Optimizing Javanese script recognition using fine-tuned ResNet-18 and transfer learning

Abstract

Javanese script, known as Aksara Jawa, is an ancient script used in historical and cultural texts. However, its complex character structure poses challenges for accurate recognition in modern digital applications. This study proposes an optimized classification approach for Aksara Jawa using a fine-tuned ResNet-18 model combined with the Adam optimization algorithm and transfer learning on the Hanacaraka image dataset. By leveraging the residual learning framework of ResNet-18, the model effectively captures deep spatial features of the script while reducing vanishing gradient issues. Fine-tuning is applied to enhance model adaptability, ensuring robust feature extraction specific to Javanese characters. Experimental results demonstrate that the fine-tuned ResNet-18 outperforms conventional deep learning architectures in recognizing Aksara Jawa characters, achieving 93% precision, 91% recall, 91% F1-score, and 91% accuracy. The study further explores the impact of hyperparameter tuning, data augmentation, and dropout regularization on model performance. The findings highlight the effectiveness of transfer learning in resource-limited scenarios, making it a feasible solution for optical character recognition (OCR) applications in Javanese script digitization. This research contributes to the preservation of cultural heritage through advancements in deep learning-based script recognition.