2021 International Conference on Emerging Smart Computing and Informatics (ESCI)

AISSMS Institute of Information Technology, Pune, India. Mar 5-7, 2021

Handwritten Digit Recognition of MNIST dataset

using Deep Learning state-of-the-art Artificial Neural
Network (ANN) and Convolutional Neural Network
2021 International Conference on Emerging Smart Computing and Informatics (ESCI) | 978-1-7281-8519-4/20/$31.00 ©2021 IEEE | DOI: 10.1109/ESCI50559.2021.9396870

(CNN)
Drishti Beohar Akhtar Rasool
Computer Science Department Computer Science Department
Maulana Azad National Institute of Technology, Maulana Azad National Institute of Technology,
Bhopal, India Bhopal, India
[email protected] [email protected]

Abstract -- Handwritten digit recognition is an intricate c) arranging mail,

assignment that is vital for developing applications, in computer
d) perusing checks and so forth.
vision digit recognition is one of the major applications. There has
been a copious exploration done in the Handwritten Character That is the reason OCR is a region of enthusiasm for a large
Recognition utilizing different deep learning models. Deep number of analysts as of late. Fundamentally, writings being
learning is rapidly increasing in demand due to its resemblance to in various dialects and contents, the efficiency of the
the human brain. The two major Deep learning algorithms recognition system to a great extent relies on the classification
Artificial Neural Network and Convolutional Neural Network strategies utilized. Consequently, the improvement of such a
which have been compared in this paper considering their feature framework was an outrageous task. Artificial intelligence has
extraction and classification stages of recognition. The models opened many doors before this AI era. Anyone who wants to
were trained using categorical cross-entropy loss and ADAM
design a classifier has to manually do the feature extraction
optimizer on the MNIST dataset. Backpropagation along with
Gradient Descent is being used to train the networks along with
and feel the machine gives the decision based on the features
reLU activations in the network which do automatic feature feed and the programming done. Earlier the algorithms like
extraction. In neural networks, Convolution Neural Network SVM gave great results but that also came with the cost of
(ConvNets or Convolutional neural networks) is one of the mathematics and consumed a lot of time. Now with the
primary classifiers to do image recognition, image classification enhancement to technology, the Deep learning algorithms in
tasks in Computer Vision. machine learning like Artificial Neural Network,
Convolutional Neural Network, and Recurrent Neural
Keywords—Handwritten digit recognition, Convolutional Neural Network gives an outstanding performance. The character
Network (CNN), Deep learning, MNIST dataset, Epochs, Hidden recognition framework utilizing deep neural networks
Layers, Stochastic Gradient Descent, Backpropagation (DNNs) for different scripts have been created with
practically ideal outcomes.
I. INTRODUCTION
At whatever point we have heard the word Handwritten II. LITERATURE REVIEW
Character Recognition, the initial term that rings a bell is OCR, Hand Digits Recognition turns out to be progressively
termed for Optical Character Recognition, it is a system of significant in the advanced world because of its actual
reading characters and manipulating them into a structure that a implementation in our every day life [1]. Recently , various
machine can interpret. This was the solitary innovation utilized recognition frameworks have been presented inside numerous
by the scientists for 3-4 decades to change over any actual applications where higher order effectiveness is required. It
record into machine editable structure. OCR is a pipeline of causes us to take care of increasingly complex issues and is
5 phases i.e stage one is Image Acquisition, after which simpler..[2] .Programmed .preparing.of .bank .checks, .the .postal
stage two of Pre-processing is cone stage three is .location .is .a .general .utilization of .hand-written .digit

Segmentation. After segmentation, the Feature Extraction of .recognition [3]. In this particular paper, .we prepared both
the data is done and the final stage i.e. Classification. OCR Artificial neural network and Convolutional neural network
follows the pipeline idea, progressive paces of each stage rely model to .recognize .written .by .hand .digits .from .0 .to .9. A .node
on the success pace of the past stage. With the headway of .in .a .neural .system .can .be .comprehended .as .a .neuron .in .the

innovation, we need the machine to perform the most extreme brain. Every. node. is associated with .different. nodes through
undertakings. There are various uses of computer vision like .weights (which are basically the edges between the nodes)

which .are .balanced in the algorithm. A value is determined for

a) interpretation of the report,
every node dependent on the feature .and. methods .of .previous
b) language translation, node. This procedure is called forward .propagation [6]. The

