基于组反馈融合机制的视频超分辨率模型  

Video super-resolution model based on group feedback fusion mechanism

作　　者：张庆武 迟小羽 朱鉴 陈炳丰[1] 蔡瑞初[1] Zhang Qingwu;Chi Xiaoyu;Zhu Jian;Chen Bingfeng;Cai Ruichu(School of Computer Science&Technology,Guangdong University of Technology,Guangzhou 510006,China;Qingdao Research Institute of Beihang University,Qingdao Shandong 266104,China)

机构地区：[1]广东工业大学计算机学院,广州510006 [2]北京航空航天大学青岛研究院,山东青岛266104

出　　处：《计算机应用研究》2022年第11期3492-3497,共6页Application Research of Computers

基　　金：国家重点研发计划资助项目(2021ZD011150);国家自然科学基金优秀青年基金资助项目(6212200101);国家自然科学基金委员会面上项目(61976052);广东省自然科学基金资助项目(2016A030310342);广东省科技计划资助项目(2016A040403078,2017B010110015,2017B010110007);广州市科技计划资助项目(201604016075,202007040005);广州市珠江科技新星资助项目(201610010101);中国高等教育学会实验室研究专项资助项目(21SYYB17)。

摘　　要：现有的许多视频超分辨率(video super-resolution,VSR)工作都集中在如何有效地对齐相邻帧以更好地融合相邻帧信息,而很少在相邻帧信息融合这一重要步骤上进行研究。针对该问题,提出了基于组反馈融合机制的视频超分辩模型(GFFMVSR)。具体来说,在相邻帧对齐后,将对齐视频序列输入第一重时间注意力模块;然后,将序列分成几个小组,各小组依次通过组内融合模块实现初步融合,不同小组的融合结果经过第二重时间注意力模块;各小组逐组输入反馈融合模块,利用反馈机制反馈融合不同组别的信息;最后将融合结果输出重建。经验证,该模型具有较强的信息融合能力,在客观评价指标和主观视觉效果上都优于现有的模型。Many existing VSR works focus on how to effectively align adjacent frames to better fuse adjacent frame information,and little research has been done on the important step of adjacent frame information fusion.To solve this problem,this paper proposed a video super-resolution model based on group feedback fusion mechanism(GFFMVSR).Specifically,after adjacent frames were aligned,it input the aligned video sequences to the first temporal attention module.Then,it divided the sequence into several groups,and each group achieved preliminary fusion through the intra-group fusion module in turn,the fusion results of different groups went through a second temporal attention module.Next,each group input the feedback fusion module group by group,and used the feedback mechanism to feedback and fuse the information of different groups.Finally,it reconstructed the fusion result output.It verifies that the model has strong information fusion ability,and it is superior to the existing models in both objective evaluation indicators and subjective visual effects.

关 键 词：视频超分辨率 时间注意力 反馈机制 分组融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]