Class-specific prototype networks for open-set recognition

Yulin Li

doi:10.1016/j.neucom.2026.134210

Class-specific prototype networks for open-set recognition

2026-06-09

0

PRE

AI

DOI：10.1016/j.neucom.2026.134210

Original

Original request for help

Save

Abstract

En 中文

Open-set recognition (OSR) aims to address the challenge of correctly identifying samples from unknown classes that are not observed during training. Although existing prototype-based OSR methods have improved model robustness to some extent, relying on a shared prototype encoder across all classes often limits the ability to capture class-specific semantic features. To overcome this limitation, we propose a Class-Specific Prototype Encoder (CSPE) framework, in which an independent lightweight encoder is assigned to each class at the end of the backbone network to learn representative prototypes. Each encoder focuses solely on mapping samples of its target class as close as possible to its corresponding prototype while pushing non-target samples away, without considering the relative positions of other class prototypes. This design enables the learned prototypes to be more compact and discriminative. Furthermore, we propose a prototype constraint mechanism that restricts each class prototype to lie on the boundary of the embedded feature space, thereby reducing the risk of unknown sample intrusion. Experimental results demonstrate that the proposed method outperforms existing prototype-based approaches on multiple benchmark datasets. The code is publicly available at https://github.com/LYLCML/CSPN .

Class-specific prototype networks for open-set recognition

Abstract

Journal

Organization

Cited Papers