Error simulation of paired comparison based scaling methods

摘要

Subjective image quality measurement usually resorts to psychophysical scaling. However, it is difficult to evaluate the inherent precision of these scaling methods. Without knowing the potential errors of the measurement, subsequent use of the data can be misleading. In this paper, the errors on scaled values derived from paired comparison based scaling methods are simulated with randomly introduced proportion of choice errors that follow the binomial distribution. Simulation results are given for various combinations of the number of stimuli and the sampling size. The errors are presented in the form of average standard deviation of the scaled values and can be fitted reasonably well with an empirical equation that can be used for scaling error estimation and measurement design. The simulation proves paired comparison based scaling methods can have large errors on the derived scaled values when the sampling size and the number of stimuli are small. Examples are also given to show the potential errors on actually scaled values (image color quality) of color image prints as measured by the method of paired comparison.