Generalizing to Out-of-Sample Degradations via Model Reprogramming.

Area of Science:

• Computer Vision
• Image Processing
• Quantum Mechanics

Background:

• Current image restoration models lack generalization to unseen degradations.
• Zero-shot methods require specific degradation priors, which are often impractical to determine.
• A need exists for restoration models with inherent generalization capabilities.

Purpose of the Study:

• Introduce the Out-of-Sample Restoration (OSR) task.
• Develop a framework enabling restoration models to handle novel, out-of-sample degradations.
• Overcome limitations of existing models in real-world, unpredictable scenarios.

Main Methods:

• Propose a model reprogramming framework leveraging quantum mechanics and wave functions.
• Decouple input images into amplitude and phase wave function terms.
• Adapt the phase term to translate out-of-sample degradations while preserving content in the amplitude term.

Main Results:

• Demonstrate the framework's effectiveness in handling diverse, out-of-sample degradations.
• Show that restoration models can generalize without fine-tuning.
• Validate the flexibility and robustness of the proposed approach through extensive experiments.

Conclusions:

• The proposed quantum-mechanics-inspired framework effectively addresses the OSR task.
• This approach enhances the generalization ability of image restoration models.
• The method offers a flexible solution for real-world image restoration challenges.