GIDDM: Generating Labels With Diffusion Model to Promote Cross-Domain Open-Set Image Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing cross-domain open-set image recognition methods often use thresholding, struggling with complex class boundaries and feature confusion.
• This leads to negative transfer effects and reduced accuracy when dealing with unknown classes.

Purpose of the Study:

• To propose a Graph Isomorphic Distillation Diffusion Model (GIDDM) for improved cross-domain open-set image recognition.
• To address limitations of threshold-based methods by learning intricate boundary relationships between known and unknown classes.

Main Methods:

• A diffusion classifier quantifies predictive uncertainty using Monte Carlo sampling and models uncertainty distributions.
• An open-set recognition framework employs knowledge distillation from a teacher (closed-set diffusion classifier) to a student classifier.
• Knowledge distillation is framed as a graph isomorphic optimization problem to ensure consistent predictive manifolds, integrated into an adversarial domain adaptation framework.

Main Results:

• The proposed GIDDM achieves state-of-the-art performance on multiple hyperspectral image (HSI) datasets.
• Demonstrated superior ability in separating known and unknown classes and aligning distributions across domains.
• Effectively mitigates negative transfer effects caused by feature confusion.

Conclusions:

• GIDDM offers a robust solution for cross-domain open-set image recognition by effectively modeling predictive uncertainty and class boundaries.
• The graph isomorphic distillation approach enhances knowledge transfer and classifier consistency.
• The method shows significant promise for real-world applications involving complex and evolving datasets.