分组mask引导的单幅图像去除雨滴

Removing raindrop from a single image guided by group mask

恢复被雨滴损毁图像是提高自动驾驶或视频监控等机器视觉在自然场景中识别性能的必要预处理任务。该任务的核心技术是对雨图中雨滴的定位和恢复被雨滴覆盖的背景信息。现有基于数据驱动的解决方案是基于雨滴定位采取固定阈值下硬掩码(hard mask)或软掩码(soft mask)引导的深度神经网络去除雨滴。考虑到雨滴形状的多样性、对背景损毁程度的模糊性,该文提出基于分组函数的自适应阈值分割算法(称为分组掩码)引导的雨滴去除算法。首先,分组掩码根据雨滴大小的多样性和雨滴对背景模糊程度的不同自适应提取雨滴信息;然后,将雨滴损毁图像和分组掩码级联作为输入,用对抗损失训练生成对抗网络(GAN)去除雨滴,输出恢复的干净背景图。综合实验可见,文中提出的图像去雨滴算法比现有的算法更具有优越性。

Restoration of rain-damaged images is a necessary preprocessing task to improve the recognition performance of automatic driving or video surveillance in natural scenes.The key technology of the task is to locate the raindrop and recover the background information covered by raindrops.The current data-driven methods is to remove raindrops based on a deep neural network guided by hard mask or soft mask with fixed threshold.Considering the diversity of the shape of raindrops and the fuzziness of background,an adaptive threshold segmentation algorithm based on grouping function(called group mask)is proposed to remove raindrops.According to the diversity of raindrop size and the blurring degree of raindrop to background,the group mask can extract the raindrop information adaptively,and then,the cascade of the image damaged by raindrop and grouping mask are used as input,and the generative adversarial network(GAN)is trained by the adversarial loss,meanwhile,the restoration of the clean background image is outputted from the network.Comprehensive experiments show that the algorithm proposed in this paper has more advantages than the existing algorithms.