PENGUJIAN KINERJA EKSPERIMENTAL RESNET-18 PADA PENGENALAN 30 KARAKTER DASAR AKSARA SUNDA: Aksara Sunda, ResNet-18, Pembelajaran Mendalam, Jaringan Saraf Konvolusional, Klasifikasi Citra, Pengenalan Karakter Tulisan Tangan

Arief Fadhiel Januarizky (Universitas Pembangunan Jaya)
Mohammad Nasucha (Universitas Pembangunan Jaya)

RABIT: Jurnal Teknologi dan Sistem Informasi Univrab
Vol 11 No 2 (2026): Juli

Sundanese script is one of the cultural heritages that holds significant historical value for the Sundanese community. However, the use of Sundanese script in daily life has gradually declined, creating a need for preservation efforts through digital technology. One potential approach is the automatic recognition of Sundanese script using deep learning techniques. The ability of deep learning to identify visual patterns in images makes it suitable for handwritten Sundanese script classification. This study aims to evaluate the performance of the ResNet-18 architecture in recognizing 30 basic Sundanese script characters from handwritten image data. The research began with the collection of a self-created dataset obtained from 10 respondents. Each respondent was asked to write 30 basic Sundanese script characters, resulting in a total of 300 image samples. The collected images then underwent preprocessing stages, including cropping, resizing, and grayscale conversion. The processed dataset was divided into three data-splitting scenarios, namely 60:40, 70:30, and 80:20 for training and testing purposes. For each scenario, ResNet-18 was trained using both pretrained and non-pretrained approaches. After the training process, the resulting weights were saved and used during the evaluation stage. Model performance was evaluated using confusion matrices and classification metrics, including accuracy, precision, recall, and F1-score. The evaluation results from each scenario were then compared to analyze the influence of training data size on classification performance. The experimental results demonstrate that ResNet-18 is capable of classifying 30 basic Sundanese script characters with satisfactory performance. The best performance was achieved using the 80:20 data-splitting scenario with the pretrained ResNet-18, obtaining an accuracy of 98.33%, precision of 98.89%, recall of 98.33%, and F1-score of 98.22%. Furthermore, the results indicate that increasing the amount of training data contributes positively to classification performance. Based on these findings, ResNet-18 can be considered an effective approach for Sundanese script recognition and has the potential to support cultural preservation efforts through deep learning-based image processing technology.