Generazione di Testo Scritto a Mano per il Riconoscimento: Da Archetipi Visivi a Modelli Auto-Regressivi

This thesis advances Handwritten Text Generation (HTG) to improve Handwritten Text Recognition (HTR) in challenging, low-resource settings. The core objective of HTG is to generate realistic text images that match user-specified styles from minimal examples by decoupling handwriting characteristics from textual content. Key Innovations: First, the research introduces novel few-shot HTG models utilizing CNNs and Transformers. A major contribution is the use of visual archetypes, which are glyph-like images that leverage geometric similarities to accurately generate rare or unseen symbols. Second, through specialized large-scale pretraining, data augmentation, and a novel autoregressive paradigm, the models successfully adapt to unfamiliar styles and generate text of any length. Third, the thesis proposes a task-specific metric, Handwriting Distance (HWD), to address the limitations of Frechet Inception Distance (FID). HWD directly measures stroke geometry and spatial configurations for a more accurate evaluation of style similarity. Application to HTR: The thesis highlights HTG's value as a synthetic data pipeline. Generating style-consistent synthetic text for pretraining significantly boosts HTR accuracy, especially when real annotations are scarce or stylistically mismatched. Validated across modern and historical manuscripts, this research establishes a powerful synergy where HTG supplies essential training data, leading to vastly improved recognition outcomes.