.png){kind=icon}
.png){kind=icon}
Return
Deep binary hyperbolic embedding for large-scale image retrieval
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.131411.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the rapid development of the web and mobile devices, the scale of multimedia data has grown exponentially. Efficiently retrieving target information from massive datasets has become an important research topic in the multimedia area. Benefiting from advantages such as fast retrieval speed and low storage cost, hashing methods have gradually become mainstream techniques in massive multimedia retrieval tasks. Most existing deep hashing methods commonly employ Euclidean space as the embedding space to measure the semantic similarity among raw samples. However, since the volume of Euclidean space grows polynomially, it struggles to effectively model latent semantic structures, especially when dealing with hierarchical semantics that are widely present in the real world. To address the above problem, by mapping the learned embeddings to hyperbolic space, a novel Deep Binary Hyperbolic Embedding (DBHE) framework is proposed to generate high-quality discrete descriptors. Specifically, based on the Poincaré ball, the feature embeddings obtained from Vision Transformer are projected into hyperbolic space, which better represents the hierarchical semantic structures among raw samples. A pairwise cross-entropy loss based on hyperbolic space is designed to encourage embeddings of semantically similar samples to stay close, while pushing those of dissimilar samples apart, thereby improving the discriminability of hash codes. Extensive experiments on three public datasets show that, compared with the Euclidean setting, our proposed DBHE achieves significant performance improvements across multiple retrieval evaluation metrics, fully demonstrating its advantages in image retrieval applications. The source code is available at https://github.com/QinLab-WFU/DBHE .
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
View more organizations