【作者】 杨泽新；

【导师】 冯玉才；

【作者基本信息】 华中科技大学 ， 计算机软件与理论， 2006， 硕士

【摘要】 随着数字图像处理技术的快速发展和Internet的普及,数字图像的应用愈加广泛,对图像进行有效管理和查询的需求也越来越迫切。基于内容的图像检索(content-based image retrieval, CBIR)技术自提出以来,得到国内外信息领域科技人员的广泛重视和研究,迅速成为多媒体、数据库技术中的研究热点之一。图像的特征提取算法是CBIR技术的一个重要研究内容。颜色是最常用的可视化特征,常见的基于颜色直方图的算法丢失了大量的有用信息例如空间信息,导致检索效果不能满足需要。为改进这些不足,设计了基于有意义区域的颜色检索算法:通过计算图像的颜色聚合度,确定图像分割方法。对聚合度较高的图像利用区域生长方法进行分割;对聚合度较低的图像进行简单的分块,每个分块作为一个区域。然后提取各区域的颜色特征并动态加权求和进行检索。为提取图像的形状特征,改进并实现了基于傅立叶半径描述子的形状检索算法。对彩色图像进行图像分隔,并通过迭代法求得自适应阈值后将彩色图转化为二值图。然后进行轮廓跟踪,再求取形状的形心对各个采样点的距离,所求得的距离值形成一个半径序列,对这个序列进行傅立叶变换,截取结果数组前面若干个元素,取模作为形状特征的描述,然后在此基础上进行形状检索。在分析常见典型CBIR系统结构优缺点的基础上,设计并实现了达梦图像数据库检索系统(DMIR)。对功能进行了合理划分,对后台数据库作了部分优化,提出了优化金字塔技术(Optimized Pyramid Technique, OPT)处理高维向量的索引问题。在该系统上的试验结果表明以上提出的两个算法在绝大多数情况下具有较高的准确度和查全率,具有一定的实用价值。更多还原

【Abstract】 With the rapid development of digital image process technique and popularization of INTERNET, Digital images are widely used in many respects. A problem appears that how to locate a desired image in a large and varied database efficiently. So the Content-Based Image Retrieval (CBIR) technique is proposed, which is widely discussed in literature and becomes a new hotspot. The method for feature extraction is the key issue in this field. In this paper, we mainly discuss algorithms for image retrieval based on color and shape features.Color histogram is widely used in CBIR. However, the classical color histogram method has some drawbacks, such as without containing information about spatial locations or distributions of pixels in an image. To overcome the limitations above, a meaningful regions-based algorithm for color image retrieval is proposed: the color cluster parameter of images are first calculated, those images of high value parameters are segmented into sub-regions, and the others are only divided into sub-blocks ,each of which is regarded as a region. Each region is given a dynamic weight, and the total similarity between two images is computed according to those regions’color histograms.In this paper, we implement and improve an algorithm based on Fourier radius descriptor. An image is segmented and transformed into binary color, and the profile is tracked. The distance of each sample point from the region center is determined, and these distance values form a list, on which a Fourier transform is carried through. The result of transform can be used to describe the shape feature and support retrieval.We propose the framework of DMIR (DM Image Retrieval) system, with optimization of the background DBMS.DMIR adopt the Optimized Pyramid Technique (OPT) to deal with the high-dimensional data spaces so that it can obtain results as quickly as possible. On this foundation, the DMIR system and some retrieval algorithms are implemented. The results of our experiments comparing with other algorithms demonstrate that the methods proposed above are of great value in practice.更多还原