Perceptual error measure and its application to sampled and interpolated single-edged images

Error metrics quantify the difference between a reproduced image and the corresponding unprocessed original image. A drawback of the commonly used metrics such as the mean square error is their poor correlation with the perceived quality of the reproduced image. We present a framework for an alternative metric that uses the distance in a perceptual space to predict the perceived impairment of reproduced images. The perceptual space is spanned by perceptual artifacts that are introduced by image-reproduction techniques. For image reproduction using sampling and interpolation it is shown how such a multidimensional space can be determined from the image. The sensory strengths of the artifacts’ periodic structure and blur are two of the orthogonal dimensions of this space. In addition, we demonstrate that, after the perceptual-error metric is calibrated to a particular observer, this metric can successfully predict experimentally determined subjective image quality of sampled and interpolated simple black-and-white images.