Unsupervised random walk manifold contrastive hashing for multimedia retrieval

Abstract

With the rapid growth in both the variety and volume of data on networks, especially within social networks containing vast multimedia data such as text, images, and video, there is an urgent need for efficient methods to retrieve helpful information quickly. Due to their high computational efficiency and low storage costs, unsupervised deep cross-modal hashing methods have become the primary method for managing large-scale multimedia data. However, existing unsupervised deep cross-modal hashing methods still need help with issues such as inaccurate measurement of semantic similarity information, complex network architectures, and incomplete constraints among multimedia data. To address these issues, we propose an Unsupervised Random Walk Manifold Contrastive Hashing (URWMCH) method, designing a simple deep learning architecture. First, we build a random walk-based manifold similarity matrix based on the random walk strategy and modal-individual similarity structure. Second, we construct intra- and inter-modal similarity preservation and coexistent similarity preservation loss based on contrastive learning to constrain the training of hash functions, ensuring that the hash codes contain complete semantic association information. Finally, we designed comprehensive experiments on the MIRFlickr-25K, NUS-WIDE, and MS COCO datasets to demonstrate the effectiveness and superiority of the proposed URWMCH method.