978-1-7281-8519-4/21/$31.00 ©2021 IEEE 542

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 16,2022 at 10:38:08 UTC from IEEE Xplore. Restrictions apply.
.last output .of .the .system .is .related .with .the .objective .output, .at
.that .point weights .are changed according to the loss function .to
.depicting .whether .the .system is .speculated .effectively .[7]. This

.procedure .is .called .back .propagation .[8]. To include complexity

and. correctness .in the .neural network, the systems .have .different
Layers. In the middle of a fully connected neural system,.there
are various. layers. that .exist, in particular .information, .output
.and .hidden .layers. Suppose we have features x1, x2, x3….. xn.

The edges from one node to the other node of the network have
weights that play the most important role in both forward and
backward propagation. In forward propagation, there are two
types of operation that happens in the hidden layer with the Fig. 2. Forward and Back propagation
feature and the weights being passed to the neuron or node. The
sum of the product of feature and weights and then applying an Whenever we have a deep Neural network or a network
activation function. Whenever we have a Neural Network which with a huge number of layers then we have a huge number of
is very deep at that time you will understand there are many weights and bias parameters as well which leads to overfitting
weights and bais parameters. In backward propagation we have the dataset problem or a particular data. In a multilevel Neural
to change the values of the previous epoch weights, this reduces Network, underfitting will never happen because we will be
the loss value. In a .completely .associated .neural .system .nodes having multiple levels that try to fit the training data perfectly.
.in .each particular layer .are .associated .with .the .nodes .and .the High variance is a problem with increasing levels in the
.layers preceding .and .succeeding .them[9]. network. We can apply regularization(L1 or L2) or Dropout
layer to decrease the overfitting problem. In a Random forest
multiple decision trees are created. Every Decision tree is
created to its depth which also leads to an overfitting problem.
Similarly, like the decision tree, we will be using a subset of
features which is regularization which improves the accuracy
of the whole model.

Fig. 1. Artificial Neural Network

In 1980-2000 researchers were not able to create a deep

neural network in an Artificial neural network. The reason is the
use of sigmoid function in every neuron ( in 1980-2000 the
ReLU was not invented). This is termed a vanishing gradient
problem in a Neural network. The Activation function(sigmoid),
when applied to the summation of the product of weights and
the features, is always ranging between 0-1 and the derivative of
the activation function ranges between 0-0.25 which gets
smaller when the layers of the neural network become deeper.
To deal with the vanishing gradient problem the use of ReLU or
other activation function which does not lead to the collapse of
the derivative is used.

‫ݓ‬௡௘௪ ൌ  ‫ݓ‬௢௟ௗ െ ߟ ‫ߜ כ‬ሺ‫ݏݏ݋ܮ‬ሻȀߜሺ‫ݓ‬௢௟ௗ Ȍ (1) Fig. 3. Graph of derivative ߜሺ‫ݏݏ݋ܮ‬ሻȀߜሺ‫ݓ‬௢௟ௗ ሻ for vanishing and exploding
gradient problem

Equation-1 Gradient Decent

When the weights assigned are large numbers then the

expected number of the derivative ߜሺ‫ݏݏ݋ܮ‬ሻȀߜሺ‫ݓ‬௢௟ௗ ሻ will be a
very large number which will result in a large variation in the
new and old values when backpropagating. Then new weight
will jump on large values over the epochs and the weights will
vary a lot with the value never converging at a point. So the
weight initialization in a Neural network is a very crucial point
Fig. 4. CNN Model Architecture
otherwise this can lead to an Exploding Gradient problem[10].

