基于超图的图片检索和标注技术研究

发布时间：2018-03-03 03:09

本文选题：图片检索　切入点：图片标注　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着社交媒体和移动多媒体的发展,社交图片网站(如Flickr)提供了大量由不同用户进行文本标注的图片。社交图片往往附带多种信息,譬如视觉特征、标签和用户等。一般仅对单一特征进行管理,或使用普通图对语义网络进行表示。然而,单一特征只能表示某方面的相关性,不能用来表示真实的语义关联。由于普通图不能够表示高维关系,故造成了信息缺失,因而需要一个框架来融合多种特征信息和表示高维关系,并能进行有效的检索和管理。不同于基于普通图模型的方法,本文使用超图模型对社交图片进行建模,因为社交图片多种信息间的关系比成对关系更加复杂。基于超图模型,本文给出了一个可扩展的交互式图片检索和标注系统HIRT;该系统利用个性化PageRank度量结点之间的相似度,并使用Top-k查询以支持相似图片检索、关键字图片检索和图片标注等功能。为了保证系统的可扩展性和高效性,本文利用批量技术、并行技术和缓冲技术提出了四种高效方法来计算转移概率矩阵,并将其存储于磁盘上的B+树中,同时开发了并行和近似的个性化PageRank算法以进一步加速Top-k查询。此外,一种基于群体计算的交互式方法也被提出以提升Top-k查询的质量。最后在大规模Flickr数据集上的大量实验评估证实,相较于现有的系统和技术,本文所提出的HIRT系统具有更好的效率和可扩展性。
[Abstract]:With the development of social media and mobile multimedia, social photo sites (such as Flickr) provide a large number of text tagged images by different users. Social pictures often carry a variety of information, such as visual features. Tags, users, etc., usually manage only a single feature, or use a common graph to represent the semantic network. However, a single feature can only represent the correlation of a certain aspect. It can't be used to represent the real semantic association. Because the common graph can't represent the high dimension relation, it causes the lack of information, so it needs a framework to combine the multi-feature information and the high-dimensional relation. Different from the method based on ordinary graph model, this paper uses hypergraph model to model social picture, because the relationship between many kinds of information of social picture is more complex than pairwise relation. This paper presents an extensible interactive picture retrieval and annotation system, which uses personalized PageRank to measure the similarity between nodes, and uses Top-k query to support similar image retrieval. In order to ensure the expansibility and efficiency of the system, this paper proposes four efficient methods to calculate the transfer probability matrix by using batch technology, parallel technology and buffer technology. It is stored in B tree on disk, and parallel and approximate personalized PageRank algorithm is developed to further speed up Top-k query. An interactive method based on group computing has also been proposed to improve the quality of Top-k queries. Finally, a large number of experimental evaluations on large-scale Flickr datasets confirm that, compared with existing systems and techniques, The proposed HIRT system is more efficient and extensible.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】