11 III March 2023

https://doi.org/10.22214/ijraset.2023.49884
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue III Mar 2023- Available at www.ijraset.com

Handwritten Digit Prediction Using CNN

Sai Sharan Alugunuri1, Kaithapuram. Vishal Reddy2, Chevvula. Shiva Kumar3, T. Bhavani Prasad4
SR University

Abstract: For many years, numerous methods have been used in extensive research on handwriting recognition. The capacity to
create an effective algorithm that can recognise handwritten digits given by users via scanner, tablet, and other digital devices is
at the core of the issue. The automatic processing of bank checks, postal addresses, and other sorts of data already makes
substantial use of handwritten digit recognition. Computational intelligence methods like artificial neural networks used by
several current systems. CNN and the MNIST data set will be used to complete this. Handwriting Recognition, Deep Learning,
CNN, and Computational Intelligence are key terms.
I. INTRODUCTION
Handwriting Recognition is a machine's ability to recognise and predict human handwritten digits. It is a difficult task for machines
because handwritten digits are not perfect and can be made in a variety of flavours. As a result, this paper presents a solution for
accurately recognising and predicting handwritten digits. The number recognition framework is simply a task that the machine must
complete in order to prepare and interpret numbers. Handwritten digit recognition interprets manually written numbers from a
variety of sources such as messages, bank checks, documents, photos, and so on, as well as in a variety of situations for web-based
handwriting recognition on PC Tablets and in vehicles. There are numerous techniques that can be used to gain recognition.
Convolutional Neural Network (CNN), Semi Incremental Method,Line and Word Segmentation, and other techniques are used.
Convolutional Neural Networks are one of the most effective and well-known methods of handwriting recognition (CNN). It's a
component of deep learning. Artificial neurons make up Convolutional Neural Networks (CNN). The number recognition model
recognises numbers from various sources using large datasets.

II. LITERATURE REVIEW

Handwriting recognition is a vast area of study. Many researchers are working on this topic to achieve the best possible results.
Various techniques were used by different researchers for image recognition and digit recognition. Now the paper proposes a
framework for annotating large scale handwritten word images with ease and speed by Kartik Dutta, Praveen Krishnan, Minesh
Mathew, and C.V. Jawahar [1]. This paper uses state-of-the-art deep neural architectures to benchmark major Indian scripts such as
Devanagri and Bangla for the tasks of word spotting and handwritten recognition. It employs Word Spotting via a CNN-RNN
hybrid network. Nikita Singh [2], This paper proposes a method for recognising handwritten Devanagari characters. The proposed
method is based on individual character classification using ANN (Artificial Neural Network). The proposed method could be
helpful for blind people to read handwritten contents. Roshan Fernandes and Anisha P Rodrigues [3], propose two techniques for
recognising handwritten Kannada script with high accuracy. Techniques : 1. Teseeracttool, 2. Convolution neural Network (CNN).
We achieved 86% accuracy with the tesseract tool and 87% accuracy with the convolution Neural Network. One of the authors
proposed using decision tree learning to classify various writing styles of identical digits. To implement the classification, several
direction features were used. That is, if the stroke direction of some digits is similar, Decision tree learning can correctly classify
them. However, as more features are added, it becomes more difficult to manage sets of possibilities.

III. FEASIBILITY ANALYSIS

There are numerous aspects of feasibility. Let it be discussed individually.

A. Technical Feasibility
The software used in this project is completely open source, and anyone can connect to it whenever they want. Aside from that, the
entire software running environment Google Colab is fully open source and easily accessible in the presence of an internet
connection. The user can also be a non-programmer and set the digit in the webcam screen and see the output by clicking the Run
button.

B. Seasonal Feasibility
This project is feasible in terms of time, which means that it was started on a specific date and completed in the allotted time. It was
an efficient effort that resulted in the project being completed on time.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2040
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue III Mar 2023- Available at www.ijraset.com

C. Economic Feasibility
Because all open source software was used, this project is economically free; therefore, no money was charged or given. Only the
study materials on the developer or designer side are not available for free. The software or programme is completely free.
IV. METHODOLOGY
A. Dataset
The first step in recognising handwritten forms was to collect a large amount of data for training. The hand printed forms and digits
database NMIST Dataset is used. The dataset we used contains 60,000 of handwritten images in size 28*28, with each number in
the image centre fitted to a 20*20 pixel box. Every image will be in Grey Scale.

B. Preprocessing
It is a procedure for converting raw data into a usable and efficient format. Pre-processing is a process that consists of various
operations that are performed on input images. Images are reshaped during this process. The rearrangement of the data's form
without changing the data's contents. Various types of arrangements are made in this process based on the parameters required to
continue the process.

C. Dataset Training
Model accuracy is measured using training and testing. The data is divided into two sets during this process: training and testing.
Data is used for training 80% of the time and testing 20% of the time. CNN has been used to train this system. Training accuracy is
0.9885, training loss is 0.04737, and testing accuracy is 0.98607, with a testing loss of 0.06844.

