基于SIFT特征向量的图像检索优化  

Optimization of SIFT-Based Image Retrieval

作　　者：肖曼玉[1] 卢江虎 谢公南[1] 

机构地区：[1]西北工业大学理学院应用数学系,西安710072

出　　处：《应用数学和力学》2013年第11期1209-1215,共7页Applied Mathematics and Mechanics

基　　金：国家自然科学基金青年科学基金(11302173)~~

摘　　要：基于SIFT(scale-invariant feature transform,尺度不变特征转换)向量的图像检索在精度和实时性方面都与使用者的心理预期有较大的偏差,该文在建树(build vocabulary tree)、检索、以及匹配度计算方面做了一些改进,在满足实时性的要求下,提高了检索精度;在建树过程中,重新定义了SIFT特征向量聚类机制,将分类和K均值聚类法结合起来代替传统的K均值聚类法;在进行图像检索时,直接利用已有欧氏距离信息,减少向量之间距离的计算,对SIFT向量统一化处理;最后通过改进单位化处理方法,克服SIFT大数据造成的误差.数值结果表明,改进后vocabulary tree的节点有更强的差异性,克服了将训练集按数量均分而不是按距离均分和直接决定树的层数的缺陷;使得检索时间很好地满足了实时性的要求;改进的单位化方法消除了SIFT大数据的误差,从而极大地提高了检索精度.In order to deal with the great discrepancy between the expectations of users and the real performance in image retrieval, some improvement on building tree, retrieval and matching methods were made with great success both in accuracy and in efficiency. More precisely, a new clustering strategy was firstly redefined during the building of vocabulary tree, which com- bined the classification and the conventional K-means method. Then a new matching method to eliminate the error caused by large-scale SIFY was introduced. What was more, a new unit mechanism was adopted to shorten the cost of indexing time. Finally, the numerical results show that an excellent performance is obtained after these improvements. A vocabulary tree with more distinguished nodes is achieved, of which the height is defmed automatically and the index accuracy is enhanced greatly. Furthermore, a faster indexing procedure is realized, of which the indexing time is much less than 1 s.

关 键 词：SIFT 图像检索 倒排文件 K均值聚类 

分 类 号：V19[航空宇航科学与技术—人机与环境工程] O343.6[理学—固体力学]