Conditional Hyper-Network for Blind Super-Resolution With Multiple Degradations.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Single-image super-resolution (SISR) methods struggle with real-world multiple degradations.
• Existing models degrade significantly with distribution shifts between training and test data.

Purpose of the Study:

• To develop a novel framework (CMDSR) that adapts to varying degradation distributions in SISR.
• To improve the robustness and performance of SISR models under diverse real-world conditions.

Main Methods:

• Propose a conditional hyper-network framework (CMDSR) utilizing a ConditionNet to extract degradation priors.
• Employ a support set to learn degradation priors and adapt a basic SR network (BaseNet) parameters.
• Introduce a task contrastive loss to refine task-level feature extraction.

Main Results:

• CMDSR demonstrates significant improvements over existing blind and non-blind SISR methods.
• The framework effectively adapts to distribution shifts without prior degradation map knowledge.
• Experiments confirm the model's superior performance on various degradation scenarios.

Conclusions:

• CMDSR offers a robust and adaptable solution for multi-degradation SISR.
• The proposed ConditionNet and task contrastive loss enhance degradation prior extraction.
• CMDSR serves as a flexible framework applicable to various SISR models.