dc.contributor.advisor |
Wijesiriwardana C P |
|
dc.contributor.author |
Karunarathne ML |
|
dc.date.accessioned |
2022 |
|
dc.date.available |
2022 |
|
dc.date.issued |
2022 |
|
dc.identifier.citation |
Karunarathne, M.L. (2022). Handwritten sinhala character recognition using deep learning [Master's theses, University of Moratuwa]. Institutional Repository University of Moratuwa. http://dl.lib.uom.lk/handle/123/20311 |
|
dc.identifier.uri |
http://dl.lib.uom.lk/handle/123/20311 |
|
dc.description.abstract |
Sinhala language is the national language in Sri Lanka. Sinhala alphabet includes 60
characters and is slightly complex compared to other languages like English. Around
25-30 researches have been done since 1990 regarding Sinhala handwritten character
recognition. Handwritten Sinhala character recognition remains mostly unsolved in
pattern recognition, due to many perplexing characters and excessive curves in Sinhala
handwriting. The existing recognizers are also unable to provide acceptable performance
for practical applications.
This research aims to enhance the performance of handwritten Sinhala character
recognition by using a new approach focused on deep neural networks, which have
recently given excellent performance in many applications. This research implements
Convolutional Neural Networks (CNNs) and Gabor initialized Convolutional Neural
Network (GCNN). In addition to that, it investigates the performance of the proposed
network architectures when introducing the dropout. To apply Gabor initialized CNN,
the effect of the parameters of the Gabor filter over the Sinhala character image dataset
is also examined. Considering the effect of the parameter on the GCNN architecture,
parameter values for the proposed GCNN architecture are determined. The training
accuracy of the first CNN method is 96.33 % and the testing accuracy is 90.14%.
According to the literature, this is the highest accuracy obtained for 60 Sinhala characters
compared with primitive methods. This accuracy is obtained with the 0.5 dropout effect.
The Gabor initialized CNN architecture provides 95.15% training and 80% testing
accuracy. Even though the training accuracy is approximately 1% less than the training
accuracy of the first CNN architecture, it converges to the results rapidly. So, it saves
time and computational cost.
Considering the results of implemented CNN architectures and Gabor initialized CNN
architecture, the best-performing architecture is selected for the Sinhala handwritten
character recognition process |
en_US |
dc.language.iso |
en |
en_US |
dc.subject |
IMAGE REPRESENTATION |
en_US |
dc.subject |
HANDWRITTEN SINHALA CHARACTERS |
en_US |
dc.subject |
DEEP LEARNING |
en_US |
dc.subject |
CONVOLUTIONAL NEURAL NETWORKS |
en_US |
dc.subject |
INFORMATION TECHNOLOGY- Dissertation |
en_US |
dc.subject |
COMPUTER SCIENCE - Dissertation |
en_US |
dc.title |
Handwritten sinhala character recognition using deep learning |
en_US |
dc.type |
Thesis-Abstract |
en_US |
dc.identifier.faculty |
IT |
en_US |
dc.identifier.degree |
Msc. in Information Technology |
en_US |
dc.identifier.department |
Department of Information Technology |
en_US |
dc.date.accept |
2022 |
|
dc.identifier.accno |
TH4821 |
en_US |