掩码对齐、加权消歧：基于掩码加权对比的可靠无监督跨模态哈希
Mask to Align, Weight to Disambiguate: Reliable Unsupervised Cross-Modal Hashing with Masked-Weight Contrast
在多模态融合前引入掩码构造部分观测交互，并用语义亲和度加权的对比目标抑制假阴性，这套「掩码+加权对比」思路可直接迁移到视觉-文本检索、图文对齐等跨模态表征学习任务中，缓解主导模态偏置与语义相近样本被误判为负例的问题。
原文依据 UWMCH introduces masking before multimodal fusion to construct partially observed interactions
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论在无监督跨模态哈希中，真实世界的多模态数据往往呈现部分对齐与语义歧义。主导模态容易使融合过程产生偏置，而语义相关的样本在对比学习中可能被误当作负例，从而导致优化不稳定。为解决这些问题，我们提出无监督加权掩码对比哈希（UWMCH）。UWMCH 在多模态融合之前引入掩码以构造部分观测的交互，促使模型学习互补语义并减少对主导模态线索的过度依赖。我们进一步设计了语义亲和度引导的加权对比目标，通过将实例级一致性与聚类共识先验相结合来降低假负例的影响。此外，融合空间的全局与局部语义几何通过聚类中心一致性（CCA）与语义结构正则化（SSR）得到稳定。在三个基准数据集上的大量实验证明了所提方法的有效性与鲁棒性。
英文原文摘要 1088 字符
In unsupervised cross-modal hashing, real world multimodal data often exhibit partial alignment and semantic ambiguity. Dominant modalities can easily bias the fusion process, while semantically related samples may be mistakenly treated as negatives in contrastive learning, leading to unstable optimization. To address these issues, we propose Unsupervised Weighted Masked Contrastive Hashing (UWMCH). UWMCH introduces masking before multimodal fusion to construct partially observed interactions, encouraging the model to learn complementary semantics and reducing over-reliance on dominant modality cues. We further develop a semantic affinity guided weighted contrastive objective to reduce the influence of false negatives by combining instance level consistency with a cluster consensus prior. In addition, the global and local semantic geometries of the fused space are stabilized via Cluster-Centroid Agreement (CCA) and Semantic Structure Regularization (SSR). Extensive experiments on three benchmark datasets demonstrate the effectiveness and robustness of the proposed method.