Abstract
新的模型和新的深度学习结构,binary map(rain streak layer + back ground layer)对雨痕定位,来处理haven rain情况(mist,overlapping rain streaks),提出含有contextualized dilated 结构
recurrent rain detection and removal network清除haven rain。
Introduction
问题:
1.雨痕与背景纹理的重叠,导致纹理细节被去除,某些区域被over-smoothing。
2.感受野的大小有限,无法获取spatial contextual information in larger ragions
贡献:
1.region-dependent rain models:binary map,pixel=1表示该为rain streak,pixel=0表示该点为background。
2.检测和去雨一体化的deep network:对rain和non-rain regions 采取不同操作
3.contextualized dilated network:空洞卷积扩大感受野,采用多路平行,卷积核各异的结构
4.recurrent rain detection and removal network。
Region-Dependent Rain Image Model
widetildeS无法用常用分布的描述 O=B+˜S(1)
Joint Rain Streak Detection and Removal
缺乏细节,如何对 F 进行conv得到R?以及后续的F,R conv得到S?