Domain Adaption in Sequence-to-Sequence Scene Text Recognition

Abstract

Domain adaption techniques such as gradually vanishing bridge (GVB), have shown promising results in image classification problems. However, their efficacy in sequence-tosequence scene text recognition (STR) is yet to be known. In this paper, we combine GVB and connectionist temporal classification (CTC) techniques in STR model to improve the text recognition performance. The proposed approach is evaluated on publicly available datasets. Experimental results show the performance gain compared with state-of-the-art approaches.