Meta-SR: A Magnification-Arbitrary Network for Super-Resolution. CVPR 2019
1 文章摘要
Single Image Super-Resolution
提出Meta-SR在单⼀模型中完成任意尺度(包括⾮整数)的超分任务
Meta-SR中的Meta-Upscale Module可替换传统放⼤模块
Meta-Upscale Module利⽤输⼊的scale factor动态预测upscale filters的权重,并以此⽣成HR image with arbitrary size
⽹络结构简单,速度快,⽅便⾼效
2 问题背景
Super-Resolution of arbitrary scale factor 被⻓久忽视
现有的⼤部分⽅法:
- 将每个scale factor单独建模训练,对新的scale factor需要重新建模
- 仅仅实现整数的scale factor,⽆法实现x1.5、x2.5……
3 原理方法
Meta-Learning
- The meta-learning,让机器学会学习,⽽不仅仅是学习⼀个函数
- The meta-learning主要应⽤于:
- few-shot/zero-shot learning
- transfer learning
- The weight prediction is one of meta-learning strategy in the neural network
- neural network的权重被另⼀个NN预测,⽽⾮直接从training dataset中学习
本⽂思路:利⽤meta-learning预测与每个scale factor相对应filters的weights