An end-to-end pipeline for historical censuses processing

Abstract Censuses are structured documents of great value for social and demographic history, which became widespread from the nineteenth century on. However, plurality formats natural variability historical data make their extraction arduous often lead to ungeneric recognition algorithms. We propose an end-to-end processing pipeline, based on optimization, in attempt reduce number free parameters. The layout analysis is semantic segmentation using neural networks a generic explicit column structure. implicit row structure deduced directly position text segments. handwritten detection complemented by intelligent framing method significantly improves quality HTR. In end, we combine several post-correction approaches, networks, language models, further improve performance. Ultimately, our flexible methods it possible accurately detect more than 98% columns 88% rows, despite lack graphical separator diversity formats. Thanks various reframing strategies, HTR results reach excellent performance 3.44% character error rate these noisy data. total, 18,831 pages were extracted 72 censuses over century. This large dataset, as well training data, made open-access released along with this article.