543

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 16,2022 at 10:38:08 UTC from IEEE Xplore. Restrictions apply.
 In the Neural Network, we select a subset of features from then be able to be plotted to give expectations to learn and
the input layer and select a subset of hidden neurons. The other adapt knowledge into how well a model is learning the issue.
neurons which are not selected in the subset are deactivated
[11]. The number of nodes in a subset count is calculated by the IV. THE RECOGNITION MODEL
use of the dropout ratio. In image classification, object Optical character recognition(OCR) is a recognition
detection, and many other data augmentation Convolutional system that has various stages. Each stage plays a very
Neural Network(Convolutional neural network) plays a very important role in the model. The stages are pipelined one after
major role. In the Convolutional neural network, the input data other Fig 7 shows the stages in the Recognition model.
is in the form of a matrix which is having values in each cell
ranging from 0-255 and either one or 3 artificial neural networks A. Image Acquisition
depending on grayscale and RGB scale respectively.[12] Image acquisition is the first stage of all the recognition
models. In this stage, the images are gathered, filtered, and
cleaned before any preprocessing is done on the images.
B. Pre-Processing
Preprocessing is a very vital operation in the image. In
Pre-processing major operations that are carried are image
cleaning to reduce the noise in the image and removing the
Fig. 5. Operations on an image using CNN model
garbage. The optimization of the image is also done in this
stage by filling the voids or holes, straightening curved lines.
The filters are applied to images and the output is also a
Different algorithms are also performed for skew correction.
matrix in a particular operation. The images go through a
The output of this particular stage is a binary image which is
pipeline of operations of convolution layers with the filter,
done by binarization and texture filtering.
pooling, fully connected layer, and applying softmax function.
The beneath figure is the complete architecture of a convolutional C. Segmentation
neural network to process an input picture and classify it based on Decomposition of an image into sub-images is
values.[13] segmentation. Segmentation is of three types line, word, and
III. SCRIPTS AND DATASETS character segmentation. When the input is an image with
multiple lines breaking that image into a single line is line
segmentation. When the input image is an image with a single
line but multiple words and words have to be segmentation is
word segmentation. Similarly, in character segmentation, the
words are segmented into words.
D. Feature Extraction
Feature Extraction is a very important stage in the
recognition model. It is a part of dimensionality reduction
techniques. In Dimensionality reduction, the input data is
converted into more simple and easy operation data. Large
datasets like MNIST are great for this step as this particular
stage optimizes the whole process of recognition. This stage
removes the redundant data by retaining the originality of the
dataset. In image processing, the feature extraction stage helps
in edge detection and many other operations. Without the
feature extraction stage, the classification of the image is a bit
more complex and time-consuming. PCA and Image pixel
Fig. 6. MNIST Dataset vector are some techniques for feature extraction.
The MNIST is a great dataset for the handwritten digit E. Classification
classification problem. The MNIST dataset is a very Classification is the decision-making stage in the pipeline
authenticated and great dataset for the students and researchers. of image recognition. The input to this stage is the output of
It has 60000 images with 10 classes (0-9) which is enormous in the feature extraction stage. For classification nowadays many
itself. Each image in the MNIST dataset is of 28 height and 28 classifiers are present like Logistic regression, random forest,
weight which make the image of 784-dimensional vectors. The K nearest neighbors (KNN), Support vector machine (SVM)
MNIST dataset is available easily on the internet. Each image in algorithm, Artificial Neural Network (Artificial neural
MNIST is a grey-scale image and the range is 0-255 which network), Convolutional Neural Network (Convolutional
indicates the brightness and the darkness of that particular pixel. neural network), and many more. For image classification,
The MNIST dataset was created by the National Institute of Deep neural network classifiers give great results i.e.
Standards and Technology(NIST). To estimate the performance Artificial Neural Network and Convolutional Neural
of a model, we split the preparation set into a training and Network. The MNIST dataset is huge and classifiers like
testing dataset. Execution on the train and testing dataset would Artificial Neural Network (Artificial neural network) and

544

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 16,2022 at 10:38:08 UTC from IEEE Xplore. Restrictions apply.
Convolutional Neural Network (Convolutional neural network) B. Convolutional Neural Network
give great accuracy on training 80% of the dataset and testing
20% of the dataset.

