Semantic Guidance for Boundary Aware Cross-Modal Hashing under Noisy Supervision

Abstract

When labels are corrupted by noise, cross-modal hashing often fails to maintain reliable semantic neighborhoods and clear decision boundaries in the Hamming space. Existing robust hashing methods mainly improve the reliability of samples or pairwise supervision, while the class-level semantics contained in multi-label annotations are not explicitly retained in conventional binary relations. We propose Semantic Guidance Boundary Aware Hashing (SGBH), which exploits label information at both class and relation levels for robust hash learning. Semantic Compass constructs class-level semantic anchors from active category names and transfers category-relevant cues to multimodal hash representations through gated cross-attention and one-way guidance. On the resulting representations, Adaptive Boundary Formation calibrates mini-batch-specific positive and negative boundaries to emphasize hard pairs, while Noise-Tolerant Semantic Alignment reduces the influence of corrupted supervision through cross-modal neighborhood reliability and graded label-overlap relations. Experiments on MIRFLICKR-25K, NUS-WIDE, and MS COCO demonstrate competitive retrieval performance across different code lengths and consistent robustness under varying levels of label noise.