
Based on the literature review above, we observe the need for further modifications and
optimization to enhance the accuracy level of the Indonesian Rupiah banknote detector.
177
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

Additionally, several studies mentioned above still rely on smartphones, which, in our view, may
not be widely adopted by the visually impaired; instead, they often use canes. Image capture is
conducted using a mobile phone camera, resulting in the partial visibility of currency details [13].
Moreover, all computations are performed on the smartphone, leading to a dependency on its
specifications. Our strategy to improve accuracy involves the utilization of the latest YOLO
algorithm and use a separate device for computing.
The YOLO (You Only Look Once) algorithm, first proposed by Redmon and his
colleagues in 2016, is an object detection algorithm based on regression. YOLO divides the input
image into grid cells of a certain size and predicts bounding boxes and class probabilities for each
grid cell. By transforming object detection into a regression problem, YOLO can detect objects
rapidly and accurately [14].
Hence, this study aims to elucidate how the utilization of CNN technology, specifically
YOLO V8, in the development of the "All-in-one smart cane" can effectively address the
challenges faced by visually impaired individuals in identifying and distinguishing Indonesian
Rupiah denominations efficiently.

Figure 1 The Detection Head for YOLOv8

The YOLO (You Only Look Once) object detection framework has undergone
continuous development and refinement, evolving from its first version (YOLOv1) to its latest
iteration, YOLOv8, introduced in early 2023. YOLOv8 encompasses five distinct models:
YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x, and it supports efficient API calls
through the Python programming language.
YOLOv8 incorporates a Decoupled-Head architecture with separate computational

178
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

branches to enhance its performance [15]. Data augmentation processes in YOLOv8 disable
Mosaic Augmentation during the final 10 epochs, effectively improving its accuracy. The
Classification loss is transformed into VFL Loss, and CIOU Loss is introduced alongside DFL
(Distribution Focal Loss) as the regression loss function. YOLOv8 also replaces IOU matching
or one-sided allocation with the Task-aligned approach.
These advancements in YOLOv8 represent a significant step forward in the field of object
detection, offering enhanced accuracy and efficiency in object recognition tasks. The introduction
of multiple model variants and API support in Python further contribute to its versatility and ease
of use in various applications. The adoption of the Decoupled-Head architecture and innovative
loss functions underscores the commitment to continually improve the YOLO framework's
capabilities. As YOLOv8 continues to evolve, it promises to be a valuable tool for researchers
and practitioners in computer vision and object detection.

2. RESEARCH METHOD

The stages undertaken in this research encompass data collection, data preparation, data
labeling, device design, and testing.

2.1 Data Collection

The data utilized to support this research comprise images of Indonesian Rupiah
banknotes issued from 2009 to 2023. These banknotes are categorized into seven denominations:
Rp 1,000, Rp 2,000, Rp 5,000, Rp 10,000, Rp 20,000, Rp 50,000, and Rp 100,000. The dataset
includes images of both sides of the banknotes, captured from various angles and rotations. In
addition to the complete banknote images, the dataset has been modified by cropping only the
nominal value areas. This modification was performed to enable the model training phase to better
recognize the distinctive characteristics of each denomination.

Figure 2 Sample data rotated and flipped from various sides.

Figure 3 Sample crop yield data from each nominal.

2.2 Data Preparation

The data employed in this research comprise multiple classes. To facilitate data
management and labeling, we opted to organize the data into seven distinct folders. Each folder
is named based on the number of samples it contains, namely, 1000, 2000, 5000, 10000, 20000,
50000, and 100000.
This categorization process was executed meticulously, ensuring that each data sample
corresponds to the quantity indicated by the folder's name. With this approach, we can readily
identify and access data with specific sample counts.
179
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

Figure 4 Folder of each image.

Figure 5 Sample dataset image.

