Self-Supervised Learning for Real-World Super-Resolution From Dual and Multiple Zoomed Observations.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Reference-based super-resolution (RefSR) is crucial for enhancing image quality on smartphones.
• Existing RefSR methods face challenges in selecting appropriate reference images and self-supervised learning.

Purpose of the Study:

• To propose a novel self-supervised learning approach for real-world RefSR using dual and multiple camera zooms on smartphones.
• To address the challenges of reference image selection and self-supervised learning in RefSR.

Main Methods:

• Leveraging telephoto images as references for super-resolving ultra-wide images in a dual zoomed super-resolution (DZSR) framework.
• Implementing self-supervised learning by using the telephoto image as supervision, with a two-stage alignment method to handle misalignments.
• Introducing local overlapped sliced Wasserstein loss for visually pleasing results and a progressive fusion scheme for multiple zoomed observations.

Main Results:

• The proposed method achieves superior quantitative and qualitative performance compared to state-of-the-art methods.
• Demonstrated effective self-supervised RefSR using readily available dual and multiple zoomed smartphone images.
• Successfully mitigated misalignment issues between different camera zoom levels.

Conclusions:

• The novel self-supervised RefSR approach effectively utilizes smartphone multi-camera systems for enhanced image super-resolution.
• The proposed methods provide a robust solution for real-world RefSR challenges, improving image quality and usability.
• This work paves the way for advanced image enhancement techniques in mobile photography.