.png){kind=icon}
.png){kind=icon}
Return
Semantic Preservation-Based Hash Code Generation for fine-grained image retrieval
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.126668.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Most fine-grained hashing methods focus solely on designing stronger feature extraction strategies to obtain fine-grained features, without considering how to preserve discriminative information during the hash transformation. This results in generated hash codes lacking sufficient semantic information, making effective retrieval in complex scenarios challenging. To address this issue, we propose a novel Semantic Preservation-Based Hash Code Generation (SPBH) method, aimed at generating highly discriminative hash codes for fine-grained retrieval. Firstly, the object feature localization module is introduced to filter out background noise and focus attention on object regions within raw images. Secondly, the multi-scale feature learning module is designed to obtain multi-scale feature representations by selecting representative local images of varying sizes within target regions, thereby enhancing the precision of fine-grained feature representation. Finally, to reduce the loss of semantic information during the hashing transformation, we transform the extracted fine-grained features into low-dimensional intermediate features and design adaptive category voting masks to enhance critical category information, thereby retaining sufficient semantic information in low-dimensional hash codes. Experiments show that our method outperforms the state-of-the-art hashing methods across four public fine-grained datasets, with the mAP of our proposed method averaging 7.67%, 5.97%, 4.97%, and 4.42% higher than the suboptimal method, respectively. Source code link: https://github.com/x-28/SPBH.git.
Keywords:
Fine-grained hashing
Fine-grained image retrieval
Learning to hash
Multiscale feature
Adaptive category voting mask
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W