基于先验知识的单幅图像雨雾去除方法

单幅图像雨雾去除是计算机视觉研究的热点之一, 现有的去雨方法往往需根据雨的特性恢复图像, 忽视了场景中的雾以及场景深处雨累积的影响。提出了一种新型单幅图像去雨雾方法, 设计了一种基于先验知识的三阶段单幅图像去雨雾框架。首先, 将深度图作为图像透射率的引导, 利用暗通道先验知识对输入图像的低频部分去雾, 然后, 采用残差网络学习高频雨痕特征, 最后, 引入条件生成对抗网络 (conditional generative adversarial network, cGAN), 对图像局部细节进行精细化修复, cGAN 对图像的空间信息和纹理细节更为敏感, 能有效恢复去雨雾过程中丢失或未能填补的细节。分别用高质量的合成雨雾数据集和真实数据集进行实验测试, 结果表明, 本文方法在峰值信噪比和结构相似度上均较基准方法优, 能将雾浓度各异和雨分布变化情况下的图像恢复至细节丰富的干净场景图。

Single image rain and haze removal is an important task in computer vision. Most state-of-the-art methods focus on the features of rain while ignoring the disturbance from haze and rain accumulation. To solve this problem, this paper proposes a novel strategy of rain and haze removal from a single image and develops a three-stage framework. First, only the low-frequency of the input image is dehazed based on dark channel prior, where a depth map is used as guidance of transmission to improve the result; Then the streaks in high-frequency can be learned through a residual network; At last, local details are refined by a conditional generative adversarial network (cGAN). A high-quality synthetic dataset and real photo data are used in testing. Experiment results show that this framework has better performance on rainy image restoration than baseline, by comparing the value of PSNR and SSIM with ground-truth images. Our method is robust in restoring clean scene from images with various haze density or rain distribution.