2018 International Conference On Advances in Communication and Computing Technology (ICACCT)

Amrutvahini College of Engineering, Sangamner, Ahmednagar, India. Feb 8-9, 2018

Deep Learning Based Indian Currency

Coin Recognition
1 2 3 4 5 6
A.U. Tajane, J.M. Patil, A.S. Shahane, P.A. Dhulekar, Dr.
 S.T. Gandhe,Dr.
 G. M. Phade
1, 2, 3, 4, 5, 6
Department of Electronics and Telecommunication Engineering, Sandip Institute of
Technology and Research Center, Nashik, Maharashtra Savitribai Phule Pune University, Pune
1 2 3
[email protected]; [email protected]; [email protected];
4 5 6
[email protected]; [email protected]; [email protected]

Abstract— Automatic coin recognition and identification to increase the performance of coin recognition system which
systemplays vital role in vending machine, slot machine and in is deep learning based by using constitutional neural network
several banking related equipment’s. Most of the existing coin (CNN) [2].
recognitions systems are based on physical parameters of the Deep Learning is a new technology in Machine Learning
coins whereas image processing methodologies relies on
research, which has been introduced with the objective of
extraction of color, shape and edge features. For recognition and
detection of Indian coin we have proposed Deep Learning moving Machine Learning closer to one of its original goals:
approach in this paper. Pretrained convolutional neural network Artificial Intelligence. In this system novel algorithm is used
i.e. AlexNet is trained by using the features such as textures, for training the machine.
colors and shape. The model is trained on more than 1600 images
and can classify images into four object categories like one, two, II. RELATED WORK
five and ten rupees coins. The trained model is tested on various Coin Recognition using Image Abstraction and Spiral
standard and own recorded datasets consist of rotational, Decomposition” is given by Abdolah Chalechale in year
translated and shifted images. The parameters used to calibrate 2007.This approach is based on finding similarities between
the performance system are recognition accuracy and response
full color and multi-component coin image. This system does
time. Obtained results shown the outperformance of proposed
methodology over conventional systems.
not need image segmentation due to which cost is reduced. A
research, based on deriving strong edges of abstract images on
Keywords—Coin recognition; Deep learning; Convolutional coin image. A set of compact, effective features is extracted
Neural network; AlexNet. using pixel decomposition on abstract image. The query set,
image database is then further used in the tests are scanned,
I. INTRODUCTION photographed, or images downloaded from the web. The result
is compared with many other well- known approaches within
Coin recognition has a tremendous range of application. For
the literature. Practical results show good improvement in the
example, vending machine, slot machine etc. There are 3
traditional ways for the coin recognition that Mechanical, Recall ratio using the proposed features [3, 4]. Discovering
Electromagnetic and image processing. Mechanical coin Characteristic Landmarks on Ancient Coins Using
recognition is normally based on the physical characteristics Convolutional Networks” is given by Jongpil Kim and
of coin structure. Such as radius, weight, width. However, if Vladimir Pavlovic Rutgers in year 2016. There system
two coins with same physical structural properties then system presents method which find characteristic landmarks and
fails is not able to compare those two coins. Whereas, in recognize ancient Roman imperial coins using deep
electromagnetic approach it deals with material properties [1]. convolutional neural networks (CNNs) combined with expert-
In image processing, the coin recognized on the basis of designed domain hierarchies. First they propose new
pattern of coin. The image based coin identification involve framework which recognize the Roman coin by exploits the
the basic two steps that is taking the picture of coin and hierarchical knowledge structure embedded in the coin
compare it with the reference coin data base. The data base domain, which then combine with the classifiers based on
consists of pictures of different coins taken at different angle. CNN. Then next formulate an optimization problem to
Firstly, using camera or scanning device the image of a coin is discover class specific salient coin regions. Analysis of
taken for recognition. FFT, DCT, edge detection, Gabor discovered salient regions confirms that they are largely
Wavelets image subtraction, decision trees, ANN, consistent with human expert annotations. Practical results
show that the proposed system is able to successfully
segmentation, SIFT etc. are used to process the image. Based
recognize the ancient Roman coins and effectively identify
on features extracted, different coins are recognized in
landmarks in a general fine-grained classification problem. For
different categories. Here we have a new and better approach
this research, we have collected a new Roman coin dataset

