【作者】 黄鹏；

【导师】 陈纯； 卜佳俊；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2008， 博士

【摘要】 随着数码技术、扫描技术和Internet的迅速发展,Web图像资源日益丰富。但由于Web数据具有多样性、复杂性和无规则性等特点,如何快速、准确地从海量Web资源中查找用户感兴趣的图像成为一个非常具有挑战性的任务。Web图像检索通过合理地组织Web图像资源,并研究高效的查询和检索方法以达到充分利用Web图像资源的目的。目前主流的Web图像检索方法大致可以分为两大类,即基于关键词的图像检索(TBIR)和基于内容的图像检索(CBIR)。Web图像主要包含两种类型的信息,一种是Web图像本身所包含的丰富的视觉信息,另一种是Web图像所在网页包含的丰富的文本信息。TBIR仅仅使用从文本信息中抽取的文本特征索引和检索图像,而CBIR仅仅使用从图像视觉信息中抽取的低层视觉特征索引和检索用户。显然,要较好地满足用户检索图像的需求,在Web图像检索过程中必须充分利用并融合上述两种不同类型的信息;此外,还需要为Web图像附加上包括图像内容的高层语义在内的各种信息,因为用户主要根据图像的高层语义特征判断图像满足自己的需要程度。但如何为Web图像附加语义信息,以及如何实现上述两种信息的融合直到今天依旧是图像检索领域中的研究难点。针对上述问题,本文首先提出了浅层语义处理技术一词汇相似性计算技术。词汇相似性计算是自然语言处理领域中语义处理的基础性研究之一,主要研究如何计算词汇之间的语义相似程度。本文的研究中将词汇相似性计算技术作为语义信息的度量手段,这使得存在于人类思维中的抽象语义信息具有了可计算性和可对比性,同时也进一步使得Web图像的文本信息和视觉信息(图像视觉信息被表示为高层语义特征,参考下文)的融合成为可能。其次,本文针对Web图像的低层视觉特征和高层语义特征之间的语义鸿沟问题,提出了一个Web图像自动加权标注模型:首先使用各种机器学习和统计技术学习从图像低层视觉特征到图像高层语义特征的映射模型;然后利用这个映射模型抽取图像的高层语义特征;最后根据Web图像文本信息和提取出的高层语义特征本身,使用词汇相似性计算技术度量抽取出的高层语义特征的质量。通过上述步骤可以将Web图像表示为带有权重的高层语义特征,同时也进一步将图像视觉信息和文本信息的融合转化为表达图像内容的高层语义特征和Web图像文本信息的融合。接着,针对Web图像包含的文本信息和从图像视觉信息中提取的高层语义特征,提出了一种具有可扩展性的Web图像检索模型。为了充分利用Web文档中的文本信息和从Web图像低层视觉特征中抽取的高层语义特征,该模型构架在贝叶斯推理网上,利用推理网内在的多信息源融合能力,将Web图像文本特征和Web图像的高层语义特征无缝地融合在一起实现Web图像检索。基于上述研究,本文设计并实现了一个Web图像检索原型系统,该系统充分利用Web图像的两类信息:从Web图像内容中提取高层语义特征,然后将它们与从Web图像文本信息中提取的文本特征融合在一起实现Web图像检索,研究结果验证了本文提出的模型在Web图像检索中的有效性。文章最后对本文的研究工作进行了总结和展望。更多还原

【Abstract】 The rapid development of digital image technology and scan technology and Internet greatly enriches accessible web image resources. Due to the diversity, complexity and irregularity of web resources, how to quickly and accurately find images of interest to users from the huge volumes of the web resources is a very challenging long-term task. To make full use of these web resources, it is necessary to do more research on web image retrieval, including how to organize them reasonably, and how to query and retrieval them effectively. Currently prevalent approaches to web image retrieval fall into two main categories: text-based image retrieval (TBIR) and content-based image retrieval (CBIR).Web images mainly contain two types of information, one is lots of visual information in image contents, and the other is lots of textual information in web pages. TBIR makes use of textual features extracted only from image textual information to index and search images, while CBIR makes use of low-level visual features extracted only from image visual information. To satisfy common user information needing, it is necessary to make full use of the above two types of information in web image retrieval. Furthermore, high-level semantic features should be extracted from image contents, because the degree of satisfaction of retrieved image is judged mainly based on image high-level semantic features. Unfortunately, the extraction of image high-level semantic features and fusion of image textual features and visual features are still a difficult task in the domain of image retrieval.To address the above issues, term similarity measure, a shallow semantic processing technology, is firstly proposed. The research of term similarity measure is one of fundamental research in the domain of natural language processing, focusing on how to quantize term semantic similarity. In this study, term similarity measurement, as the metric form of semantic information, make it possible to compute and compare abstract semantic information existing in human thinking. Furthermore, it is the precondition of the fusion of textual web image information and visual web image information presented by image high-level semantic features.Secondly, a automatic weighted-annotation model for web image is proposed to address the issue of "semantic gap" existing in image low-level visual features and image high-level semantic features: Firstly, it struggles for learning the mapping from image low-level visual feature to high-level semantic feature by means of machine learning and statistical technology; Secondly, the learned map is used to extract high-level semantic feature from image contents; Finally, the quality of extracted high-level features is measured as term similarity based on web textual information and high-level features, resulting the weighted image high-level semantic features which in turn changes the fusion of visual and textual information into the fusion of textual information and image high-level semantic information.After then, we propose a scalable model for web image retrieval. To make full use of textual features extracted from web pages and high-level semantic features extracted from image contents, in this study, the proposed image retrieval model is based on Bayesian inference network. The image textual features and high-level semantic features can be integrated into web image retrieval seamlessly with the help of inference network which has an inherent fusion capability of multiple information sources.Based on the above work, a web image retrieval prototype system is designed and implemented. This system makes full use of two types of web image information as follows. Firstly, extracting high-level semantic features from image contents, then in image retrieval they are integrated with textual features extracted from web textual information. The research results demonstrate the usefulness of the proposed model in web image retrieval.Finally, conclusions and future work are presented.更多还原