Abstract
Background: Previously we estimated acuity thresholds for Sloan letters and six groups of Chinese characters (CC1∼CC6, 2–4∼16–18 strokes) based on six subjects' 109,200 trials (VSS05). This data set allowed a detailed analysis of human errors to reveal the psychophysical mechanisms underlying recognition of Chinese characters and English letters near acuity. Methods: Confusion matrices (CMs) were constructed for seven groups of stimuli using results from experimental sessions with 50∼60% correct rates (∼8,000 trials each). The off-diagonal elements (incorrect responses) were fit with a difference of moments (DoM) model, in which a) a letter or character was described by its lower-order geometric moments (e.g., ink pixel number (0th), mean (1st), variance (2nd), skewness (3rd), kurtosis (4th)) in horizontal, vertical, and diagonal directions; b) the perceptual distance between two letters (D) was determined by the sum of weighted differences of corresponding moments; and c) confusability (C) was defined as C=exp(−D). Results: Fitting data with a 5-moment (0th plus 2nd and 3rd H and V moments) model produced correlations from 0.94 (CC1) to 0.61 (CC6) (mean r=0.76). Adding 4th-order H and V moments (7-moment model) slightly improved correlation (r=0.78). Adding six diagonal moments (13-moment model) improved fitting by ∼10% for most groups and by 20% for CC6 (r=0.84). DoM model fitting was better than fitting by template model (r=0.41) and blurred template model (r=0.56). Conclusions: Recognition of near acuity Chinese characters and English letters is most likely based on global stimulus properties, as represented by the lower-order moments in our model.