Human Systems Integration Division @ NASA Ames

The study of subjective visual quality, and the development of computed quality metrics, require accurate and meaningful measurement of visual impairment. A natural unit for impairment is the JND (just-noticeable-difference). In many cases, what is required is a measure of an impairment scale, that is, the growth of the subjective impairment, in JNDs, as some physical parameter (such as amount of artifact) is increased. Measurement of sensory scales is a classical problem in psychophysics. In the method of pair comparison, each trial consists of a pair of samples and the observer selects the one perceived to be greater on the relevant scale. This may be regarded as an extension of the method of forced-choice: from measurement of threshold (one JND), to measurement of the larger sensory scale (multiple JNDs). While simple for the observer, pair comparison is inefficient because if all samples are compared, many comparisons will be uninformative. In general, samples separated by about 1 JND are most informative. We have developed an efficient adaptive method for selection of sample pairs. As with the QUEST adaptive threshold procedure[1], the method is based on Bayesian estimation of the sensory scale after each trial. We call the method Efficient Adaptive Scale Estimation, or EASE ("to make less painful"). We have used the EASE method to measure impairment scales for digital video. Each video was derived from an original source (SRC) by the addition of a particular artifact, produced by a particular codec at a specific bit rate, called a hypothetical reference circuit (HRC). Different amounts of artifact were produced by linear combination of the source and compressed videos. On each pair-comparison trial the observer selected which of two sequences, containing different amounts of artifact, appeared more impaired. The scale is estimated from the pair comparison data using a maximum likelihood method. At the top of the scale, when all of the artifact is present, the scale value is the total number of JNDs corresponding to that SRC/HRC condition. We have measured impairment scales for 25 video sequences, derived from five SRCs combined with each of five HRCs. We find that EASE is a reliable method for measuring impairment scales and JNDs for processed video sequences. We have compared our JND measurements with mean opinion scores for the same sequences obtained at one viewing distance using the DSCQS method by the Video Quality Experts Group (VQEG), and we find that the two measures are highly correlated. The advantages of the JND measurements are that they are in absolute and meaningful units and are unlikely to be subject to context effects. We note that JND measurements offer a means of creating calibrated artifact samples, and of testing and calibrating video quality models.