Page Level Input for Handwritten Text Recognition in Document Images

Important information present all over the world in the form of libraries need to be digitized. An end-to-end handwritten text recognition system covers every aspect, from capturing an image as input to predict text written inside it. In the present study, we have presented a document image-based handwritten text recognition system with three parts as: page level image as input with multiple lines, recognition of text of input image and attention level of text. This model is trained and tested on various split ratios of labelled words of the JAM handwriting dataset. The recognition model evaluation is performed on two publicly available datasets that are JAM handwriting dataset and handwritten short answer scoring dataset. We have used popular CTC word beam search technique by Scheidl et al. [1] in recognition phase, and attention of text has been incorporated using established sentic data by Ref. [2]. Our findings at recognition level result in 8.99% validation character error rate on JAM handwriting dataset and 35.92 BLEU score on the handwritten short answering scoring dataset. The post recognition findings suggest text attention role in acceptance of recognized words useful for real life business documents.