MAT-AGCA: Multi Augmentation Technique on small dataset for Balinese character recognition using Convolutional Neural Network
Ni Putu Sutramiani, Nanik Suciati, Daniel Siahaan
Abstract
The lontar manuscript is an ancient Balinese cultural heritage written using Balinese characters on palm leaves. The recognition of Balinese characters in lontar is challenging because it has noise and limited data availability. To solve these problems, data augmentation is needed to increase the variety and amount of data to improve recognition performance. In this study, we collected Balinese character images from 50 lontar manuscript writers. We proposed MAT-AGCA that combines Adaptive Gaussian Thresholding and Convolutional Autoencoder for data augmentation. Based on experiments using InceptionResnetV2, DenseNet169, ResNet152V2, VGG19, and MobileNetV2, our proposed method achieved the best performance with 96.29% accuracy.