Image restoration and enhancement with physically-based cues

Abstract

Image restoration and enhancement are one of the subjects widely studied in computer vision for many years. Recently, even more research has been proposed with the rise of deep learning and generative models. In this dissertation, we focus on single image completion and single image reflection removal problem. The image completion method inpaints the specific area of the image that is assigned by a user. The reflection removal technique automatically detects and removes the reflection area when images were taken with a glass. Therefore, this dissertation contains the subject from the classical single image completion method to the GAN-based reflection removal method that automatically detects and removes the unnecessary reflection area. In particular, to make the best use of deep learning-based methods, we utilize a physically-based rendering for generating training images and validate its effects. At the end of this dissertation, we propose shape-focused guidance for learning the properties of reflection removal better with physically-based rendered reflection training images and validate its effect both numerically and visually.