Subjects examined image pairs consisting of an original and a reproduction created using either JPEG compression or digital halftoning. Subjects marked locations on the reproduction that differed detectably from the original. We refer to the distribution of error marks by the subjects
as image distortion maps.The empirically obtained image distortion maps are compared to the visible difference calculated using three color difference metrics. These are color distortions predicted by the widely used mean square error (point-by-point MSE) computed in RGB values,
the point-by-point CIELAB ΔE color difference formula (CIE, 1971), and S-CIELAB, a spatial extension of CIELAB that incorporates spatial filtering in an opponent colors representation prior to the CIELAB calculation (Zhang & Wandell, 1996).For halftoned reproductions the RMS, CIELAB,
and S-CIELAB error sizes correlated with the locations marked by subjects reasonably well, given the freedom to select a threshold level separately for each image. The S-CIELAB metric had the most consistent threshold levels across images; the RMS metric had the least consistent levels. For
JPEG reproductions, all three metrics provided poor predictions of subjects' marked locations.