Deep Stacked Hierarchical Multi-patch Network for Image Deblurring

与之前方法pipeline的对比：

(a) (b) 缺点：Increasing the network depth for very low-resolution input in multi-scale approaches does not seem to improve the deblurring performance. 总的来说就是在a的这种pipeline中增加后面小图NN的depth没有太大提升。

( c ) proposed 优点：

1. As the inputs at different levels have the same spatial resolution, we can apply residual-like learning which requires small filter sizes and leads to a fast inference
2. We use a (Spatial Pyramid Matching) SPM-like model which is exposed to more training data at the finest level due to relatively more patches available for that level. 可以理解为 multiple patches 让网络获得更多的数据。

本文contribution:

1. 分patch的多级model
2. 针对增加深度无法提升模型performance的问题，提出了一个新的stack方法。

Method:

其中每一个encoder decoder的细节如下：

每一个encoder-decoder的排列方式有多种，文章实验了两种：
DMPHN：

VMPHN：

每一次loss只在lv1阶段的output做L2 loss。
每一个sub-model 不共享参数。