Fig. 9. CNN Dimensionality reduction in each stage

Convolutional neural networks are one of the best models

in deep learning algorithms. Convolutional neural networks
are spectacular for image classification in computer vision.
The main components in Convolutional neural networks are
Fig. 7. Character Recognition Stages
feature extraction and classification. Deep learning algorithms
V. CLASSIFIERS are designed in a way that has similarities with the human
brain. As we need several pictures to learn a particular object
A. Artificial Neural Network: in the same way the Convolutional neural network is trained
with the images. Unlike the human brain, their world only
understands the numbers so to feed the Convolutional neural
networks with those training images an image is represented
as a 2D matrix (also pixels). Like an Artificial neural network,
the Convolutional neural network doesn’t have similar
architecture. In the Convolutional neural network, the layers
of the model have architecture such that they are in #
dimensions i.e. width, height, and depth. Like an Artificial
neural network in a Convolutional neural network in the
convolution layer, all the layers are not connected with the
ones in the next layer. The final result in the Convolutional
Fig. 8. ANN Network
neural network is a vector with dimensions that has a value of
The artificial neural network speaks to Artificial Neural the probability scores. The first layer is the convolutional
Networks. It’s a computational model. That relies upon layer following the pooling layer and the last Fully-connected
structures and components of normal neural frameworks. layer. In a Convolutional neural network the feature extraction
Regardless of the way that, the structure of the Artificial neural is followed as a Feature map is produced with the use of two
network is impacted by a movement of information. Therefore, components the first one is a filter and the second is the kernel
the neural framework changes relied upon data and output. It in the first layer. In the Convolutional layer, the method of
imitates the human cerebrum's capacity to recognize designs. convolution is done by multiplication at every cell and then
Neural networks are fundamentally utilized for arrangement summing up the result onto the feature map. To make our
issues. The feed-forward technique got popular with the output nonlinear we use an activation function(ReLU,
Artificial neural network model also called multilayer Sigmoid). Stride is the size of the step the convolution filter
perceptron. The node or neuron in the network or the Artificial moves each time. To prevent the feature map from shrinking
Neural Network model is connected with the edge which has as its size tend to be smaller than the input we use padding.
their respective weights. The weights are the key to classifying As the input in the Convolutional neural network is mostly
the class of the given input. Figure 8 shows the network is large and the result is a vector that shows a need which
broadly classified into three classes of input layers which take dimensionality reduction. In a convolutional neural network,
the input i.e. features from the fourth stage of the recognition it is handled by adding a pooling layer in between the layers.
model and then passes it onto the hidden layer which may have The pooling layers is a great way to control overfilling in a
1 or more layers and each layer depending upon the complexity Convolutional neural network model. There is two type of
of the dataset. The number of layers and neurons in each layer pooling - max, min. The classification stage of the
artificial neural network be too high or too low. Too large Convolutional neural network model is handled by the Fully-
several layers and neurons can lead to an overfitting problem Connected layer. The features being extracted by the
which can be handled using normalization or dropout. convolution layer and pooling layer are passed on to the
Backpropagation is a great technique by Jeffrey Hinton which Fully-Connectedwhich performs classification. Fully-
opens a lot of doors in the field of deep learning. It lets us have a connected usually accepts 1-D data. The last layer of the
deep neural network. The Gradient descent technique is used to Convolutional neural network is very similar to the Artificial
update the weight of the edges and bais when backpropagating.

