基于局部关键点特征的视频近重复检测算法研究

发布时间：2018-05-28 17:06

本文选题：近重复检测 + 视频检索　；参考：《复旦大学》2012年硕士论文

【摘要】：视频是信息最丰富的载体,随着网络技术的突飞猛进,人们的交互和信息传递方式也从传统的书信和纸张慢慢迁移到图片、语音和视频中。随着数字视频技术和网络传输技术的发展,加上现代社交网络的日新月异,网络上数字视频的数据量呈现指数级增长。因此如何对这些网络上的海量视频进行检索、管理和版权保护就成了现在一个比较大的问题。其中一个比较突出的问题就是这些网络视频中存在着大量的近重复(near-duplicate)的视频。这些大量重复的视频不仅浪费了比较宝贵的存储空间,而且对于后期的检索和维护也带来诸多不便。对于视频的近重复检索不仅具有深远的学术价值,而且能带来可观的经济价值,可以提高网络视频的管理、检索和浏览效率,还可以提供新闻图像的跟踪、版权保护,甚至现在商业的搜索引擎都已经提供基于图片和视频的近重复检索技术的图像和视频搜索。本论文首先会介绍下现有框架各个部分算法的发展历史,然后会另辟蹊径,把这些算法组合成一个新的框架,并应用其来解决视频的近重复检索问题,最后把本文的算法框架和现有的算法做个比较,并阐明各自的优缺点。本论文的主要工作如下：首先简单介绍一下现有视频近重复检测理论的基本框架、研究现状、研究方向等等,并从视频近重复检测理论的研究框架中截取部分进行比较详细的阐述；比如视频的帧、关键帧、镜头和场景分割等等。挑选出比较经典的算法进行重点介绍,并说明其存在的缺陷。然后对于采用的关键点特征(key point feature)做一个重点描述。包括其算法框架、发展的历史、有哪些经典的关键点特征的提取算法,以及其存在的缺陷和待改进和提高的地方。局部特征是孤立的特征,其拥有良好的抗干扰能力,但是其失去了特征与特征之间的全局性,所以后来的研究者有很多是针对局部特征的组织全局联系的。论文的第三部分会对这一块的研究工作做个简单的回顾。最后利用局部关键点特征和新的特征构造和组织框架,提出了一种新的基于局部关键点特征的视频近重复检测算法。并将现有的算法跟以往的经典算法做了比较。
[Abstract]:Video is the most informative carrier. With the rapid development of network technology, people's interaction and information transfer from traditional letters and paper to pictures, voice and video. With the development of digital video technology and network transmission technology, and the rapid development of modern social networks, the amount of digital video data on the network has increased exponentially. Therefore, how to retrieve, manage and protect the massive video on these networks has become a big problem. One of the more prominent problems is the existence of a large number of near-repeated near-dual video in these network videos. These repeated videos not only waste valuable storage space, but also bring much inconvenience to the later retrieval and maintenance. The near-repeated retrieval of video not only has profound academic value, but also brings considerable economic value. It can improve the efficiency of network video management, retrieval and browsing, and also provide news image tracking and copyright protection. Even commercial search engines now offer image and video search based on near-repeated retrieval of images and videos. This paper first introduces the development history of each part of the existing framework algorithms, and then will find a new way to combine these algorithms into a new framework, and use them to solve the problem of near-repeated video retrieval. Finally, the algorithm framework and existing algorithms are compared, and their advantages and disadvantages are clarified. The main work of this thesis is as follows: firstly, the basic frame, research status and research direction of the existing video near-repetition detection theory are briefly introduced. And from the research framework of the theory of video near-repetition detection, the interception part is described in detail, such as the frame of video, key frame, shot and scene segmentation and so on. Select a more classical algorithm to focus on the introduction, and explain its shortcomings. Then the key point features are described. Including its algorithm framework, history of development, what are the classic key point feature extraction algorithm, as well as its shortcomings and areas for improvement and improvement. Local feature is an isolated feature, which has good anti-jamming ability, but it loses the global character between feature and feature. The third part of this paper will do a simple review of this piece of research work. Finally, using the local key point feature and the new feature structure and organization framework, a new video near-repetition detection algorithm based on the local key point feature is proposed. The existing algorithms are compared with the previous classical algorithms.
【学位授予单位】：复旦大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41

【相似文献】