HANDWRITTEN DIGITS RECOGNITION
Loading...
Date
2022-06-10
Journal Title
Journal ISSN
Volume Title
Publisher
UNIVERSITY of M'SILA
Abstract
The main purpose of this thesis is to build a method for automatic recognition of handwritten
digits. To accomplish the recognition task, first, the data set is presented as input, this is
followed by a preprocessing phase, where the image undergoes various operations such as noise
reduction, normalization, smoothing, and skeletonization. The result of this stage can be given
as input to the stage of automatically extracting feature maps from images using 3 layers of
Convolutional Neural Networks (CNN), and then fed to 4 layers of FC ANN, where these
images are classified into ten classes (0 to 9). The accuracy of the final model is important
because more accurate models make better decisions. Our results showed a classification
accuracy of 99.78% and a loss value of 1.82% using the MNIST data set.
Description
Keywords
Convolutional Neural Networks; MNIST; features extraction; preprocessing; Artificial Neural Networks.