各向同性的迭代量化哈希算法

Isotropic Iterative Quantization Hashing

准确有效的哈希算法是实现海量高维数据近邻检索的关键.迭代量化哈希(Iterative Quantization,ITQ)和各向同性哈希(Isotropic Hash,Iso Hash)是两种知名的编码方法.但是ITQ算法对旋转矩阵施加的约束过于单薄,容易导致过拟合;而Iso Hash算法缺乏对哈希编码的更新策略,降低了编码质量.针对上述问题,提出了一种各向同性的迭代量化哈希算法.该方法采用迭代的策略,对编码矩阵和旋转矩阵交替更新,并在正交约束的基础上增加各向同性约束来学习最优旋转矩阵,最小化量化误差.在CIFAR-10、22K Label Me和ANN-GIST-1M基准库上与多种方法进行对比,实验结果表明本文算法在查准率、查全率以及平均准确率均值等指标上均明显优于对比算法.

Hashing is a key technique to achieve fast nearest neighbor search in high-dimensional,massive datasets.Among various methods,Iterative Quantization（ ITQ） and Isotropic Hash（ Iso Hash） are probably the most popular ones due to their high retrieval accuracy. However,as the constraints imposed on the rotation matrix are too weak,the optimization problem in ITQ is severely under-deterministic and therefore easy to cause over-fitting. In Iso Hash,the isotropic projection matrix is updated in a manner that is completely independent of the binary hash codes,and thereby the quality of the produced hash codes may be depressed. To address these issues,this paper proposes an isotropic iterative quantization hashing method,which extends the formulation of ITQ by incorporating properly the isotropic prior adopted in Iso Hash. In our method,the hash code matrix and rotation matrix are updated alternately in an iterative fashion. Experiments are conducted on three benchmark datasets,CIFAR-10,22 K Label Mel and ANN-GIST-1M. The results showthat the proposed method performs better than the competing methods in terms of precision,recall and m AP.