摘要
结合互联网图像搜索的特点,提出了一种适用于大数据集基于电路网络的线性多图融合框架.根据核矩阵对应图模型的马尔可夫网络特征,分析了电路网络描述马尔可夫网络和多马尔可夫网络融合的可行性.分析了单源单地电路网络和谱聚类在描述流形上的相似关系,并受此启发找到了对应于等周分割模型的单地多源电路网络模型,给出了它的希尔伯特空间解释.分析了电路网络和经典排序算法的联系,进而提出了一种新的快速迭代算法.在以上工作的基础上,提出了多图融合的电路网络模型,该模型可以方便的利用k近邻信息和相关信息.在多模态图像搜索的应用里,比较了单地多源电路网络模型和流形排序以及其他多图融合、相关性分析模型;实验结果验证了融合模型和快速算法的有效性.
Characteristics of the internet image search were considered, and a linear multi-graph fusion framework for large scale datasets based on circuit networks was proposed. According to the Markov network characteristic of the kernel matrix＇s corresponding graph model, the feasibility of using circuit networks to describe both the Markov network and the fusion of Markov networks was analyzed. The similarity of the single-source single-ground circuit network to spectral clustering in describing the manifold was analyzed, inspired by which the single-ground multi-source circuit network corresponding to the isoperimetric partitioning model was found, and its explanation of Hilbert space was given. Relations between circuit networks and classical ranking algorithms were analyzed, and a novel fast iteration algorithm was proposed. The circuit network model for the multi-graph fusion was proposed, which can incorporate the k-nearest neighbor information and the correlation information easily. In the multi-modality image search application, the single-ground multi- source circuit network, manifold ranking, and other multi-graph fusion and correspondence analysis models were compared; experimental results showed the efficiency of the fusion model and the fast algorithm.
关键词
图融合
电路网络
多模态
图像搜索
graph fusion
circuit network
multi-modalitw image search
