Single Image Super-Resolution Using Local Geometric Duality and Non-Local Similarity

Area of Science:

• Computer Vision
• Image Processing

Background:

• Single image super-resolution (SR) is crucial for enhancing image detail.
• Estimating high-resolution (HR) images from low-resolution (LR) inputs presents challenges in reliability and robustness.

Purpose of the Study:

• To propose a novel single image SR method.
• To improve the reliability and robustness of super-resolved image estimation.

Main Methods:

• Exploiting local geometric duality (GD) and non-local image similarity as priors.
• Generalizing soft-decision interpolation into an adaptive GD (AGD)-based local prior.
• Developing local non-smoothness detection and directional standard-deviation-based weights for adaptive prior design.
• Combining AGD prior with a variational-framework-based non-local prior.
• Speeding up the algorithm using fast GD matrices construction via selective pixel processing.

Main Results:

• The proposed method effectively constrains super-resolved results using image priors.
• Adaptive weight design enhances the AGD prior's performance.
• The integration of local and non-local priors leads to improved SR quality.
• Experimental results demonstrate superior performance compared to state-of-the-art SR algorithms.

Conclusions:

• The novel single image SR method effectively utilizes local geometric duality and non-local similarity.
• The proposed adaptive priors and fast computation contribute to robust and efficient super-resolution.
• The method shows significant improvements over existing state-of-the-art techniques.