DAN: a Segmentation-free Document Attention Network for Handwritten Document Recognition

Unconstrained handwritten text recognition is a challenging computer visiontask. It is traditionally handled by a two-step approach, combining linesegmentation followed by text line recognition. For the first time, we proposean end-to-end segmentation-free architecture for the task of handwrittendocument recognition: the Document Attention Network. In addition to textrecognition, the model is trained to label text parts using begin and end tagsin an XML-like fashion. This model is made up of an FCN encoder for featureextraction and a stack of transformer decoder layers for a recurrenttoken-by-token prediction process. It takes whole text documents as input andsequentially outputs characters, as well as logical layout tokens. Contrary tothe existing segmentation-based approaches, the model is trained without usingany segmentation label. We achieve competitive results on the READ 2016 datasetat page level, as well as double-page level with a CER of 3.43% and 3.70%,respectively. We also provide results for the RIMES 2009 dataset at page level,reaching 4.54% of CER. We provide all source code and pre-trained model weights athttps://github.com/FactoDeepLearning/DAN.