978-1-5386-0926-2/18/$31.00 ©2018 IEEE 130

where all coins are annotated and consist of obverse (head) C. Neural Network
and reverse (tail) images [5]. 1. Deep neural network
III. DEEP LEARNING APPROACH Deep neural networks have a multiple hidden layer and the
one hidden layer is known as shallows and each of the hidden
A. Deep learning layer is used to perform the different task that means we can
Deep learning is algorithm of machine learning that able to extract the different features of the given image. And
in the deep neural network output is convolve to the input of
teaches computers to do a specific job that is object
the next layer.
identification or the image recognition which is learn by
example. Deep learning is plays vital role in technology like 2. Non-Deep neural network
driverless cars, teaches them to identify a stop sign or the It contains the only one hidden layer and that is used for the
unwanted disturbance. It is the solution to voice control in
training it can be supervised or unsupervised.
electronics devices like phones, tablets, TVs, and hands-free
speakers. As Deep learning is novel it has different D. Convolutional neural network
advantages. Its achieving results that were not possible by Convolutional Neural Network (CNN) is the network which
using pervious technology. consists of more than one convolutional layer and followed by
In deep learning, a computer model learns to perform the more than one fully connected layer as in a standard
classification tasks directly from images, text, or sound. Deep multilayer network format. The advantage of the 2D structure
learning models can achieve state-of-the-art accuracy, for an input image is taken in the convolutional neural
sometimes exceeding human-level performance. Models are networks. The feature extraction is achieved with the help of
trained by using a large set of labeled data and neural network connections and the weights and followed by the different
architectures that contain many layers [6, 7]. polling methods. The approach of deep learning is
implemented using a convolutional neural network. The
B. Artificial neural network convolution neural network has a variety of applications that is
Artificial neural networks (ANNs) have the concept which detection of the different object and in the object and image
comes from biological concept of the animal brains and processing also. They learn directly from the given image
Neurons. it does not require any programming these type of database, it does not need any manual feature extraction [8].
network is obtained with the help of example. it has a
1. Architecture
tremendous ability of the extract of patterns which are
complex in nature that are not seen by the human or any other A CNN is made of a number of convolutional and
computer system. it can also derive the meaning of the much sampling layers and then fully connected layers. The input
more complicated and the unsecured data. image to the convolutional layer is of size lies between the
1. A Simple Neuron 100KB to 1MB whereas the pixel size of the image represents
The structure of an artificial Neuron shows that it has number the height and width of the image. The convolutional layer
of the input with one output. There are two operating modes of will have a different multiple layer whereas each convolution
the simple Neuron that are the mode while it trains and the layer has different properties and functions. The training
user mode. In the first mode of operation the Neuron is get architecture of CNN consist of the 3 fully connected layers
trained to perform a specific task for the particular input. The and five convolutional layers that means total eight training
using mode is somewhat different than the above in this when layers. In this Layer 1 is a convolutional layer with the output
the input is detected at the input then the output becomes the of 55*55*96 size of the image. The total number of Neurons
current output. If the input pattern does not belong in the input present in this is 290400. The figure below shows CNN
patterns, the firing rule is used to determine whether to fire or consisting of convolutional and sampling layers [9, 10].
not. Figure 3.1 shows the concept of simple Neuron. E. AlexNet
AlexNet is the type of network which is based on the or run
with GPU support and it is convolutional neural network
written with CUDA., which competed in the Visual
Recognition Challenge 2010 challenge of classifying millions
of images into 1000 unique categories. Image net dataset is a
very good data set for Deep Learning due to its size of 1.2
million images and 22000 categories which is lacking in
traditional datasets like CIF AR 10/100, MNIST and CalTech-
101. The error compatibility of the AlexNet is 15.3%, more
Fig 3.1: Concept of Neuron
than 10.8 percentage points ahead of the runner up. AlexNet
was invented by the Supervision group, consisting of Alex

131
 Fig.3.2 Architecture of AlexNet

Krizhevsky, Geoffrey Hinton, and Ilya Sutskever. The Compares with the database provided by the user and gives
architecture of AlexNet is shown in Figure 3.2. AlexNet precise output to the display.
consist of eight learning layers, five convolutional and three
fully connected layers. The output of the final fully connected Camera
layer is a soft max regression which converts the weights to (227*227) pix Processor Display
probability distributions of the given 100classes. The input
images are of 224*224*3 resolution which feed to the
convolutional layer which filters the image using 96 kernels of
size 11*11*3 at a stride of 4 pixels [11, 12]. Keypad
All the layers that is third, fourth and last convolutional layers Fig 4.1: Block diagram
are connected in cascading without any intermediate like
pooling.
V. DATASET
IV. METHODOLOGY Developed algorithms have been tested on two types of
datasets. The detail description of these datasets are given in
We have divided the proposed system into two different following subsections.
phases namely software implementation of project on the
A. Standard Dataset
MATLAB and the implementation using a controller. In this
we have been used a raspberry pi as a controller board through We have collected standard dataset from various URL which
that we can control the operation of the system. In the contains hundred images of each object category such as one,
hardware implementation of the project firstly we captured the two, five and ten rupees coins. The resolution of all images is
snapshot of the given inserted coin then it will be process 256x256 with memory size of around 100KB to 1MB.
using the raspberry pi controller. System overview is shown in
figure 4.1. B. Own Dataset
In this system after the completion of software Own dataset captured by ourselves then datasets are used for
implementation we deploy that system in real time system the training purpose in this type of datasets we can take the
such as vending machines. The block gives us the full different types of images of coin that is tilted image and the
information how we deploy that in real time system. The images in various angles. The sample of own recorded
system proposed in this paper consists of different blocks that datasets containing one rupees, two rupees, five rupees and ten
are camera, controller such as raspberry pi and keyboard, rupees are shown in subsequent figures 5.1,5.2,5.3 and 5.4.
display etc.
Camera is used for the image acquisition in this with the
help of camera we capture the image of the coin with the help
of this we can capture the tilted image and the full image of
the coin. SONY IMX219 8MP sensor which has a property of
Wide life camera traps. Then that output of the camera is fed
to the controller block the controller block is nothing but the
raspberry pi controller which controls the whole system. The Fig.5.1: One Rupees Coin
output of camera is fetched by the controller and the controller
resize that images in a standard size that is 227*227 and

132
 Fig.5.2: Two Rupees Coin

Fig.6.2 AlexNet Identified TWO Rupee Coin

Fig.5.3: Five Rupees Coin With the dataset of five rupees and ten rupees coins AlexNet is
trained and testing results are shown in figure 6.3 and figure
6.4.

