| 摘要 |
Reduction of film grain in digitized photographs and motion picture footage is an important problem with applications in such areas as special effects compositing, archival restoration, and fixed-bandwidth lossy compression. Current techniques widely used in practic e (such as median filtering) are shown, both theoretically and practically, to be significantl y suboptimal for this problem, especially for colour images. Previous work on the subject of film grain removal in particular and image noise reduction in general is extensively examined. Bayesian estimation is then proposed as a unifying model for noise reduction, and several classical techniques are rederived in a Bayesian framework. The dimensionality reduction techniques needed to make Bayesian estimation practical are discussed, the failure of certain classica l colour noise reduc- tion algorithms is explained, and the form of an ideal grain reduction technique is postulated. Finally, an algorithm based on these ideas is proposed and tested. This algorithm out-performs earlier techniques by exploiting correlations between the colour channels of an image. |
