SIA OpenIR  > 机器人学研究室
基于统计分析的视频去噪与表达建模
Alternative TitleVideo denoising and representation based on statistical analysis
沈贵萍1,2
Department机器人学研究室
Thesis Advisor唐延东 ; 韩志
Keyword噪声建模 前景建模 低秩分解 稀疏表达 主动轨迹
Pages93页
Degree Discipline模式识别与智能系统
Degree Name博士
2018-11-27
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本文以视频为研究对象,应用分析统计理论,对视频去噪与表达建模的理论方法开展研究。视频的噪声成分,有些是由应用环境和硬件技术决定的,是不可避免的。噪声会给计算机视觉的应用(如目标检测、跟踪、行为识别)带来很多不利的影响,同时降低了其在工业、医用及军事等各应用领域的性能和效率。此外视频的巨大数据内存量给数据的运输和传送带来不便,而且视频帧在记录有用的信息时,经常会遇到运动模糊、遮挡等原因造成目标结构信息的缺失。这对于后期的分析和处理带来困扰。这些问题的解决,将会促进计算机视觉应用的发展,同时为计算机视觉相关算法的研究提供一新的指导方向。本论文主要研究内容包括视频去噪和表达建模。针对视频的复杂混合噪声分布提出一个鲁棒性好的噪声模型。并借鉴视频去噪模型思想,提出一个新的视频背景重建模型,对前景建模,利用张量分解技术恢复视频背景部分。最后我们从视觉中层应用目标跟踪出发,对视频数据利用Gabor小波稀疏表达,提出了基于主动轨迹特征的视频表达建模方法。可以为前景目标生成一个可描画性和可跟踪性的草图描述。实验验证了主动轨迹特征模型对于视觉的各层应用都有很好的贡献。论文中主要创新点和贡献如下:1. 实际视频噪声是由多种复杂统计分布未知的数据组成。我们发现,目前的主流噪声模型基于弱假设等先验知识,对噪声具有选择性,对混合复杂的噪声鲁棒性差。基于此,我们将噪声分布统分为稠密型和稀疏型,统一建模,使其具有有效的泛化能力。2. 针对背景重建受光线变化、背景噪声、前背景纹理相似等影响的问题,以本文提出的噪声模型为基础,利用前景清晰的光滑结构特征为先验知识,引入马尔科夫随机场的上下文语义约束,结合张量技术表达潜在空间结构的特征,提出了一新的背景重建模型,该前景模型具有实际的物理意义解释。3. 从视频数据的稀疏表达出发,利用Gabor小波元素学习出一组线性表示的基元字典,结合主动曲线的多层语义结构,将其拓展到时间维度,得到基于主动轨迹特征模型的视频结构特征表达。并开发一个新的基于低秩特性的外观评分系统,使得该模型表达更精确和简洁。且该模型能够很好的描述视频部件的结构信息,并对视频中存在遮挡等信息缺失的情况,亦能恢复出其结构信息。4. 利用主动轨迹特征可进一步生成视频的草图表达,并可分别用在视频的低、中和高层的相关应用。草图模型可用在中层的目标跟踪、检测等方法;还可以合成为视频部件的共享草图,应用在视频高层行为识别上。
Other AbstractThe research works in this paper contain Video noise modeling, denoising and representation. Some of the noise components of video are determined by the application environment and hardware technology, which are inevitable. Video noise can affects some applications of computer vision (such as target detection, tracking and behavior recognition) and bring a lot of adverse effects. It often causes bad performance of computer vision systems and fails in some tasks. For video, its huge data memory brings inconvenience to the data transmission. In addition, the missing of object structure information caused by occlusion brings confusion to the later analysis and processing. The solution of these problems will promote the development of computer vision application. The main research content of this thesis includes video denoising and representation. It proposed a robust noise model for the complex mixed noise distribution of video. A new video background reconstruction model is proposed referring to video denoising model. The background of video is constructed by tensor decomposition. Finally, based on object tracking applied in middle level vision, we take sparse expressed of video data by Gabor wavelet. We proposed a video representation method by the active trace, and the experiments verified that the active trajectory feature model had a good contribution to the application of each layer of vision. The main innovations and contribution in this thesis are as follows: 1. For video noise,the actual video noise is composed of a variety of complex unknown statistical distribution data. We found that the current mainstream noise models are based on assuming or prior knowledge. Since existing noise models have selectivity to the noise, the noise denoising effect on mixed complex robustness is poor. Based on these experiences, we separate the noise distribution into dense and sparse statistical types, and model the noise uniformly, which has the generalization ability at present. 2. To handle the influence of the illumination changing, background noise and the texture similar on background reconstruction, based on the proposed noise model in forward, we propose a novel foreground model of video. In the model, we take the foreground as the noise part, use its clear prior knowledge of continuous structure features by introducing the context semantic constraint of markov random field. We also combine the tensor technology to exploiting its potential spatial structure features, and the foreground model has practical physical significance. 3. For the sparse expression of video data, we learn a set of linear primitive dictionary with a set of the Gabor wavelet elements. As multi-layer semantic structure of active curve in the practical application. We get a video structural feature expression based on the active trace model by extending the active curve to the time dimension. Besides, we developed a new score system based on appearance information by low rank award, which makes the model more precise and concise. It can describe the structure information of video components very well, and it can also recover the structure information of video, which has the absence of information such as occlusion. 4. The active trace feature can generates the video sketch representation. It also can be used in video low, middle and high level, respectively. The sketch model can be used to the application of object tracking, detection and other methods; it can also be synthesized as a shared sketch of video components and applied it to video high-level behavior recognition.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/23651
Collection机器人学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
沈贵萍. 基于统计分析的视频去噪与表达建模[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
Files in This Item:
File Name/Size DocType Version Access License
基于统计分析的视频去噪与表达建模.pdf(14089KB)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
[沈贵萍]'s Articles
Baidu academic
Similar articles in Baidu academic
[沈贵萍]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[沈贵萍]'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.