Patch loss: A generic multi-scale perceptual loss for single image super-resolution

Tai An, Binjie Mao, Bin Xue, Chunlei Huo, Shiming Xiang, Chunhong Pan

Abstract: Highlights•A plug-and-play multi-scale perceptual loss for single image super-resolution (SISR) is proposed and explained in principle.•Highly customizable image patch kernel (IPK) and feature patch kernel (IPK) are proposed for extracting vectors of different sizes.•Multi-scale parallel branches and parameter-free adaptive patch-wise attention are proposed for refining edges and textures at different scales.•Our work is the first to systematically compare the performance of mainstream SISR methods on a broad range of perceptual metrics.•The code will be available at https://github.com/suanmd/Patch-Loss-for-Super-Resolution.