On Recognizing Texts of Arbitrary Shapes with 2D Self-Attention

Junyeop Lee, Sungrae Park, Jeonghun Baek, Seong Joon Oh, Seonghyeon Kim, Hwalsuk Lee

Abstract: Scene text recognition (STR) is the task of recognizing character sequences in natural scenes. While there have been great advances in STR methods, current methods which convert two-dimensional (2D) image to one-dimensional (1D) feature map still fail to recognize texts in arbitrary shapes, such as heavily curved, rotated or vertically aligned texts, which are abundant in daily life (e.g. restaurant signs, product labels, company logos, etc). This paper introduces an architecture to recognizing texts of arbitrary shapes, named Self-Attention Text Recognition Network (SATRN). SATRN utilizes the self-attention mechanism, which is originally proposed to capture the dependency between word tokens in a sentence, to describe 2D spatial dependencies of characters in a scene text image. Exploiting the full-graph propagation of self-attention, SATRN can recognize texts with arbitrary arrangements and large inter-character spacing. As a result, our model outperforms all existing STR models by a large margin of 4.5 pp on average in "irregular text" benchmarks and also achieved state-of-the-art performance in two "regular text" benchmarks. We provide empirical analyses that illustrate the inner mechanisms and the extent to which the model is applicable (e.g. rotated and multi-line text). We will open-source the code.