《Multi-Scale Residual Learning-using a Cycle Spinning CNN for Single Image De-Raining》

de-raining跟超分一样，应该都属于low-level vision的问题。之前的博客《论文阅读笔记之——《Moiré Photo Restoration Using Multiresolution Convolutional Neural Networks》》介绍的去摩尔纹应该也算是同类型的问题。

一开始我认为，这类型的任务跟去噪任务非常像。

Single image de-raining is an extremely challenging problem since the rainy image may contain rain streaks which may vary in size, direction and density.

作者提到，之前的方法的一个缺陷是One of the major limitations of these approaches is that they do not consider the location information of rain drops in the image.（为啥需要考虑雨滴的位置呢？）the proposed Uncertainty guided Multi-scale Residual Learning (UMRL) network attempts to address this issue by learning the rain content at different scales and using them to estimate the final derained output.

作者在introduction认为their algorithm tends to introduce some artifacts in the final de-rained images.那么所谓的位置信息，其实是语义信息，这样的话，其出发点跟SFTGAN是相类似的。比如在人脸、花的地方，会引入artifacts

we make use of the observation that rain streak density and direction does not change drastically with different scales. Rather than relying on the rain density information (i.e. heavy, medium or light) present in the rainy image, we develop a method in which the rain streak （条纹） location information is taken in to consideration in a multiscale fashion to improve the de-raining performance.

也是采用Unet architecture with skip connections的结构作为base model

cycle spinning framework was originally proposed to remove the artifacts introduced by orthogonal wavelets in image de-noising. Similar to wavelets, deep learning-based methods also introduce some artifacts near the edges of the de-rained images。Cycle spinning has been successfully applied to reduce the artifacts introduced near the edges in many applications including image de-blurring and de-noising.

Unlike many deep learning-based methods that directly estimate the de-rained image from the noisy observation, we take a different approach in which we first estimate the rain streak component ˆr (i.e. residual map) and then use it to estimate the de-rained image as ˆx = y − ˆr.

the residual map and its corresponding confidence map were able to capture the regions where there is high probability of incorrect estimates.