Development of high quality source driven image inpainting frameworks employing conventional and generative modelling approaches

Abstract

In this digital era, images have become a convenient mode of communication newlineowing to advancements in technology. Billions of images get captured and shared in a newlineday leading to overabundance of electronic information. Editing tools play a significant newlinerole in rendering the best form of the image before unveiling to a spectrum of viewers. newlineImage inpainting is one such powerful editing technique which witnessed a gradual newlineevolution over the past decade. Image inpainting is a process where missing portions of newlinean image are reconstructed in a visually pleasing manner by referring to the known newlineregions of the image. The ability to pull out undesirable objects from an image has newlineushered inpainting into the fields of vintage photo restoration, film editing, satellite newlineimagery, not to mention basic photo editing. Embracing generative models to the newlineinpainting realm took by storm due to its unmatched capability. newlineThe thesis begins with an extensive review of the existing inpainting algorithms newlineand those were classified in a novel perspective of source of reference. The information newlinewhich is referred to, while reconstructing an image, is a critical factor of inpainting newlinealgorithms. While performing image inpainting, the information required for predicting missing region is gathered from either the host image itself, or external image sources or newlineeven guided by user inputs. Depending on the source of reference, two new categories newlinewere introduced, introspective and extrospective. The thesis further proposes newlineimplementation of an adaptive patch-based exemplar model under the introspective newlinecategory, which is suitable for reconstructing smooth images. As an extension to this newlinemodel, the thesis also explores a hybrid model under introspective category, combining newlinediffusion with the adaptive patch-based model, which efficiently handles structure and texture reconstruction simultaneously. The research further focuses on deep learning-based inpainting models under extrospective category.