.png){kind=icon}
.png){kind=icon}
Return
Contrast-based unsupervised hashing method with margin limit
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-023-16572-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The unsupervised hash image retrieval method based on contrastive learning has been widely recognized and concerned because it can make better use of unlabeled datasets to retrieve images. However, in traditional contrastive learning method the negative samples are push to extremely distant positions. The method allows the model to learn only suboptimal solutions, resulting in over-optimization problems. To solve over-optimization problem, the margin limit mechanism is introduced into the loss function of contrastive learning and an unsupervised hashing method based on margin limit is proposed. Specifically, the margin is used as a limiting boundary to control whether negative samples participate in loss calculations, avoiding pushing negative samples to extreme distances during training to reduce over optimization problems. At the same time, structural similarity is used to generate binary hash codes to preserve the likeness of deep features. Finally, experimental results on three benchmark datasets show that our proposed framework is significantly superior compared to existing state-of-the-art hashing methods.
Keywords:
Deep learning
Image retrieval
Metric learning
Image hash
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W