Analysis of No-Reference Metrics for Image and Video Quality of Consumer Applications
Margaret Pinson · 2020
This paper analyzes the performance of eight no-reference (NR) metrics. Seven assess image quality (BRISQUE, CurveletQA, IL-NIQE, NIQE, OG-IQA, QAC and SSEQ) and one assesses video quality (VIIDEO). The challenge we address in this paper is moving from research silos to a broadly applicable metric. Our analyses use six new subjective datasets that characterize modern cameras and high performing networks. Five datasets were designed around consumer applications and no-reference metric development. The sixth dataset was designed for full-reference metric analyses; we present a technique to modify older datasets for NR metric development. Our analyses show a need for more research and development. The NR metrics were inaccurate for consumer applications.