基于E^2LSH的特定人物快速检索方法

E^2LSH-based Fast Specific Person Retrieval

特定人物检索对视频监管和视频搜索具有重要意义。使用PCA-SIFT(Principal Compo-nents Analysis-Scale Invariant Feature Transform)表征图像内容,并引入E2LSH(Exact EuclideanLocality-Sensitive Hashing)来构建索引文件,提出了一种基于E2LSH的特定人物快速检索方法。新方法首先对特征进行降维映射,并根据位置敏感哈希函数的运算结果构建索引文件,然后再用点点对称OOS(One-to-One Symmetric)的匹配策略查询最近邻点,实现特定人物的快速检索。实验结果表明,与传统方法相比新方法不但提高了人物检索精度,而且检索时间也大大减少,同时,对大规模数据库有较好的适应性。

The specific person retrieval is very important to video supervision and search.In this paper,images are characterized using PCA-SIFT（Principal Components Analysis-Scale Invariant Feature Transform） and E2LSH（Exact Euclidean Locality-Sensitive Hashing） is introduced to construct index,and hence a fast specific person retrieval method based on E2LSH is proposed.Firstly,features are mapped to low dimensional space,index files are constructed according to the result of LSH,and then,the nearest neighbors are searched using OOS（One-to-One Symmetric） matching to achieve fast specific person retrieval.Experimental results show that compared to traditional methods,this novel method can increase retrieval accuracy,decrease retrieval time significantly and adapt to large scale datasets with satisfactory performance.