Evaluating Ancient Sundanese Glyph Recognition Using Convolutional Neural Network
Erick Paulus, S Hadi, Mira Suryani, I Putu Gede Eka Suryana, YD Simanjuntak · Journal of Physics Conference Series · 2019
Abstract Handwriting recognition is still a real challenge in classification tasks. Not as in modern documents, the isolated glyph images in ancient document have various random noises, non uniform background color, and smudge. Convolution neural network (CNN) is one of successful method in pattern recognition and machine learning to classify the objects. The evaluation of some CNN architectures with several different convolution layers classifying the isolated glyph image are presented in this paper. The experimental study is tested on 60 classes of glyph from the ancient Sundanese dataset that published in ICDAR 2017. Beside, the batch normalization is also investigated to measure the performance of the learning process. The results shown that the recognition rate was affected by multi convolutional layers, multi fully connected layer and batch normalization. Based on the experimental study, model 8C2F could achieve 86.15% of recognition rate.