.png){kind=icon}
.png){kind=icon}
Return
MAFH: Multilabel aware framework for bit-scalable cross-modal hashing
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2023.110922.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep hashing techniques have received considerable attention and are widely used in the field of crossmodal retrieval. The training mode of deep cross-modal hashing models has also gradually converged. However, the commonly used training mode ignores noisy labels from different modality original datasets and can only generate fixed-length hash codes, resulting in poor robustness and flexibility of the training generated hash models. Therefore, to effectively optimize the training mode, we propose a multilabel aware framework for bit-scalable deep cross-modal hashing (MAFH) with the following contributions. First, we introduce a label network module to indirectly used the hash representation of multilabels to more accurately supervise the model training process across all modalities. Second, we design a scalable coding strategy to freely manipulate the length of the hash code, which makes the trained model flexible to adapt to different scenarios. Third, we propose a layered multilabel similarity algorithm to preserve more complete feature information, which avoids the loss of discriminative feature information during model training. The training mode of MAFH can effectively enhance model robustness, flexibility and accuracy. We selected thirteen representative hash methods to compare with the proposed method, and experimental results on four public datasets show that the proposed method achieves good performance. https://github.com/x-28/MAFH.git.(c) 2023 Published by Elsevier B.V.
Keywords:
Deep cross-modal hashing
Cross-modal retrieval
Multilabel similarity
Bit-scalable coding strategy
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W