Implementasi Convolutional Neural Network Untuk Pengenalan Tulisan Tangan Akasara Sunda Ngalangéna

Azizah, Rahma Nur
Avianto, Donny

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 6 No 7 (2025): December 2025

Efforts to preserve the Sundanese script as a cultural heritage face challenges in the digital era, one of which is the limited resources for pattern recognition. This research aims to develop an effective custom Convolutional Neural Network (CNN) model for the classification of handwritten Sundanese script. Facing the constraint of no available public dataset, this study utilizes a primary dataset (Swaraksara Dataset) created by the author, consisting of 6,500 handwritten images evenly distributed across 13 classes (combinations of the "Na" script with rarangkén). The methodology applied includes a comprehensive data preprocessing stage, covering grayscale conversion, resizing to 200x200 pixels, normalization, and data augmentation techniques to prevent overfitting. The custom CNN architecture was designed with five convolutional layers (filters 32 to 512) and the Adam optimizer. The experimental results show that the optimal configuration was achieved with a learning rate of 0.001 and 50 training epochs, resulting in very high model performance. In the evaluation using test data, the model achieved an accuracy of 99.54% with a loss value of 0.0175. The optimal performance of this model is driven by the quality of the primary dataset supported by comprehensive image preprocessing stages, thus ensuring clean, uniform, and significantly noise-free data input. Analysis of the confusion matrix and learning curves also confirmed the model's excellent generalization ability with no indications of overfitting. This model has been successfully implemented in the "Swaraksara" web application as a Sundanese script recognition system.