Blind Image Quality Assessment designed by learning-based attributes selection

摘要

With the rapid growth of image processing technologies, objectivernImage Quality Assessment (IQA) is a topic where considerablernresearch effort has been made over the last two decades.rnIQA algorithms based on image structure have been shown torncorrelate well with Mean Opinion Scores (MOS). No-Referencern(NR) image quality metrics are of fundamental interest as theyrncan be embedded in practical applications This paper deals withrna new NR-IQA metric based on natural scenes statistics. It proposesrnto model the best correlated statistics of seven well knownrnno-reference image quality algorithms by a MultiVariate GaussianrnDistribution (MVGD). A part of LIVE database is used withrnthe associated DMOS to fit the MVGD model, namely Model ImagernQuality Index (MIQI). Hence, the quality of a distorted imagernis given by the DMOS that maximizes the multivariate Gaussianrnprobability density function. Experimental results demonstraternthe method effectiveness for a wide variety of distortions.