基于改进的残差网络的单幅图像去雨方法

A Single Image Derain Method Based on Improved Residual Network

摘要: 雨水是室外比较常见的天气现象。在雨水条件下，光的散射和能见度下降，导致计算机视觉算法会受到不同程度的干扰。为了进一步降低雨水给计算机视觉算法带来的影响，本文提出一种基于改进的残差网络的单幅图像去雨方法。该方法结合了图像处理传统方法和深度学习的方法，首先通过传统方法对有雨图像进行稀疏化，将稀疏化后的图像作为输入，由残差网络学习稀疏化后的图像特征，接着对网络进行训练，最终的网络可以有效的去除雨线。实验结果表明：本文提出的方法可以有效地去除雨线。同时，残差网络学习稀疏化的图像特征，可以大大提升网络的训练效率。

Abstract: Rain is a relatively common weather phenomenon outside. Under the condition of rain, the scattering and visibility of light decrease, which leads to the interference of computer vision algorithm to different degrees. In order to further reduce the influence of rain on computer vision algorithm, this paper proposes a single image derain method based on improved residual network. This method combines the traditional image processing method and the deep learning method. Firstly, the image with rain is sparsely processed by the traditional method, and the sparse image is taken as the input. The image features after sparsity, are learned by the residual network, and then the network is trained, so that the final network can effectively remove rain lines. The experimental results show that the proposed method can effectively remove the rain line. At the same time, the residual network learning sparse image features can greatly improve the training efficiency of the network.