【作者】 徐涛；

【导师】 于俊清；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2013， 硕士

【摘要】 随着多媒体与互联网技术的高速发展，以图像和视频为代表的多媒体信息呈爆炸性增长。图像特征多种多样，并且大多数特征向量可以达到上百维，甚至上千维，因此实现高维特征向量的存储与索引成为实现基于内容的海量视频检索的关键技术。面对图像检索中检索实时性的需求和维度灾难，如何利用足球视频的特征，设计高效的索引方法，实现足球视频图像的高效检索是摆在研究者面前的难题。在分析足球视频的特点的基础上，对特征生成的参数进行了调整，设计了结合非对称距离计算、残差量化编码和倒排索引的索引方法多特征分层索引，以及基于多特征的候选集排序算法。依据足球视频图片库，选取了SA-VLAD、BOC和镜头类型三种特征作为图像的描述信息，并通过实验选取了合适的参数。针对原始残差量化倒排索引只能检索单特征的问题，提出了多个特征分层索引。根据镜头类型分割倒排链表，实现了基于多特征的分层索引机制，加快了候选集过滤速度，提高了查询准确率。在候选集排序阶段，根据镜头类型，设计了基于多特征的相似度计算方法。实验结果表明，基于多特征的分层索引提高了查询性能。但是索引占用空间大的问题还未得到很到的解决。如何结合语义特征，进一步提高准确率也将是下一步的研究重点。更多还原

【Abstract】 Multimedia information especially video is growing explosively with the fastdevelopment of Internet and multimedia technology. Because of the variety of imagefeatures, which could reach several hundred dimension, or even thousands of dimension,Storing and indexing the high-dimensional feature vectors come to be the keytechnologies of the content-based video retrieval. To resolve the problem of curse ofdimension, taking advantage of the characteristics of soccer video and design efficientindexing method has become a great challenge for all the researchers.The residual quantization mechanism, which combines the asymmetric distance, theresidual quantization and inverted indexing, is improved after analyzing the characteristicsof soccer video. For soccer video, SA-VLAD, BOC and shot type are selected fordescribing the information of images. The appropriate parameters of SA-VLAD, BOC arechosen through experiences. In order to deal with the problem that the raw residualquantization indexing mechanism can only deal with the single feature, the indexingalgorithm based on multiple features is proposed. The algorithm could speed up filteringcandidates. When ranking the candidate set, a new score formula is proposed, whichwould make use of the type of soccer video shoot fully.Experiments show that the multiple-features presentation and the indexing algorithmbased on multiple features could reach higher precious rate. However, there are still manythings to solve the problem that the index occupies much larger space. The combination ofsemantic layer features and indexing algorithm need further analysis.更多还原