Estimation of structural information content in images

We address the problem of formulating a measure of image information content that is consistent with human perception of information present in a given image. When presented with an image, humans can assess the amount of interesting structural details present as well as the quality of an image in terms of sharpness and noise level. This assessment can be performed independent of any reference image or prior knowledge of ground truth. The challenge is to formulate measures that are sensitive to structurally significant intensity variations in images but which can also reject noise and clutter in a way similar to humans. It is shown that conventional measures known in literature for evaluating single images (no ground truth) fail to reject noisy images. The limitations of the published methods based on statistics, edges and entropy lead us to define a new technique based on an extension of Shannon's entropy measure and multiresolution representations. This implicitly postulates a model of perceived structures that is able to reject noise while giving high scores for sharp, clean natural images.