Efficient Poisson denoising for photography

In general, image sensor noise is dominated by Poisson statistics, even at high illumination level, yet most standard denoising procedures often assume a simpler additive Gaussian noise, which is in fact a poor approximation. Fortunately, Poisson noise can under some circumstances be simplified via variance stabilizing methods, such as the Anscombe transform, which is well known to statisticians, medical imaging specialists and astronomers. However, in order to use such a procedure effectively, the actual photon count needs to be known and not simply an illumination intensity, which is the main reason why such procedures are not frequently used in the image processing community. In this article, we propose to use Poisson distribution characteristics to estimate the photon count from relative illumination data, under simple hypotheses. This allows us to use variance-stabilizing methods on standard digital photographs. Thanks to this, the noise becomes close to additive Gaussian and standard filtering methods become significantly more effective. As an example we exhibit the level of improvement that can be achieved using the bilateral filter.