Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering

Previous Articles     Next Articles

Research on Cumulative Product Quantization Method for Image Retrieval

DU Danlei 1,LUO Entao  2,TANG Yayuan  1,LEE Yenchun  3   

  1. (1.School of Electronics and Information Engineering,Hunan University of Science and Engineering,Yongzhou 425100,China; 2.School of Information Science and Engineering,Central South University,Changsha 410083,China; 3.Chaoyang University of Technology,Taichung 41349,Taiwan,China)
  • Received:2015-04-10 Online:2015-10-15 Published:2015-10-15

面向图像检索的累加乘积量化方法研究

杜丹蕾1,罗恩韬2,唐雅媛1,李延浚3   

  1. (1.湖南科技学院电子与信息工程学院,湖南 永州 425100; 2.中南大学信息科学与工程学院,长沙 410083; 3.朝阳科技大学,中国台湾 台中 41349)
  • 作者简介:杜丹蕾(1981-),女,讲师、硕士,主研方向:图形图像处理;罗恩韬,副教授、博士;唐雅媛,讲师、硕士;李延浚,副教授、博士。
  • 基金资助:
    湖南省科技厅科技计划基金资助项目(2014FJ6095);湖南省教育厅高校优秀青年基金资助项目(14B070);湖南省教育厅科学研究基金资助项目(湘财教指[2011]91号);永州市指导性科技计划基金资助项目(永科发[2013]17号)。

Abstract: For solving the problem that the classic Product Quantization(PQ) method is restricted on data’s independence,a Cumulative PQ(CPQ) method is proposed in this paper.Orthogonal decomposition is executed on the high-dimensional feature vectors to obtain independent sub-spaces of feature vectors,and decomposes every subspace again according to the compression efficiency,and obtains dependent sub-sub-spaces of feature vectors,uses Cumulative Quantization(CQ) method to quantify the vectors sub-sub-spaces,and uses PQ method to quantify the vectors from sub-spaces.The new method reduces the impact of data’s independence on accuracy of quantization,under the premise of maintaining the compression efficiency.Experimental results show that the new method has small code error compared with classical PQ and Cartesian K-means(CKM) methods,and high recall rate in the application of image retrieval.

Key words: image retrieval, feature extraction, encoding, Product Quantization(PQ), Asymmetric Distance Computation(ADC)

摘要: 针对经典的乘积量化方法易受数据相互依赖关系限制的问题,提出一种累加乘积量化方法。对高维特征向量进行正交分解,得到相互独立的特征向量子空间,依据压缩效率要求,对各特征向量子空间进行进一步分解,得到相互不独立的特征向量次子空间,对次子空间采用累加量化方法进行编码,对子空间采用乘积量化方法进行编码,在保障压缩效率的前提下降低数据相互依赖关系对量化精度的影响。实验结果表明,与经典的乘积量化方法和笛卡尔K-均值方法相比,该方法的编码误差较小,在图像检索应用中的查全率较高。

关键词: 图像检索, 特征提取, 编码, 乘积量化, 非对称距离计算

CLC Number: