高维多模态数据规模不断扩大，数据存储和检索面临新挑战。研究多模态数据稀疏化降维及哈希技术，降低存储空间，提高检索速度，已引起学术和产业界的高度关注。传统稀疏化降维及哈希技术处理多模态数据时，与检索应用分割开来，忽略多模态数据间语义关联对降维的影响。本课题研究，如何结合多模态数据检索需求，实现检索引导的多模态数据稀疏化降维及哈希技术。首先研究建立多模态数据语义关联模型、检索序列模型、同模态与不同模态数据相似保持模型，然后将多个模型统一在一个框架中。执行多模态数据稀疏化降维及哈希投影时，该框架既可以保持同模态数据的全局、局部及语义相似特性，又可以保持不同模态数据间与检索密切相关的语义关联、检索序列和语义相似特性，从而有利于提高稀疏化处理后多模态数据的检索性能和检索效率。

Multi-modal data storage and retrieval become much challenging, since these data are high dimensional and their scale is expanding constantly. Studies on the sparse dimensionality reduction and hashing techniques of multi-modal data are helpful to decreasing the data storage space and improving the retrieval speed, thus, have attracted many attentions in academic and industrial circles. Traditional techniques perform sparse dimensionality reduction and hashing on multi-modal data independently from the retrieval requirement, thus, they ignore the impacts of the semantic correlation among heterogeneous modal data on the sparse dimensionality reduction and hashing. This research studies how to combine the multi-modal data retrieval requirement with data dimensionality reduction and hashing, and makes the techniques retrieval steered. This project first studies the establishment of multimodal data models, such as the semantic correlation model, the retrieval sequence model, the intra-modality and inter-modality similarity model, and then unifies these models in one framework. When performing sparse dimensionality reduction and hashing projection on multi-modal data, the framework not only preserves the global, the local and the semantic similarity properties of homogeneous data, and but also preserves the semantic correlation, the retrieval list and inter-modality similarity of heterogeneous data, thus improves retrieval performance and efficiency.