Codebook-softened product quantization for high accuracy approximate nearest neighbor search

被引:4
作者
Fan, Jingya [1 ]
Pan, Zhibin [1 ,3 ,4 ]
Wang, Liangzhuang [1 ]
Wang, Yang [2 ]
机构
[1] Xi An Jiao Tong Univ, Fac Elect & Informat Engn, Xian 710049, Shaanxi, Peoples R China
[2] Xian Univ Technol, Dept Informat Sci, Xian 710048, Shaanxi, Peoples R China
[3] Zhengzhou Xinda Inst Adv Technol, Zhengzhou 450002, Peoples R China
[4] Xi An Jiao Tong Univ, Res Inst, Hangzhou 311215, Zhejiang, Peoples R China
基金
中国国家自然科学基金;
关键词
Approximate nearest neighbor search; Product quantization; Vector quantization; High-dimensional space; Soften codebooks; BINARY QUANTIZATION; NORMALITY; CODES;
D O I
10.1016/j.neucom.2022.08.002
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Product quantization (PQ) is a fundamental technique for approximate nearest neighbor (ANN) search in many applications such as information retrieval, computer vision and pattern recognition. In the existing PQ-based methods for approximate nearest neighbor search, the reachable best search accuracy is achieved by using fixed codebooks. The search performance is limited by the quality of the hard codebooks. Unlike the existing methods, in this paper, we present a novel codebook-softened product quantization (CSPQ) method to achieve more quantization levels by softening codebooks. We firstly analyze how well the database vectors match the trained codebooks by examining quantization error for each database vector, and select the bad-matching database vectors. Then, we give the trained codebooks bbit freedom to soften codebooks. Finally, by minimizing quantization errors, the bad-matching vectors are encoded by softened codebooks and the labels of best-matching codebooks are recorded. Experimental results on SIFT1M, GIST1M and SIFT10M show that, despite its simplicity, our proposed method achieves higher accuracy compared with PQ and it can be combined with other nonexhaustive frameworks to achieve fast search. (c) 2022 Elsevier B.V. All rights reserved.
引用
收藏
页码:107 / 116
页数:10
相关论文
共 41 条
[1]   Optimized residual vector quantization for efficient approximate nearest neighbor search [J].
Ai, Liefu ;
Yu, Junqing ;
Wu, Zebin ;
He, Yunfeng ;
Guan, Tao .
MULTIMEDIA SYSTEMS, 2017, 23 (02) :169-181
[2]   Quarter-Point Product Quantization for approximate nearest neighbor search [J].
An, Shan ;
Huang, Zhibiao ;
Bai, Shuang ;
Che, Guangfu ;
Ma, Xin ;
Luo, Jie ;
Chen, Yu .
PATTERN RECOGNITION LETTERS, 2019, 125 :187-194
[3]   The Inverted Multi-Index [J].
Babenko, Artem ;
Lempitsky, Victor .
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2015, 37 (06) :1247-1260
[4]   Additive Quantization for Extreme Vector Compression [J].
Babenko, Artem ;
Lempitsky, Victor .
2014 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), 2014, :931-938
[5]   In defense of Nearest-Neighbor based image classification [J].
Boiman, Oren ;
Shechtman, Eli ;
Irani, Michal .
2008 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, VOLS 1-12, 2008, :1992-+
[6]   Approximate Nearest Neighbor Search by Residual Vector Quantization [J].
Chen, Yongjian ;
Guan, Tao ;
Wang, Cheng .
SENSORS, 2010, 10 (12) :11259-11273
[7]  
DAGOSTINO RB, 1971, BIOMETRIKA, V58, P341, DOI 10.1093/biomet/58.2.341
[8]   Compact Hash Codes for Efficient Visual Descriptors Retrieval in Large Scale Databases [J].
Ercoli, Simone ;
Bertini, Marco ;
Del Bimbo, Alberto .
IEEE TRANSACTIONS ON MULTIMEDIA, 2017, 19 (11) :2521-2532
[9]  
Friedman J. H., 1977, ACM Transactions on Mathematical Software, V3, P209, DOI 10.1145/355744.355745
[10]   Optimized Product Quantization [J].
Ge, Tiezheng ;
He, Kaiming ;
Ke, Qifa ;
Sun, Jian .
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2014, 36 (04) :744-755