SIA OpenIR  > 机器人学研究室
基于在线低秩表达的视频恢复算法研究
Alternative TitleVideo Restoration Based on Online Low-rank Representation Learning
高振远
Department机器人学研究室
Thesis Advisor韩志
Keyword在线鲁棒主成分分析 低秩稀疏分解 视频恢复 非局部先验 稀疏噪声
Pages73页
Degree Discipline模式识别与智能系统
Degree Name硕士
2020-05-26
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract近年来,在线低秩表达方法由于其优异的低秩结构挖掘能力、高运行与存储效率,并且可以建模动态变化数据,现已成为当下研究的热点。本论文针对具有共享性或者重复性特质的视频数据,基于在线低秩表达模型与算法。从视频数据的表达形式以及恢复算法方面入手,提出了具有创新性的模型算法。论文的主要内容和创新点如下(1)现存在线鲁棒主成分分析模型仅仅只能建模视频的全局时间冗余性,因此不能很好表达具有动态移动物体的视频中的低秩结构,而且会导致前景和噪声同时被移除的问题。针对上述问题,提出了非局部在线鲁棒主成分分析模型,模型利用视频中基于图像块的非局部低秩特性,即数据中的时空冗余性,来准确表达视频中的低秩结构。为了求解非局部在线鲁棒主成分分析模型,首先提出了基于高斯聚类的非局部方法用于图像块分组,解决了现存非局部方法不能直接应用至在线鲁棒主成分分析的问题。同时,为了更加准确地估计子空间并利用样本的重要性信息,提出了基于样本重要性的加权鲁棒主成分分析模型和在线加权投影。经过充分的实验验证了所提算法在视频恢复应用中的有效性和优越性。(2)针对非局部在线鲁棒主成分分析模型对于具有复杂场景变化的视频表达及恢复能力不足的问题。提出了基于层级子空间更新的非局部在线鲁棒主成分分析模型,从而可以同时达到高效的视频在线低秩表达和恢复的目的。为了提升模型在应对复杂视频场景下的表达能力,通过显式建模子空间的变化,提出三层级子空间更新机制:加方向子空间更新、重新估计子空间和子空间初始化机制。为了达到实际场景对运行效率的要求,提出算法的加速版本并利用并行计算机制进一步对算法加速。数值实验结果表明,该模型能够有效完成复杂变化视频低秩表达与视频恢复的任务。实验结果显示了模型的有效性和鲁棒性,同时,所提算法具有更高的效率与纹理保持能力。(3)现有的视频恢复方法仅仅利用一种尺度的关联性质,从而限制了算法恢复性能,通过考虑这两种尺度的低秩性质,提出了基于整体关联性和非局部关联性的视频恢复算法。首先,利用视频帧的整体关联性把被噪声污染的视频分解为整体低秩成分和稀疏余项成分。然后,对于余项视频部分其相邻帧存在非局部关联性,利用基于k维树的非局部技术组成低秩图像块组,并通过低秩分解模型去除图像块噪声。最后,整合整体低秩部分与处理后的余项部分,从而得到准确的视频恢复结果。另外,采用在线模型来达到动态分层的目的。对比实验验证了所提算法在视频脉冲噪声去除问题上的优异性能,尤其在大噪声情况下具有更强的鲁棒性。
Other AbstractIn recent years, online low-rank representation methods have become a hot research topic due to their advantages such as high efficiency, storage savings, and the ability to model dynamically changing data. In this dissertation, by exploiting the shared or repetitive characteristics in image and video data, the main research works focus on the following two low-rank theory related issues: video data online low-rank representation learning, and the recovery algorithms. Some novel models and algorithms are proposed based on the online low-rank representation. The main contributions of this dissertation are summarized as follows: (1) For dealing with the problem that existing online low-rank representation can only model the global temporal redundancy of the video, which can not well deal with the videos dynamically moving objects and will cause the foreground and noise to be removed at the same time. Non-local online robust principal component analysis model is proposed, which can utilize the non-local low-rank characteristics of image patches in the video and accurately represent the low-rank structure in the video. In order to solve the non-local online robust principal component analysis model, a novel non-local method based on Gaussian clustering is first proposed for image patches grouping, which can solve the problem that existing non-local methods cannot be directly applied to online robust principal component analysis. At the same time, in order to more accurately estimate the subspace and use the importance information of samples, a weighted robust principal component analysis model based on sample importance and online weighted projection are proposed. Numerical experiments validated the effectiveness and superiority of the proposed algorithm in video recovery applications . (2) To solve the problem that the non-local online robust principal component analysis model has insufficient ability to represent and restore videos with complex scene changes. A non-local online robust principal component analysis model based on hierarchical subspace update scheme is proposed, which can simultaneously achieve the goal of efficient online low-rank video representation and recovery. In order to improve the representation ability of the model in dealing with complex video scenes, by explicitly modeling the changes of subspaces, a three-level subspace update mechanism is proposed: the addition of subspace updates, re-estimation of subspace and subspace initialization mechanism. In order to meet the requirements of running efficiency of real application, an accelerated version of the algorithm is proposed and the parallel computing technique is used to further accelerate the algorithm. Numerical experimental results show that the model can effectively complete the tasks of low-rank representation and video restoration of complex changing videos. The experimental results demonstrates the effectiveness and robustness of the model. At the same time, the proposed algorithm has higher efficiency and texture preservation ability. (3) Concerning the problem that existing video restoration methods only utilize one of these correlation property, which limits the performance of video restoration algorithms. By considering these two low rank properties, a new video restoration algorithm based on global correlation and non-local correlation of video data was proposed. Firstly, the long-term global correlation of video frames was used to decompose the video corrupted by noise into global low-rank components and sparse residuals. Secondly, for residual part, there is non-local correlation between adjacent frames, non-local technique based on k-dimensional tree was utilized to form a group of similar patches, then low-rank decomposition model was used to process a group of image patches so that noise can be removed to obtain a clean image patch structure. Finally, the global low-rank part was added to the processed residual part to obtain a clean image estimate. In addition, the online model is used to achieve the purpose of dynamic global low-rank decomposition. The comparison experiment verifies the excellent performance of the proposed algorithm on the problem of video impulse noise removal, especially in high noise scenario.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/27127
Collection机器人学研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
高振远. 基于在线低秩表达的视频恢复算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2020.
Files in This Item:
File Name/Size DocType Version Access License
基于在线低秩表达的视频恢复算法研究.pd(2628KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[高振远]'s Articles
Baidu academic
Similar articles in Baidu academic
[高振远]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[高振远]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.