【作者】 邓娟；

【导师】 杨家明；

【作者基本信息】 东华大学 ， 计算机应用技术， 2005， 硕士

【摘要】 低层视觉特征提取、高维数据索引机制和相关反馈方法是面向大规模图像库基于内容检索的三个关键问题。真实地反映图像内容的低层特征是图像检索精确度得以提高的必备条件,有效的高维索引机制是面向大规模图像库的检索能达到实时性要求的关键技术,而相关反馈方法是利用人机交互方式动态地捕捉图像检索中高层语义和低层特征关系的一个重要途径。 论文探讨了颜色特征和形状特征提取方法。针对颜色直方图无法捕捉颜色组成之间的空间关系的缺陷,提出了基于主色分块的颜色特征提取方法。首先通过颜色直方图确定图像的主色,再根据熵最大化离散策略提取图像颜色空间分布信息得到颜色特征表示,最后定义了相似性测度。图像分割是解决目标检测、特征提取等问题的关键所在,而形状特征提取的首要任务是获得目标图像边界。利用本文提出的基于颜色基元和基于自适应网络模型两种图像分割法将目标与背景分开,利用Canny边缘算子描述目标对象轮廓得到形状特征表示。 对数据空间的精确描述及对数据空间的有效划分是高维索引机制中的关键问题。论文提出了基于格矢量量化的层状索引机制。该方法先采用高斯混合模型模拟特征空间数据分布,利用改进的BP算法神经网络对特征空间中的数据点进行归类。然后采用格矢量量化高斯类,这样就能充分利用单一高斯类中各维之间的相关性,对数据向量构造更加精确的近似表示。在划分特征空间、格矢量量化高斯类的基础上,利用三层树状结构来组织数据构造索引机制。并给出了索引结构的维护方法。和现有的索引机制相比,基于格矢量量化的层次型索引机制可以显著减少检索时需要访问数据向量的次数,取得了更好的检索性能。 相关反馈方法研究中涉及两个关键问题,一是如何从反馈样本中综合多特征表达用户查询概念。二是如何结合正、负反馈样本调整特征分量的权值。论文提出了基于多特征融合的相关反馈方法。通过对更多还原

【Abstract】 Visional feature extraction, high dimensional indexing mechanism and relevance feedback are three important issues in content-based image retrieval. Low-level features which reflect the image content is the necessity for improvement in the image retrieval performance, the Relevance feedback techniques are important approaches closing up the semantic gap between high-level concepts and low-level features in image retrieval effectively, and efficient indexing schemes for high-dimensional data are required for real-time retrieval in large-scale image database.Research on the feature extraction in color and shape has been done. The color histogram does not contain the information of spatial distribution of colors across an image. To deal with this problem effectively, a color feature extraction approach based on domain color partition is proposed in the dissertation. Color histogram is used to select the domain colors of the image and the maximum entropy algorithm is applied to extract the region information of selected colors. At last, similarity measurement is defined. Image segmentation is very important in object detection, feature extraction processing. Object outline detection must be done before shape features are extracted. Based on texture element and based self-adaptive network model, two methods are proposed to segment images and detect objects contour. Then Canny contour depiction is introduced to describe the shape.Accurate estimate of data distribution and efficient partition of data space are key problems in high-dimensional indexing schemes. In this dissertation, a hierarchical indexing scheme based on the lattice vector quantization (LVQ) indexing method is proposed for higher performance. Gaussian mixture distribution is used to model the feature space and lattice vector quantizers are trained to partition simplified Gaussian clustering. On the foundation of feature space partition and Gaussian clustering lattice vector quantization, hierarchical structure is used to organize the data. At last, maintenance of the indexing is introduced. Experiments on a large real-world dataset demonstrate a remarkable reduction of the amount of accessed vectors in k - NNsearches and a better performance is achieved compared with existing indexing schemes.Extracting multi-feature suitable to representing query concepts of users from feedback samples and adjusting of vector element’s weights by combining the negative samples and positive samples are two key problems for relevance feedback techniques. In the dissertation, a relevance feedback method based on multi-feature extraction (BMFE) is proposed for content-based image retrieval. By analyzing the relevance of the negative example and positive, the global dispersion formulation is designed. Lagrange multipliers are used to resolve the dispersion matrix. The BMEF approach not only dynamically adjusts the vector element’s weight, but also reflects the concepts of the users.At last, the prototype content-based image retrieval system programmed with Java is introduced in this dissertation. The system is possessed with feature extraction, high-dimension indexing schemes and relevance feedback using the skills proposed in the forecited sectors.更多还原