Single image super-resolution based on mapping-vector clustering and nonlinear pixel-reconstruction

Single image super-resolution (SR) aims to estimate high-resolution (HR) images from low-resolution (LR) ones with high reconstruction quality and low time cost. In this paper, we propose a new pixel-wise SR method by combining a classification tree of external samples and a learning stage of nonlinear pixel-reconstruction mapping-kernels. To obtain more reasonable sample sets for mapping-learning, we classify external samples by clustering the mapping-vector of LR-HR patch pairs based on fractional-norm. Then, a decision-tree branched by lightweight networks is learned to choose one reasonable class for each testing LR patch. In our mapping-learning stage, pixel-wise nonlinear mappings are represented as several full connected networks, which can provide satisfying generalization ability for LR patch reconstruction. Therefore, our method can simultaneously guarantee the reconstruction quality and execution speed at the stage of sample representation and mapping. Experiment results demonstrate the effectiveness of our approach on quantitative visual quality assessment and time cost. •To obtain more reasonable sample sets for mapping-learning, we classify external samples by clustering the mapping-vector of LR-HR patch pairs based on fractional-norm.•A decision-tree branched by lightweight networks is learned to choose one reasonable class for each testing LR patch.•In our mapping-learning stage, pixel-wise nonlinear mappings are represented as several full connected networks, which can provide satisfying generalization ability for LR patch reconstruction.