From Textline to Paragraph: A Promising Practice for Chinese Text Recognition

摘要

Although handwritten Chinese text recognition (HCTR) has achieved tremendous progress in the past decades, the traditional document analysis system suffers from two main problems: (1) The annotation of position and transcript at line level is costly to obtain; (2) The framework consists of several separately trained modules, and it's difficult for the complex system to get satisfying results. Therefore, handwritten paragraph recognition attempts to incorporate the textline segmentation and recognition into a complete network. However, large character set and great insufficient training samples make it troublesome for handwritten Chinese paragraph recognition (HCPR). In this paper, a novel framework is proposed for HCPR. To make the training process faster and more stable, we put forward the Multi-Dimensional LSTM Convolutional Attention (MLCA) recognition framework. A new writing-style-aware image synthesis method is utilized as well to overcome the problem of data insufficiency. We conduct several experiments on the ICDAR-2013 competition dataset and the corresponding corrupted dataset. From the compelling results, we can draw an encouraging conclusion that it would be a promising trend to move from HCTR to HCPR for Chinese document analysis system.