Detection of Javanese Characters Using YOLO11 Deep Learning Approach for Digital Cultural Heritage Preservation

Abstract

Javanese script represents a significant cultural heritage of the Indonesian archipelago that faces extinction threats due to Latin alphabet dominance and minimal users capable of writing with this traditional script. This research aims to develop a Javanese character detection system using You Only Look Once version 11 (YOLO11) algorithm to support cultural preservation efforts through efficient digitalization. The research methodology employs an experimental approach with deep learning, where the Javanese script dataset consisting of 20 basic characters plus background class was obtained from Kaggle and preprocessed using Roboflow with data augmentation techniques. The YOLO11 model was implemented with SGD optimizer, 640px image size, and trained for 500 epochs to achieve optimal convergence. YOLO11 architecture integrates advanced components such as C3K2 blocks, Spatial Pyramid Pooling-Fast (SPPF), and Cross-scale Pixel Spatial Attention (C2PSA) to enhance multiscale feature extraction capabilities. Model performance evaluation utilized confusion matrix with accuracy, precision, recall, and F1-score metrics. Research results demonstrate that the YOLO11 model achieved an overall accuracy of 81.00% with macro-averaged precision of 86.28%, macro-averaged recall of 87.25%, and macro-averaged F1-score of 86.41%. Model performance distribution shows 7 classes with high performance (F1-score ≥ 90%), 9 classes with medium performance (80-90%), and 4 classes with low performance (<80%). The "nga" class achieved perfect performance of 100%, while the "ha" class showed the lowest performance with an F1-score of 68.09%. This research successfully improved accuracy compared to previous methods using backpropagation neural networks (74%) and conventional backpropagation (59.5%), although challenges remain in detecting characters with similar shapes and handling background class. The main contribution is the first implementation of YOLO11 for Javanese script detection, opening opportunities for developing more efficient and accurate ancient literature digitalization systems.