摘要
为了提升雨天条件下图像成像质量,文中提出了一种兼具位置感知和密度感知的两阶段深度学习去雨网络。在第一阶段,通过多尺度渐近注意子网络来定位不同雨条纹的分布;第二阶段在前一阶段生成的注意力图的引导下,通过多尺度残差子网络将这些特征组合在一起。这两个阶段的子网络共同作用,从而较好地完成了对不同雨条纹的联合检测和去除过程。实验结果表明,提出的方法在合成数据集和真实图像上都能表现出较其他算法更优的去雨性能。
In order to improve the imaging quality under rain conditions,this paper proposes a two-stage deep learning deraining network that is both location-aware and density-aware.In the first stage,the distribution of different rain streaks is localized by a multi-scale asymptotic attention sub-network.In the second stage,these features are combined by a multi-scale residual sub-network guided by the attention map generated in the previous stage.These two stages of sub-networks work together,leads to better completion of the joint detection and removal process of different rain streaks.Experiment results show that the proposed method can show better rain removal performance than other algorithms on both synthetic datasets and real images.
作者
马悦
MA Yue(Shaanxi University of Chinese Medicine,Xianyang 712046,Shaanxi Province,China)
出处
《信息技术》
2021年第10期132-136,143,共6页
Information Technology
关键词
图像去雨
两阶段
多尺度学习
注意力机制
image deraining
two-stage
multi-scale learning
attention mechanism
