Perceptually-based Modelling for Image Composite Harmonisation

Dolhasz, Alan (2021) Perceptually-based Modelling for Image Composite Harmonisation. Doctoral thesis, Birmingham City University.

Alan Dolhasz PhD Thesis published_Final version_Submitted May 2021_Final Award Jul 2021.pdf - Accepted Version

Download (133MB)


The field of image synthesis is concerned with generation of novel image content. Image compositing, the process of combining elements from existing image data into a seamless whole, is a common approach to image synthesis, employed in application domains such as visual effects in film, architectural visualisation, or augmented reality.

This thesis combines perceptual modelling with recent advances in machine learning in order to produce generalisable models of subjective visual realism in the context of digital image compositing. To achieve this, subjective visual realism in image composites is first modelled as a function of controllable local image transformations, applied to introduce composite-like distortions. These models are then validated and used to produce just-noticeable differences, describing average transformation magnitudes required for humans to distinguish such processed objects as unrealistic. The resulting models are then evaluated in the context of visual attention and refined in an image-wise fashion, before being approximated and generalised using deep learning techniques, particularly self-supervised transformation equivariant representation learning. The resulting models are subsequently shown to outperform baselines in an auxiliary task - image composite harmonisation, indicating that models trained on perceptual data are capable of generalising to related tasks.

Item Type: Thesis (Doctoral)
May 2021Submitted
21 July 2021Accepted
Uncontrolled Keywords: Image composite, harmonisation, deep learning, perception, realism
Subjects: CAH10 - engineering and technology > CAH10-03 - materials and technology > CAH10-03-06 - others in technology
CAH24 - media, journalism and communications > CAH24-01 - media, journalism and communications > CAH24-01-05 - media studies
Divisions: Doctoral Research College > Doctoral Theses Collection
Faculty of Computing, Engineering and the Built Environment > School of Computing and Digital Technology
Depositing User: Jaycie Carter
Date Deposited: 04 Jul 2022 13:44
Last Modified: 22 Mar 2023 12:00

Actions (login required)

View Item View Item


In this section...