TASuRe: Text Aware Super-Resolution
Elena Filonenko, Alexander Filonenko, Kang-Hyun Jo · 2022
Recognition of text on low-resolution (LR) images is a challenging task. Traditional interpolation methods, as well as general super-resolution approaches, do not recover the shape of text character robustly. In this work, we propose a text-aware super- resolution neural network called TASuRe. Text awareness is interwoven in the proposed network that contains a text rectification part and text recognition auxiliary module. The training procedure is built around character shape restoration by adding a binary mask to the input image and using a specialized loss that penalties the network for missing gradients on the border of characters. Experiments on the real LR images have shown that the proposed network can deal with hard cases better than convolutional competitors.