Deep Hashing Based on the von Mises-Fisher Distribution for Fast Large-Scale Image Retrieval

Wenfeng Zhang, Xueling Li, Lei Huang, Wei Hu, Qibing Qin

Keywords: Image retrieval, deep hashing, von Mishes-Fisher distribution, vMF-Hash loss

Abstract: Deep hashing has become a pivotal technology in image retrieval, benefiting from advancements in deep learning and the computational advantages of hashing methods. Existing methods either rely on Euclidean distance, ignoring directional similarity and leading to retrieval errors in high-dimensional scenarios, or use deterministic spherical projections that neglect feature probability distributions, making them susceptible to noise interference. To address these challenges, we propose von Mises-Fisher Deep Hashing (vMF-DH), which introduces the von Mises-Fisher (vMF) distribution to map features onto the unit hypersphere. This approach models directional distributions (cosine similarity) instead of relying on Euclidean distances, leveraging the maximum entropy property of the vMF distribution to enhance both adaptability and robustness. Additionally, we design the vMF-Hash loss function, which regulates feature clustering through the vMF concentration parameter. This ensures the generation of binary hash codes that are compact within classes, well-separated across classes, and highly discriminative. Extensive results on multiple benchmark datasets show that vMF-DH outperforms current state-of-the-art deep hashing methods, demonstrating superior performance in terms of retrieval accuracy and robustness.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 22874