代碼：https://github.com/SeungjunNah/DeepDeblur_release
代碼解讀https://blog.csdn.net/wm6274/article/details/71194225
數據集下載鏈接在上面的github裏面
we propose a multi-scale convolutional neural network that restores sharp images in an end-to-end manner where blur is caused by various sources.（網絡結構）
we present multi-scale loss function that mimics（模擬） conventional coarse-to-fine approaches.（loss改進）
propose a new large-scale dataset that provides pairs of realistic blurry image and the corresponding ground truth sharp image that are obtained by a high-speed camera（數據集）
Thus, some approaches tried to parametrize blur models with simple assumptions on the sources of blurs.（傳統的方法中，通過估算blur kernel或者類似之前做過的，估算blur level map（SRMD操作）可是經過實驗後，個人覺得這種方式不靠譜）However, these blur kernel approximations are still inaccurate, especially in the cases of abrupt motion discontinuities and occlusions. Note that such erroneous （錯誤的） kernel estimation directly affects the quality of the latent image, resulting in undesired ringing artifacts.
Unlike other approaches, our method does not estimate explicit blur kernels. Accordingly, our method is free from artifacts that arise from kernel estimation errors.（不去估計level map就可以減少artifacts的引入）
Since no pairs of real blurry image and ground truth sharp image are available for supervised learning, they commonly used blurry images generated by convolving synthetic blur kernels.Thus, CNN-based models are still suited only to some specific types of blurs, and there are restrictions on more common spatially varying blurs.（這裏就闡述了CNN method實際上是對特定退化模型的擬合，泛化能力差）
These are mainly due to the use of simple and unrealistic blur kernel models. Thus, to solve those problems, in this work, we propose a novel end-to-end deep learning approach for dynamic scene deblurring.
1、we propose a multi-scale CNN that directly restores latent （潛伏的） images without assuming any restricted blur kernel model.Especially, the multi-scale architecture is designed to mimic （模擬） conventional coarse-to-fine optimization methods.
2、we train the proposed model with a multi-scale loss that is appropriate for coarse-to-fine architecture that enhances convergence greatly. In addition, we further improve the results by employing adversarial loss
3、new realistic blurry image dataset
好下面來看看這篇論文的method部分
首先作者改進了residual block（這樣的一個改進好像跟超分的EDSR還是RRDB一模一樣？所以不能叫改進？）
網絡結構如下圖所示
Coarsest level network（最粗層網絡）
The first convolution layer transforms 1/4 resolution, 64 × 64 size image into 64 feature maps. Then, 19 ResBlocks are stacked followed by last convolution layer that transforms the featuremap into input dimension.
At the end of the stage, the coarsest level latent sharp image is generated. Moreover, information from the coarsest level output is delivered to the next stage where finer scale network is.
Since the sharp and blurry patches share low-frequency information, learning suitable feature with upconvolution helps to remove redundancy. the upconvolution feature is concatenated with the finer scale blurry patch as an input.
Finer level network
Finer level networks basically have the same structure as in the coarsest level network. However, the first convolution layer takes the sharp feature from the previous stage as well as its own blurry input image, in a concatenated form.
雖然論文的結果視覺效果沒有《 論文閱讀筆記之——《Deep Stacked Hierarchical Multi-patch Network for Image Deblurring》》這篇博客裏面的好，但是顯然後者的multi-scale是直接用了前者的