Perceptual quality assessment of real-world images and videos

Abstract

The development of online social-media venues and rapid advances in technology by camera and mobile device manufacturers have led to the creation and consumption of a seemingly limitless supply of visual content. However, a vast majority of these digital images and videos are often afflicted with annoying artifacts during acquisition, subsequent storage, and transmission over the network. All these factors impact the quality of the visual media as perceived by a human observer, thereby compromising their quality of experience (QoE).

This dissertation focuses on constructing datasets that are representative of real-world image and video distortions as well as on designing algorithms that accurately predict the perceptual quality of images and videos. The primary goal of this research is to design and demonstrate automatic image and continuous-time video quality predictors that can effectively tackle the widely diverse authentic spatial, temporal, and network-induced distortions -- contrary to all present-day algorithms that operate on single, synthetic visual distortions and predict a single overall quality score for a given video.

I introduce an image quality database which contains a large number of images captured using a representative variety of modern mobile devices and afflicted with a widely diverse authentic image distortions. I will also describe the design of an online crowdsourcing system which aided a very large-scale image quality assessment subjective study. This data collection facilitated the design of a new image quality predictor that is founded on the principles of natural scene statistics of images in different color spaces and transform domains. This new quality method is capable of assessing the quality of images with complex mixtures of distortions and yields high correlation with human perception.

Pertaining to videos, this dissertation describes a video quality database created to understand the impact of network-induced distortions on an end user's quality of experience. I present the details of a large-scale subjective study that I conducted to gather continuous-time ground truth QoE scores on a collection of 180 videos afflicted with diverse stalling events. I also present my analysis of the temporal variations in the perceived QoE due to the time-varying video quality and present insights on the impact of relevant human cognitive aspects such as long-term and short-term memory and recency on quality perception. Next, I present a continuous-time objective QoE predicting model that effectively captures the complex interactions between the aforementioned human cognitive elements, spatial and temporal distortions, properties of stalling events, and models the state of any given client-side network buffer. I also show how the proposed framework can be extended by further supplementing with any number of additional inputs (or by eliminating any ineffective ones), based on the information available at the content providers during the design of adaptive stream-switching algorithms. This QoE predictor supports future research in the design of quality-aware stream-switching algorithms which could control the position, location, and length of stalls, given a network bandwidth budget and the end user's device information, such that the end user's QoE is maximized.