Improvement of Character Segmentation Using Recurrent Neural Networks and Dynamic Programming

Abstract

A common characteristic of all the existing online handwritten text recognition algorithms is that the character segmentation process is closely related to the recognition process. There are different approaches to segment data but all of them don't give absolutely correctly segmentation results due to specifics of handwriting data input. In this paper, we present a new approach for character segmentation improvement in online handwriting recognition which is based on using recurrent neural networks and dynamic programming. Due to online handwritten text is a sequence of points we propose to use Bidirectional Long Short-Term Memory (BLSTM) for classification of decoder outputs and dynamic programming for interpretation of classification results. Experimental evaluation shows the effectiveness of a proposed approach in increasing of segmentation quality.