545

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 16,2022 at 10:38:08 UTC from IEEE Xplore. Restrictions apply.
neural network. The last layer is strongly connected and is The epochs value is chosen as 10 with the batch size value of
called a fully-connected layer. 200 for 60000 images being trained. For calculating the loss
the categorical cross-entropy which is a logarithmic function
VI. RESULTS is used and optimization is done by the ADAM i.e. Adaptive
A. Implementation using Artificial Neural Network: Moment Estimation algorithm for modifying the values of the
weights and bias in the backpropagation. And the value of
The Digit recognition of the MNIST dataset consists of 0-9 Baseline error was achieved as 1.31%.
digits which act as classes in classification. The PyCharm IDE
(Integrated Development Environment) has built-in developer B. Implementation using Convolutional Neural Network
tools and is a customizable and cross-platform IDE. PyCharm is (Convolutional neural network or ConvNet):
used with the latest stable version of Python3.7. As discussed The Digit recognition of the MNIST dataset consists of 0-
above in the recognition model we have 5 stages: the Data 9 digits which act as classes in classification. The PyCharm
acquisition is already implemented as MNIST is a very reliable IDE (Integrated Development Environment) has built-in
dataset. In the Image processing phase in the Artificial Neural developer tools and is a customizable and cross-platform IDE.
Network to make all the images uniform for reducing the PyCharm is used with the latest stable version of Python3.7.
complexity of the dataset. The loading of the data is done by the As discussed above in the recognition model we have 5
python library Numpy which is a fundamental package for stages: the Data acquisition is already implemented as MNIST
scientific computing in python). As mentioned earlier the Model is a very reliable dataset. The Convolutional Neural Network
of Artificial Neural Network has 3 layers Input, hidden, and is Not as simple as the Artificial Neural Network to be
output layer. The input to the next layer is the output of the trained. Like Artificial Neural Network had the number of
previous one. In the Neural Network, the size of the input image neurons in the input layers as the number of pixels in the
is equal to the number of neurons in the input layer. In the image(i.e. image size) here we have a 2-D matrix of the
dataset description, we mentioned it as 28x28 which is 784 network.
pixels. The output of every layer is calculated with the help of
the activation function which in our model is the ReLU
activation function. The number of neurons in the hidden layer
is kept the same as that of the input layer. The number of classes
in the MNIST dataset is 0-9 which is 10 classes so the output
layer consists of 10 neurons for 10 classes.

TABLE I. RESULTS OF ARTIFICIAL NEURAL NETWORK WITH LEARNING

RATE-0.2 AND BAIS-0.5

epoch Learn correct Learn wrong Test correct Test

wrong

1 0.96403 0.03592 0.95740 0.0426

2 0.96710 0.03290 0.96000 0.04000

Fig. 10. Confusion matrix depicting true labels and predicted labels
3 0.97073 0.02927 0.96250 0.03750
As described earlier the convolutional neural network has
4 0.97388 0.02612 0.96370 0.03630 3 layers the convolutional layer the pooling layer and the
fully-connected layer. The preprocessing part is done by the
5 0.97395 0.02605 0.96150 0.03850 convolutional layer by applying numerous filters which
enhances the image for the next layer for better segmentation
6 0.97526 0.02473 0.96290 0.03710 and feature extraction. The input to this layer is in the form of
a matrix with 3 dimensions are the height of the image, width
7 0.97413 0.02587 0.96080 0.03920
of the image, and height of the image which can have binary
values i.e 0 or 1. The size of the image here is passed as a
8 0.97853 0.02147 0.96340 0.03660
parameter. These are hyper-parameters so usually we take the
number of filters as 32,64 and so on and sizes of filters as 3x3,
9 0.97616 0.02383 0.96160 0.03840
5x5, etc. The deciding factor of the number of parameters
being learned is by filter size and the number of filters and the
10 0.9802 0.01980 0.96380 0.03620
model learns the value of the filter itself.
The next layer which handles the segmentation and the
feature extraction is the MaxPooling2D. Each stride is taken
We used the softmax activation function to get the and applied with the Max Pooling function to evaluate the
probabilistic values of the output, which makes it easier to max as the name suggests. This layer takes care of the
choose the maximum value from the given output of the classes. overfitting problem in the model by regularization or Dropout

546

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 16,2022 at 10:38:08 UTC from IEEE Xplore. Restrictions apply.
ratio. Here we used the dropout ratio of 20%(here the value of all the more; however execution is less, and in the
drop out ratio p) of some particular randomly selected neurons convolutional neural organization, the underlying two layers
in the layer to reduce the overfitting problem (regularization is for example. The convolution layer and max-pooling layer are
another way to avoid overfitting). The learning rate is minimal handling the picture with the assistance of an artificial neural
as it will help us get the global minimum. The preparation network pipelined by the unequivocally associated layers and
dataset is coordinated as a 3-dimensional exhibit of occurrence, several learnable boundaries are less yet the execution is
picture width, and picture tallness. For a multi-layer perceptron better. Fit and evaluate the model can be a choice. The
model, we ought to decrease the photos down into a vector of particular model fits more than 10 epochs which revive every
pixels. For the present circumstance, the 28×28 measured 200 pictures. In the approval, dataset test data is used. Input to
pictures will be 784-pixel input esteems For this situation the one line for each preparation epoch a verbose estimation of 2
28×28 sized pictures will be 784-pixel input values. is used.

