Semester

Spring

Date of Graduation

2023

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Xin Li, PhD.

Committee Member

Donald Adjeroh, PhD.

Committee Member

Gianfranco Doretto, PhD.

Committee Member

Mohamed Hefeida, PhD.

Committee Member

Stephanie Schuckers, PhD.

Abstract

The challenge of restoring real world low-quality images is due to a lack of appropriate training data and difficulty in determining how the image was degraded. Recently, generative models have demonstrated great potential for creating high- quality images by utilizing the rich and diverse information contained within the model’s trained weights and learned latent representations. One popular type of generative model is the generative adversarial network (GAN). Many new methods have been developed to harness the information found in GANs for image manipulation. Our proposed approach is to utilize generative models for both understanding the degradation of an image and restoring it. We propose using a combination of cycle consistency losses and self-attention to enhance face images by first learning the degradation and then using this information to train a style-based neural network. We also aim to use the latent representation to achieve a high level of magnification for face images (x64). By incorporating the weights of a pre-trained StyleGAN into a restoration network with a vision transformer layer, we hope to improve the current state-of-the-art in face image restoration. Finally, we present a projection-based image-denoising algorithm named Noise2Code in the latent space of the VQGAN model with a fixed-point regularization strategy. The fixed-point condition follows the observation that the pre-trained VQGAN affects the clean and noisy images in a drastically different way. Unlike previous projection-based image restoration in the latent space, both the denoising network and VQGAN model parameters are jointly trained, although the latter is not needed during the testing. We report experimental results to demonstrate that the proposed Noise2Code approach is conceptually simple, computationally efficient, and generalizable to real-world degradation scenarios.

Download

Share

COinS