STSR: Sequence-Based Scene Text Image Super-Resolution

Abstract

When faced with low-resolution scene text images, the text recognizers often perform unsatisfactorily. Therefore, the super-resolution can be used as the preprocessing work to improve the accuracy of text recognition. However, current existing image super-resolution approaches mostly aim to recover the details of images while ignoring the text feature in text images. Besides, there is no specially designed loss functions for text images super-resolution. In this study, we propose a Sequence-based Scene Text Image Super-Resolution Network (STSR), in which the Bi-directional Text Sequential Block (Bi-TSB) is designed to capture the text sequential features. Meanwhile, the text sequential loss is proposed to supervise the text readability of the reconstructed image. The experimental results indicate that STSR outperforms the baseline methods in terms of SSIM and recognition accuracy in TextZoom dataset.