Single image super resolution based on multi-scale structure and non-local smoothing

Abstract In this paper, we propose a hybrid super-resolution method by combining global and local dictionary training in the sparse domain. order to present differentiate feature mapping different scales, set is trained multiple structure non-linear function used choose appropriate initially reconstruct HR image. addition, introduce Gaussian blur LR images eliminate widely but inappropriate assumption that low resolution (LR) are generated bicubic interpolation from high-resolution (HR) images. deal with blur, iteratively updated K -means principal component analysis (K-PCA) gradient decent (GD) model effect during down-sampling. Compared state-of-the-art SR algorithms, experimental results reveal proposed can produce sharper boundaries suppress undesired artifacts of blur. It implies our could be more real applications HR-LR relation complicated than interpolation.