Presenter

Mr. Adriaan Barri - ETRO-VUB [Email]

Abstract

Objective measures to automatically predict the perceptual quality of images or videos can reduce the time and cost requirements of end-to-end quality monitoring. For reliable quality predictions, these objective quality measures need to respond consistently with the behavior of the Human Visual System (HVS). In practice, many important HVS mechanisms are too complex to be modeled directly. Instead, they can be mimicked by machine learning (ML) systems, trained on subjective quality assessment databases and applied on predefined objective quality measures for specific content or distortion classes. On the downside, ML systems are often difficult to interpret and may even contradict the input objective quality measures, leading to unreliable quality predictions. In addition, there is the problem of insufficient training data. Current image databases for training ML-based objective quality measures are typically obtained through subjective experiments in a controlled environment. However, since subjective quality experiments are costly and time consuming, the available data for training only represents a small fraction of content that can occur in practice.
To address problems inherent to traditional ML for quality assessment, I developed a customized system called the Locally Adaptive Fusion (LAF). By imposing strict regulations on the ML behavior, such as an interpretable weighting mechanism, LAF is able to yield more reliable quality predictions than conventional ML. To address the problem of insufficient training data, I developed an uncertainty model for objective quality measures using nonparametric Bayesian regression. More precisely, I extended the Gaussian process framework with three important features: support for heteroscedastic noise, monotonicity, and asymptotic constraints. This resulted in a novel regression model called Heteroscedastic Gaussian Processes with iterative Noise Updates (HGPNU). Combining the LAF system with a HGPNU-based uncertainty model allows a complementary training on large image or video databases without subjective quality annotations.

The presentation that I am going to give at the public doctoral defense introduces the field of quality assessment, and describes the theory and applications of the LAF system and HGPNU model for automatically predicting the perceptual quality of images and videos.

Short CV

Master in Mathematics, VUB, 2010