Fig.5.4: Ten rupees coin

VI. RESULTS
In the first phase we have trained AlexNet in a such a way that
it is able to identify the different coins and then we kept one-
rupee coin in front of camera then the trained network
identified the given coin as one rupee and two rupee which is
shown in figure 6.1 and figure 6.2. Fig.6.3 AlexNet Identified FIVE Rupee Coin

Fig.6.1 AlexNet Identified ONE Rupee Coin Fig.6.4 AlexNet Identified TEN Rupee Coin

133
 VII. CONCLUSION AND FUTURE SCOPE [3] Abdolah Chalechale, “Coin Image Analysis, Semantic Evaluation and
Efficiency Assessment”, pp.01-04, 2007.
Conventional image processingbased methods requires a [4] Kiran Kale, Sushant Pawar and Pravin Dhulekar, “Moving Object
lot of training images in order to achieve the higher accuracy Tracking Using Optical Flow and Motion Vector Estimation” Published
or the better performance whereas comparatively deep in IEEE 4th International conference on reliability, Infocom technologies
and optimization (ICRITO 2015), 2-4 Sept 2015.
learning need lesser data for training which ultimately [5] Preeti Kale, G. M. Phade, S. T. Gandhe, Pravin A. Dhulekar,
increases the speed of operation and reduces response time. In “Enhancement of old images and documents by Digital Image Processing
our preliminary coin recognition experiments, we trained an Techniques” Presented in IEEE International Conference on
AlexNet on a pre-prepared dataset of images, which Communication, Information & Computing Technology (ICCICT),
confirmedthat the amount of training data was enough. The Mumbai, India. 16-17 January 2015.
[6] Kaiping Wei, Bin He, Fang Wang, Tao Zhang, Quanjun Ding “A Novel
performance of the best model showed that 1000 -coin images Method for Classification of Ancient Coins Based on Image Textures”,
are needed in the training phase but significantly all the ranges Second Workshop on Digital Media and its Application in Museum &
of the coin should be considered in that. The deep learning Heritage, pp.63-66, 2007.
architecture that we implemented proved to have substantial [7] P.A. Dhulekar, S.T. Gandhe, Anjali Shewale, Sayali Sonawane, Varsha
Yelmame, “Motion estimation for human activity surveillance”, IEEE
advantages: these include an immediate response to the user, International Conference on Emerging Trends & Innovation in ICT
and the ability to use GPUs for training, validation and testing (ICEI), Pages 82-85, 2017.
of the NN in order to classify models quickly. We developed [8] P.A. Dhulekar, Priyanka Aher, Swapnali Choudhari, Yogita Khairnar,
an application for with the real-time basis that provide a “Design of IR based image processing technique for digital camera
visualization of the results based on an image of a coin, deactivation”, IEEE International Conference on Global Trends in Signal
Processing, Information Computing and Communication (ICGTSPICC),
obtained using the device’s camera. The proposed work can be Pages 83-86, 2016.
extended by deployment of developed algorithms in real time [9] Hussein R. Al-Zoubi, Member, “E cient Coin Recognition Using a
embedded systems. Statistical Approach”, pp.01-04, 2010.
[10] Priyanka Borude, Dr.S.T. Gandhe, Pravin A. Dhulekar, G.M. Phade,
“Identification and Tracking of Facial Features” Published in Elsevier
REFERENCES International Conference on Advances in Computing, Communication
and Control (ICAC3 '15),1st-2nd April 2015.
[1] Nicola Capece, Ugo Erra, Antonio Vito Ciliberto “Implementation of a [11] G. M. Phade, Prerna D. Uddharwar, Pravin A. Dhulekar, Dr. S. T.
Coin Recognition System for Mobile Devices with Deep Learning”, 12th Gandhe “Motion Estimation for Human-Machine Interaction” Presented
International Conference on Signal-Image Technology & Internet-Based in IEEE International Symposium on Signal Processing and Information
Systems, pp.187-192, 2016. Technology (ISSPIT), Noida, India, 15-17 Dec 2014.
[2] Jongpil Kim and Vladimir Pavlovic Rutgers, “Discovering Characteristic [12] Adnan Khashman, “Intelligent Coin Identification System”, International
Landmarks on Ancient Coins Using Convolutional Networks”, 23rd Symposium on Intelligent Control Munich, Germany, October 4-6, 2006.
International Conference on Pattern Recognition (ICPR) Cancún Center,
Cancún, México, December 4-8, 2016

134