Learning Super-Resolution Jointly From External and Internal Examples.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single image super-resolution (SR) is an ill-posed problem requiring regularization.
• Existing methods rely on external datasets or internal image patterns for learning image priors.
• A need exists for methods that leverage both external and internal information adaptively.

Purpose of the Study:

• To propose a joint super-resolution (SR) method that adaptively combines external and internal SR approaches.
• To introduce novel loss functions for both external (sparse coding) and internal (epitomic matching) image priors.
• To develop an adaptive weighting mechanism to balance the contributions of different priors.

Main Methods:

• Developed a joint SR framework integrating sparse coding-based external priors and epitomic matching-based internal priors.
• Defined two distinct loss functions tailored for external and internal feature learning.
• Implemented an adaptive weighting strategy based on reconstruction errors to dynamically balance prior contributions.

Main Results:

• Extensive experiments demonstrate the superior performance of the proposed joint SR method compared to state-of-the-art techniques.
• The method effectively reconstructs high-resolution (HR) images from low-resolution (LR) inputs.
• Subjective evaluation studies confirmed the visual quality improvements achieved by the proposed approach.

Conclusions:

• The proposed joint SR method offers an effective way to regularize the ill-posed SR problem by adaptively fusing external and internal image priors.
• The adaptive weighting mechanism ensures optimal utilization of different prior types based on reconstruction fidelity.
• This approach represents a significant advancement in single image super-resolution.