D. CNN Modelling

Convolutional Pooling
Input Fully-Connected
Layer Layer
Layer
Feature Extraction

Output

Classification
Deep Learning includes Convolutional Neural Networks (CNN). To this day, CNN is a very efficient and effective way to achieve
handwritten recognition. Convolutional Neural Networks are used to extract image features using multiple layers of filters.
V. WORKING

Start

Input

Resize the image

CNN Model

Prediction

Output

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2041
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue III Mar 2023- Available at www.ijraset.com

In practise, input is provided via a graphical user interface (GUI). We've now created a new file for the GUI in which we've built an
interactive window with buttons for drawing digits on the canvas and recognising the digits. The Tkinter library was used to create
the GUI. Python. It enables you to quickly and easily create GUI applications. After receiving the input, it goes through the
prediction process. The given input is advanced to resize in a specific format in order to obtain the actual prediction. The resized
image is then moved down for the Prediction model, which extracts the given input features. It occurs. Following that, based on the
estimated significance of the set of input variables, the modelling generates a prediction representing the probability of the target
variable. The graphical user interface application displays the expected results. The output will contain the expected digits with
position.

VI. UML DIAGRAM

VII. ANALYSIS

Recognition Accuracy(%) and Total Time Elapsed

Model Layer k s d p i/p o/p r 8-16-32 6-12-24 12-24-32 8-16-24 8-24-32 12-24-
28
Layer 1 5 2 2 2 28 14 5
93.76% 84.76% 98.76% 94.08% 96.12% 98.08%
Case 1 Layer 2 5 2 1 2 14 7 9 (20 s) (42 s) (45 s) (46 s) (42 s) (44 s)

Layer 3 5 2 1 2 7 4 25

Layer 1 5 2 2 2 28 14 5
96.04% 88.91% 99% 93.08% 96.12% 98.48%
Case 2 Layer 2 5 2 1 2 14 7 9 (37 s) (27 s) (37 s) (37 s) (37 s) (17 s)

Layer 3 5 2 1 2 7 4 17

Layer 1 5 2 2 2 28 14 5
98.96% 86.88% 99.7% 98.72% 99.28% 99.60%
Case 3 Layer 2 5 2 1 2 14 7 13 (27 s) (27 s) (29 s) (39 s) (31 s) (53 s)

Layer 3 5 2 1 2 7 4 29

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2042
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue III Mar 2023- Available at www.ijraset.com

VIII. CONCLUSION
To recognise the various handwritten digits, the project presented a Convolutional Neural Network approach. This project involves
the classification of digits. The project is completed using a traditional neural network. The accuracy we obtained was greater than
90%. The project provides the highest accuracy for text with the least amount of noise. The accuracy is entirely dependent on the
dataset; as the data set grows, so will the accuracy. If we try to avoid cursive writing, we will get the best results. We intend to
expand this study in the future so that different embedding models can be considered on a wide range of datasets. No one will write
on paper or with a pen in the future because everything is based on technology. In that scenario, they wrote on touch pads so that the
inbuilt software could automatically detect the text they were writing and convert it to digital text, making searching and
understanding much easier.

REFERENCES
[1] Kartik Dutta, PraveenKrishnan, Minesh Mathewand C.V.Jawahar, “Towards Spotting and Recognition of Handwritten Words in Indic Scripts,” 2018.
[2] Nikita Singh,“ An Efficient Approach for Handwritten Devanagari Character Recognition based on Artificial Neural Network,” 2018.
[3] Roshan Fernandes, Anisha P Rodrigues, “Kannada Handwritten Script Recognition using Machine Learning Techniques” IEEE, 2019
[4] https://www.ijrte.org/wp-content/uploads/papers/v9i5/E5287019521.pdf
[5] https://www.diva-portal.org/smash/get/diva2:1293077/FULLTEXT02.pdf
[6] https://doi.org/10.22214/ijraset.2022.42062
[7] https://en.m.wikipedia.org/wiki/Handwriting_recognition
[8] https://www.academia.edu/41447877/Handwritten_Digit_Recognition_using_Artificial_Neural_Network
[9] https://globaljournals.org/GJCST_Volume18/3-Handwritten-Digit-Recognition.pdf
[10] https://www.irjet.net/archives/V9/i6/IRJET-V9I6208.pdf
[11] https://www.researchgate.net/publication/302871740_Handwritten_Digit_Recognition_Using_Convolutional_Neural_Networks
[12] https://www.researchgate.net/publication/326408524_Handwritten_Digit_Recognition_Using_Machine_Learning_Algorithms
[13] https://thedatafrog.com/en/articles/handwritten-digit-recognition-scikit-learn/
[14] https://www.ijraset.com/research-paper/handwritten-digit-recognition
[15] https://doi.org/10.22214/ijraset.2022.42062
[16] Handwritten Digit Recognition with Pattern Transformations and Neural Network Averaging | SpringerLink
[17] Handwritten Digit Recognition Using Logistic Regression, SVM, KNN and CNN Algorithms | JOURNAL OF OPERATING SYSTEMS DEVELOPMENT &
TRENDS (stmjournals.com)
[18] 2106.12614.pdf (arxiv.org)
[19] (PDF) Handwritten Digit Recognition Using Machine Learning Algorithms (researchgate.net)
[20] Microsoft Word - 63. anmhwr-paper3-final-IJCSIT
[21] Handwritten Digit Recognition Using Convolutional Neural Networks in Python with Keras - MachineLearningMastery.com
[22] BDNet: Bengali Handwritten Numeral Digit Recognition based on Densely connected Convolutional Neural Networks - ScienceDirect
[23] Handwritten Digit Recognition With Machine Learning Algorithms (dergipark.org.tr)

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2043