2.3 Data Labeling

The labeling process was carried out manually, commencing with the identification of the
object to be detected, which in this case, is currency within the images. Each image was analyzed
individually. The regions containing currency within the images were recognized and enclosed
within bounding boxes to distinguish them from the background. This process was executed
meticulously to ensure that the desired currency objects were accurately encompassed within the
bounding boxes.
After identifying and enclosing the currency objects within bounding boxes, the
subsequent step involved labeling these objects. The labels assigned had to accurately represent
the class of the detected object, which, in this context, is "currency." Each bounding box within
the image was labeled as "currency" to signify that the object within it was the currency to be
detected.
The outcome of the labeling process resulted in annotation files in the .txt format,
conforming to the format required by YOLOv8. Each .txt file contained the coordinates of the
bounding box and the label "currency" for each detected currency object within the image. This
format was utilized in training the YOLOv8 object detection model.

180
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

Figure 6 Dataset labeling process.

Following the labeling process, the next step involved data preprocessing. At this stage,
the dataset was divided into three folders: "train," "test," and "valid." The division was carried out
using a scheme of 70% for the "train" folder, 20% for the "valid" folder, and 10% for the "test"
folder. Data within the "train" folder, known as the training set, was used for training the model
using the pre-trained YOLOv8 model. Subsequently, the data in the "valid" folder, referred to as
the validation set, was employed to evaluate the performance of the created model. Lastly, the
data within the "test" folder, designated as the testing set, served the purpose of testing the
performance of the developed model.

2.4 Design and Device Development

2.4.1. Physical Design
The design of this device encompasses the use of a conventional folding cane for the
visually impaired, onto which an acrylic case is mounted at the top to accommodate electronic
components. The case is made transparent to allow visualization of the electronic devices
contained within.
2.4.2 Electronic Components
● 10,000mAh 5V, 2.4A Power Bank: Utilized to provide power to the entire system and
attached auxiliary devices.
● NVIDIA Jetson Nano: A mini-computer equipped with AI processing capabilities, used
for object recognition and navigation.
● 720p 30fps HD Webcam: Employed to capture images and videos of the user's
surrounding environment.
● Micro Speaker: Used as an audio output for recognizing Indonesian Rupiah currency
denominations.

Figure 7 Jetson Nano: The mini-computer

181
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

2.4.3 3D Visualization
To streamline the design process and visualize how all components would be integrated
into the casing, we employed 3D design software. This aided in spatial planning and establishing
connectivity among the components.

Figure 8 3D Model Design

2.5 Device Testing

Performance testing was conducted to evaluate the effectiveness of the "All-in-One Smart
White Cane," specifically employing the YOLOv8 (You Only Look One Level 8) algorithm. This
assessment aimed to gauge the device's ability to facilitate the mobility of visually impaired
individuals. Key parameters measured in these tests encompassed the exceptionally quick
response time required by the YOLOv8 algorithm within the device to detect objects in the user's
vicinity, along with the accuracy of object recognition. The results derived from these tests not
only provide insights into the device's overall performance but also highlight the YOLOv8
algorithm's efficiency, particularly its swift response time, in enhancing the mobility and safety
of users with special needs.

Figure 9 Webcam Detects Nominal Money

3. RESULT AND DISCUSSION

In the Rupiah banknote nominal detection system, various adjustments are applied to the
dataset, including contrast adjustments and changes in object orientation variations. This process
indirectly results in an increased dataset following the adjustments. Contrast adjustment is
performed using histogram equalization. The results and discussion in the journal can elaborate
on how these adjustments impact the performance of the detection system, enhance accuracy, and

182
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

provide detailed insights into the effects of changes in orientation variations on detection
outcomes.

