Attention-enhanced multi-scale residual network for single image super-resolution

Yubin Sun, Jiongming Qin, Xuliang Gao, Shuiqin Chai, Bin Chen

Abstract: Single image super-resolution (SISR) has important applications in many fields. With the help of this technology, the broadband requirement of image transmission can be reduced, the effect of remote sensing observation can be improved, and the location of lesion cells can be accurately located. Convolutional neural networks (CNNs) using multi-scale feature extraction structure can gain a large amount of information from a low-resolution input, which is helpful to improve the performance of SISR. However, these CNNs usually treat different types of information equally. There is a lot of redundancy in the information obtained, which limits the representation ability of the networks. We proposed an attention-enhanced multi-scale residual block (AMRB), which increases the proportion of useful information by embedding convolutional block attention module. Furthermore, we construct an attention-enhanced multi-scale residual network based on one time feature fusion (OAMRN). Extensive experiments illustrate the necessity of the AMRB and the superiority of proposed OAMRN over the state-of-the-art methods in terms of both quantitative metrics and visual quality.