多视角自适应哈希学习及其应用

With the rapid development of internet technology, the scale of the image data is experiencing explosive growth. It is the core problem for the storage and management of massive image data to construct efficient indexing mechanism and implement quick query. Hash learning based nearest neighbor search is the most critical technique to carry out the efficient indexing of massive image data. However, massive image data contains some characteristics such as multiple views, diverse semantics, and heterogeneous structures, which makes the performance of current hash-based search methods unsatisfactory. The purpose of this project is to investigate multi-view adaptive hash learning. By analyzing the topological structure and the information discrepancy of multi-view data, we first construct the information metric model for hash function and design the variable bit allocation strategy, then we propose an adaptive quantization based hashing method, which can better preserve the similarity of binary codes. Second, through capturing the relevance and complementarity among different views and leveraging the local construction consistency of manifold embedding, we propose a unified similarity metric learning model and then present a framework of multi-view manifold hash learning, which can further improve the search accuracy. Third, with collaboratively modeling the structure of the sample's nearest neighbor subspace in multiple hash tables, we establish an interpretably complementary criterion for multiple hash tables and then propose a framework of multi-view complementary hash table construction. The research fruits of this project can promote the theoretical development of hash learning and facilitate its application in massive image retrieval and other fields.

随着网络技术的快速发展，图像数据规模急剧增加。建立高效的索引机制并实现快速查询已经成为海量图像数据存储和管理的核心问题。基于哈希的最近邻搜索方法是实现海量图像数据高效索引的关键技术。然而，海量图像数据的多视角、多语义、非结构等特性使得现有的哈希搜索方法难以获得理想的检索性能。本项目以多视角自适应哈希学习为研究对象，通过分析多视角数据的拓扑结构和信息差异，构建哈希函数信息度量模型并设计可变哈希比特分配策略，进而提出自适应哈希量化编码方法，以更好地保持编码相似性；通过挖掘多视角数据之间的关联性和互补性，利用流形嵌入的局部重构一致性，建立联合相似性测度学习模型，进而提出多视角流形哈希学习框架，以提高检索精度；通过对多哈希表样本近邻空间结构进行协同建模，建立可解释的多哈希表互补准则，进而提出多视角互补哈希表构造框架。本项目的研究成果将推进哈希学习技术的理论发展，促进其在海量图像检索中的推广应用。

本项目以海量图像数据的高效索引和快速查询为应用背景，充分利用海量图像数据中的多视角、多语义和非结构等特性，构建自适应哈希比特量化策略，进而研究多视角流形哈希学习方法和多视角互补哈希表构造方法，从基础理论和实际应用两个层面提高哈希最近邻搜索的整体性能，促进哈希学习技术在信息检索、数据挖掘、模式识别、推荐系统、以及社交网络分析等多个领域的推广应用。在国家自然科学基金的支持下，本项目组累计发表学术论文51篇，其中国际顶级会议论文26篇，国际知名期刊论文25篇；申请国家发明专利19项；获得国家自然科学奖二等奖1项、陕西省自然科学奖一等奖1项。在项目实施过程中，负责人获得“陕西省青年科技标兵”称号、首届“西安之星”（科技领域）奖、“陕西省中青年科技创新领军人才”称号。此外，在人才培养方面，培养博士生2人，硕士生15人。