Motion compensated super-resolution of video by level sets evolution

In this contribution, we present an algorithm for motion compensated video super-resolution that combines the models we have developed for regularized image up-sampling and image reconstruction from irregularly-spaced samples. The information from several frames of a video sequence is used to obtain a high-resolution version of one of them. The problem is formulated in a variational framework as the minimization of an energy functional containing two kinds of terms. One relates to the error resulting from the approximation of the low-resolution images from the high-resolution one. A novel image formation model that accounts for the characteristics of the capturing and display devices is used for this term. The other term is a regularization functional measuring the correspondence of the HR image to a model related to our a priori knowledge of image characteristics. For this purpose we use a total variation (TV) prior. A continuous-space spline function models the high-resolution image, allowing to exactly solve a continuous-space defined problem in a discrete framework. The energy functional is minimized using the level-set representation. Preliminary results show the validity of the formulation and of the selected models.