Cosine Similarity as a Distance Metric for Javanese Script Image Recognition Classification

Abstract

Javanese character (Hanacaraka) recognition presents significant challenges due to the intricate patterns and variations in character features. Addressing these issues is crucial for digitizing cultural heritage and supporting educational applications. This study aims to evaluate the effectiveness of cosine similarity as a distance metric for classifying Javanese characters, comparing its performance against traditional Euclidean and Manhattan distance metrics. The research used a feature-extraction technique based on the histogram of oriented gradients and evaluated cosine similarity across different classification models. Model performance was assessed using precision, recall, F1-score, and accuracy metrics. The results showed that cosine similarity, when combined with a support vector machine, achieved an accuracy of 99.84%, significantly outperforming other distance metrics. When applied to another classification model, cosine similarity improved accuracy to 90%, demonstrating its robustness in handling complex patterns. Parameter optimization was performed using a grid-based search, and model reliability was assessed through cross-validation. Compared with previous studies that primarily relied on deep learning, this research offers an alternative method that balances efficiency and accuracy while maintaining high interpretability. The findings establish a new benchmark for Javanese character recognition and highlight the potential of cosine similarity in broader applications. Future research can expand this study by incorporating more diverse feature extraction techniques, larger datasets, and hybrid approaches to further enhance recognition performance.