Toward Real-World Super Resolution With Adaptive Self-Similarity Mining

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Super-resolution (SR) methods struggle with real-world image degradations due to dataset limitations and model complexity.
• Existing supervised SR approaches often fail to generalize across diverse degradation scenarios.

Purpose of the Study:

• To develop a novel SR approach integrating sample-adaptive properties with large-scale data knowledge.
• To enhance fully-supervised SR models for robust real-world degradation handling in a plug-and-play manner.

Main Methods:

• Uniting internal learning (image self-similarity) and external learning (large-scale data) via an unrolled optimization process.
• Employing an attention-based weight-updating method to guide self-similarity mining.
• Utilizing data augmentations and exponential moving average strategy for efficiency.

Main Results:

• The proposed method demonstrates superior performance over existing approaches in both qualitative and quantitative evaluations on real-world degraded images.
• Augmented SR models effectively handle diverse real-world degradation scenarios.

Conclusions:

• The integration of internal and external learning offers a promising direction for improving SR model robustness.
• The developed method provides a plug-and-play enhancement for existing SR models, broadening their applicability to real-world images.