SPEAKERS: Edul N. Dalal, John C. Handley, Wencheng Wu, Jing Wang ABSTRACT: The experimental design of paired comparisons is often used in image quality evaluations. Psychometric scale values for quality judgments are modeled using Thurstone's Law of Comparative Judgment in which distance in a psychometric space of image quality is a function of the probability of preferring one over the other. Two treatments with probability ½ of preference have distance zero in psychometric space. The greater the probability of preferring one treatment over another, the greater the distance in a psychometric space. The transformation from psychometric space to probability is a cumulative probability distribution such as cumulative normal (Thurstone-Mosteller) or logistic (Bradley-Terry). The major drawback of a complete paired comparison experiment is that every treatment is compared to every other, thus the number of comparisons grows quadratically with the number of treatments. This is especially burdensome for hardcopy image quality work where prints need to be prepared and evaluated. We ameliorate this difficulty by performing paired comparisons in two stages. The first stage estimates precise anchors in the psychometric scale space by using many observations across the entire scale. Anchors are then chosen as treatments extremes of the scale plus those treatments with scale values at least one but no more than two JNDs (just noticeable differences) apart. This guarantees that other treatments are within one jnd of an anchor. We also choose the minimum number of anchors such that this condition holds. In the second stage of the design, we compare new treatments to the anchors in a forced choice experiment. The number of comparisons is now linearly rather than quadratically proportional to the number of treatments. It is known that paired comparisons with Thurstone-Mosteller and Bradley-Terry models correspond to probit and logistic regression, respectively, where the model matrix is a matrix of contrasts and the parameter vector is a vector of scales.  Both of these are special cases generalized linear models (GLM). In the anchored paired comparison model, we have a GLM where the regression equation has a constant offset vector determined by the anchors. The result of this formulation is a straightforward statistical model easily analyzed using any modern statistics package with GLM estimation. This enables model fitting and diagnostics. To test this approach, this psychophysical method was applied to two different types of quality evaluations of color pictorial hardcopy images: perceived color accuracy, and overall preference.  The results were found to be compatible with complete paired comparison experiments, but with significantly less effort for large sample sets.