TABLE II. RESULTS OF CONVOLUTIONAL NEURAL NETWORK WITH

BASELINE ERROR OF 0.91%

Epoch Loss Acc Val-loss Val-acc

1/10 0.2754 0.9231 0.1339 0.9600

2/10 0.1089 0.9684 0.0935 0.9717

3/10 0.0710 0.9794 0.0866 0.9743

4/10 0.0496 0.9854 0.0732 0.9766

5/10 0.0358 0.9900 0.0634 0.9798

Fig. 11. Accuracy and Loss graph

6/10 0.0257 0.9933 0.0597 0.9826

7/10 0.0192 0.9955 0.0626 0.9802

VII. CONCLUSION
Convolutional Neural Network and Artificial Neural
8/10 0.0142 0.9969 0.0612 0.9817 Network both are trained and tested with the MNIST dataset.
Both the models were trained with an 80% dataset and 20%
9/10 0.0107 0.9981 0.0573 0.9825 was tested. The models used the ReLU activation function for
the backpropagation algorithm. And softmax activation
10/10 0.0081 0.9985 0.0558 0.9829 function for the probabilistic values of the output. The
training was done with the categorical cross-entropy loss. For
further optimization, the ADAM optimizer was used for
The discussion above in the classifiers section about decreasing the loss. The critical point noticed during the
Convolutional Neural Network that the dimensionality reduction training in hardware purpose was that the average error of
is done the input layer was a 3D matrix and with Max Pooling Convolutional neural networks is less than artificial neural
layer it is converted to a 2D matrix and the final layer or the networks on the CPU. Convolutional Neural networks gave
output layer is feed with the 1D matrix. To convert the 2D better performance for image classification. The average
matrix to 1D form we use the Flatten function. The output layer baseline error for the Artificial Network was 1.31% and for
is fed with 128 neurons after flattening and an activation Convolutional Neural Network was 0.91%. This clearly
function which in our case is the ReLU function. The output shows the advantage of the Convolutional Neural Network
layer has 10 neurons for 10 classes with the softmax activation over the Artificial Neural Network. The disadvantage of the
function for the probabilistic value of the output. The Convolutional Neural Network over the Artificial Neural
logarithmic capacity which is likewise called unmitigated cross- Network is that Convolutional Neural Network takes more
entropy in Keras is utilized as the misfortune capacity and time and CPU power. For reducing the time we can use GPU
ADAM (Adaptive Moment Estimation) advancement over a CPU for better performance. Since Convolutional
calculation is performed to get familiar with the various loads neural networks could be successfully utilized in more
and their inclinations. For assessment of the model, the standard impressive and cutting edge computers. This above-examined
error is determined utilizing the precision metric on the test strategy could be probably the best element for all the bodies,
pictures. Furthermore, the estimation of Baseline error was working in the recognition of manually written digits. This
accomplished as 0.91%. It is seen that the gauge mistake on leads us to believe that the Convolutional neural network’s
account of convolutional neural organizations is not exactly the strategy is significantly better than other techniques.
benchmark blunder in a convolutional neural network. The vital
contrast in both is the number of layers and learnable ACKNOWLEDGEMENT
boundaries. On account of a fake neural organization, all various I want to pay extraordinary appreciation and warmth to
layers are emphatically associated, and learnable boundaries are my Research Work Guide, Dr. Akhtar Rasool, who allowed

