present
Measuring visual quality blind: No-reference video quality estimation and some
applications
Amy Reibman, IEEE Distinguished Lecturer
AT&T Research
Date:
Location: Imaging Science Building (No. 76) Room. 1275 at R.I.T.
Time:
RSVP: Dhiraj Joshi (dhirajjoshi16@gmail.com) for
pizza count
Abstract
Digital image processing is everywhere today: digital photography, digital radiology, digital cinema, video conferencing, and streaming video on the web. An accurate method to estimate the quality of images and video is necessary so that algorithms can be optimized, products can be benchmarked, video outages can be detected, and service-level agreements can be written. Unfortunately, the complexity of the human visual system makes accurate assessment challenging. No-reference (NR) image and video quality estimators (QE) are more widely applicable than their full- and reduced-reference counterparts. The fundamental challenge of the socalled "blind" NR QE is to distinguish desired signal from impairment without having access to the original unprocessed images. To achieve this, NR QE rely on decoded images, the compressed bit stream, and assumptions about the impairments and about the signal itself. In this talk, I'll provide provide a high-level overview of NR QEs. I'll start by describing a 4-step framework for NR QE, comprised of measurement, linearization, pooling, and mapping to subjective quality. Each step incorporates models and information about human perception and human preferences. I'll conclude by presenting a sampling of our recent contributions to image and video quality estimation, including catastrophic outage detection, measuring video quality inside the network, and QE for acquisition of repurposed content.