基于内容的图像检索研究

【作者】 王井龙；

【导师】 赵春霞；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2003， 硕士

【摘要】 大量多媒体化信息的产生使得基于内容的检索技术成为研究的热点。基于内容的图像检索(Content Based Image Retrieval，简称CBIR)，是指直接根据描述图像内容的各种特征进行检索，目前主要集中于底层特征的相似度匹配的研究，包括颜色、纹理、形状等。 颜色是图像内容组成的基本要素，是人识别图像的主要感知特性之一。基于颜色直方图的检索方法是使用最为普遍的方法，该过程主要涉及图像颜色空间的选择、颜色空间的量化、颜色直方图定义以及直方图空间中距离的定义等。本文提出了一种基于分块策略的检索方法，取得了较好的效果。 纹理是图像中一个重要而又难于描述的特征，通常采用统计方法、结构方法以及频谱方法进行分析。灰度共生矩阵法是一种有效的统计方法，本文实现了从灰度共生矩阵抽取统计特征来进行基于纹理的检索。 基于图像内物体形状的检索是基于内容检索当中最具挑战性的问题之一，因为寻找符合人眼感知特性的形状特征并不是一件简单的工作。对于形状的描述通常从边界和区域两方面进行，其中傅立叶描述符和不变矩是两种成功的表达方法。本文通过试验验证了七个不变矩对平移、旋转、缩放等的不变性。 最后介绍了典型的基于内容的图像检索系统的结构，通过综述指出了目前研究中存在的主要问题并指出了今后的研究方向。更多还原

【Abstract】 The Creation of large multimedia information makes the content-based retrieval technology become a research focus. Content Based Image Retrieval (CBIR) means retrieving images based on the image content description features directly. The present research on CBIR focuses on the similarity-matching of lower-features, such as color, texture and shape.Color is basal element of image content and is one of the main sensing features for recognizing images. Color histogram is used widely for retrieving images, and the processing mostly concerns the choice and quantization of color space, the definition of color histogram and the distance of two points in color histogram space. This paper presents a method based on dividing-parts strategy and getting better retrieval results.Texture is an important image feature that is hard to describe. The analysis methods mostly include statistic method, structure method and frequency method. Gray scale co-existing matrix is an effective method for describing texture. This paper gives the procedure for retrieving images based on the characters extracted form it.Retrieval based on the object shape in an image is one of the most challenging problems in CBIR, because looking for the shape characters according with the human vision perception is not a simple work. Boundary and region are two aspects for describing shape. Most successful expression methods are Fruior Descriptors and Invariant Moments. The experiments show that seven Hu-Moments are invariant to displacement, rotation and scaling.Finally, we introduce the typical system framework of CBIR. Based on the survey, some open issues, which need to be researched further, are presented, and the future research trends are also given.更多还原