【AAAI2021】EfficientDeRain: Learning Pixel-wise Dilation Filtering for High-Efficiency Single-Image Deraining
对于图像去雨任务，作者指出当前方法的问题：1）the rain model assumptions adopted by many existing methods may be limited, which do not well represent and reflect the real-world rain patterns. 2) Most existing methods are computationally expensive, either requiring complex iterative optimization to find the optimal solution, or constructing the deraining framework in multiple stages with recurrent or progressive refinement steps involved.
为此，作者的目标是构建高效并泛化性强的去雨模型，具有两个特点：1）the proposed method is model-free, which makes no assumptions of how the rain is generated. 2) the proposed method follows a single-stage and does not require either iterative optimization or progressive refinement.
该模型的总体架构如下图所示：
Kernel prediction network: 指的是上图中 U-Net 结构的网络，作者指出， the network can predict spatial-variant kernels that adapt to the rain streaks with different thickness and strength while preserving the object boundary. 如图中（c）部分所示，从R1到R6，雨的强度逐渐减弱，但是网络仍然可以较好的去除雨条。
Multi-dilated image filtering and fusion： 为了更好的适应多尺度的雨条，作者又应用了 dilated conv 。在论文中使用了4个尺度的卷积，这样就得到 4 张去雨后的图像，最后又使用一个3X3的卷积层将4张图融合得到最终的结果。
RainMix : 为了降低合成雨图和真实雨图间的 gap，作者又设计了 RainMix 算法，随机从雨图中采样，构建新样本。个人感觉这个过程有些像AugMix 。
