A human vision system model for objective image fidelity and target detectability measurements

An algorithm is described that accurately predicts human perceptibility of differences between two image sequences, across a broad range of signal types (e.g., consumer video, mammography), difference types (e.g., DCT quantization noise, presence/absence of target of interest), and tasks (e.g., subjective quality rating, target detection). The algorithm, Sarnoff's Just-Noticeable Difference (JND) Vision Model, is based on known physiology and psychophysics of vision, and is calibrated using simple psychophysical tasks of low contrast sine grating detection and sine grating contrast discrimination. Model outputs are derived from psychophysical units of JNDs, in which one JND is defined as a difference between two signals that is at the threshold of detectability. Model results are presented showing excellent predictive performance across a broad range of conditions, as well as some results in which attentional effects demonstrate the need for additional model development.