Example-based Antialiasing

This paper presents an example-based method for removing undesirable aliasing in images. The method involves two stages: a learning phase, in which a pair of images, with one image is an antialiased version of the other, is presented as "training data"; and an application phase, in which the learned antialiasing kernel is applied to new aliasing images in order to create an image with antialiased edges. Our framework makes use of two techniques primarily: antialiasing recovery, to learn local appearance changes from the aliasing to the antialiased; and texture synthesis, to find patch correspondences between the training and the new aliasing images. Then we selectively change the local appearance of the new aliasing image so that it has similar antialiased edges to the training antialiased image. We shall demonstrate that our approach works well on a number of examples.