Single Image Super-Resolution Using Gaussian Process Regression With Dictionary-Based Sampling and Student- ${t}$...

Area of Science:

• Computer Vision
• Machine Learning
• Statistical Modeling

Background:

• Gaussian Process Regression (GPR) is a powerful tool for non-linear mapping.
• Current GPR applications in Super-Resolution (SR) face challenges with high computational cost and unsuitable Gaussian likelihoods for SR observation models.

Purpose of the Study:

• To address the limitations of GPR in Super-Resolution (SR).
• To propose a novel GPR-based SR method that enhances computational efficiency and model accuracy.

Main Methods:

• Implemented dictionary-based sampling (DbS) to create compact training subsets, reducing computational complexity.
• Utilized student-t likelihood, validated through statistical tests, as a more appropriate observation model for SR reconstruction.

Main Results:

• The proposed GPR-based SR method significantly reduces computational complexity compared to traditional approaches.
• Student-t likelihood proved more effective for the SR observation model.
• Experimental results demonstrated superior performance over existing methods, yielding enhanced texture details.

Conclusions:

• The novel GPR-based SR method effectively overcomes computational and modeling limitations.
• The combination of DbS and student-t likelihood offers a promising direction for advanced image super-resolution.
• The method produces high-quality SR images with improved detail rendition.