We consider the problem of enhancing the resolution of video through the addition of perceptually plausible high frequency information. Our approach is based on a learned data set of image patches capturing the relationship between the middle and high spatial frequency bands of natural images. By introducing an appropriate prior distribution over such patches we can ensure consistency of static image regions across successive frames of the video, and also take account of object motion. A key concept is the use of the previously enhanced frame to provide part of the training set for super-resolution enhancement of the current frame. Our results show that a marked improvement in video quality can be achieved at reasonable computational cost.
|Title of host publication||Proceedings Ninth International Conference on Artificial Intelligence and Statistics 2003|
|Number of pages||8|
|Publication status||Published - 2003|