Fast Single Image Super-Resolution Using a New Analytical Solution for l2 - l2 Problems.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Single image super-resolution (SR) aims to reconstruct a high-resolution image from a low-resolution version, often degraded by blurring, decimation, and noise.
• Current SR algorithms employ diverse strategies for decimation and blurring, including gradient methods and splitting-based approaches that separate up-sampling and deconvolution.

Purpose of the Study:

• To develop a novel, fast single image super-resolution approach.
• To simultaneously handle decimation and blurring operators by exploiting their frequency domain characteristics.
• To provide an efficient analytical solution for l2-regularized quadratic models, applicable to various image priors.

Main Methods:

• The proposed method analyzes decimation and blurring operators in the frequency domain, enabling simultaneous treatment.
• An analytical solution is derived and efficiently implemented for Gaussian priors or other l2-regularized quadratic models (l2-l2 optimization).
• The framework is flexible, accommodating generic image priors and learning-based approaches, and can be integrated into existing splitting methods for non-Gaussian priors.

Main Results:

• The developed SR scheme demonstrates flexibility with various priors, including generic and learning-based ones.
• For non-Gaussian priors, the analytical solution significantly reduces the computational cost of existing splitting-based algorithms.
• Simulations on diverse images confirm the effectiveness and speed of the proposed SR approach compared to existing techniques.

Conclusions:

• The proposed frequency-domain approach offers a fast and effective solution for single image super-resolution.
• The method's ability to handle decimation and blurring simultaneously, coupled with its flexibility in regularization, advances SR techniques.
• Integration of the analytical solution into existing frameworks offers substantial computational savings, enhancing the practicality of advanced SR methods.