547

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 16,2022 at 10:38:08 UTC from IEEE Xplore. Restrictions apply.
me this opportunity to do this research paper. His crucial help [10] Study and Observation of the Variations of Accuracies for Handwritten
made it possible to accomplish the objective. I might also want Digits Recognition with VariousHidden Layers and Epochs using
Convolutional Neural Network Rezoana Bente Arif; Md Abu Bakr
to thank different writers of reference materials referenced in the Siddique; Mohammad.Mahmudur.Rahman… Khan.; Mahjabin
reference segment for their commendable research. Rahman. O,2019
[11] Beyond human. Recognition: A.Convolutional neural network-based
REFERENCE framework for .handwritten character recognition Li Chen;
[1] M .Nagu, N. V.Shankar, and K.Artificial neural network apurna, "A novel Song.Wang; Wei Fan; Jun Sun; Satoshi Naoi,2020
method for Handwritten Digit Recognition with Neural Networks," 2011. [12] Dhanya Sudarsan; Shelbi Joseph, “A Novel Approach for Handwriting
[2] Y .LeCun, B .E .Boser, J .S .Denker, D .Henderson, R .E .Howard, W .E Recognition in Malayalam Manuscripts using Contour Detection and
.Hubbard, et al., "Handwritten digit recognition with a backpropagation Convolutional Neural Nets”,2020.
network," in Advances in neural information processing systems, 1990, [13] Nanehkaran, Y.A., Zhang, D., Salimi, S. et al. “Analysis and
pp .396-404 . comparison of machine learning classifiers and deep neural networks
[3] A .Ashworth, Q .Vuong, B .Rossion, M .Tarr, Q .Vuong, M .Tarr, et al., techniques for recognition of Farsi handwritten digits”. J Supercomput
"Object Recognition in Man, Monkey, and Machine," Visual Cognition, (2020).
vol .5, pp .365-366, 2017 . [14] S. Oktaviani, C. A. Sari, E. Hari Rachmawanto and D. R. Ignatius
[4] J . Janai, F . Güney, A . Behl, and A . Geiger, "Computer Vision for Moses Setiadi, "Optical Character Recognition for Hangul Character
Autonomous Vehicles: Problems, Datasets, and State-of-the-Art," arXiv using Artificial Neural Network," 2020 International Seminar on
preprint arXiv:1704.05519, 2017. Application for Technology of Information and Communication
(semantic), Semarang, Indonesia, 2020.
[5] K .Islam and R .Raj, "Real-Time (Vision-Based) Road Sign Recognition
Using an Artificial Neural Network," Sensors, vol .17, p .853, 2017 . [15] R. Sharma, B. Kaushik, and N. Gandhi, "Character Recognition using
Machine Learning and Deep Learning - A Survey," 2020 International
[6] D .Arpit, Y.Zhou, B.Kota, and V.Govindaraju, "Normalization
Conference on Emerging Smart Computing and Informatics (ESCI),
propagation: A parametric technique for removing internal covariate shift
Pune, India, 2020.
in deep networks," in International Conference on Machine Learning,
2016. [16] P. Gupta, S. Deshmukh, S. Pandey, K. Tonge, V. Urkunde and S. Kide,
"Convolutional Neural Network-based Handwritten Devanagari
[7] I .Patel, V.Jagtap, and O.Kale, "A Survey on Feature Extraction Methods
Character Recognition," 2020 International Conference on Smart
for Handwritten Digits Recognition," International Journal of Computer
Technologies in Computing, Electrical and Electronics (ICSTCEE),
Applications, vol .107, 2014.
Bengaluru, 2020.
[8] I . H . Witten, E . Frank, M.A.Hall, and C . J . Pal, Data Mining: Practical
[17] P. Dhande and R. Kharat "Recognition of cursive English handwritten
machine learning tools and techniques: Morgan Kauf Artificial neural
characters" 2017 International Conference on Trends in Electronics
network, 2016.
and Informatics (ICEI) pp. 199-203 2017.
[9] Study and Observation of the Variations of Accuracies for Handwritten
[18] Shalini Puri and Satya Prakash Singh "An efficient Devanagari
Digits Recognition with Various Hidden Layers and Epochs using Neural
character classification in printed and handwritten documents using
Network Algorithm Md.Abu Bakr Siddique; Mohammad Mahmudur
SVM" Procedia Computer Science vol. 152 pp. 111-121 2019.
Rahman Khan; Rezoana Bente Arif; Zahidun Ashrafi,2018
[19] J. Schmidhuber "Deep learning in neural networks: an overview"
Neural Networks vol. 61 pp. 85-117 2015.

548

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 16,2022 at 10:38:08 UTC from IEEE Xplore. Restrictions apply.