A Self-Attention Based Model for Offline Handwritten Text Recognition

EasyChair Preprint 7037

Abstract

Offline handwritten recognition is an important part of handwritten document analysis systems and has received a lot of attention from numerous researchers for decades. In this paper, we present a self-attention-based model for offline handwritten text line recognition. The proposed model consists of three main components: a feature extractor by CNN, an encoder by BLSTM and a self-attention module, and a decoder by CTC. The self-attention module is complementary to RNN in the decoder and helps it to capture long-range, multi-level dependencies across the input sequence. According to the extensive experiments on the two datasets of IAM Handwriting, and Kuzushiji, the proposed model achieves similar or better accuracy than the state-of-the-art models. The self-attention map visualization shows that the self-attention mechanism helps the decoder capture the dependencies between different positions of arbitrarily distance in the input sequence.

Keyphrases: BLSTM, CNN, CTC, Handwritten Text Recognition, multi-head, self-attention

@booklet{EasyChair:7037,
  author    = {Nam Tuan Ly and Trung Tan Ngo and Masaki Nakagawa},
  title     = {A Self-Attention Based Model for Offline Handwritten Text Recognition},
  howpublished = {EasyChair Preprint 7037},
  year      = {EasyChair, 2021}}