LAPAR: Linearly-Assembled Pixel-Adaptive Regression Network for Single\n Image Super-Resolution and Beyond

Abstract

Single image super-resolution (SISR) deals with a fundamental problem of\nupsampling a low-resolution (LR) image to its high-resolution (HR) version.\nLast few years have witnessed impressive progress propelled by deep learning\nmethods. However, one critical challenge faced by existing methods is to strike\na sweet spot of deep model complexity and resulting SISR quality. This paper\naddresses this pain point by proposing a linearly-assembled pixel-adaptive\nregression network (LAPAR), which casts the direct LR to HR mapping learning\ninto a linear coefficient regression task over a dictionary of multiple\npredefined filter bases. Such a parametric representation renders our model\nhighly lightweight and easy to optimize while achieving state-of-the-art\nresults on SISR benchmarks. Moreover, based on the same idea, LAPAR is extended\nto tackle other restoration tasks, e.g., image denoising and JPEG image\ndeblocking, and again, yields strong performance. The code is available at\nhttps://github.com/dvlab-research/Simple-SR.\n