Orientational Distribution Learning With Hierarchical Spatial Attention for Open Set Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open set recognition (OSR) is crucial for reliable AI systems, aiming to distinguish known from unknown classes.
• Current deep learning OSR methods often use distance-based losses in single feature spaces, neglecting spatial distribution impacts.
• This limits the discriminative power of decision boundaries.

Purpose of the Study:

• To introduce Orientational Distribution Learning (ODL) with hierarchical spatial attention for improved OSR.
• To enhance the discriminability of decision boundaries by optimizing the orientation of feature representations.
• To leverage multi-layer and multi-approach features for richer representations.

Main Methods:

• Orientational Distribution Learning (ODL) optimizes feature spatial distribution orientationally.
• A hierarchical spatial attention mechanism captures global feature space dependencies.
• A composite feature space integrates multi-layer and multi-approach features.
• A decision-level fusion combines composite and naive feature spaces.

Main Results:

• ODL significantly improves the discriminability of decision boundaries in OSR tasks.
• The hierarchical spatial attention effectively captures spatial relationships for better feature representation.
• Integrating composite and naive features leads to more comprehensive classification results.
• ODL achieves state-of-the-art performance on multiple benchmark OSR datasets.

Conclusions:

• ODL offers a novel and effective approach to enhance open set recognition.
• Optimizing feature orientation and incorporating spatial attention are key to improving OSR reliability.
• The proposed method demonstrates superior performance and robustness across various datasets.