Date of Publication

9-1-2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science

Subject Categories

Computer Sciences

College

College of Computer Studies

Department/Unit

Software Technology

Thesis Advisor

Joel Ilao

Defense Panel Chair

Arnulfo Azcarraga

Defense Panel Member

Joel Ilao
Macario Cordel II
Conrado Ruiz Jr.
Kai-Lung Hua

Abstract/Summary

Deep learning networks are often difficult to train if there are insufficient image samples. Gathering real-world images tailored for a specific job takes a lot of work to perform. This dissertation explores techniques for synthetic image generation and virtual environments for various image enhancement/ correction/restoration tasks, specifically distortion correction, dehazing, shadow removal, and intrinsic image decomposition. First, given various image formation equations, such as those used in distortion correction and dehazing, synthetic image samples can be produced, provided that the equation is well-posed. Second, using virtual environments to train various image models is applicable for simulating real-world effects that are otherwise difficult to gather or replicate, such as dehazing and shadow removal. Given synthetic images, one cannot train a network directly on it as there is a possible gap between the synthetic and real domains. We have devised several techniques for generating synthetic images and formulated domain adaptation methods where our trained deep-learning networks perform competitively in distortion correction, dehazing, and shadow removal. Additional studies and directions are provided for the intrinsic image decomposition problem and the exploration of procedural content generation, where a virtual Philippine city was created as an initial prototype.

Keywords: image generation, image correction, image dehazing, shadow removal, intrinsic image decomposition, computer graphics, rendering, machine learning, neural networks, domain adaptation, procedural content generation.

Abstract Format

html

Language

English

Format

Electronic

Physical Description

256, [4] leaves

Keywords

Computer graphics; Rendering (Computer graphics); Machine learning; Neural networks (Computer science)

Upload Full Text

wf_yes

Embargo Period

9-1-2023

Download

Share

COinS