Unsupervised Deep Hashing with Dynamic Pseudo-Multi-Labels for Image Retrieval

Hashing has received a lot of attention in large-scale image retrieval due to its high retrieval accuracy and speed. Unsupervised deep hashing methods with pseudo-labels suffer from suboptimal performance due to low clustering accuracies, which result in unreliable generated pseudo-labels, and are highly sensitive to the number of clusters. To tackle these challenges, we propose an unsupervised deep hashing image retrieval method with dynamic pseudo-multi-labels (UDHPM). Specifically, UDHPM designs a dynamic pseudo-multi-label generation network by employing a soft clustering model to maximize the approximation to the real data distribution and continuously optimizing the pseudo-multi-labels in an end-to-end manner to provide reliable supervised information. Furthermore, UDHPM preserves the category information of dynamic pseudo-multi-labels by applying Kullback–Leibler (KL) divergence to the hashing network. In experiments on three benchmark datasets, UDHPM significantly improves retrieval performance over existing state-of-the-art unsupervised deep hashing methods. UDHPM also exhibits a degree of robustness to the number of clusters.