Online semantic embedding correlation for discrete cross-media hashing

The rapid expansion of multimedia data has generated an urgent need for efficient retrieval methods. While batch-based cross-modal hashing has advanced precision in retrieval, it becomes inefficient for online streaming data, increasing computation and storage costs. Additionally, existing online methods often overlook the interdependencies among multiple labels in multimodal data, limiting their ability to generate highly discriminative hash codes. To address these issues, we propose a new online hashing method known as Online semantiC Embedding correlAtion for discrete cross-media hashiNg (OCEAN). OCEAN directly extracts key feature information from multimodal data and uses a normalized label inner product to connect the supervised information accumulated over all rounds, embedding rich semantics into hash codes while reducing computational and storage needs. An asymmetric strategy is introduced to enhance class information embedding, circumventing optimization issues from discrete constraints. Furthermore, OCEAN employs an adaptive label association strategy to dynamically learn label correlations, strengthening the semantic depth of supervised information. An online discrete iterative optimization strategy also helps create concise hash codes with improved discriminative power. Experiments on three benchmark databases show that OCEAN outperforms previous methods, offering superior scalability, efficiency, and search performance. Codes are available at https://github.com/nufehash/OCEAN.