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.

残差网络（ResNets）最近在单幅图像超分辨率方面取得了巨大的成功，但其身份连接导致了混合层次的特征表示的积累，并且不利于优化信息流。为了解决这些问题，提出了一种基于广义残差（GenRes）模块的 Resnet 架构（RiRSR）中的新型 Resnet 来重建高分辨率图像。具体来说，网络由包含多个 GenRes 子块的 Resnet 块中的一些 Resnet 组成，用于特征提取，其中 GenRes 子块结合了残差和非残差流，保留了身份快捷连接的好处，提高了特征表达能力。此外，整合了局部和全局残差学习以缓解信息流并加速收敛。最后，多尺度训练策略用于进一步提高重建性能并减少模型参数的数量。实验结果表明，所提出的 RiRSR 可以在整数和非整数尺度上以更高的精度提高重建性能（比 ResNets 提高 0.1 dB 以上），并且在平衡模型参数、重建速度和重建质量方面具有竞争力。