Stats-aware-GAN: Domain specific statistics matching in generative adversarial networks

Kamath, Nidhish Ganesh

Permalink

https://hdl.handle.net/2142/124435 Copy

Description

Title

Stats-aware-GAN: Domain specific statistics matching in generative adversarial networks

Author(s)

Kamath, Nidhish Ganesh

Issue Date

2024-05-02

Director of Research (if dissertation) or Advisor (if thesis)

Lazebnik, Svetlana

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Industrial Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Generative Adversarial Networks
• Distribution matching in GANs
• Domain specific statistics in generative modeling
• Statistical distributions in generated radiology images
• GANs
• Generative image modeling
• Computer vision for medical images
• Medical image generation

Abstract

State-of-the-art GANs such as StyleGAN2 have been regarded as being capable of mimicking a real data distribution and are able to produce photorealistic images when trained using large datasets of images such as faces of people, animals or natural scenes. However, in the case of images that are specific to a particular engineering domain like radiology or material science, StyleGAN2-generated images, though still seemingly realistic to the untrained eye, have been shown not to follow known distributions of key statistics specific to the domain. This non-adherence of important statistical distributions renders the generated images unsuitable for downstream tasks in that domain. Our work aims to find a way to bridge this gap, i.e., aid GANs (particularly StyleGAN2) in correctly modeling domain specific statistics in images. We study the use of various regularizing loss functions formulated with the goal of either improving distributional similarity or fooling a two sample test. Particularly, we explore using the Bhattacharyya distance and Kullback-Leibler divergence for measuring distributional similarity and smoothened formulations of the Friedman-Rafsky and k-Nearest Neighbor two sample tests, and report our findings.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Nidhish Ganesh Kamath

Owning Collections