Recognizing Handwritten Chinese Texts with Insertion and Swapping Using a Structural Attention Network

Shi Yan, Jin-Wen Wu, Fei Yin, Cheng-Lin Liu

Abstract: It happens in handwritten documents that text lines distort beyond sequential structure because of in-writing editions such as insertion and swapping of text. This kind of irregularity can not be handled using existing text line recognition methods that assume regular character sequences. In this paper, we regard this irregular text recognition as a two-dimensional (2D) problem and propose a structural attention network (SAN) for recognizing texts with insertion and swapping. Particularly, we present a novel structural representation to help SAN learn these irregular structures. With the guidance of the structural representation, SAN can correctly recognize texts with insertion and swapping. To validate the effectiveness of our method, we chose the public SCUT-EPT dataset which contains some samples of text with insertion and swapping. Due to the scarcity of text images with text insertion and swapping, we generate a specialized dataset which only consists of these irregular texts. Experiments show that SAN promises the recognition of inserted and swapped texts and achieves state-of-the-art performance on the SCUT-EPT dataset.

External IDs:dblp:conf/icdar/YanWYL21