Scratch-Trained CRNNs vs. Pretrained Vision Transformers: A Controlled Benchmark of Augmentation and Model Scale in Low-Resource Chagatai HTR

Digitizing the historical Chagatai manuscripts of the Gunnar Jarring Collection is severely hindered by cursive orthographic complexity and extreme data scarcity. This study establishes a rigorous empirical HTR benchmark, comparing two distinct paradigms under identical, highly constrained conditions: a convolutional recurrent neural network (CRNN) trained from scratch and pretrained Vision-Transformer-based architectures (TrOCR Base and Large). To eliminate intra-page data leakage, a strict page-level 80/10/10 partitioning protocol is enforced across the 2531 manually annotated text lines derived from 230 pages. To guarantee statistical reliability, all configurations are evaluated across three independent random seeds. The fine-tuned TrOCR Large model with architecture-specific data augmentation achieved the highest accuracy, yielding a Character Error Rate (CER) of 9.82 ± 0.05% and a Word Error Rate (WER) of 33.27 ± 0.77%, outperforming the from-scratch CRNN baseline (18.07 ± 0.79% CER) by 8.25% absolute. Our decoupled evaluations reveal that while pretraining provides a robust prior that stabilizes learning, spatial augmentation acts as a crucial regularizer. Our decoupled evaluations reveal that the performance advantage of fine-tuned Vision Transformers over scratch-trained CRNNs stems primarily from the stabilizing prior of pretrained transfer learning, while spatial augmentation acts as a crucial regularizer under extreme data scarcity. Character-level error taxonomies identify systemic failure modes, including deletion biases in Transformers and insertion-heavy “diacritic babbling” in connectionist temporal classification models.