DBPNet: A dual-branch pyramid network for document super-resolution

Convolutional neural networks (CNN), aiming to preserve the structural and texture in-formation lost in the initial low-resolution (LR) images, has been widely used to improve the quality of LR images. Most of the existing super-resolution methods focus on image super-resolution reconstruction in natural scenes, while a few super-resolution methods for document images focus on text data. However, most CNN mod-els focus more on the reconstruction quality of the entire document and ignore the feature difference between the text region and the image region which both exist in the document image. To address this issue, this paper proposes a dual-branch pyramid network (DBPNet) for document super-resolution by taking into consideration about the texture difference between the text region and the image region. DBPNet consists of a region segmentation module (RSM), two parallel pyramid SR branches (PPSRB), and text regions, with a pyramid edge restoration module (ERM) and a region fusion module (RFM) for re-gion stacking. Furthermore, to verify that the model can be adapted to super-resolution reconstruction of document images with more features, we construct a new document super-resolution (SR) dataset namely DSR2021. It contains paired LR and HR images involving English, Chinese, Japanese and Korean languages. Experiments on two document image datasets (DSR2021 and Text330) demonstrate that our method outperforms several state-of-the-art methods quantitatively and qualitatively.(c) 2022 Elsevier B.V. All rights reserved.