Skip to Main Content
In this paper we use stochastic sampling to obtain images from the probability distributions induced by well-known image coders SPIHT, JPEG, JPEG2000, and JPEG-LS. Our aim is to obtain actual samples from the typical sets that these coders form in the space of images. We utilize mathematical measures that allow us to quantify the role edges play in representing the resulting sample images. We qualitatively and quantitatively compare these images with a basic form of natural image definition as well as with the standard test images Lenna and Barbara. Our results indicate that today's state-of-the-art image coders are applicable to a variety of localized image regions but they do not significantly model or take advantage of edges that mark the boundaries of different regions.