基于双倍比特量化与分段哈希索引的军事图像过滤  

Large-scale Military Image Filtering System Based on Double-bit Quantization and Index Hashing

作　　者：李雯[1] 邓涵 许玉珍 Li Wen;Deng Han;Xu Yuzhen(Institute of Information and Telecommunication,National University of Defense Technology,Wuhan 430010,China;Institute of Information Engineering,Chinese Academy of Science,Beijing 264200,China;Beijing Institute of Structure and Environment Engineering,Beijing 100076,China)

机构地区：[1]国防科学技术大学信息通信学院,武汉430010 [2]中国科学院信息工程研究所,北京100049 [3]北京强度环境研究所,北京100076

出　　处：《航天控制》2019年第4期59-65,共7页Aerospace Control

摘　　要：二进制码占用存储空间少且易于进行距离度量,因此很多研究者提出二进制量化方法把浮点型特征量化为二进制码,以实现大规模数据环境下的快速最近邻查询。但是,二进制量化会损失原始特征的信息量,使原始特征之间的相似性不能完全保持,导致查询精度降低。针对这一问题,提出双倍比特量化与分段哈希的近似查询索引。首先,设计了一种双倍比特量化方法,通过把特征的每一维数据量化为2个比特二进制码,增加特征之间的区分性;然后,针对双倍比特量化的二进制码提出双倍比特分段哈希索引,通过对二进制码分段并建立哈希索引的方式,提高查询速度。据此,设计了基于双倍比特量化与分段哈希索引的大规模军事图像过滤系统。实验表明,相比于Faster R-CNN+CNNH+MIH系统,本文方法可以使军事图像过滤精度提升5.4%。Due to the advantages of less storage space and easy distance measurement of binary codes,the binary quantization therefore is proposed by many researchers,which is used to quantize the floating-point features into binary codes for the purpose of achieving fast nearest neighbor queries under large-scale data context.However,the binary quantization,by losing the information of the original features,is hard to completely keep similarities among the original features and then leads to decrease the query accuracy.To further reduce quantization loss and improve the accuracy and speed of image filtering,this double-bit quantization and index hashing method is proposed in this paper.Firstly,a novel double-bit quantization(DBQ)is designed to assign more bits to each dimension for higher retrieval accuracy.Then,a double-bit index hashing(DBIH)is presented to efficiently establish index binary codes generated by DBQ.Accordingly,the large-scale military image filtering system is designed in this paper,which is based on DBIH.The experiment results show that the proposed method can improve the filtering precision of military image by 5.4%compared with the precision of faster R-CNN+CNNH+MIH system.

关 键 词：双倍比特量化 分段哈希索引 军事图像过滤 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]