Document Segmentation under Limited Data Conditions: Transfer Learning and Few-Shot Approaches

This thesis addresses the problem of document segmentation under limited data conditions, focusing on industrial and historical documents. Document layout analysis is essential for converting unstructured visual information into machine-readable formats, supporting applications ranging from industrial automation to cultural heritage preservation. However, real-world scenarios often involve scarce annotated data and high variability in document layouts, which challenges traditional segmentation approaches that rely on large labeled datasets. To overcome these limitations, this thesis explores transfer learning and few-shot learning strategies, supported by the creation of two novel datasets specifically designed for industrial and historical documents. A comprehensive analysis of low-resource document layout segmentation highlights how annotation scarcity and structural diversity affect model generalization. Extensive experiments demonstrate that domain-specific pre-training enhances performance more effectively than generic large-scale pre-training, while a novel few-shot learning framework achieves accurate segmentation using only a small number of annotated examples. The proposed methods also exhibit strong cross-domain generalization, addressing the challenges posed by variability and low-resource scenarios. Overall, the thesis shows that effective document segmentation can be achieved with minimal annotation effort, bridging the gap between theoretical advances and real-world applications and contributing to both industrial efficiency and cultural heritage preservation.