HANDWRITTEN TEXT RECOGNITION FOR CHAGATAI MANUSCRIPTS: COMPARING CRNN TRAINING FROM SCRATCH AND TRANSFORMER-BASED TRANSFER LEARNING

Abstract

The digitization of archival Chagatai manuscripts is limited by the scarcity of annotated corpora and the complexity of Arabic-based cursive scripts. This study evaluates two Handwritten Text Recognition (HTR) approaches: a Kraken-based CRNN trained from scratch and a transformer-based TrOCR model using transfer learning. Experiments were conducted on 2531-line images from the Jarring Collection with page-level data splitting to prevent data leakage. Architecture-specific augmentation strategies were applied to address data scarcity. The TrOCR model fine-tuned on augmented data achieved the best performance, reaching a Character Error Rate (CER) of 10.6% and a Word Error Rate (WER) of 35.59%. Data augmentation improved recognition performance for both evaluated architectures, with the augmented configurations achieving lower CER and WER than their respective baselines. The results indicate that transfer learning provided a more effective initialization strategy for the investigated Chagatai corpus under limited data conditions. The study establishes an initial experimental benchmark for Chagatai manuscript recognition.