增强型残差量化的图像视觉特征不完全检索方法

Enhanced Residual Vector Quantization-based Non-exhaustive Retrieval for Image Visual Features

针对图像视觉特征的快速检索问题,提出了一种增强型残差量化的不完全检索方法。建立在增强型残差量化的基础上,提出利用多层低复杂度的码书构建包含较大规模倒排列表的多维倒排索引结构,使得只需根据图像视觉特征的量化编码就可以将其快速地插入到倒排索引结构中。此外,结合倒排索引结构,设计了一种不完全检索方法和图像视觉特征之间近似距离的计算方法。通过在公开数据集进行实验和性能对比,所提出不完全检索方法较典型的三种不完全检索方法具有更好的检索精度和检索效率。

A non-exhaustive retrieval method based on enhanced residual vector quantization is proposed for rapid retrieval among image visual features. Built on enhanced residual vector quantization, a multi-inverted index composed of large number of inverted lists is presented, which is constructed by several codebooks of low complexity. Then, a visual feature can be quickly inserted into the index according to its quantization codes. By combining with the multi-inverted index, a non-exhaustive retrieval method is designed, also, the method of computing the approximate distance between image visual features is shown. The experimental results on public datasets show that the proposed non-exhaustive retrieval method outperform 3 existing typical methods over retrieval accuracy and efficiency.