A Deep Pyramid Attention Network for Single Image Super-resolution
Garas Gendy, Hazem Mohammed, Nabil Sabor, Guanghui He
Abstract
The pyramid attention (PA) network is a new structure developed for digital image processing. This network was designed to extract long-range features at different locations and scales. Recently, PA architecture has been introduced in single image super-Resolution (SISR) to improve the model's ability to benefit from data's self-similarity. However, the effects of location and number of PA on extracting the self-similarity are not explored. In this paper, a Deep Pyramid Attention Network (DPANet) is proposed for SISR based on exploring the PA block. This is performed by studying the effect of varying the number of PA blocks and their locations on the model performance. Moreover, the effect of the residual scale on the PA's performance is studied. Evaluated based on five benchmark datasets, we concluded that using five PA blocks without down-scale residual interchanging with Resblocks in the network achieves significantly better results compared to the state-of-the-art methods. In addition, our model achieves superior visual quality and accuracy.