0% found this document useful (0 votes)
61 views7 pages

An In-Depth Deep Learning Approach To Handwritten Digits Recognition

Due to the variations in human handwriting, computerized handwritten digit recognition is a challenging task. This abstract describes a system that identifies handwritten digits in images and documents using Convolutional Neural Networks built with PyTorch. In order to solve a variety of practical problems, this technology is crucial in applications like check processing, postal sorting, and number plate recognition.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
61 views7 pages

An In-Depth Deep Learning Approach To Handwritten Digits Recognition

Due to the variations in human handwriting, computerized handwritten digit recognition is a challenging task. This abstract describes a system that identifies handwritten digits in images and documents using Convolutional Neural Networks built with PyTorch. In order to solve a variety of practical problems, this technology is crucial in applications like check processing, postal sorting, and number plate recognition.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

An in-Depth Deep Learning Approach to Handwritten


Digits Recognition
1 2
A.Leela Vathi V.Jyothsna Rani
Sr. Asst Professor CSE Department B. Tech III Year, Department of AIML,
Sri Vasavi Engg. College, Tadepalligudem. Sri Vasavi Engg. College, Tadepalligudem.

3 4
A.Devi Siva Prasad S.Sai Nikitha
B. Tech III Year, Department of AIML, B. Tech III Year, Department of AIML,
Sri Vasavi Engg. College, Tadepalligudem. Sri Vasavi Engg. College, Tadepalligudem.

5 6
D.Vijaya Lakshmi K.Renuka
B. Tech III Year, Department of AIML, B. Tech III Year, Department of AIML,
Sri Vasavi Engg. College, Tadepalligudem. Sri Vasavi Engg. College, Tadepalligudem.

Abstract:- Due to the variations in human handwriting, These network architectures are motivated by the
computerized handwritten digit recognition is a complicated structure of the mammalian visible machine,
challenging task. This abstract describes a system that mainly the concept of a receptive area, which changed into
identifies handwritten digits in images and documents first observed by means of D. H. Hubel and his colleagues.
using Convolutional Neural Networks built with
PyTorch. In order to solve a variety of practical No longer like traditional synthetic neural networks
problems, this technology is crucial in applications like (ANNs), CNNs display off a comparable architectural
check processing, postal sorting, and number plate framework but feature with wonderful traits. Each layer in a
recognition. The abstract compares different machine CNN includes neurons that aren't absolutely interconnected;
learning and deep learning algorithms, such as Support rather, they're related to nearby receptive fields. Education
Vector Machine, Multilayer Perceptron, and of the network is facilitated via the software of a fee feature,
Convolutional Neural Network, based on their which always refines community normal performance via
performance, accuracy, and training times. The results evaluating the output with the favored cease end result. This
are presented visually for easy comprehension through iterative way is enabled through the usage of gradient
Matplotlib-generated plots and charts, providing descent and backpropagation algorithms, main to the
insightful information into the state of handwritten digit persevering with adjustment of shared weights and biases
recognition and opening the door for improvements in within the receptive fields.
this crucial area of artistic endeavor.
The primary reason of this text is to investigate how
Keywords:- Deep Learning, Convolutional Neural hidden layers in CNNs have an effect on handwritten digit
Network(CNN), Support Vector Machine(SVM),MINIST reputation specifically. On the changed national Institute of
Dataset. requirements and era (MNIST) dataset, an expansion of
convolutional neural community algorithms have been
I. INTRODUCTION completed as a manner to accomplish this using
TensorFlow, a neural network library built on Python. The
Convolutional Neural Networks (CNNs) are getting evaluation of the outputs produced via numerous combos of
increasingly popular as a flexible tool for visual information hidden layers within the CNN structure serves due to the
evaluation because of the combination of deep mastering fact the number one awareness of this have a examine.
into a spread of domain names. A wide range of programs, Trying out come to be finished using the beforehand set of
such as robotics, item detection, facial popularity, video rules, while training became carried out the use of stochastic
evaluation, photograph segmentation, and natural language gradient descent and backpropagation. This take a look at
processing, have observed use for CNNs. Inside the areas of combines deep mastering's expanding capabilities with
speech popularity, handwritten digit popularity, regression CNNs' profound have an effect on on duties like
evaluation, unsolicited mail filtering, subject matter handwritten digit popularity and different visible data
categorization, and photograph type, they showcase analysis.
tremendous human-stage accuracy. Deep Convolutional
Neural Networks (CNNs) development is largely
responsible for this accomplishment.

