Pengenalan Tulisan Tangan Aksara Bali Menggunakan Algoritma CRNN

Abstract

This research proposes the development of a handwritten Balinese script recognition system using the Convolutional Recurrent Neural Network (CRNN) algorithm as a vital step in preserving this intangible cultural heritage from the threat of obsolescence in the digital era. Given the inherent complexity and variability of handwritten Balinese script, which distinguishes it from printed text, a deep learning approach is essential. CRNN was chosen for its ability to integrate the strengths of CNN in extracting spatial visual features with the power of RNN (specifically BiLSTM) in modeling sequential dependencies. Primary handwritten data was meticulously validated by a Balinese language expert, then processed through grayscale conversion, pixel normalization, and resizing for standardization. The model was constructed with convolutional layers, recurrent BiLSTM layers, and a Connectionist Temporal Classification (CTC) transcription layer, which is effective in translating sequential features into character labels. Performance evaluation of the model using the Character Error Rate (CER) on separate test data showed an average accuracy of 89.9%. These results significantly affirm the great potential of CRNN in supporting the digitalization efforts of Balinese script, as well as facilitating its broader integration into modern environments.