Statistical analysis of subjective preferences for video enhancement

摘要

Traditional Thurstone scaling (1927) constructs a perceptual scale from pairwise comparisons without providing statistical inferences. We show that subjective preferences for moving video using pairwise comparisons can be analyzed to construct a perceptual scale and provide the statistical significance of preference differences. Two statistical methods (binary logistic regression and linear regression) are described. Data sets from two studies are used to demonstrate the perceptual scale construction from the traditional Thurstone method and from the described statistical methods. Both the studies showed videos on two side-by-side TVs. Four enhancement levels (Off, Low, Medium and High) were applied to the videos using a commercial device. Subjects made pairwise comparisons to indicate their preference of one video over another. The perceptual scales constructed from the three methods were comparable, except when there were cells missing from the preference matrix. Binary logistic regression easily permitted modeling of additional factors, such as side bias. Video quality can be systematically assessed using pairwise comparisons and statistical methods that permits construction of a perceptual scale and provide statistical significance for the compared levels.