Coarse-to-Fine Learning for Single-Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Single-image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs.
• Existing SR methods often struggle with balancing detail generation and artifact suppression.
• Example-learning methods excel at detail but introduce aliasing, while reconstruction-based methods preserve edges but lack fine details.

Purpose of the Study:

• To develop a robust and efficient coarse-to-fine framework for single-image super-resolution (SR) reconstruction.
• To leverage the complementary strengths of example-learning and reconstruction-based SR algorithms.
• To improve the visual quality of SR images while maintaining computational efficiency.

Main Methods:

• A coarse stage utilizes correlative neighbor regression for dictionary atom mapping to generate an initial SR estimate with low computational cost.
• A fine stage employs a regularization term integrating local structure, nonlocal self-similarity, and collaborative representation for enhanced detail.
• The framework combines learned exemplars and reconstruction principles for superior SR performance.

Main Results:

• The proposed coarse-to-fine SR method achieves high-quality image reconstruction.
• Experimental results demonstrate superior performance compared to state-of-the-art SR techniques.
• The method effectively suppresses aliasing artifacts while generating plausible fine details.

Conclusions:

• The developed coarse-to-fine framework offers an effective solution for single-image super-resolution.
• The integration of complementary SR approaches leads to improved image quality and artifact reduction.
• The SR method is computationally efficient in both initial estimation and enhancement stages.