Dual-Encoder Transformers have become a potent vision-language architectural framework that entails inserting images and texts into an overlapping semantic space. They facilitate useful multimodal dialogue, which is essential in applications like cross-modal retrieval, captioning and recommendation systems. Nevertheless, the current dual-encoder methods typically use global-level alignment and ignore fine-scale interactions between image segments and text segments. This restricts them in capturing subtle semantic links resulting in cross-modal representations that are mismatched or partial. In order to address these drawbacks, this paper suggests a new Hierarchical Cross-Modality Attention Distillation (HCAD) model. HCAD presents a multi-level distillation procedure joining both the local (object-word) and global (scene-sentence) features of the two encoders. Using hierarchical patterns of attention, the framework improves semantic correspondence and provides powerful fine-grained alignments in multimodal embedding spaces. The suggested approach can be successfully utilized to the field of multimodal medical report retrieval, where the accurate correspondence of medical images and diagnostic text is important to provide clinical decision support and share knowledge. Experimental findings indicate that HCAD is much more effective in retrieval accuracy, fine-grained matching and semantic robustness than the traditional dual-encoder models. The results indicate that it can be used to promote real-world vision-language multimodal tasks with more realistic and interpretable cross-modal alignment.
Tandi et al. (Thu,) studied this question.