这一篇是CVPR16 Kim的VDSR，通过VERY DEEP的简单模型，又快又好地解决了SR问题，成为暂时这个问题上的标杆模型。
Abstract
Our final model uses 20 weight layers. By cascading small filters many times in a deep network structure, contextual infor- mation over large image regions is exploited in an efficient way. With very deep networks, however, convergence speed becomes a critical issue during training. We propose a sim- ple yet effective training procedure. We learn residuals only and use extremely high learning rates (104 times higher than SRCNN [6]) enabled by adjustable gradient clipping.
Introduction
Single image super-resolution(SISR):upsampling方法，而后neighbor embedding,如今用CNN； SRCNN的limitation: 1.relies on context of small image regions; 2.only works for single scale； VDSR的主要优点有：1.通过small size kernel but very deep, to obtain a large context(receptive region) 2.Convergence very fast by residual-learning and BN high learning rate 3.Multi-Scale Factor,把多个scale的SR融合进一个网络模型
Methodology
20层CONV+BN+RELU，L2 LOSS, HIGH LEARN RATE WITH RESIDUAL LEARNING AND ADJUSTABLE WEIGHT CLIPPING.
Experiment
- THE DEEPER THE BETTER ON PSNR/SSIM
- RESIDUAL LEARNING WORKS
- MULTI-SCALE MODEL BETTER THAN SINGLE SCALE ON LARGE SCALE
