Two-stage Cross Modal Hashing Retrieval Based on Composite Quantization and Cauchy Distribution

Abstract

With the popularity of Internet applications and the rapid growth of multimodal data, cross-modal retrieval has become one of the key technologies in related fields. Because of the need to handle multi-modal data, cross-modal retrieval consumes a large amount of storage space and retrieval time. In this paper we propose a two-stage cross-modal hash retrieval method based on composite quantization and Cauchy distribution, which is able to learn both hash codes and quantization codes in an end-to-end framework and utilize the two codes successively using a two-stage retrieval mode, thus achieving a reduction in retrieval time while improving retrieval accuracy. Besides, to address the problem of unreasonable distribution of multi-modal data hash codes in Hamming space, cross-modal association loss is improved by using the Cauchy distribution, which can generate higher quality hash codes, resulting in smaller hash code distances for semantically similar objects and larger hash code distances for semantically dissimilar objects, thus improving the overall retrieval effect. The experimental results on MIR-Flickr-25K, MS COCO and IAPR TC-12 dataset demonstrate that our method can effectively improve the cross-modal retrieval performance.