基于感兴趣区域的图像检索

【作者】 高和蓓；

【导师】 胡学龙；

【作者基本信息】 扬州大学 ， 计算机应用技术， 2006， 硕士

【摘要】 基于区域的图像检索技术(RBIR)是基于内容图像检索(CBIR)的一个重要研究方向。利用图像分割技术把图像分成多个区域,用区域特征集表示和索引图像。在一定程度上实现了对象层次的检索,减小了图像底层特征和高层语义之间的语义鸿沟。但是这些分割出来的区域中只有一部分是用户想要得目标区域。为了提高系统的检索效率,进一步获得语义信息,必须获得查询过程中用户感兴趣的区域。本文分析和概括了图像检索的基本原理、关键技术和检索结果的评价方法,并且对基于区域检索的三个关键步骤进行了研究,提出了一种利用颜色和数学形态学的分割方法。利用图像的颜色特征把图像分成多个区域,然后根据数学形态学对分割后的图像的区域进行合并,去除过分割的区域,合并对象层次上的区域目标。通过对Zernike矩性质的研究,对分割后图像中的各个区域提取了6阶6重矩,结合区域的颜色、中心、轮廓、尺寸,构成了区域的特征向量。用这些特征向量来表示区域。为了提高检索效率,采用阶梯式匹配的方法,进行基于感兴趣区域的图像匹配。相关反馈是最近图像检索中比较重要的研究方法,它最初发源于文本文档检索,这一技术试图填补底层图像特征和高层语义内容之间的空白。从模式分类和机器学习的角度,以支持向量机(SVM)为分类器,进行相关反馈。对每个检索结果,让用户标定正确结果和错误结果,并用标定的结果形成正例集合和反例集合,训练SVM分类器作为模型,并根据学习所得模型进行检索。结果表明该方法可以有效地检索出更多的相关图像,并且在小样本的情况下具有很好的推广能力。本文建立了一个图像检索的原型系统,该系统可以完成基于底层特征的全局图像检索,如基于直方图、颜色对、Tamura纹理、不变矩、Hu矩等特征的图像检索,以及本文采用的基于Zernike矩特征的感兴趣区域的检索。更多还原

【Abstract】 Region-based image retrieval (RBIR) is an important research aspect of CBIR. Region-based retrieval applies image segmentation to decompose an image into several regions, and uses region visual features to represent and index the image. It can reduce the gap between low-level feature and high-level semantic features and more close to the human perception. These regions not only include relevant objects, but also irrelevant image areas. The irrelevant areas limit the effectiveness of system. To overcome this limitation, the systems must be able to determine similarity based on relevant regions alone. It is called region of interest (ROI).The paper analyzes and summarizes the fundamental, key techniques and performance evaluation of CBIR. An image segmentation method using image color and mathematical morphology is developed in the thesis. It divide image into several regions, then, combinate these regions based on mathematical morphology in order to reduce over-segmentation. Some regions in level of object are shaped.In the research on Zernike moment character, 21 moments are extracted from regions of divided image. These moments combine with the features of color, center, size, contour, moments to represent and index the image. In order to improve the efficiency of retrieval, step match is adopted through ROI-based image retrieval.Relevance feedback is one of important study aspects. It is first introduced in document retrieval, in order to fill the blank between low-level features and high-level semantic features in CBIR, relevance feedback is adopted by CBIR. In the view of pattern classification and machine learning, a classifier is constructed by support vector machine (SVM) for relevance feedback. Users can label the right results and wrong results after one query. Right set and wrong set are formed by these labeled results, which are used for constructing SVM classifier model. Next retrieval is based on this model for finding more relevant images efficiently. It is showed that it has better generalization ability in small sample problem.A prototyping system of CBIR is built in the thesis which can be used for CBIR based on low level visual character, for example, histogram, color pair, Tamura texture, moment invariants, Hu moment feature, and the Zernike moment feature that is adopted in the thesis for ROI based image retrieval.更多还原