论文阅读笔记之——《CRRN: Multi-Scale Guided Concurrent Reflection Removal Network》

In this paper, we propose the Concurrent Reflection Removal Network (CRRN) to tackle this problem in a unified framework.

Reflections observed in front of the glass significantly degrade the visibility of the scene behind the glass. By obstructing （阻塞）, deforming or blurring the background scene, reflections cause many computer vision systems likely to fail. Reflection （反射） removal aims at enhancing the visibility of the background scene while removing the reflections.

contributions

1、we combine the two separate stages (gradient inference and image inference) in one unified mechanism to remove reflections concurrently. 梯度推理和图像推理

2、propose a multi-scale guided learning network to better preserve the background details, where the background reconstruction in the image inference network is closely guided by the associated gradient features （梯度特征） in the gradient inference network.

3、We design a perceptually motivated loss function, which helps suppress the blurry artifacts （抑制模糊伪影） introduced by the pixel-wise loss functions, and generate better results.

作者构建了一个数据集Reflection Image Dataset （RID），如下所示

Concurrent Reflection Removal Network (CRRN) with a multi-task learning strategy

For the gradient inference network (GiN), the input is a 4-channel tensor, which is the combination of the input mixture image and its corresponding gradients（输入混合图像及其相应的梯度）;it estimates ∇B （the gradient information） to extract the image gradient information from multiple scales and guide the whole image reconstruction process.

结构如下图所示

Multi-scale representations have shown to be effective in the extraction of image details for reflection removal [28] and other inverse imaging problems. To make full use of the multi-scale information of the decoder part in GiN, the output of each
transposed convolutional layers of GiN is concatenated with the output of transposed convolutional layers in IiN at the
 same level, which is illustrated in Figure 1.

在去摩尔纹和这篇论文中，都提及到，multi-scale有利于恢复图像。在不同的尺度上去除

从下面的效果看来，一般般