- 总体概括：该论文主要是在SRResNet网络的基础之上进行了改进，因为ResNet主要是针对分类或者检测问题来说效果比较好，所以，对本文想对原有的残差网络进行修改，从而提升效果。并且，该论文实现了两种模型，一种是Single-Scale Model,即EDSR,另外一种是Multi-Scale Model，即MDSR。
- 论文中的一些改进内容
-
残差块：(1)We remove the batch normalization layers from our network. Since batch normalization layers normalize the features, they get rid of range flexibility from networks by normalizing the features, it is better to remove them.(2)Furthermore, GPU memory usage is also sufficiently reduced since the batch normalization layers consume the same amount of memory as the preceding convolutional layers.
-
- Single-scale model（单尺度模型）：因为feature maps的数量增加到一定的数量之后，训练就会变的很不稳定，所以，本文中增加了一个residual scaling，（In each residual block, constant scaling layers are placed after the last convolution layers. These modules stabilize the training procedure greatly when using a large number of filters.）。当对该模型进行训练的时候，上采样因子分别为x3和x4，所以我们初始化参数使用的是预训练x2的网络，这种预训练的措施加速了训练的过程，并且提升了实验效果，从下面右图可以看出来，采用上面的措施收敛的会更快。
- Multi-scale model(多尺度模型)：一开始每个尺度都有两个独自的残差块，之后经过若干个残差块，最后再用独自的升采样模块来提高分辨率。文中作者设置 B=80，F=64。
- Geometric Self-ensemble：这是一个很神奇的方法，测试时，把图像90度旋转以及翻转，总共有8种不同的图像，分别进网络然后变换回原始位置，8张图像再取平均。这个方法可以使测试结果有略微提高。Note that geometric self-ensemble is valid only for symmetric downsampling methods such as bicubic downsampling.
3.训练细节：（1）For training, we use the RGB input patches of size 48×48 from LR image with the corresponding HR patches. We augment the training data with random horizontal flips and 90 rotations.（2）We train our networks using L1 loss instead of L2.（3）scalespecific residual blocks and upsampling modules that correspond to different scales other than the selected one are not enabled nor updated.（不启用也不更新对应于除所选尺度之外的不同尺度的尺度特定残差块和上采样模块）
