Pseudo-label driven deep hashing for unsupervised cross-modal retrieval

With the rapid development of big data and the Internet, cross-modal retrieval has become a popular research topic. Cross-modal hashing is an important research direction in cross-modal retrieval, due to its highly efficiency and small memory consumption. Recently, many unsupervised cross-modal hashing methods achieved great results on cross-modal retrieval tasks. However, how to narrow the heterogeneous gap between different modalities and generate more discriminative hash codes are still the main problems of unsupervised hashing. In this paper, we propose a novel unsupervised cross-modal hashing method Pseudo-label Driven Deep Hashing to solve aforementioned problems. We introduce clustering into our modal to obtain initialized semantical information called pseudo-label, and we propose a novel adjusting method that uses pseudo-labels to adjust joint-semantic similarity matrix. We construct a similarity consistency loss function that focuses on the heterogeneity gap between different modalities, and a real values and binary codes fine-tuning strategy for closing the gap between real value space and Hamming space. We conduct experiments on five datasets including three natural datasets which have larger inter-class distances and two medical datasets which have smaller inter-class distances, the results demonstrate the superiority of our method compared with several unsupervised cross-modal hashing methods.

Automated summary

With the rapid development of big data and the Internet, cross-modal retrieval has become a popular research topic. Cross-modal hashing is an important research direction in cross-modal retrieval, and recent unsupervised methods have achieved great results. However, narrowing the heterogeneous gap between different modalities and generating more discriminative hash codes remain the main challenges. In this paper, we propose a novel unsupervised cross-modal hashing method called Pseudo-label Driven Deep Hashing to address these challenges. Experimental results demonstrate the superiority of our method compared to several unsupervised cross-modal hashing methods.