<title>Prediction and measurement of high quality in still-image coding</title>

Abstract

In image coding applications, quantitative quality metrics which approximate the perceived quality of an image can be used for evaluating the performance of coders, designing new coders and optimizing existing coders. In this paper, we consider images of high quality levels for which most of the errors are in the threshold region of perception, and perceptible errors are sparsely distributed. Two different methodologies are used; first, we use an objective picture quality scale (PQS) which combines partial measures, denoted distortion factors, of random as well as structured perceived distortions due to coding. We also consider an alternative approach, applicable at high quality, that is based on a multi-channel vision model, the visible differences predictor (VDP) proposed by Scott Daly. The VDP produces a probability detection map identifying regions in which errors are sub-threshold, threshold and supra-threshold. For the Phe PQS, since distortion factors due to structured errors and errors in the vicinity of high contrast image transitions are most important at high quality, these two factors are analyzed to compute spatial distortion maps of their contributions. This paper compares the spatial distortion maps produced by both methods for high quality still images to experimental observations. Global values for the quality have been obtained by integrating the factor images obtained using PQS and yield a correlation of 0.9 with mean opinion of score values for JPEG, subband and wavelet coded images.