Distortion Criteria of the Human Viewer

Abstract
The ultimate judges of the quality of reproduced pictures are human observers. In many practical situations it is desirable to have objective measures which closely mirror the performance of the human viewer; today we lack accurate measures. To study this problem, a set of five pictures was used in which distortions of different types were introduced. The resulting series of pictures was rated by viewers by means of an impairment rating test. These ratings were then compared with various objective measures of the error. The objective measure is sought which is most closely monotonically related in a smooth manner to the subjective measure. The average spread of the subjective deviations about a polynomial regression of the subjective measure onto the objective measure is used as the criterion. This technique is used to study both linearly filtered error measures and picture dependent nonlinear weightings of the error. The procedure has enabled a crude quantative model of the human viewer to be determined.