首页|一种结合增强对齐与多注意力融合的偏振视频超分辨方法

一种结合增强对齐与多注意力融合的偏振视频超分辨方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
在偏振成像探测中,偏振多通道视频序列给数据传输和处理带来挑战,同时探测精度又依赖高分辨率图像,为此超分辨率计算成像就成为偏振成像探测系统的关键技术.从视频运动特征入手,提出一种结合增强对齐与多注意力融合的偏振视频超分辨率重建方法.采用三级级联结构实现特征对齐,每一级特征由上一级特征下采样两倍得到,并采用光流引导的可变形卷积进行特征对齐;在融合阶段结合包括时间注意力、空间注意力、通道注意力在内的多种注意力机制,以充分利用通道间依赖性和时空相关性来提升超分辨率重建效果.所提方法与VSR-DUF、RBPN、EDVR、RLSP、RSDN、RRN、DARN、MSAWN等视频超分辨率方法在0°、45°、90°、135°偏振角视频上进行定性和定量对比实验,该实验表明该算法峰值信噪比和结构相似度指标高于其他算法,并且视频帧的高频纹理细节的修复更加丰富,接近偏振成像系统的原始高清图像帧.
A polarization video super-resolution method combining enhanced alignment and multi attention fusion
In polarization imaging detection,polarization multi-channel video sequence brings challenges to data trans-mission and processing,while detection accuracy depends on high-resolution images.Therefore,super resolution compu-ting imaging becomes the key technology of polarization imaging detection system.A polarization video super-resolution reconstruction method is proposed based on enhanced alignment and multi-attention fusion.A three-level cascade structure is adopted to achieve feature alignment.Each level of features is obtained by twice downsampling of the upper level fea-tures,and feature alignment is performed by using deformable convolution guided by optical flow;In the fusion stage,multiple attention mechanisms,including temporal attention,spatial attention and channel attention,are combined to make full use of inter channel dependency and spatiotemporal correlation to improve the effect of super-resolution recon-struction.qualitative and quantitative comparative experiments were conducted between the method proposed and other video super-resolution methods,such as VSR-DUF,RBPN,EDVR,RLSP,RSDN,RRN,DARN,MSAWN,on 0°,45°,90°,135°polarized videos.Experiments show that the method outperforms other algorithms in terms of peak signal-to-noise ratio and structural similarity metrics,and that the high-frequency texture detail restoration of video frames is richer and closer to the original HD image frames from the polarization imaging system.

imaging detectionvideo super-resolutiondeep learningpolarization videosattention mechanism

徐国明、金苗、梁栋、马健、王峰、李毅

展开 >

安徽大学互联网学院,安徽合肥 230039

安徽大学农业生态大数据分析与应用技术国家地方联合工程研究中心,安徽合肥 230601

偏振光成像探测技术安徽省重点实验室,安徽合肥 230031

安徽文达信息工程学院智能技术研究所,安徽合肥 231201

展开 >

成像探测 视频超分辨率 深度学习 偏振视频 注意力机制

国家自然科学基金国家自然科学基金安徽省重大专项安徽省自然科学基金安徽省自然科学基金陆军装备部预研子课题(十三五)安徽省高等学校自然科学研究重点项目

6190611862273001202003A060200161908085MF2082108085MF230KJ2019A0906

2024

光学技术
北京兵工学会 北京理工大学 中国北方光电工业总公司

光学技术

CSTPCD北大核心
影响因子：0.441
ISSN：1002-1582
年,卷(期)：2024.50(3)
  • 25