Multi-scale dual regression networks for single-image super-resolution

The implementation of neural networks with deep convolution has risen in single-image super-resolution reconstruction (SISR) networks due to their deep architecture, which has now gained significant performance. However, the currently proposed super-resolution reconstruction network is still insufficient for the recovery of complex scene images. Because complex scenes generally contain rich detail information, a strong detail extraction capability of the reconstruction network is required for recovering such images. In this study, we describe a multi-scale dual regression network (MDRN) to recover details in high-resolution images by extracting feature information at different scales so that various texture details in complex scenes can be reconstructed. Specifically, the MDRN extracts deep features by stacking residual efficient attention blocks on one hand, and on the other hand, a multi-scale efficient attention module (MREAB) is designed to enhance the network’s ability to characterize the information at different scales, which allows our network to not only focus on the large-scale variations in the image but also to capture more texture details. Quantitative and qualitative evaluations on five benchmark datasets show that the proposed algorithm performs well compared to state-of-the-art super-resolution methods.