Figure 10 Adjusting Contrast And Variability of Image Orientation

Figure 11 Training results using the YOLOv8 model

In this research, training was conducted using a pre-trained model from YOLOv8s with
the configuration of epochs = 30 and optimizer = SGD. The image above illustrates the training
results using the YOLOv8 model. It is evident that the recall metric shows improvement starting
from epoch 20 onwards, while precision remains relatively stable between epochs 20 to 40.
In this study, the performance evaluation of the Rupiah currency detection model using
YOLOv8 is elucidated through various metrics. Box Loss measures the accuracy of predicting
bounding boxes, crucial in determining the location and size of the detected currency
denomination. Class Loss assesses the model's ability to predict object classes, representing
different currency denominations in this context. DFL Loss gauges how well the model predicts
the probability distribution of objects, particularly in detecting the likelihood of accurately
identifying currency denominations. Graphs illustrate that the values of Box Loss, Class Loss,
and DFL Loss tend to decrease with increasing epochs, indicating an improvement in the model's
performance in predicting the location, type, and probability of currency detection. Precision and
recall also exhibit improvement with increasing epochs, indicating that the model continues to
learn and enhance the accuracy of object detection. Additionally, the values of mAP50 and
mAP50-95 consistently show improvement in detection performance, even for small-sized
objects or in complex environments. These findings depict that YOLOv8 holds practical potential
for assisting visually impaired individuals in accurately identifying currency denominations with
high precision.

183
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

Figure 12 Precision-Recall (PR) Curve

The Precision-Recall (PR) curve is a visual tool used for evaluating the performance of a
classification model, especially when dealing with imbalanced datasets. The PR curve depicts
how precision and recall change as the model's decision threshold is adjusted. In this study, the
PR Curve on the validation set demonstrates an overall average value of 99%, with the highest
values observed in classes 10,000, 100,000, 20,000, 5,000, and 50,000 at 99.5%, and the lowest
value in the 2,000 class at 97.4%.
These evaluation metrics provide insights into the model's ability to effectively classify
and recognize different Indonesian Rupiah currency denominations. The PR curve, in particular,
illustrates the model's precision and recall performance across various currency classes,
highlighting its robustness and accuracy in handling imbalanced datasets.

4. CONCLUSION

The YOLOv8 model proves to be well-suited for training on Indonesian Rupiah currency
datasets and can be implemented into a white cane device, creating a novel product that aids
visually impaired individuals in recognizing currency denominations by providing auditory
feedback. With the utilization of adequate datasets and careful training, this model delivers
accurate results. However, while YOLOv8 excels in currency detection, it still faces challenges
when dealing with backgrounds that closely resemble the currency's color. This suggests that
further improvements in image processing or pre-detection processing may be necessary to reduce
identification errors. The performance of the YOLOv8 model in terms of accuracy can be
enhanced through various methods such as fine-tuning, data augmentation, or exploring more
advanced model architectures. Continuous testing and model refinement will remain interesting
areas of research.

184
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

The integration of YOLOv8 into a white cane device opens up possibilities for greater
independence and empowerment for the visually impaired community. As technology continues
to advance, refining this solution will contribute to improving the quality of life for individuals
with visual impairments.

REFERENCES

[1] W. Artini, W. Artini, T. D. Gondhowiarjo, T. Rahayu, and Y. D. Lestari, “Impacts of

