没有合适的资源?快使用搜索试试~ 我知道了~
温馨提示
内容概要:本文提出了一种用于视频超分辨率(VSR)的一阶段时序可变形对齐网络(TDAN)。传统方法通常依赖于光流来对齐参考帧和支持帧,这可能导致图像伪影并影响最终重建的高分辨率(HR)视频帧质量。TDAN则在特征级别上自适应地对齐参考帧和支持帧,无需计算光流。通过从参考帧和支持帧中提取特征,动态预测卷积核的偏移量,TDAN可以将支持帧转换为与参考帧对齐的状态。实验结果表明,TDAN在多个基准数据集上优于现有方法,特别是在处理复杂运动场景时表现出色。 适用人群:从事计算机视觉、深度学习研究的科研人员,尤其是专注于视频处理和超分辨率领域的研究人员和技术开发者。 使用场景及目标:①解决视频超分辨率任务中因相机或物体运动导致的帧间不对齐问题;②提高视频超分辨率模型的性能,特别是在处理复杂运动场景时;③减少传统基于光流的方法中可能出现的图像伪影。 其他说明:TDAN的设计具有较强的泛化能力,不仅适用于视频超分辨率任务,还可以扩展到其他视频修复任务如视频去噪、视频去模糊和视频帧插值等。此外,TDAN采用自监督训练方式,无需额外标注数据,降低了训练难度。实验部分详细对比了TDAN与其他先进方法在不同退化条件下的表现,并通过消融研究验证了各模块的有效性。未来工作将集中在构建更大规模的高分辨率视频数据集,以进一步提升TDAN的性能。
资源推荐
资源详情
资源评论
























TDAN: Temporally Deformable Alignment Network for Video Super-Resolution
Yapeng Tian
1
, Yulun Zhang
2
, Yun Fu
2,3
, and Chenliang Xu
1
1
Department of Computer Science, University of Rochester
2
Department of Electrical and Computer Engineering, Northeastern University
3
College of Computer and Information Science, Northeastern University
Abstract
Video super-resolution (VSR) aims to restore a photo-
realistic high-resolution (HR) video frame from both its
corresponding low-resolution (LR) frame (reference frame)
and multiple neighboring frames (supporting frames). Due
to varying motion of cameras or objects, the reference frame
and each support frame are not aligned. Therefore, tempo-
ral alignment is a challenging yet important problem for
VSR. Previous VSR methods usually utilize optical flow be-
tween the reference frame and each supporting frame to
wrap the supporting frame for temporal alignment. There-
fore, the performance of these image-level wrapping-based
models will highly depend on the prediction accuracy of
optical flow, and inaccurate optical flow will lead to arti-
facts in the wrapped supporting frames, which also will be
propagated into the reconstructed HR video frame. To over-
come the limitation, in this paper, we propose a temporal
deformable alignment network (TDAN) to adaptively align
the reference frame and each supporting frame at the fea-
ture level without computing optical flow. The TDAN uses
features from both the reference frame and each supporting
frame to dynamically predict offsets of sampling convolu-
tion kernels. By using the corresponding kernels, TDAN
transforms supporting frames to align with the reference
frame. To predict the HR video frame, a reconstruction net-
work taking aligned frames and the reference frame is uti-
lized. Experimental results demonstrate the effectiveness of
the proposed TDAN-based VSR model.
1. Introduction
The goal of video super-resolution (VSR) is to recon-
struct a high-resolution (HR) video frame from its cor-
responding low-resolution (LR) video frame (reference
frame) and multiple neighboring LR video frames (support-
ing frames). HR video frames contain more image details
Walk
HR Bicubic
DUF [9] Ours
Figure 1. VSR results for a frame in the walk sequence. We can
find that our method can restore more accurate image structures
and details than the recent DUF network.
and are more preferred to humans. Therefore, the VSR tech-
nique is desired in many real applications, such as video
surveillance and high-definition television (HDTV).
To super-resolve the LR reference frame, VSR will ex-
ploit both the LR reference frame and multiple LR support-
ing frames. However, the LR reference frame and each sup-
porting frame are likely not fully aligned due to the motion
of camera or objects. Thus, a vital issue in VSR is how to
align the supporting frames with the reference frame.
Previous methods [3, 17, 27, 32, 22] usually exploit op-
tical flow to predict motion fields between the reference
frame and supporting frames, then wrap the supporting
frames using their corresponding motion fields. Therefore,
the optical flow prediction is crucial for these approaches,
and any errors in the flow computation or the image-level
wrapping operation may introduce artifacts around image
structures in the aligned supporting frames.
To alleviate the above issues, we propose a temporally
deformable alignment network (TDAN) in this paper that
performs one-stage temporal alignment without using opti-
cal flow. Unlike previous optical flow-based VSR methods,
our approach can adaptively align the reference frame and
1
arXiv:1812.02898v1 [cs.CV] 7 Dec 2018
资源评论


码流怪侠

- 粉丝: 4w+
上传资源 快速赚钱
我的内容管理 展开
我的资源 快来上传第一个资源
我的收益
登录查看自己的收益我的积分 登录查看自己的积分
我的C币 登录后查看C币余额
我的收藏
我的下载
下载帮助


最新资源
- 基于AT89S52单片机的数字温度计设计.doc
- 2023年初级通信工程师考试试题及答案.doc
- 软件项目立项评审报告模版.doc
- 项目管理-系列2-3:配置管理实用手册vss.docx
- 家里怎么安装网络.pdf
- 工业机器人遥操作控制系统设计嵌入式伺服控制器设计样本.doc
- 精品课程网络教学资源和硬件环境.pdf
- 发电系统安全控制.doc
- 网络诈骗常见手段及防范对策.doc
- 计算机组成原理习题)(1)解析.doc
- 电子商务和税收筹划的关系是什么.doc
- 再生资源回收体系建设项目管理手册.doc
- 国家网络安全法考试试题与答案.pdf
- 软件测试实验报告.doc
- 数据库课程设计宾馆客房信息管理系统.doc
- 网络推广最高效率的方案.doc
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈



安全验证
文档复制为VIP权益,开通VIP直接复制
