Non-parametric regression for patch-based fluorescence microscopy image sequence denoising

We present a non-parametric regression method for denoising fluorescence video-microscopy volume sequences. The designed method aims at using the 3D+t information in order to restore acquired data contaminated by Poisson and Gaussian noise. We propose to use a variance stabilization transform to deal with the combination of Poisson and Gaussian noise. Consequently, we further propose an adaptive patch-based framework able to preserve space-time discontinuities and reduce significantly noise level using the 3D+t space-time context. This approach lead to an algorithm whose parameters are calibrated and then ready for intensive use. The performance of the proposed method are then demonstrated on both synthetic and real image sequences using quantitative as well as qualitative criteria.