A No-Reference Metric for Perceived Gaussian Blur

摘要

Objective metrics to predict the perceived strength of artifacts in displayed video are useful to improve the overall image quality. Since the original image is usually not available in real-time applications, a no-reference metric should be used. In this paper, our approach to a no-reference blur metric for a subset of images of the LIVE database is discussed. The contrast sensitivity function is used to design six spatial band-pass filters with which the still images are filtered. Subsequently, the standard deviation and the corresponding variance in the band-pass filtered images are calculated and used as input to a linear regression model to fit the weight of the standard deviation and variance for each frequency band to the subjective blur scores. The resulting metric yields a Pearson correlation coefficient of 0.881 with the logarithm of the subjective blur scores. The metric is proven reasonably stable against different content by randomly selecting a subset of the images for training and testing the model.