Deep cross-visual semantic hashing with self-calibrated collaborative attention

Deep hashing has garnered considerable attention due to its remarkable retrieval efficiency and low storage cost, particularly in visual retrieval scenarios. However, current deep hashing methods generally integrate hash coding into a single-stream architecture, which limits the discriminative power of learned visual features and yields suboptimal hash codes. Additionally, over-reliance on semantic labels shared across samples fails to fully exploit the intrinsic semantic correlations between labels and corresponding visual features. To address these issues, we propose a deep cross-visual semantic hashing (DCvSH) method for image retrieval. First, we develop a visual image feature decoupling encoding network that leverages a self-calibrated collaborative attention mechanism to disentangle common and specific semantics across related images. These decoupled features are fed into a shared decoder for image reconstruction, yielding discriminative visual feature representations. Second, we construct a cross-visual semantic representation learning network with a two-level multi-layer perceptron to capture the underlying relationships between semantic label encodings and visual feature embeddings, while a hypergraph structure is introduced to preserve pairwise similarity relationships. Experimental results on the CIFAR-10, NUS-WIDE, and MIRFLICKR datasets demonstrate consistent improvements, with average mean average precision (mAP) scores reaching 0.895, 0.874, and 0.881 at different code lengths, respectively. Notably, DCvSH outperforms other baselines across all evaluation metrics.