IJISRT23DEC250 www.ijisrt.com 573


Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Applications of Handwritten Digit Recognition: España-Boquera and their performed crew delivered a
paradigm-moving technique with the creation of a hybrid
 Optical Character Recognition (OCR): OCR technology Hidden Markov model (HMM) tailored for recognizing
is widely used to convert handwritten text and numbers unconstrained offline handwritten texts. Their progressive
into machine-readable and editable text, making it technique elegantly amalgamated Markov chains and
essential for digitizing documents, historical records, and Multilayer Perceptrons, heralding a structural improve
handwritten notes. within the discipline of optical character recognition.
 Postal Services: Handwritten digit recognition is crucial
in the postal industry for automatically reading and III. EXISTING METHODOLOGY
sorting mail, particularly when dealing with handwritten
addresses and postal codes. Presently used system: A manual Vector device (SVM)
 Banking and Finance: In the financial sector, is employed in the handwritten digit popularity approach
handwritten digit recognition is employed for processing presently in use. SVM is a hard and fast of guidelines for
handwritten checks, recognizing the amounts, and controlling gadgets this is commonly used for class
extracting account numbers, contributing to efficient and obligations. It's miles used on this context to find out
accurate financial transactions. handwritten digits. SVM operates through identifying the
 Number Plate Recognition: This application is handiest hyperplane that efficiently divides various digit
extensively used in traffic management and surveillance instructions in the feature area. This division allows the
for identifying and tracking vehicles by recognizing class of handwritten digits with accuracy. This records may
handwritten license plate numbers, which is crucial for be protected in your paper e-book in case you need to pay
law enforcement and security purposes. interest at the vital approach hired on your studies for
handwritten digit reputation.

IV. PROPOSED SYSTEM

Proposed system unquestionably we use a


convolutional neural network cnn as the main component of
our technology in our proposed system for reading
handwritten digits cnns are a particular kind of deep learning
model that performs exceptionally well at image recognition
tasks these networks are built with multiple layers that are
intended to autonomously acquire and extract important
features from input images making them remarkably
effective at correctly identifying handwritten digits in-depth
analyses of this systems architecture and its outstanding
capability to recognize handwritten digits will be covered in
Fig 1 Number Plate Recognition your paper publication.

II. LITERATURE SURVEY  Advantages

Early strides in man or woman recognition research  Powerful function Extraction: CNNs are extremely good
marked full-size milestones, with Grimsdale's seminal at routinely figuring out pertinent features from enter
paintings in 1959 serving as a foundational catalyst. The snap shots.
early 1960s witnessed the emergence of the analysis-via-  2) High Accuracy: They have got a demonstrated song
synthesis approach, pioneered with the aid of Eden in 1968, document of completing photo reputation obligations
which played a pivotal role in formally organising that each with a high degree of accuracy.
one handwritten characters are composed of a finite set of  3) Spatial Hierarchies :CNNs are exquisite for figuring
schematic functions. This perception would subsequently out complicated patterns in handwritten digits due to the
function a cornerstone in the improvement of syntactic fact they can seize spatial hierarchies in statistics.
person recognition methodologies.  4) They reduce the requirement for guide function
engineering, enabling extra computerized and powerful
Collaboratively, Gaurav, Bhatia, and their buddies reputation.
undertook a comprehensive exploration of pre-processing  5) Scalability CNNs are adaptable for exceptional
strategies for man or woman popularity. Their research packages, which makes them beneficial for digit
spanned various picture kinds, encompassing handwritten reputation.
paperwork, files with difficult colours, and complicated
backgrounds. The have a look at encompassed important V. DESCRIPTION OF DATASET
factors consisting of skew detection, image enhancement,
noise reduction, normalization, and segmentation. The changed country wide Institute of requirements
Importantly, it underscored the importance of integrating a and generation dataset is understood via the abbreviation
couple of strategies to optimize consequences. MNIST dataset.

