Single-Image Super-Resolution Using Active-Sampling Gaussian Process Regression.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Gaussian Process Regression (GPR) is effective for example-based image super-resolution (SR).
• High computational cost limits GPR applicability with large datasets.

Purpose of the Study:

• To address the computational limitations of GPR in SR.
• To propose a novel, efficient SR method with high image quality.

Main Methods:

• Introduced Active-Sampling GPR (AGPR), an example learning-based SR method.
• Employed an active learning strategy to select informative training samples for GPR.
• Implemented an accelerating scheme using a pre-learned projection matrix to reduce time complexity.

Main Results:

• AGPR significantly improves computational efficiency compared to standard GPR.
• The method maintains or enhances the quality of reconstructed images.
• Demonstrated superior performance in producing sharper edges and finer details.

Conclusions:

• AGPR offers a more computationally efficient and effective solution for image super-resolution.
• The active learning strategy and acceleration scheme enhance GPR's practical applicability in SR.