Impaired Vision and Eye Diseases on Vision-Related Quality of Life in Indonesia,”
Makara Journal of Health Research, vol. 21, no. 3, Dec. 2017, doi:
10.7454/msk.v21i3.7612.
[2] J. Wang, S. Wang, and Y. Zhang, “Artificial intelligence for visually impaired,” Displays,
vol. 77, p. 102391, Apr. 2023, doi: 10.1016/j.displa.2023.102391.
[3] A. Hermawan, L. Lianata, Junaedi, and A. R. K. Maranto, “Implementasi Machine
Learning Sebagai Pengenal Nominal Uang Rupiah dengan Metode YOLOv3,” SATIN -
Sains dan Teknologi Informasi, vol. 8, no. 1, pp. 12–22, Jun. 2022, doi:
10.33372/stn.v8i1.816.
[4] K. M. Azhar, I. Santoso, and Y. A. A. Soetrisno, “Implementasi Deep Learning
Menggunakan Metode Convolutional Neural Network Dan Algoritma Yolo
Dalam Sistem Pendeteksi Uang Kertas Rupiah Bagi Penyandang Low Vision,”
Transient: Jurnal Ilmiah Teknik Elektro, vol. 10, no. 3, pp. 502–509, Sep. 2021, doi:
10.14710/transient.v10i3.502-509.
[5] F. Andika and J. Kustija, “Nominal of Money and Colour Detector for the Blind People,”
IOP Conference Series: Materials Science and Engineering, vol. 384, p. 012023, Jul.
2018, doi: 10.1088/1757-899x/384/1/012023.
[6] A. D. K. Zulfiansyah, H. Kusuma, and M. Attamimi, “Rancang Bangun Sistem
Pendeteksi Keaslian Uang Kertas Rupiah Menggunakan Sinar UV dengan Metode
Machine Learning,” Jurnal Teknik ITS, vol. 12, no. 2, Sep. 2023, doi:
10.12962/j23373539.v12i2.118320.
[7] A. A. J. V. Priyangka and I. M. S. Kumara, “Classification Of Rice Plant Diseases Using
the Convolutional Neural Network Method,” Lontar Komputer : Jurnal Ilmiah Teknologi
Informasi, vol. 12, no. 2, p. 123, Aug. 2021, doi: 10.24843/lkjiti.2021.v12.i02.p06.
[8] A. Darmawan, I. G. N. G. A. Widyadhana, and E. H. Binugroho, “Implementasi Metode
Deep Learning Pada Prototipe Validator Uang Rupiah,” Sebatik, vol. 26, no. 2, pp. 535–
542, Dec. 2022, doi: 10.46984/sebatik.v26i2.2101.
[9] A. K. E. Lapian, S. R. U. A. Sompie, and Pinrolinvic D. K. Manembu, “You Only Look
Once (YOLO) Implementation For Signature Pattern Classification,” Jurnal Teknik
Informatika , vol. 16, no. 3, pp. 337–346, Jul. 2021.
[10] S. Kudalkar, P. Patil, and N. Shirdhone, “Fake Currency Detection Using Image
Processing,” AIJR Publisher, Jun. 2022. Accessed: Nov. 26, 2023. [Online]. Available:
http://dx.doi.org/10.21467/preprints.388
[11] A. Bhatia, V. Kedia, A. Shroff, M. Kumar, B. K. Shah, and Aryan, “Fake Currency
Detection with Machine Learning Algorithm and Image Processing,” in 2021 5th
International Conference on Intelligent Computing and Control Systems (ICICCS), May
2021. Accessed: Nov. 26, 2023. [Online]. Available:
http://dx.doi.org/10.1109/iciccs51141.2021.9432274
[12] A. Antre, O. Kalbhor, P. Jagdale, and G. Dhanne, “Fake Currency Detection Using
Convolution Neural Network,” International Research Journal of Modernization in
Engineering Technology and Science, vol. 5, no. 4, Apr. 2023, doi:
10.56726/irjmets35901.
[13] G. Khoharja, Liliana, and A. N. Purbowo, “Aplikasi Deteksi Nilai Uang pada Mata Uang
Indonesia dengan Metode Feature Matching,” Jurnal Infra, vol. 5, no. 1, pp. 51–55, 2017.
185
 Techno.COM, Vol. 23, No. 1, Februari 2024: 176-186

[14] Y. Su, B. Cheng, and Y. Cai, “Detection and Recognition of Traditional Chinese
Medicine Slice Based on YOLOv8,” in 2023 IEEE 6th International Conference on
Electronic Information and Communication Technology (ICEICT), Jul. 2023. Accessed:
Nov. 26, 2023. [Online]. Available:
http://dx.doi.org/10.1109/iceict57916.2023.10245026
[15] G. Wang, Y. Chen, P. An, H. Hong, J. Hu, and T. Huang, “UAV-YOLOv8: A Small-
Object-Detection Model Based on Improved YOLOv8 for UAV Aerial Photography
Scenarios,” Sensors, vol. 23, no. 16, p. 7190, Aug. 2023, doi: 10.3390/s23167190.

186