IJISRT23DEC250 www.ijisrt.com 574


Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
It's far a collection of 60,000 tiny rectangular grayscale  NumPy –
images, each measuring 28 through 28, of handwritten NumPy stands for Numerical Python. We use the
single digits among zero and 9. NumPy library to work with arrays. It can also be used to
work in the domain of linear algebra, matrices and Fourier
The challenge is to region a handwritten digit transform. NumPy was created by Travis Oliphant in 2005.
photograph into certainly one of ten training that correspond It’s an open source project and can be used freely.
to integer values from zero to 9, inclusively.
 PIL –
It's far a dataset that is frequently used, very well PIL is a Python Imaging Library also known as Pillow
understood, and, for the most component, "solved." The that provides extensive file formatting for an image which is
exceptional fashions are deep gaining knowledge of a handwritten vector taken on canvas in GUI ,and also it
convolutional neural networks, which on the preserve out gives an efficient internal representation, and fairly powerful
take a look at dataset have an errors charge among 0.Four% image processing capabilities Pillow is used to represent
and zero.2% and a classification accuracy of over 99% images in Python and provides a number of Predefined
methods.
Handwritten person popularity is a sizable location of
research that previously contained specific techniques of  Scikit-learn -
implementation that SciKit-learnprovides numerous built-in machine
learning algorithms and models, called estimators and also
datasets which we can work with . Each estimator can be
fitted to some data using its fit method.

 Loading Datasets:
For the training and testing data, we will be using a
Dataset which can be used from keras module named as
datasets. This specific dataset is the MNIST data that
contains around sixty thousand images for the training data
and another ten thousand images for testing data confined to
a dimension of 28X28.

 Creating a Model:
We are creating a model using a convolutional neural
Fig 2 MINIST Dataset
network to recognize handwritten digits. We are using Three
dense layers namely input hidden and output layers.
VI. METHODOLOGY
 Training the Model:
 Importing the Libraries:
We are Training the model using 3 dense layers, one
Importing Libraries is an important aspect in working
using input layer, one using hidden layer and the last one
with python modules having specific functionality for every
using the output layer. The 3 dense layers take 128, 128, 10
library thus it can make developers job more efficient. It's a
parameters each. We flatten the pixels of the image in the
set of predefined codes, that can be called while we are
input layer. Then we apply the activation functions to the
making use of them without having to do it yourself.
values in the hidden layer. The output layer gives the result
Different libraries have different restrictions on fair use, but
as a prediction of a digit.
this is a code that was designed to be used by others, instead
of just standing alone.
 Testing the Model:
Once the model is trained, we can use the accuracy
The libraries used in this code are –
function and f1 score and some metrics to estimate the
performance of the model. We should have accuracy as high
 PyTorch-
as possible and loss as less as possible to get the desired
PyTorch is a free and open-source machine learning
output (with accuracy being close to 1 and loss being close
library for Python. It is used for applications such as natural
to 0).
language processing and computer vision. PyTorch provides
a robust library of modules and makes it simple to define
 Getting the Output:
new custom modules, allowing for easy construction of
We take a input from the GUI provided which is a
elaborate, multi-layer neural networks.
canvas so that user can draw with mouse stroke and that is
used by our model. And before giving it as input to model
 Matplotlib – the image is flatten into 1d array of 784 columns in a binary
An object-oriented plotting library.A procedural format and thus used for classification.
interface is provided by the companion pyplot module,
which may be imported directly.Matplotlib was initially
written by John D. Hunter (1968-2012) and is now
developed and maintained by a host of others.

