Date of Award

2015

Publication Type

Doctoral Thesis

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

Keywords

Full-Reference Image Quality, Image Quality Assessment, Image Saliency, No-Reference Image Quality, Perceived Sharpness Evaluation, Training-free Blind Image Quality

Supervisor

Jonathan Wu

Rights

info:eu-repo/semantics/openAccess

Abstract

Objective assessment of image quality is the process of automatic assignment of a scalar score to an image such that the rating or score corresponds to the score provided by the Human Visual System (HVS). Despite extensive studies since the last two decades, it remains a challenging problem in image processing due to the presence of different types of distortions and limited knowledge of the HVS. Existing approaches for assessing the perceptual quality of images have relied on a number of methodologies that directly apply known properties of the HVS, construct hypotheses considering the HVS as a blackbox and use hybrid approaches that apply both of the techniques. All of these methodologies have relied on different types of visual features for Image Quality Assessment (IQA). In this dissertation, we have studied the problem of different types of IQA from the feature extraction point of view and showed that effective combinations of simple visual features can be used to develop IQA approaches having competitive performance with the state-of-the-art. Our work is divided into four parts each having the final goal to bring about performance improvement in the areas of Full-Reference (FR) and No-Reference (NR)-IQA. We have gradually moved from FR to NR-IQA in the works presented in this dissertation. First, we propose improvements in two existing FR-IQA techniques by introducing changes in the features used. Next, we propose a new FR-IQA technique by extracting image saliency as global features and combining them with the local features of gradient and variance to improve the performance. For NR-IQA, we propose a novel technique for sharpness detection in natural images using simple features. The performance of this method provides improvement over the existing methods. After working with the specific purpose NR-IQA, we propose a general purpose technique using suitable features such that no training with pristine or distorted images or subjective quality scores is required. This technique, despite having no reliance on training, provides competitive performance with the state-of-the-art techniques. The main contribution of the dissertation lies in identification and analysis of effective features and their combinations for improving three different sub-areas of IQA.

Share

COinS