Javanese Script Detection Using ResNet18 with Hyperparameter Optimization of Chimpanzee Leader Election Optimization Algorithm

The Javanese script (Aksara Jawa), a vital component of Indonesia's cultural heritage, faces significant challenges in digital preservation due to declining literacy and the labor-intensive nature of manual transcription of manuscripts. This research addresses the need for an automated recognition system by proposing a novel approach that integrates a ResNet18-based transfer learning model with the Chimpanzee Leader Election Optimization (CLEO) algorithm for hyperparameter tuning. The objective is to develop a robust and accurate Javanese script detection model that overcomes the sensitivity of deep learning architectures to hyperparameter configurations. The methodology employs 5-fold stratified cross-validation, where CLEO optimizes five key hyperparameters: learning rate, weight decay, dropout, batch size, and backbone freezing strategy, using a population of five chimps over four iterations per fold. The optimal configurations are then used to train the final models with aggressive data augmentation and early stopping. Experimental results demonstrate outstanding performance, with an average test accuracy of 94.05 % and a low standard deviation of 0.0048 across all folds. Per-class analysis reveals that most characters achieve precision, recall, and F1 Scores above 0.95, although challenges remain in distinguishing visually similar characters, such as ya and ga. The proposed CLEO-optimized ResNet18 framework provides an effective solution for Javanese script recognition, contributing significantly to digital preservation efforts and establishing a methodological foundation for applying metaheuristic optimization to other low-resource script recognition tasks.