VDSR--Accurate Image Super-Resolution Using Very Deep Convolutional Networks

论文及MATLAB代码：(https://cv.snu.ac.kr/research/VDSR/)

Abstract
受VGG网络启发，使用非常深的网络(研究目标：网络的深度对于准确度的影响；实施：使用3*3的filter，最终达到16-19层深)

Introduction
SRCNN的三个限制
first, it relies on the context of small image regions;
second, training converges too slowly;
third, the network only works for a single scale.

解决办法
1、 use large receptive field takes a large image context into account.
2、 residual-learning CNN and extremely high learning rates.

3、We propose a single-model SR approach.
Scales are typically user-specified and can be arbitrary including fractions.
放大倍数不在是唯一的。

Contribution
提出非常深的高度准确的卷及网络
（1）网络深所带来的问题：使用小的学习率，收敛很慢
使用大的学习率导致梯度爆炸
如何解决：residual-learning and gradient clipping
（2）extend our work to cope with multi-scale SR problem in a single network

2. Related Work
2.1
VGG网络说明网络深，准确度高；但是SRCNN，深层网络性能不如浅层（ They conclude that deeper networks do not result in better performance）
作者给出自己的分析：
（1）首先通过实验表明网络深比浅层，性能更好
（2）training：分析图像的高频信息和低频信息；给出SRCNN的功能分析；

3.1
除了第一层和最后一层外，有d层，filter大小3364；

4.2重点研究residuals

使用zero-padding，没说具体取值；

Once image details are predicted, they are added back to
the input ILR image to give the final image (HR).

和pipeline 不一样？？

3.2 training
High Learning Rates for Very Deep Networks
SRCNN深层不如浅层的原因：一个原因可能是未收敛，学习率太小；
学习率大，导致vanishing/exploding gradients

使用 adjustable gradient clipping
Gradient clipping

• 经常用于训练RNN；但是，作者认为，限制于训练CNN
• 常用的处理梯度的方法：clip individual gradients to the predefined range [−θ,θ].
• 如果使用高的学习率， θ is tuned to be small to avoid exploding gradients

4.1 感受野的论述
stacking many such layers leads to filters that become increasingly global

本文使用的receptive field 均是3*3；
网络深度和感受野大小关系：(2D + 1) × (2D + 1)

为什么要用大的receptive field
A large receptive field means the network can use more context to predict image details。
为什么要用深层网络：
1、 receptive field 更大
2、very deep networks can exploit high nonlinearities.
因此， very deep networks significantly improve SR performance.

4.2. Residual-Learning

未完待续