From Layout Analysis to Article Segmentation: A Pipeline for Swedish 19th Century Newspaper Digitization

Historical newspapers are important sources for cultural memory, offering insights into the social, political contexts of their time. However, unlocking this information remains challenging due to complex layouts, degraded material, and the lack of annotated data. This thesis investigates the use of a modular pipeline to improve document understanding in 19th-century Swedish newspapers, consisting of Document Layout Analysis, Text Line Segmentation, Optical Character Recognition, and Article Segmentation. Motivated by the goal to make historical content more accessible, the study combines computer vision and natural language processing techniques to process scanned historical pages into article-level data. Each stage is designed to operate sequentially. The system demonstrates strong performance across all stages by combining lightweight object detection models, transformer-based Optical Character Recognition, and unsupervised semantic clustering. The pipeline proved effective in handling variation across historical layouts and print conditions, with Optical Character Recognition reaching competitive accuracy and article segmentation benefiting from fused embeddings and adaptive clustering. The findings of this study offer valuable information for researchers working with historical document collections, especially newspapers, and highlight the potential of modular pipelines as a practical foundation for future research in historical newspaper understanding.