【作者】 罗锋；

【导师】 李胜利；

【作者基本信息】 华中科技大学 ， 计算机软件与理论， 2008， 硕士

【摘要】 随着计算机技术的发展和网络带宽的提高,Web上图像资源变的越来越丰富,它们被大量的内嵌在网页中,构成了一个庞大的“Web图像数据库”。Web图像检索致力于解决从纷繁复杂的Web上,帮助用户快速的检索到需要的信息。而目前Web图像检索的瓶颈问题是如何提高检索效率和如何准确的标识图像的语义。基于文本的图像检索(Text-Based Image Retrieval,TBIR)是当前商业图像搜索引擎所采用的主要方式,它面临的主要问题是只利用了Web图像的文本信息来间接地检索图像,没有利用图像本身的内容信息;基于内容的图像检索(Content-Based Image Retrieval,CBIR)则是当前图像检索学术研究领域的主流方式,它面临着主要问题是“语义鸿沟”的问题,即图像的底层视觉特征不能有效的描述其高层语义。根据EMD(the Earth Mover’s Distance)算法的近似匹配算法,提出了Web图像的内存索引方法,此方法主要把高维的图像特征降维为一维的加权平均中心,并以此建立平衡二叉搜索树内存索引。并把索引常驻内存,有效的减少了磁盘I/O的访问开销,显著提高了系统的检索速度。通过改进系统的检索模式,提出了全局检索模式。此模式先基于KNN(K-Nearest Neighbor)的范围查找,过滤掉许多对查询结果没有影响的聚类中心,然后EMD算法匹配找到与样例图像最相似的K个聚类中心,能够用更少的时间检索出比分层检索模式更好的查询结果。针对Web图像的多模特性,提出了基于图像内容和图像文本信息的融合聚类方法。此方法的核心思想是在聚类过程中同时利用Web图像的文本信息和内容特征,实现相互作用或关联以缩小图像的“语义鸿沟”,建立文本关键字和图像内容特征的联系。采用此方法明显提高了图像语义标识的准确度,使得聚类时能够把相似的Web图像尽可能的分到同一类中,从而达到提高检索准确度的目的。通过在VAST(VisuAl & SemanTic image search)系统上的测试分析,证明Web图像的内存索引方法能够在保证系统查准率的前提下,将检索时间减少到原来的1/3左右。采用融合聚类方式,也达到了比较好的检索效果,相对于顺序检索的查准率达到了98.1%。更多还原

【Abstract】 With the development of computer technology and the improvement of network bandwidth, there are more and more Web images because of the rich resources. Most of Web images are embedded in pages, so they constitute a huge "Web image database." Web image retrieval helps users to quickly access to the information which they needed on the complex Web environment. The bottlenecks of current Web image retrieval are how to increase efficiency and how to annotate image of semantics.Text-Based Image Retrieval (TBIR) is the main technology in the current commercial image search engine, which depends on the text only to indirectly retrieve Web images. In contrast, Content-Based Image retrieval (CBIR) has recently receveived a great deal of interest in the research community, the major charllenge of which is the semantic gap problem, i.e. the gap between the low-level visual features and the high-level semantic concepts.We propose the memory indexing algorithm of Web images, on the basis of the approximation algorithm for the Earth Mover’s Distance (EMD). Down through Mitigating the Problem of High Dimension by the weighted average centers, the balanced binary searching tree by memory indexing is built. The index are stored in memory,in order to effectively decrease frequent visits of disk I/O, and significantly improve the speed of the system retrieval. By improving the system retrieval model, the global retrieval model is proposed. First query is based on the scope of the K-Nearest Neighbor (KNN) algorithm. Many of the cluster centers, which do not affect the query results, are filtered to reduce the number of matching operations. Second, EMD algorithm is used to find the K cluster centres, which are similar to the sample image, with less time to get better results than hierarchical retrieval model.Because of the Web images have multi-modal characteristics obviously, based on the content features and textual features of images of multi-modal integration clustering method is proposed. The key idea is to using the content features and textual features while in the process of clustering. So it can simultaneously leverage all types of data which are related to Web image, explore their mutual reinforcement, and construct the association between textual features and content features to bridge the semantic gap.Using this method significantly improve the accuracy of annotation of the images, making the similar Web images put into the same cluster as far as possible, in order to improve accuracy of the retrieval.Based on the test analysis in the VisuAl & SemanTic image search (VAST) system, it proves that memory indexing method of Web iamges spends only 1/3 around time than the original retrieval, under the premise of high precision. The integration of clustering achieves a relatively good retrieval results, in relation to sequence retrieval scheme with the precision of 98.1 percent.更多还原