【作者】 叶航军；

【导师】 徐光祐；

【作者基本信息】 清华大学 ， 计算机应用技术， 2003， 博士

【摘要】 相关反馈方法和高维数据的索引机制是面向大规模图像库基于内容检索的两个重要问题。相关反馈方法是弥合图像检索中高层语义和低层特征之间语义间隔的一个重要途径，而有效的高维索引机制则是面向大规模图像库的检索能够达到实时性要求的关键技术。对数据分布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。论文提出了基于矢量量化的精确索引机制。该方法使用更适合真实图像库的高斯混合模型描述数据分布，并训练优化的矢量量化器划分数据空间。和现有的精确索引机制相比，可以显著减少检索时需要访问数据向量的次数。在矢量量化索引机制的基础上，论文提出了基于分层结构的近似索引机制。该方法在矢量量化索引机制的基础上使用了概率近似最近邻的方式进行检索，这种两种索引机制相结合的方法取得了比单独的索引机制更好的性能。相关反馈方法研究中的一个关键问题是如何从数目很少并具有不对称性的反馈样本中抽取出最适合表达用户查询概念的特征。论文提出了基于特征子空间分析的相关反馈方法。该方法同时利用图像特征向量在特征子空间内的距离和到特征子空间的距离，并使用两阶段判别式分析的方法求解出适合形成图像检索决策准则的特征。该方法所使用的特征判别既能较好的处理正负样本的不对称性，又能更充分的反映用户的查询概念。针对线性方法对真实图像库中样本分布描述能力的不足，论文研究了基于非线性特征抽取的相关反馈方法。通过对二次型距离的推广，提出了二分量判别函数，使用非线性的方法抽取反映散度判据和距离判据的特征，具有比线性方法具有更高的检索精度。核函数方法是解决非线性问题的通用方法。针对现有基于核函数的相关反馈方法处理样本不对称性时的缺陷，提出了基于超球面的支持向量分类方法。该方法能够同时利用正负反馈样本构造分类器，并能更好的适应两类样本的不对称性，进一步提高检索精度。更多还原

【Abstract】 Relevance feedback techniques and high dimensional indexing schemes are two significant research issues for content-based image retrieval in large-scale image database. 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.Accurate estimate of data distribution and efficient partition of data space are key problems in high-dimensional indexing schemes. In this dissertation, an indexing scheme using vector quantization (VQ) is proposed for exact nearest neighbor (NN) searches. The VQ-based approach assumes a Gaussian mixture distribution, which fits real-world image data reasonably well. After estimating this distribution through EM method, this approach trains optimized vector quantizers to partition data space. Experiments on a large real-world dataset demonstrate a remarkable reduction of the amount of accessed vectors in exact NN searches compared with existing indexing schemes.Based on the VQ-based indexing method, a hierarchical indexing scheme is proposed for higher performance. This approach integrates VQ-based indexing structures with approximate NN searches and performs probabilistic approximate NN searches on approximate vectors. Experiments show the presented hierarchical indexing scheme outperforms the original VQ-based indexing method and probabilistic approximate NN searches. Both presented approaches support quadratic-form distance metric and can integrate with relevance feedback techniques for practical large-scale image retrieval systems.<WP=6>Extracting feature suitable to representing query concepts of users from small and asymmetric feedback samples is a key problem for relevance feedback techniques. In this dissertation, a relevance feedback method using feature subspaces analysis (FSA) is proposed for content-based image retrieval. This approach considers relevance feedback as a classification problem of two-class. Both distance in feature subspace and distance from feature subspace are employed and a two-stage discriminant analysis method is adopted to make the decision rule for image retrieval. The employed feature criterions treat asymmetry between positive and negative samples and represent query concepts of users reasonably well. Experiments on a large database of real-world images demonstrate the FSA approach achieves more stable and precise retrieval results compared with existing linear approaches.The linear approaches do not fit the data distribution of real image database very well. A novel approach of binary component discriminant function (BCDF) is proposed by generalizing the original quadratic-form distance metric. The BCDF approach nonlinearly extracts features by scatter criterion and distance criterion. Experiments show that the BCDF approach outperforms linear approaches and has good computation efficiency. The kernel function method is a general approach dealing with nonlinearity and existing kernel-based relevance feedback methods don’t deal with asymmetry between positive and negative samples very well. A novel approach name hyper-sphere support vector classification (HS-SVC) is proposed to address the sample asymmetry. This method learns a classifier by positive and negative samples and simultaneously considers the asymmetry between them reasonably well. Experiments show a notable precision improvement from other kernel-based methods.更多还原