Page 311 - 《软件学报》2021年第10期

软件学报 ISSN 1000-9825, CODEN RUXUEW E-mail: jos@iscas.ac.cn
Journal of Software,2021,32(10):32833292 [doi: 10.13328/j.cnki.jos.006018] http://www.jos.org.cn
©中国科学院软件研究所版权所有. Tel: +86-10-62562563


基于双注意力残差循环单幅图像去雨集成网络 

张学锋, 李金晶

(安徽工业大学 计算机科学与技术学院,安徽 马鞍山 243000)
通讯作者: 李金晶, E-mail: 2424899546@qq.com

摘 要: 降雨会严重降低拍摄图像质量和影响户外视觉任务.由于不同图像中,雨的形状、方向和密度不同,导致单
幅图像去雨是一项困难的任务.提出一种新的基于双注意力的残差循环单幅图像去雨集成网络(简称 RDARENet).
在网络中,因为上下文的信息对于去除雨痕十分重要,所以首先采用多尺度的扩张卷积网络去获得更大的感受野.雨
痕信息可以认为是多个雨层特征的叠加,为了更好地提取雨痕的特征和恢复背景图层信息,运用了通道和空间注意
力机制的残差网络.通道注意力能够反映不同雨层的权重,而空间注意力则通过相邻空间特征之间的关系增强区域
的表征.随着网络的加深,防止低层信息的丢失,采用级联的残差网络和长短时间记忆网络,将低层特征信息传递到
高层中去,逐阶段地去除雨痕.在网络的输出部分,采用集成学习的方式,将每个阶段的输出结果通过门控网络加权
相加,得到最终的无雨图像.实验结果表明,去雨和恢复纹理细节的效果都得到较大提升.
关键词: 单幅图像去雨;双注意力机制;残差网络;门控网络
中图法分类号: TP391

中文引用格式: 张学锋,李金晶.基于双注意力残差循环单幅图像去雨集成网络.软件学报,2021,32(10):32833292. http://
www.jos.org.cn/1000-9825/6018.htm
英文引用格式: Zhang XF, Li JJ. Single image de-raining using a recurrent dual-attention-residual ensemble network. Ruan Jian
Xue Bao/Journal of Software, 2021,32(10):32833292 (in Chinese). http://www.jos.org.cn/1000-9825/6018.htm

Single Image De-raining Using a Recurrent Dual-attention-residual Ensemble Network

ZHANG Xue-Feng, LI Jin-Jing
(School of Computer Science and Technology, Anhui University of Technology, Maanshan 243000, China)
Abstract: Rain streaks can severely degrade the quality of captured images and affect outdoor vision. However, due to non-uniform in
shape, direction, and density of rain in different images, it is a difficult task to remove rain from a single image. This study proposes a
single image de-raining using an ensemble recurrent dual-attention-residual network, called RDARENet. In the network, as contextual
information is very important for the process of rain removal, a multi-scale dilated convolution network is firstly adopted to acquire large
receptive field. Rain streaks can be regarded as the accumulation of multiple rain streaks layers, the residual of the channel attention and
spatial attention mechanisms are used to extract the features of the rain streaks and restore the background layer information. The channel
attention can assign different weights to rain streaks layers, and the spatial attention enhances the representation of the area through the
relationship between adjacent spatial features. With the deepening of the network, to prevent the loss of low-level information, a cascaded
residual network and a long-term memory network are used to transfer low-level feature information to the high-level and remove rain
streaks stage by stage. In the output of the network, ensemble learning method is adopted to weight the output of each stage through the
gated network, and add to get the clean image. Extensive experiments demonstrate that the effect of removing rain and restoring texture
details is greatly improved.
Key words: single image de-raining; dual-attentionmechanism; residual network; gated network

 基金项目: 安徽省教育厅重大课题(KJ2017ZD05); 安徽高校协同创新项目(GXXT-2019-008)
Foundation item: Major Issues of the Anhui Provincial Department of Education (KJ2017ZD05); Anhui University Collaborative
Innovation Project (GXXT-2019-008)
收稿时间: 2019-08-11; 修改时间: 2019-11-24, 2020-01-04; 采用时间: 2020-02-07