Improving methods for perceptual image quality assessment

• Author / Creator

  Kamballur Kottayil, Navaneeth

• Image quality assessment (IQA) algorithms aim to simulate human judgment of visual quality on an image. These algorithms are essential components of every multimedia pipeline. IQA is divided into full reference(FR) or no reference (NR) depending on the presence or absence of a pristine image while the image is being judged. Traditional FR and NR-IQA algorithms show high performance on conventional images. However, they fail to generate good results on newer modalities of multimedia data like HDR images and 3D textured meshes. Furthermore, these IQA algorithms limit themselves to low-level features and do not incorporate the effect of image content on human judgment of visual quality. In this thesis, we explore and extend the current IQA capabilities to address these issues. We focus on adaptation of IQA to newer modalities of multimedia content, incorporation of scene level knowledge and integration of low-level features. The goal of this thesis is to advance the IQA algorithms to perform on a larger range of multimedia content by accounting for all available data features and applying latest Machine Learning technologies.The new modalities of multimedia content that we explore in this thesis are High Dynamic Range (HDR) images, and 3D textured meshes (tex-meshes). HDR images pose challenges to conventional IQA methods because of the much larger range of perceptual effects shown by them. This leads to the failure of statistics-based approaches followed by NR-IQA techniques. To account for perceptual effects of HDR, we designed machine learning models that can determine human visual sensitivity from subjective image quality data, without going through psycho-visual experiments. Using this model, we developed a blind noise estimation and quality assessment algorithm for HDR images. Next, we addressed the lack of research into the perceptual effects of texture in tex-meshes. We performed subjective experiments to model the effects of texture compression and incorporated our results into existing research into 3D mesh quality assessment. Furthermore, we design two algorithms that show better correlation to human judgments on quality compared to the existing FR-IQA. The first algorithm is a content-specific IQA performance enhancer, which can be applied to any IQA. The second algorithm is a new full reference algorithm that integrates more low-level features and color elements to improve IQA accuracy. Finally, we performed a case study that analyzed the changes in gaze behavior of humans with the level of familiarity of task. We show statistically significant differences in gaze behavior dependent on familiarity. We validate all of our proposed algorithms by comparing the predictions of our algorithms with human opinions. We observed a high degree of correlation between the human and algorithms scores.

• Subjects / Keywords

  • image quality, deep learning, perceptual quality

• Graduation date

  Fall 2018

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/R35D8NX0J

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.