基于双线性迭代量化的哈希图像检索方法

Hashing image retrieval based on bilinear iterative quantization

针对迭代量化哈希算法未考虑高维图像描述符中呈现出的自然矩阵结构,当视觉描述符由高维特征向量表示并且分配长二进制码时,投影矩阵需要昂贵的空间和时间复杂度的问题,提出一种基于双线性迭代量化的哈希图像检索方法。该方法使用紧凑的双线性投影而不是单个大型投影矩阵将高维数据映射到两个较小的投影矩阵中;然后使用迭代量化的方法最小化量化误差并生成有效的哈希码。在CIFAR-10和Caltech256两个数据集上进行实验,实现了与最先进的八种哈希方法相媲美的性能,同时具有更快的线性扫描时间和更小的内存占用量。结果表明,该方法可以减轻数据的高维性带来的影响,从而提高ITQ的性能,可广泛服务于高维数据长编码位的哈希图像检索应用。

Iterative quantization hashing algorithm does not consider that most of high-dimensional visual descriptors for images exhibit a natural matrix structure.When high-dimensional feature vectors represented the visual descriptors and assigned long binary codes,a random projection matrix requires expensive complexities in both space and time.To address the above issues,this paper proposed a hashing image retrieval method based on bilinear iterative quantization,which mapped the high-dimensional into two smaller projection matrices using compact bilinear projections instead of a single large projection matrix.Then it minimized the quantization error in an iterative quantization way and generate effective hash code.Experiments on two datasets,CIFAR-10 and Caltech256.This method achieved comparable retrieval accuracy to the state-of-the-art 8 hashing methods while having orders of magnitude faster linear scanning time and smaller memory footprint.The results show that the proposed method can reduce the impact of the high dimensionality of data and improve the performance of ITQ,the algorithm can widely serve the hashing image retrieval application of high-dimensional data and long coded bits.