A perceptual adaptive image metric for computer graphics

Abstract

This paper presents two points: a new simple color vision model and an adaptive way to compute an image metric based on a vision model. Metrics are very useful in computer graphics. Applications include perceptually-based rendering or image comparison for photorealism. Usual vision model-based metrics make an expensive use of memory and cpu resources, mainly for two reasons. First, the vision model is a pipeline of non linear functions applying on a multi-scale decomposition of the image. Second, the model is computed for every single pixel of the picture. In this paper, we designed a very simple mono-scale vision model taking into account many perceptual issues like masking effects and adaptation. We also propose an adaptive approach of distance computation : the image plane sample scheme is designed to be denser when distance variation is greater. This method is usable with any vision model and only uses two parameters, making it very easy to con gure. By combining it with our simple vision model, it computes a difference map interactively for 512x512 pictures.