IJISRT23DEC250 www.ijisrt.com 575


Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Steps for Digit Recognition  Segmentation:

 Data Collection and Preparation:  If you're working with photos containing multiple digits,
you could want to section the picture to isolate character
 gather a dataset of photos containing digits. digits. You may use techniques like contour detection,
Commonplace datasets consist of MNIST, USPS, and connected factor evaluation, or sliding home windows.
SVHN.
 split the dataset into education and testing units.  Feature Extraction:

 Preprocessing:  Feature extraction entails reworking the photograph into


a hard and fast of relevant features that can be used for
 picture resizing: Resize all images to a consistent, type. Commonplace techniques encompass:
workable length (e.G., 28x28 pixels for MNIST).  Histogram of orientated Gradients (HOG): Describes the
 Grayscale conversion: Convert color snap shots to distribution of local gradients within the image.
grayscale to simplify the problem.  local Binary styles (LBP): Captures texture styles inside
 Noise reduction: observe filters or strategies to lessen the image.
noise in the pictures.  Scale-Invariant feature rework (SIFT): Detects keypoints
 evaluation enhancement: alter photo evaluation to make and their descriptors.
the digits more distinguishable.  Convolutional Neural Networks (CNNs): may be used to
 Normalization: Scale pixel values to a common range analyze features at once from the image information.
(e.G., [0, 1] or [-1, 1]).
 Classification:

 Observe a class algorithm or version to predict the digit


primarily based on the functions extracted in the
preceding step. Common type techniques include:

 k-Nearest Neighbor (k-NN)


 Support Vector machine (SVM)
 Random Forests
 Neural Networks

 System Design

Fig 3 Steps for Digit Recognition

Fig 4 System Design

IJISRT23DEC250 www.ijisrt.com 576


Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
The enter layer consists of 28 with the aid of 28 pixel layer is another hidden layer that connects every neuron
pix that means that the network is ready with 784 neurons to from the preceding layer to the following to mitigate
procedure this input statistics these enter pixels are overfitting dropout regularization is used at completely
grayscale with a value of 0 representing white and 1 linked layer 1 this technique randomly deactivates a few
representing black in this CNN version there are 5 hidden neurons at some point of education thereby enhancing the
layers the primary hidden layer known as convolution layer communitys performance and robustness the output layer
1 is accountable for extracting features from the input facts consists of ten neurons every liable for figuring out the digit
it achieves this thru convolution operations on small from zero to nine an activation characteristic together with
localized regions the usage of filters carried out to the softmax is employed inside the output layer to enhance
preceding layer this sediment incorporates more than one model performance classifying the output digit by way of
feature maps with learnable kernels and rectified linear assigning the highest activation price
gadgets the kernel length determines the clear outs locality
relu is employed as an activation feature on the cease of VII. RESULTS AND DISCUSSION
every convolution layer and in the absolutely related layer to
beautify version performance the subsequent hidden layer Our studies in hand-written digit popularity has yielded
referred to as pooling layer 1 reduces the statistics output especially promising outcomes. The model continually
from the convolution layer thereby decreasing the wide achieves high accuracy in identifying and predicting
variety of parameters and computational complexity in the handwritten digits, tested by sturdy performance metrics like
version distinct sorts of pooling can be used including max precision, recollect, and F1-score. Moreover, it excels in
pooling min pooling average pooling and l2 pooling in this presenting clear and informative graphical representations,
case max pooling is utilized to down sample each function making it user-pleasant. This era holds super potential for
map convolution layer 2 and pooling layer 2 have analogous practical applications in check processing, postal offerings,
features to convolution layer 1 and pooling layer 1 differing and schooling, promising efficiency upgrades and errors
in particular in their function maps and kernel length after reduction. This paves the manner for in addition studies in
the pooling layer a flatten layer is employed to transform the actual-time recognition and superior features, culminating in
second function map matrix right into a 1d feature vector a significant development in hand-written digit reputation,
facilitating processing by means of fully connected layers imparting valuable contributions to a couple of fields of
the absolutely related layer also referred to as the dense utility.

