The Best of Both Worlds: A Framework for Combining Degradation Prediction with High Performance Super-Resolution...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Current blind super-resolution (SR) methods either ignore degradation information or use suboptimal networks.
• Existing approaches face limitations in leveraging full degradation details or employing advanced SR architectures.

Purpose of the Study:

• To develop a unified framework for blind super-resolution (SR) that integrates any degradation prediction mechanism with any deep SR network.
• To enable high-performance non-blind SR architectures to effectively handle blind SR tasks.

Main Methods:

• A novel framework combining a lightweight metadata insertion block with degradation prediction mechanisms.
• Implementation of contrastive and iterative degradation prediction schemes compatible with advanced SR networks like RCAN and HAN.
• Testing the framework on images with complex degradations including blurring, noise, and compression.

Main Results:

• The proposed framework allows non-blind SR architectures to achieve performance rivaling or exceeding state-of-the-art blind SR networks.
• Demonstrated compatibility with high-performance SR networks and various degradation prediction schemes.
• Successfully performed blind SR on images with combined blurring, noise, and compression artifacts.

Conclusions:

• The framework offers a versatile and effective solution for blind super-resolution by bridging the gap between degradation prediction and advanced SR models.
• This work establishes a new baseline for blind SR on complex degradation pipelines, paving the way for future research.