Multidimensional modeling of image quality

In this paper, multidimensional models of image quality are discussed. In such models, alternative images, for instance, obtained through different processing or coding of the same scene, are represented as points in a multidimensional space. The positioning is such that the correlation between geometrical properties of the points and the subjective impressions mediated by the corresponding images is optimized. More specifically, perceived dissimilarities between images are monotonically related to interpoint distances, while the strengths of image quality attributes (such as perceived noise and blur or image quality) are, for instance, monotonically related to point coordinates along specified directions. The goal of multidimensional models is to capture subjective impressions into a single picture that is easy to interpret. We apply multidimensional models to two existing data sets to demonstrate that they indeed account very well for experimental data on image quality. The program XGms is introduced as a new interactive tool for constructing multidimensional models from experimental data. Although XGms is introduced here within the context of image-quality modeling, it is also potentially useful in other applications that rely on multidimensional models.