MBMR-Net: multi-branches multi-resolution cross-projection network for single image super-resolution

Deep convolutional neural networks (CNNs) have achieved significant developments in the field of single image super resolution (SISR) due to their nonlinear expression ability. However, existing architectures either rely on the representations learned from a single scale or extract deep features by cascading multiple resolutions, which unuse or underutilize the interdependence between low-resolution (LR) images and high-resolution (HR) images. In view of this trait, we propose a deep network called the multi-branches multi-resolution cross-projection network (MBMR-Net), which has multiple parallel branches, and cross-projection is performed between multiple branches to exchange information. Then, we introduce a novel attention unit that integrates second-order channel attention with spatial attention to better fuse information from multiple resolutions. Moreover, in terms of the characteristics of the model, we devise a loss function for enhancing the restoration of high-frequency details while ensuring the content information. Extensive quantitative and qualitative evaluations on benchmark datasets illustrate the effectiveness of our method and its competitive performance over state-of-the-art methods.