PENERAPAN MODEL HYBRID CNN-BILSTM UNTUK KONVERSI TULISAN TANGAN KE DOKUMEN DIGITAL

Abstract

Digital transformation through Handwritten Text Recognition (HTR) technology based on deep learning can optimize the efficiency of processing public administration documents. In regional government agencies, processing handwritten files manually causes severe bottlenecks and human error. This study aims to analyze the effect of hyperparameter configurations and evaluate the quantitative performance of a hybrid CNN-BiLSTM architecture with Connectionist Temporal Classification (CTC) Loss using real handwritten administrative document datasets from the Department of Cooperatives and SMEs in Makassar City. The CNN module acts as a spatial visual feature extractor, while BiLSTM processes sequence transitions bidirectionally across 128 time-steps. The dataset consists of 400 line images divided into an 80:20 ratio for training and testing. The experimental results demonstrate that the hybrid CNN-BiLSTM model achieved a character accuracy of 27.03% (CER 72.97%, WER 98.75%), Precision of 0.1919, Recall of 0.1781, and F1-Score of 0.1747, outperforming single baseline models (CNN 11.45% and BiLSTM 8.20%). The model achieved 100% exact match transliteration on clean background documents. The main factors limiting recognition include cursive handwriting, sharp slant angles, and local non-standard acronyms. In conclusion, the hybrid CNN-BiLSTM architecture effectively integrates spatial feature extraction and sequential contextual modeling, providing a reliable backend foundation for automating regional public document digitization.