【作者】 李海霞；

【导师】 孟祥增；

【作者基本信息】 山东师范大学 ， 教育技术学， 2004， 硕士

【摘要】 自然语言是表达思想的有效工具，利用自然语言表达图像的语义是一种简洁、有效的方法。基于自然语言的图像检索就是根据用自然语言描述的图像语义和图像内容对图像进行检索。图像的语义指图像表达的主题意义，图像的内容主要指图像的内容特征和低层的视觉特征，包括图像的颜色、背景、主体以及主体的视觉特征，如主体的颜色、形状、纹理、位置、大小、方向等。 查询语句的关键语义信息提取技术和图像相似性度量技术是基于自然语言的图像检索技术的关键问题。本文主要针对这两个问题，研究如何提取语句的关键语义信息，以及恰当地对图像进行相似性度量。 本文在研究汉语语句分词特点及方法的基础上，引用已有的词库及包含分词函数的动态链接库对语句进行自动分词，并进一步提取语句的关键语义信息，而且将关键语义信息中的颜色名转换为可量化的HSI数值，为后面进行图像相似性度量打下基础。 对图像内容的不同属性，采用不同方法计算其相似度。对于图像主色调，计算相似度时综合考虑两图像中主色调的相对频率及其每对主色调的相似度。对于图像主体大小，根据主体面积与图像面积之比及检索目标要求，采用不同公式(对应大、中、小)计算其相似度。对于主体位置，采用相对位置归—化距离计算相似度。对于主体方向，采用主体的方向角差计算相似度。整体相似度采用层次化递归方法计算。 为了衡量图像描述的复杂程度，引入了信息测度的概念。图像信息测度也采用层次化递归的方法计算，即根据“图像属性的表征方法”，采用逐层逐级递归计算。如何计算图像某一子属性的信息测度是一个值得探讨的问题，本文对Shannon概率信息测度方法做了改进，并用于对检索结果的排序。 本文在上述工作的基础上设计了一个基于自然语言的图像检索系统。实验结果较为理想。论文最后总结了本文的工作，并提出了进一步的研究探索方向。更多还原

【Abstract】 Natural language is an effective tool to express ideas. It is a sententious and effective method to express the semantic information of images. Natural language-based image retrieval (NLBIR) is a technique for searching images on the basis of image semantic features and the content of image .The image semantic features mean what the image express, while the image content features are the low-level visual features, which include image color, image background, image bodies which are described by body direction, body size and body position etc.There are two key issues in the NLBIR. One is how to extract key semantic information from the Chinese sentences. The other is how to match query and stored images in away that reflects human similarity judgement.On the basis of researching the word segmentation characteristics and methods of the Chinese language, the paper refers to the word-stocks and Dynamic-Link Libraries in order to realize word segmentation of Chinese sentence automatically, and then extract key semantic information from the sentence. What’ s more the paper transforms the color words in the key information into HIS values that can be measured. The HIS values are the basis of calculating the color similarity of the query and stored images.According to different attributes, different similarity measuring methods are adopted. When calculating main colors similarity of two images, the frequencies of colors are considered. The object size of one image is expressed as proportion of the area of object and the area of image. There are three kinds of object size such as large, middle, small. The paper takes different method to gain the object size similarity according to their different character and target image. The object position similarity is calculated on the basis of the distance of two positions. Some method is taken to make the distance less than 1.0. The object direction similarity is calculated on the basis of objects’ angledifference. The paper takes climax to calculate the whole similarity of two images because we take hiberarchy to express one image.To scale the complexity of one image’ s description, the paper introduces a concept that is the information measurement. We also take climax to calculate the information measurement of one image. How to calculate the information measurement of one image’ s attributes is a question that is value to discuss. The paper improves the Shannon method, and uses it in the compositor of retrieval outcomes.In similarity measurements mentioned above, one image searching system that is on the basis of natural language comprehension is designed. The system obtains perfect outcome by experiment proved. In the end, the further research direction is pointed out.更多还原