一种由粗到细的图像逐级去雨算法

A gradual image deraining algorithm from coarse to fine

雨水会影响到图像的清晰程度和细节信息,从而对后续视觉任务处理造成不便,进而导致许多计算机视觉系统瘫痪。现有的图像去雨算法虽能去除雨,但是这些算法通常假设雨的排列稀疏,并且在下落的方向和形状上有相似的特征,从而只能在特定的情况下工作。为了解决这一问题,提出了一种由粗到细的深度学习图像去雨算法,该方法通过多阶段逐级的处理方式来实现雨痕的反复处理,直到雨痕被清除干净。此外,由于空间背景信息对图像去雨的影响,还开发了一种密集连接的膨胀卷积块来处理不同大小的雨带。定性和定量结果分析表明,提出的方法在Rain100L和Rain100H两个基准数据集上表现出了超越其他算法更先进的性能,在PSNR和SSIM指标上比第二名的算法提升了4%左右。

Rain will affect the clarity and details of the image,which will cause inconvenience to the follow⁃up visual task processing and lead to the paralysis of many computer vision systems.Although the existing image rain removal algorithms can remove rain,they usually assume that the arrangement of rain is sparse,and they have similar characteristics in the direction and shape of falling,so they can only work under certain conditions.In order to solve this problem,a deep learning image rain removal algorithm from coarse to fine is proposed,which can deal with rain marks repeatedly through multi⁃stage and step⁃by⁃step processing until the rain marks are removed.In addition,for the influence of spatial background information on image rain removal,a densely connected expansion convolution block is developed to deal with different sizes of rain bands.The qualitative and quantitative results show that the proposed method outperforms other algorithms in Rain100L and Rain100H benchmark datasets,and the algorithm comparing with the second place in PSNR and SSIM index is improved by about 4%.