面向海量图像搜索的高维索引结构与快速检索算法研究

With the rapid development of Internet and multimedia technology, multimedia information especially image is growing explosively, then people's demands of retrieving their interested images rapidly from huge image databases are becoming rather urgently. In order to provide retrieval service accurately and quickly, how to index large scale image in limited memory has become the key problem in the field of image retrieval. In this project, hashing-based rapid indexing approach for large scale and high dimensional image is proposed to meet real-time retrieval. In view of image features' space distribution characteristics, adaptive soft assignment based image descriptor constructing approach is presented to improve discrimination. Take the advantage of quick location of hashing, indexing method for large scale image features is proposed. Therefore, according to the distribution characteristics of each image feature component, adaptive quantization and encoding approach is proposed to resolve the problem of storing large scale image features. By adding attribution information to feature vectors in indexing structure, a pre-sorting approach is presented to reduce the scale of candidate retrieval set and improve image query speed. The performance of proposed approaches will be tested and evaluated on a prototype system and real video search engine.The results of this project will effectively improve the retrieval speed of the image search engine.

随着互联网和多媒体技术的高速发展，以图像为代表的多媒体信息呈爆炸性增长，如何从海量的图像中快速搜索到自己感兴趣的图像内容，如何在有限的计算机内存中构建高维索引已经成为图像搜索引擎必须解决的关键问题。本项目针对图像搜索引擎的实际需求，提出基于哈希映射的大规模高维图像特征快速索引方法；针对图像特征的空间分布特点，提出通过自适应软分配方法构造图像描述符以提高其区分度；利用哈希映射的快速查找特点，实现大规模图像特征索引结构的快速构建方法；针对图像特征各维度分量的分布特点，研究自适应特征量化和编码方法，解决大规模图像特征在内存的存储问题；通过在索引结构中为特征向量增加归属信息来对其进行预排序，降低查询候选集规模，提高图像查询速度。本项目的研究成果将有效地提升图像搜索引擎的检索速度。

随着互联网和多媒体技术的高速发展，以图像为代表的多媒体信息呈爆炸性增长，如何从海量的图像中快速搜索到自己感兴趣的图像内容，如何在有限的计算机内存中构建高维索引已经成为图像搜索引擎必须解决的关键问题。本项目针对图像搜索引擎的实际需求，提出基于哈希映射的大规模高维图 像特征快速索引方法；针对图像特征的空间分布特点，提出通过自适应软分配方法构造图像描述符以提高其区分度；利用哈希映射的快速查找特点，实现了大规模图像特征索引结构的快速构建方法；针对图像特征各维度分量的分布特点，研究自适应特征量化和编码方法，解决大规模图像特征在内存的存储问题；通过在索引结构中为特征向量增加归属信息来对其进行预排序，降低查询候选集规模，提高图像查询速度。本项目申请专利2项，软件著作权1项，发表高水平论文7篇，培养博士生2名，硕士生5名。本项目的研究成果将有效地提升图像搜索引擎的检索速度。