A probabilistic pairwise-preference predictor for image quality
Amy R. Reibman, Kenneth E. Shirley, Chao Tian · 2013
Current image quality estimators (QEs) compute a single score to estimate the perceived quality of a single input image. When comparing image quality between two images with such a QE, one only knows which image has a higher score; there is no knowledge about the uncertainty of these scores or what fraction of viewers might actually prefer the image with the lower score. In this paper, we present a Probabilistic Pairwise Preference Predictor (P4) that estimates the probability that one image will be preferred by a random viewer relative to a second image. We train a multilevel Bayesian logistic regression model using results from a large-scale subjective test and present the degree to which various factors influence subjective quality. We demonstrate our model provides well-calibrated estimates of pairwise image preferences using a validation set comprising pairs with 60 reference images outside the training set.