Joint recognition of text and layout in historical Russian documents

In this paper, we evaluated the Document Attention Network (DAN), the first end-to-end segmentation-free architecture on Historical Russian Documents. The DAN model jointly recognizes both text and layout from whole documents, it takes whole documents from any size as an input and output the text as well as logical layout tokens. For comparison purposes, we conduct our experiments on Digital Peter dataset as it has been recognized at line-level. Dataset consists of documents of Peter the Great manuscripts; ground truths are represented according to a sophisticated XML schema which enables an accurate detailed definition of layout and text regions. We achieved good results at page-level: 18.71 % for Character Error Rate (CER), 39.7 % for Word Error Rate (WER), 14.11 % For Layout Ordering Error Rate (LOER), and 66.67 % for mean Average Precision (mAP).