A user who enlarges an image in a graphics package, such as Photoshop, often finds that while the image is larger, it also looks too smooth. This happens because traditional methods for enlarging images, such as bicubic interpolation, are unable to reconstruct the high-frequency image content necessary for a sharp image.  In this talk, I will describe how to build a statistical model of high-resolution images and show how this model can be used to estimate a high-resolution image from a lower-resolution input image.  To do so, I will show how to construct the model by using the choosing the proper state representation and clique potentials for a Markov Random Field.  I will also discuss the efficiency and efficacy of various algorithms for finding the high-resolution image, given a low-resolution input image.

Joint Work with Bryan C. Russell and William T. Freeman