Implementasi Augmentasi Data untuk Meningkatkan Akurasi Klasifikasi Citra Aksara Jawa Berbasis Deep Learning

Abstract — Javanese script, as part of Indonesia's cultural heritage, faces the threat of extinction due to declining interest among the younger generation. This study developed a deep learning-based Javanese script recognition system using the ResNet-50 architecture with transfer learning to overcome dataset limitations. The Indonesian Local Script Characters dataset, consisting of 20 basic script classes (11,657 images), was divided into training (84.3%), validation (14.9%), and testing (0.8%). Extreme data augmentation was applied, including color jitter, texture transforms, Gaussian noise, rotation (±15°), and horizontal flip to improve model generalization. A head-to-head experiment was conducted by comparing the model without augmentation (Model A) and the model with extreme augmentation (Model B). Model A achieved a validation accuracy of 99.88% but showed extreme overfitting with a test accuracy of only 75.00% (a gap of 24.88%). Model B with augmentation achieved a validation accuracy of 99.19% and a test accuracy of 92.00%, demonstrating superior generalization with a minimal gap of 7.19%. Model B also showed an F1-score of 91.63% with 13 out of 20 classes achieving a perfect score (100%). The web application implementation using Streamlit enables real-time predictions with an inference time of less than 1 second per image. The research results prove that extreme augmentation not only improves accuracy but also functions as implicit regularization that prevents overfitting. This research surpasses previous studies.