作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

基于投影残差量化哈希的近似最近邻搜索

杨定中,陈心浩   

  1. (中南民族大学实验教学中心,武汉 430074)
  • 收稿日期:2014-12-01 出版日期:2015-12-15 发布日期:2015-12-15
  • 作者简介:杨定中(1978-),男,实验师、硕士,主研方向:移动搜索,多媒体处理技术;陈心浩,副教授。
  • 基金资助:
    中南民族大学中央高校基本科研业务费专项基金资助项目(CZQ12018)。

Approximate Nearest Neighbor Search Based on Projected Residual Quantization Hashing

YANG Dingzhong,CHEN Xinhao   

  1. (Experimental Teaching Center,South-Central University for Nationalities,Wuhan 430074,China)
  • Received:2014-12-01 Online:2015-12-15 Published:2015-12-15

摘要: 针对投影哈希中投影误差较大,二进制编码时原始信息丢失严重等问题,提出一种近似最近邻搜索方法。该方法通过多阶段量化策略减少编码过程中的投影及量化误差。在每阶段训练时,对前一阶段的量化残差采用投影、按维度训练码书及量化、反投影等运算生成各阶段的子量化器。子量化器按投影后数据的维度提供多个哈希函数,最终的哈希函数由各阶段哈希函数共同构成。在最近邻搜索时,给二进制编码加上权重以便对搜索结果进行重排,提高搜索精度。实验结果表明,基于投影残差量化哈希的近似最近邻的搜索性能优于当前主流的哈希方法。

关键词: 投影残差量化哈希, 大规模搜索, 近似最近邻搜索, 编码权重, 多阶段量化

Abstract: In order to solve the problems of projection errors of the projection hash and serious loss of original information in the process of binary encoding,the projection residual quantization based hash algorithm is proposed.This method reduces the projection and quantization errors by the multi-stage quantization strategy.In each stage,the projection,training dimension codebook,quantization and the quantification of residual error back projection methods are adopted for building the sub-quantizer based on quantification results of residual error in the previous stage.According to its dimension,each sub-quantizer supplies several hash functions.The final hash function is made up of the hash function in each stage.To boost the search accuracy,the weighted binary encoding method is employed to rearrange the search results.Experimental results demonstrate that the approximate nearest neighbor search based on projected residual quantitative hash considerably outperforms other existing hash methods.

Key words: projected residual quantization hashing, large-scale search, approximate nearest neighbor search, coding weight, multi-stage quantification

中图分类号: