Learning Uniform Hyperspherical Centers for Open and Closed Set Recognition
IEEE Access, vol.13, pp.200114-200124, 2025 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 13
- Publication Date: 2025
- Doi Number: 10.1109/access.2025.3635908
- Journal Name: IEEE Access
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Page Numbers: pp.200114-200124
- Keywords: classification, deep learning, neural collapse, Open set recognition, uniformly distributed centers
- Eskisehir Osmangazi University Affiliated: Yes
Abstract
Open-set recognition remains a challenging problem, particularly when traditional closed-set classifiers are unable to generalize to unseen classes. In this paper, we propose a unified framework that leverages hyperspherical embeddings with learnable class centers for both open-set and closed-set recognition. Each class is represented by a center point uniformly distributed on the surface of a hypersphere, and training samples are encouraged to form compact clusters around their respective centers. Unlike previous methods that constrain features to the hypersphere boundary, we adopt a Euclidean distance-based formulation to improve flexibility and generalization. Our approach jointly optimizes class centers and feature representations, eliminating the reliance on predefined center locations. Additionally, we introduce a mechanism to incorporate background or unknown samples during training to further enhance open-set robustness. Extensive experiments on multiple benchmarks demonstrate that our method outperforms existing approaches, achieving state-of-the-art accuracy in both open-set and closed-set settings.