.png){kind=icon}
.png){kind=icon}
Return
Predictive Completion Enhanced Deep Hashing With Auxiliary Code Guidance for Incomplete Cross-Modal Retrieval
.png){kind=icon}
PRE
AI
Z
Z
DOI：10.1109/tbdata.2026.3689012.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep cross-modal hashing has gained considerable attention in recent years due to its high efficiency and low memory consumption. However, most of them need to assume the availability of complete multimodal data, overlooking the issue of missing modalities caused by data collection failures or transmission interruptions. In addition, the limited representational capacity of low-bit hash codes restricts their applicability in resource-constrained environments. To overcome these issues, in this paper, we propose a novel deep hashing framework, called predictive completion enhanced deep hashing with auxiliary code guidance (PCEH-ACG), for incomplete cross-modal retrieval. Specifically, we employ a pre-trained CLIP model as an encoder to extract semantic features from both image and text modalities, thereby enhancing cross-modal consistency. Then we adopt a bidirectional prediction network that efficiently completes missing modality features from available ones using a variational inference mechanism. A residual contrastive network is introduced to further enhance the discriminative ability and align the completed multi-modal features, thereby mitigating distribution shifts caused by missing data. Furthermore, high-bit auxiliary codes are considered as semantic teachers during the hashing stage. Through knowledge distillation, they guide low-bit hash codes to learn richer semantic representations, effectively reducing information loss caused by compression. Extensive experiments on three benchmark datasets demonstrate that the proposed PCEH-ACG method consistently outperforms several state-of-the-art cross-modal hashing methods under various incompleteness rates.
Keywords:
Deep hashing
incomplete cross-modal retrieval
variational prediction
residual contrastive network
knowledge distillation
compression
Journal
I
IF:
5.7
Papers:
834
Citations:
3.0K
Organization
Cited Papers