This paper studies the performance enhancement of Historical Handwriting Recognition (HTR) by applying TrOCR, a state-of-the-art transformer-based HTR model, to a 16th-century Latin manuscript by Rudolf Gualter. Specifically, we apply image preprocessing and various data augmentation techniques (including four novel augmentation techniques that consider the characteristics of historical handwritings), and evaluate an ensemble learning method that leverages the strengths of augmented models. As a result, a single augmented model (Elastic) achieves a character error rate (CER) of 1.86, while a top-five model voting ensemble achieves a CER of 1.60, representing a 42% relative performance improvement over the previous best-performing model and a 50% relative performance improvement over the existing TrOCR_BASE. This demonstrates the importance of domain-specific augmentation and ensemble strategies.
