Javanese Script Text Image Recognition Using Convolutional Neural Networks
Goldy Najma Adli Kesaulya, Ama Fariza, Tita Karlita · 2022 International Electronics Symposium (IES) · 2022
Javanese script is one of the traditional Indonesian scripts used on the island of Java, which is currently starting to decline in its usages due to the difficulty in learning the many, complex, and similar forms between characters. This can have an impact on the preservation of the declining Javanese script culture, which if left ignored will cause the loss of one of the characteristics of Indonesia. This study aims to overcome these problems by creating an application that can work well in recognizing images of Javanese script text. The method used in this study is Convolutional Neural Network (CNN) with a new approach to transfer learning using the pre-trained model ResNeXt. The dataset used for model training stage is obtained by merging Javanese script images dataset publicly available on Kaggle website with manually collected images and by applying image augmentation techniques on the collected dataset. The recognition process used in this application is based on Optical Character Recognition (OCR) with preprocessing, segmentation, feature extraction and classification, and post-processing stages. The best CNN model testing results were obtained in an experiment with a combination of using Adam optimizer, learning rate of 0.0001 without a scheduler, and freezing first 4 layers of the 10-layer CNN ResNeXt model, with a testing accuracy value of 98.19% Meanwhile the OCR processes in the mobile application are tested on 11 images captured from Javanese Pepak book achieving an average character error rate of 38.09%.