Fig 5 Outputs Obtained

IJISRT23DEC250 www.ijisrt.com 577


Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 6 Outputs Obtained

Fig 7 Output Obtained

IJISRT23DEC250 www.ijisrt.com 578


Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
VIII. CONCLUSION REFERENCES

In summary, the research presented in this article [1]. Nasser Omer Sahel Ba-Karait and S.M. Shamsuddin,
explores the interesting field of coding using neural “Handwritten digits popularity the use of Particle
networks (CNN) and the MNIST dataset. Our study aims to Swarm Optimization”, 2nd Asia global conference on
show the effect of number variation and configuration of the Modelling & Stimulation.
hidden layer in the CNN architecture on number recognition [2]. Bi Ran and Leong Tze Yuny, “Hand written digit
accuracy and performance. recognition and its improvement”, to be had at cs2306-
system getting to
Throughout the experiment, we found that different know.Googlecode.Com/svn/trunk/CS2306S.../a.Pdf
secret methods differ in terms of accuracy and failure. More [3]. Ishani patel, Virag Jagtap and Ompriya Kale, “A
importantly, we achieved the highest accuracy of 99.21% in Survey on characteristic Extraction strategies for
Case 2, which demonstrates the ability to achieve higher Handwritten Digits popularity”,global journal of
accuracy in machine code recognition. computer applications quantity 107 – No 12, December
2014.
Instead, we note that the release process plays an [4]. Gaurav Kumar, Pradeep Kumar Bhatia, “Neural
important role in reducing the loss, with the lowest total test community based Technique for popularity of
around 0.026303 in Case 2. The reduction is expected to textual content pictures”, global journal of Pc
result in better resolution and the ability to handle noise, applications quantity 62– No.14, January 2013
further strengthening the performance of the CNN. [5]. Malothu Nagu, N. Vijay Shankar and k.Annapurna, “a
unique Approach for Handwritten Digit popularity
Our findings highlight the importance of optimizing with Neural Networks”,Global journal of laptop
the hidden layer and practical use of the output layer for science and records Technologies, Vol. 2 (4), 2011.
coding in CNNs. As technology continues to advance, this [6]. Jagyanseni Panda, Manaswinee M. Panda,
research is leading to a broader goal of accurate and Aryapriyanka Samal and Niva Das, “Odia Handwritten
efficient digital recognition systems, with applications Digit reputation using single Layer Perceptron”,
including behavioral recognition: fixing eyes on machine worldwide journal of Electronics and Communique
vision. Engineering and technology(IJECET) volume five,
Trouble 4, April (2014)
Looking forward, our research paves the way for future [7]. Viragkumar N. Jagtap and Shailendra ok.
research on the impact of different latent methods and batch Mishra,“rapid green Artificial Neural network for
size on overall classification accuracy. This will give us a Handwritten Digit Popularity”,international magazine
better understanding of the interaction between network of computer technological know-how and Statistics
architecture and typing, providing a better understanding for technologies, Vol. Five , 2014.
the development of powerful and efficient systems in this [8]. B. Yegnanarayana, “artificial Neural community”, jap
area. financial system Version

ACKNOWLEDGMENTS

The genuine essence of achieving a goal is attained


through a combination of excellence and unwavering
discipline. I couldn't have successfully completed my task
without the invaluable cooperation, encouragement, and
support from various individuals. I would like to seize this
opportunity to convey my deep gratitude to my mentor,
Senior Assistant Professor A. Leelavathi, for her esteemed
guidance and unwavering encouragement, particularly
during challenging times. Her insightful suggestions
broadened my perspective and steered me towards success
in this endeavor. I am profoundly appreciative of her
guidance and have gained a wealth of knowledge under her
leadership.

I would also like to extend my heartfelt thanks to Dr.


Loshma Gunisetti, the Head of the Artificial Intelligence and
Machine Learning Department at Sri Vasavi Engineering
College, Pedatadepalli, for her invaluable support that
played a pivotal role in making me a consistent performer.

IJISRT23DEC250 www.ijisrt.com 579

You might also like