Samuel Smith

Date of Award

6-13-2024

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2024

Department

Computer Science and Engineering

First Advisor

Ying Liu

Abstract

Single image super-resolution (SR) involves taking a given low-resolution (LR) image and generating a corresponding high-resolution (HR) image. This is a core task in the computer vision field due to its multitude of applications ranging from helping current-day issues of storage and transfer of data to restoration of low-resolution images. The current strategies, however, struggle to reach the quality needed for their widespread use and are often too resource-intensive for the average consumer. While other lightweight SR techniques exist with different techniques, like Cascading Residual Networks [3], their success comes at the cost of expensive technology inaccessible to most situations. With this in mind, we propose a model built to be a lightweight SR network while still improving the quality of the image. This was achieved by using an existing lightweight SR network, SwinIR [9], as the generator. To increase the quality, we used a Generative Adversarial Network (GAN) based on SR360 [3] with a lightweight discriminator, MobileViT [11] to improve upon the quality without sacrificing the lightweight nature of SwinIR. This resulted in a generator that is lightweight and gives high-quality results that was submitted to the NTIRE 2024 Image Super-Resolution (x4).

Download

Share

COinS