Boosting Single Image Super-Resolution Learnt From Implicit Multi-Image Prior.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single Image Super-Resolution (SISR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs.
• A key challenge in SISR is maintaining and enhancing continuous, sharp edges in the reconstructed HR images.
• Existing methods often struggle with edge artifacts and detail preservation.

Purpose of the Study:

• To propose a novel approach for SISR that explicitly incorporates boundary information.
• To leverage multi-view observations, specifically light field (LF) data, as an implicit prior for edge enhancement.
• To improve the performance of various SISR networks, particularly in preserving edge details.

Main Methods:

• Developed a framework that utilizes multi-image priors encoding scene disparity and boundary structure.
• Employed light field (LF) data as an effective multi-image prior during the training phase.
• Introduced a hybrid loss function considering content, structure, variance, and disparity from 4D LF data.

Main Results:

• The proposed method successfully supervises SISR networks for edge-preserving reconstruction.
• The general training scheme enhances the performance of various SISR architectures, especially along object edges.
• Experiments demonstrated a consistent performance gain of approximately 0.6 dB across different backbone SISR models.

Conclusions:

• The integration of an implicit boundary prior learned from multi-view data significantly mitigates challenges in SISR.
• The proposed method offers a general training scheme that boosts SISR performance without altering network architectures.
• This approach effectively improves edge preservation and overall image quality in super-resolved images.