Boosting “shotgun denoising” by patch normalization

A recent seminal paper on the absolute bounds of image de-noising [1] proposes a patch denoising method effectively realizing the minimal mean square error, given all the known image patches. It is extremely important to reach these absolute limits, but they require processing a limitless database. In the above mentioned paper this database had 10 billion patches. In this paper we demonstrate that by factorizing the patch space the method can be sped up by a factor of more than a thousand, while maintaining the theoretical claim that the method is optimal. Using the method on real images demonstrates its potential to beat the state of the art, as it performs better on difficult patches.