Recognition of Baybayin (Ancient Philippine Character) Handwritten Letters Using VGG16 Deep Convolutional Neural Network Model

Abstract

We proposed a system that can convert 45 handwritten baybayin Philippine character/s into their corresponding Tagalog word/s equivalent through convolutional neural network (CNN) using Keras. The implemented architecture utilizes smaller, more compact type of VGG16 network. The classification used 1500 images from each 45 baybayin characters. The pixel values resulting from the resized characters (50x50 pixels) of the segmentation stage have been utilized for training the system, thus, achieving a 99.54% accuracy. To test the developed recognition system, 90 handwritten baybayin characters were captured in real time using its 1080P Full-HD web camera. Next, the system will classify the test sample. Lastly, the corresponding Tagalog word output is shown on the screen for the user. The overall accuracy for the testing phase is 98.84%. As a result, the proposed system will find possible applications in character extraction in documents and any related translation of handwritten document to structural text form.

Convolutions vs. Sequences: Understanding performances of neural-based methods for automatic Baybayin script recognition

“…[6] Another study used 45 classes of Baybayin Character with a total of 108,000 images contained in the dataset that was collected on 90 different people to draw the Baybayin Script. 50 iterations with a batch size of 32 to perform training and testing [4]. Using DCNN (Deep Convolutional Neural Network) model with VGG16 as their architecture and a module that is image processing using OpenCV library for their character recognition.…”

“…To segment their data to training and testing, the researchers used 80% as their training data and 20% for their testing, their result of their training and testing with the accuracy of 99.54% and an overall rating of 98.84%. [4] Another set researchers used SVM (Support Vector Machine) model to train using 1000 Images of Baybayin Words. Using Words to train their model gave them 97.9% recognition accuracy.…”