Multi-Label Contrastive Semantics Preserving Based Cross-Modal Hashing
Xinghui Zhu, Zeqian Yi, Nanxin Ouyang, Hongyan Zhang, Zhuoyang Zou, Zhengming Yi · 2023
Thanks to the efficient storage and retrieval capabilities provided by hashing techniques, coupled with the highly discriminative feature extraction capabilities of deep neural networks, the field of deep cross-modal hashing has made significant progress in realizing efficient cross-modal retrieval tasks in recent years. Nonetheless, a majority of current supervised cross-modal hashing techniques typically rely on a solitary label to gauge the semantic similarity between cross-modal paired instances. This approach tends to overlook the fact that numerous cross-modal datasets encompass extensive semantic information across multiple labels. To solve this challenge, we propose a novel multi-label contrastive semantics preserving based cross-modal hashing (MCSPH). MCSPH firstly utilizes the multiple labels of instances and calculates the semantic similarity matrix of raw data through cosine similarity, and secondly uses supervised contrastive loss to make our method have better performance. Thorough experimental evaluations conducted on two benchmark datasets substantiate the superiority of the MCSPH method over prominent baseline approaches.