Picture Quality Prediction Based on a Visual Model

Distortion measures are developed for the purpose of predicting the subjective quality of moving monochrome television pictures. The need for such measures is particularly recognized in the area of digital picture coding. Subjectively relevant distortion measures that mirror viewers' assessments of picture quality would make the task of designing and optimizing coding schemes considerably easier. The distortion measures are based on a spatiotemporal model of threshold vision that incorporates the filtering and masking processes. The visual filtering is carried out by parallel excitation and inhibition paths, each of which is separately linear but which combine in a nonlinear way to take account of the adaptation with background luminance. The masking is in the form of a point-by-point weighting of the filtered error based on the amount of spatial and temporal activity in the immediate surround. The processed error averaged over the picture is then used as a prediction of picture quality. Three Classes of distortion measure are considered: 1) the raw error measures that have been used in the past; 2) the filtered error measures where the filtering properties of vision are taken into account; and 3) the masked error measures where the masking processes are also included. It is shown that the filtered error measures are better predictors of picture quality than the raw error measures. The masked error measures lead to further improvements but only if local rather than global averaging procedures are used. It is postulated that this is because viewers tend to base their quality ratings on critical areas rather than on the whole picture.