Text Line Segmentation in Historical Newspapers

This paper deals with page segmentation into individual text lines used as an input of a line-based OCR system. This task is usually solved in one step which directly identifies text lines in whole documents. However, a direct approach may jeopardize the reading order of the lines and thus deteriorate the overall transcription result. We propose a novel approach which decomposes this problem into two steps: text-block and text-line segmentation. The particular tasks are handled by algorithms based on fully convolutional neural networks. The proposed method is evaluated on two standard corpora, Europeana and RDCL 2019, and on a novel dataset created from data available in Porta fontium portal. This dataset is freely available for research purposes.