Discriminability metric based on human contrast sensitivity

Abstract

We evaluated a metric for predicting the discriminability of different digitized versions of alphanumeric characters. The metric is based on the assumption that there exists a visual filter such that discriminability is monotonic with the contrast energy in the visually filtered difference between stimuli. To test this hypothesis, we presented two same or different digital versions of a master character and asked subjects to indicate whether the characters were the same or different in a forced-choice procedure with feedback. The filtered contrast energy difference was calculated by convolving the difference between stimulus pairs with filters derived from published human contrast sensitivity functions, following an initial nonlinear transformation of stimulus intensity, and summing the squared result. For some types of stimulus difference, such as contrast quantization errors and Gaussian blurring, performance on discrimination tasks is monotonically related to the contrast energy of the filtered difference vector. The results are consistent with the hypothesis that there exists a single psychometric function that can predict the discriminability of different digitized versions of characters when displayed on various devices.