Combining Visual and Textual Features for Semantic Segmentation of Historical Newspapers

The massive amounts of digitized historical documents acquired over the last decades naturally lend themselves to automatic processing and exploration. Research work seeking automatically process facsimiles extract information thereby are multiplying with, as a first essential step, document layout analysis. If identification categorization segments interest in images have seen significant progress years thanks deep learning techniques, many challenges remain among others, use finer-grained segmentation typologies consideration complex, heterogeneous such newspapers. Besides, most approaches consider visual features only, ignoring textual signal. In this context, we introduce multimodal approach for semantic newspapers that combines features. Based on series experiments diachronic Swiss Luxembourgish newspapers, investigate, predictive power their capacity generalize across time sources. Results show consistent improvement models comparison strong baseline, well better robustness high material variance.