Product Quantization to Reduce Entropy of Labels for Fast and Accurate Image Retrieval
Product Quantization to Reduce Entropy of Labels for Fast and Accurate Image Retrieval
复制标题
DOI：
--
复制
发表时间：
2021-12
期刊：
影响因子：
--
通讯作者：
Fuga Nakamura;Ryosuke Harakawa;M. Iwahashi
中科院分区：
文献类型：
--
作者：
Fuga Nakamura;Ryosuke Harakawa;M. Iwahashi
乘积量化是一种流行的从大规模数据库中快速检索图像的技术。PQ方法将图像特征量化为短码，并利用基于短码的查找表实现快速检索。虽然标签的熵(即用于检索的基本事实)对检索性能至关重要，但现有的PQ方法只关注量化误差。为了提高检索性能，提出了一种新的PQ方法，该方法通过降低标签的熵来提高检索性能。我们假设每个训练样本的正确标签都是已知的；然后，我们训练代码，以便我们可以最小化标签误差以及量化误差，以降低标签的熵。这使得在给出查询(即，其标签未知的图像)时能够快速而准确地检索。
Product quantization (PQ) is a popular technique for fast image retrieval from a large-scale database. PQ methods quantize image features into short codes and realize fast retrieval using lookup tables based on the codes. Although the entropy of labels (i.e., ground truths for retrieval) is crucial for the retrieval performance, existing PQ methods focus only on the quantization errors. This paper proposes a novel PQ method that reduces the entropy of labels to improve the retrieval performance. We assume that correct labels for each training sample are known; then, we train the codes so that we can minimize the label errors as well as the quantization errors to reduce the entropy of labels. This enables fast and accurate retrieval when queries (i.e., images whose labels are unknown) are given.