【作者】 张海龙；

【导师】 胡迎松；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2007， 硕士

【摘要】 随着多媒体技术和因特网的飞速发展,各种图像数据大量涌现。图像检索已成为重要研究课题。基于内容图像的检索直接对图像内容进行分析,抽取表达图像内容的低层特征(颜色、纹理、形状、空间特征),形成表达图像内容的特征向量。执行相似度匹配算法,找到最近似的图像。相关反馈技术通过用户在检索过程的反馈判断图像是否相关,不断减小查询结果的范围,直至定位到目标图像。图像检索系统的性能评价指标包括查准率和查全率等。语义图像检索试图提取图像的语义特征。语义层次化模型分为低层特征层、对象层和语义概念层。图像语义表示有文本表示法、传统知识表示法和描述定义语言法。图像语义提取对应特征提取、对象识别和基于领域知识的语义提取三个关键处理过程。语义分析方法包括低层特征自动提取语义、建立关键字语义网络和基于语义向量。语义标注有人工、自动、半自动、分层标注和同义词聚类。语义网络表示成一个三元组。关联度表示语义网络语义的丰富程度。语义网络的逻辑结构是一组图像结点与关键字结点连接构成的网络。权值是图像与关键字的匹配程度。改进后的语义网络采用关键字的聚合和图像的归类,对语义网络进行化减和图像自动语义标注。实现一个基于语义网络的图像检索实验系统。存储结构采用关系数据库模型,设计实体-联系图和关系表。包括图像表、注释表、图像注释表,特征索引表和低层特征表。改进的语义网络的数据库表在图像表和注释表中各增加一个关联数字段。语义网络的进化必须完成三个步骤:计算并导入图像低层特征;语义网络的初始化;语义网络的训练,建立关联。实验系统采用语义检索、低层特征检索和混合检索三种查询。设计四组实验比较不同查询方式下的查全率和查准率,并分析该系统的性能。更多还原

【Abstract】 As the highly development of multi-media and Internet, lots of image data comes out. The image retrieval becomes an important research task. The content-based image retrieval tries to analyze the image content directly in order to extract the low-level features (color, texture, shape, spatial feature) which describe the image content, forming a vector which does for the image. Then the most similar image is dug out through running the similarity match arithmetic. By the relevant feedback technology, users try to evaluate if or not the checked-out picture is suitable, narrow the retrieval field and try again until the most similar picture is seeked out. The recall rate and precision rate is two parameters to evaluate the quality of the image retrieval system.The semantic image retrieval intends to extract the semantic feature of a picture. The semantic hierarchy model is composed of low-level layer, object layer and concept layer. The semantic expression of the image includes text-based method, traditional knowledge method and language description method. The image semantic extract has three key processes which are feature extract, object recognition and field-knowledge-based semantic extract. The semantic analyse method includes extracting semantic automatically, setting up semantic network and basing on semantic vector. A picture can be labeled by hand, semi-automatically, by divided layer and by words clustering.The semantic network can be taken as a three-element group. The relevancy is to show how informative the semantic network is. The logic structure of a semantic network is like a net, in which a group of pictures combined with a group of keywords. How a picture matches a keyword is judged by weight. An improved semantic network uses keyword clustering and image classification in order to make the network less complex and label the picture automatically.An image retrieval experiment system based on semantic network is built up, using relation database model to store the network. Both the entity-relation model and the relation table are designed. The relation table includes five tables which are image table, annotation table, image-annotation table, low-level-feature indexing table and low-level feature table. The database table of an improved semantic network adds a conjoint field both to image table and annotation table. The evolution of the semantic network has to accomplish three steps: the first step is to calculate and input the low-level feature of image; the second step is the initialization of the semantic network; the third step is to train the semantic network. The experiment system uses three search methods, each of which uses semantic retrieval, low-level retrieval and mixed retrieval separately. Four rounds of experiment are done to compare the recall rate and precision rate in different ways and analyse the performance of this system.更多还原