GUSL: a green learning approach to single-image super-resolution

This work proposes a Green U-Shaped Learning (GUSL) method for single image super-resolution. As an implementation of green learning, GUSL is distinguished by high interpretability and notable efficiency, with comparable evaluation performance in PSNR and SSIM and close visualization to an entry-level deep-learningbased method. GUSL is a multi-level approach that progressively predicts residual from the deep level of global long-range structure to the local short-range texture. Benefiting from the multilevel diagram, GUSL can decompose the overall recovery into cascaded conditional recovery with finer management and supervise local texture residual recovery by the global structure.