Abstract
Open-set recognition (OSR) is designed to classify the seen classes and identify the unseen classes as unknown. However, existing open-set classifiers rely on deep networks trained using supervised learning techniques for known classes present in the training samples. This factor tends to specialize the learned features of the model toward known classes, making it challenging to differentiate between unknown classes. In this study, we propose a novel open-set recognition framework based on contrastive learning with an unknown score. First, we propose a novel training framework based on contrastive learning to learn more informative features and preserve beneficial information to separate unknown from known. Second, we propose an unknown score function based on multi-layer features to detect unknown samples by considering the difference between the known and unknown classes of intermediate-layer features. Based on extensive experiments conducted on multiple benchmark datasets, the proposed method has been demonstrated to outperform existing methods and achieve state-of-the-art results.
Original language | English (US) |
---|---|
Article number | 111926 |
Journal | Knowledge-Based Systems |
Volume | 296 |
DOIs | |
State | Published - Jul 19 2024 |
Externally published | Yes |
All Science Journal Classification (ASJC) codes
- Software
- Management Information Systems
- Information Systems and Management
- Artificial Intelligence
Keywords
- Classification
- Contrastive learning
- Open-set recognition
- Unknown score