An Efficient Scene Text Spotter with a Feature-Level Super-Resolution Module

End-to-end scene text spotting, which aims to jointly detect and recognize text in an image, plays an essential role in numerous practical applications. However, low-resolution and arbitrary-shaped text are still two challenges to be addressed to-wards accurate text spotting. To solve these two problems, in this paper, we present an end-to-end accurate text spotting network, which adopts super-resolution from feature-level for text texture enhancement. In detection stage, we adopt a segmentation-based detection manner using anchor-free mechanism without the restriction by pre-designed anchors to generate accurate polygonal proposals of text instances, especially for the arbitrary-shaped scene text. In recognition stage, we design a feature-level super-resolution module, which uses masked pooled text feature as input to focus on reconstructing the text region efficiently and thus provides robust feature representation, to improve the accuracy of recognition results by enhancing text feature representation. Experiment results demonstrate that the proposed method not only outperforms existing methods on the accuracy of low-resolution text spotting, but also achieves a certain extent accuracy improvement when spotting arbitrary-shape scene text.