Reintroducing the baybayin script through e-learning and deep neutral networks / by James Arnold E. Nogra

Thesis (DIT) -- Cebu Institute of Technology - University, College of Computer Studies, March 2020

Includes bibliographical references.

The Philippine Congress has signed the House Bill 1022 which declares that the Baybayin script will be considered as the national writing system of the country. The Department of Education, National Commission for Culture and Arts, and some other organization have vowed to reintroduce this writing system back. This handwriting system is not taught in the schools anymore so an easier training solution is needed. To be able to make this transition easier and faster, a neural network is proposed to convert hand-drawn characters to its corresponding English alphabet syllabication.
The type of neural network used in this study is the Long Short-Term Memory (LSTM) network and Convolutional Neural Network (CNN). To gather the necessary number of training images for the neural network, 25 people were employed to draw the characters. All of them used an Android mobile app that will let them draw their own version of a character based on three characters presented to them. All of the contributed handwritten characters are manually viewed to ensure the quality of the sample data. In order to hasten up the learning process, a convolutional neural network is designed to check the classification of hand-drawn characters. For the CNN, nine models were designed to check which is the best for this type of character recognition. By introducing this script to the people through an Electronic Learning (e-Learning) mobile application, the interaction needed between students and teachers will be reduced. The proposed Learn Baybayin e-Learning app is divided into five stage groups for introduction, three writing stages, and reading. In the reading and writing stages, the app will upload the hand-drawn character by the user and then a convolutional neural network online will return the rating of that drawn character.
After testing five different LSTM networks, the model with 512 and 256 units in the hidden layers and 128 units in the dense layer is the best model for classifying hand-drawn Baybayin characters. All of the models are tested up to 50 epochs. Using 8500 images for training and 1200 images for validation, this model has achieved a 95.6% training accuracy and 92.9% validation accuracy. For the CNN, it has found out that the best neural network for this type of classification is composed of three convolutional layers with 32 channels, 64 channels, and 128 channels respectively using 3x3 filters. The final model has also three max-pooling layers right after each convolution layer with 2x2 size and two fully connected layers at the end. The number of output of this neural network model is 63 which is the same as the total number of Baybayin characters. The model yields a 94% accuracy rate using the validation data. The other 8 CNN models also did well with accuracy rates ranging from 57% to 92%. The eLearning app also shows promising results. After five months in the Android PlayStore, a total of 524 people used the app. For all the five stage groups, the users got an average score of 54.75%. This means that these users know how to read and write most of the Baybayin characters. A usability measurement tool was emailed to the users of the app in which 102 responded. The usability review has 10 measurement items in which users can respond from1 (strongly disagree) to 5 (strongly agree). The average for all the measurement items gathered from the user is 4.3 which indicates that they agreed in almost every aspect of the e-Learning app.