Image Super-Resolution Based on Generalized Residual Network

Residual networks (ResNets) have recently achieved great success for single-image super-resolution, but its identity connection leads to accumulation of a mixed levels of feature representations, and it is not conductive to optimize the information flow. To address these issues, a novel Resnet in Resnet architecture (RiRSR) based on generalized residual (GenRes) module is proposed to reconstruct the high-resolution image. Specifically, the network is composed of some Resnet in Resnet blocks containing multiple GenRes sub-blocks for feature extraction, where the GenRes sub-block combines the residual and non-residual streams, which retains the benefits of identity shortcut connections, improving features expressivity. Moreover, both local and global residual learning are integrated to ease the information flow and accelerate convergence. Finally, multi-scale training strategy is utilized to further boost reconstruction performance and reduce the number of model parameters. Experimental results demonstrate that the proposed RiRSR can improve the reconstruction performance (more than 0.1 dB than ResNets) with higher accuracy for the integer and non-integer scales, and is competitive in balancing model parameters, reconstruction speed and reconstruction quality.