Date of Award

6-12-2024

Document Type

Thesis - SCU Access Only

Publisher

Santa Clara : Santa Clara University, 2024

Department

Computer Science and Engineering

First Advisor

Ying Liu

Abstract

Stereo image super-resolution extracts features from two stereoscopic images to enhance the quality of both views. The project goal was to develop a state-of-the-art stereo image super-resolution model to compete in the NTIRE 2024 Stereo Image Super-Resolution competition. While single image super-resolution models are prominent, the field of stereo imaging lacks the same advancements.

By taking an existing high-performance model, we aimed to surpass current baselines by implementing up-to-date resolution strategies. This paper proposes a model that builds upon the NAFSSR architecture by implementing BSRN approaches to single image super-resolution. By modifying the NAFBlock contained within the NAFSSR architecture, we developed a model that provided competitive results. We judged our approach using PSNR and SSIM metrics.

Our results did not achieve state-of-the-art performance, indicating that while integrating advanced single image super-resolution techniques into stereo models shows promise, it is insufficient on its own to surpass existing baselines. The constraints on model size and computational complexity imposed by the competition, along with resource and timeline limitations, impacted our ability to fully exploit these techniques. This work provides valuable insights and a foundation for future research to continue advancing stereo image super-resolution technologies.

Download

Share

COinS