面向图像检索的互补哈希表构造方法研究

Multiple hash table method based on locality sensitive hashing can significantly boost the efficiency and performance of image retrieval, and thus becomes the key technique in these applications, meanwhile attracting much attention in recent years. Despite the progress of the research on multiple hash tables, it still needs further study to discover the determinant factors that have great impacts on multi-table search performance; Moreover, images are usually associated with different visual features and semantics. However, state-of-the-art multi-table methods ignore the information from multiple views. To address these problems, this project will propose the principles for complementary hash tables, and study construction methods for complementary hash tables using multi-view information in image retrieval. At the theoretical level, it will analyze the distribution characteristics of nearest neighbors of high-dimensional data, study the key factors for complementary hash tables, and build the principles of complementarity and optimal configuration in complementary hash tables. At the technical level, it will construct complementary hash tables by adaptively fusing multiple visual features, which can improve the robustness and performance of multiple hash tables under the unsupervised scenario; it also will construct complementary hash tables based on the sparse association between semantic labels and hash tables, which can boost the semantic discriminative capability of hash tables and the power of query-adaptive search. Finally, this project will build the construction system and techniques for complementary hash tables, aiming to improve the overall performance of image retrieval.

基于局部敏感哈希的多哈希表方法能够显著提高图像检索的效率和整体性能，是海量图像检索等应用的关键技术之一，同时也是目前基于哈希的图像检索的研究热点之一。然而，目前多哈希表方法中影响其整体检索性能的因素还有待深入研究；同时，图像关联的多种视觉和语义特征也未能在其构造中有效综合利用。针对上述问题，本项目将建立互补哈希表的构造准则，结合图像检索，系统地研究集成多视角信息的互补哈希表方法。在理论层面上，分析高维数据最近邻分布特性，研究影响互补哈希表性能的主要因素，建立多哈希表的互补性准则和配置优化准则。在技术层面上，自适应融合图像多种视觉特征建立互补哈希表，提高非监督情形下多哈希表的鲁棒性和检索整体性能；基于语义标签和哈希表的稀疏关联构造语义适配的互补哈希表，提高哈希表对语义近邻的分辨能力及多哈希表自适应查询能力。本项目最终将针对图像检索建立互补哈希表的构造体系和方法，有效提高图像检索的整体性能。

基于局部敏感哈希的多哈希表方法能够显著提高图像检索的效率和整体性能，是海量图像检索等应用的关键技术之一，同时也是目前基于哈希的图像检索的研究热点之一。然而，目前多哈希表方法中影响其整体检索性能的因素还有待深入研究；同时，图像关联的多种视觉和语义特征也未能在其构造中有效综合利用。针对上述问题，本项目从互补哈希表构造准则、多特征互补哈希表构造方法以及互补哈希表精准排序方法等三个方面开展哈希最近邻搜索的研究。分析高维数据最近邻分布特性，建立结构敏感的哈希量化策略；研究多哈希表互补机制，探索互补哈希表集成策略；综合利用时序图像数据特征，研究时空敏感的图像哈希方法；研究图像多视觉特征的融合机制，提出高效的多特征多哈希表构造方法；利用查询样例的数据分布特性，设计哈希表查询自适应机制；动态融合图像多种视觉特征，建立多哈希表精准排序策略。项目在执行期间就互补哈希表在 国际上发表了一系列原创成果，并在网络安全等方面开展了应用示范。