A nonlinear myriad filter for a recursive video enhancement using a robust SRR based on stochastic regularization

In practical applications, a number of linear filtering theory, such as median (L1) and mean (L2), are limited to the cases of Gaussian noise, presenting serious performance degradation in the presence of non-Gaussian noise. Due to the registration error and system noise, the real noise model contaminating the observed images is unknown and usually non-Gaussian noise. Hence, SRR (Super Resolution Reconstruction) algorithms based on median and mean filter may degrade the reconstructed image sequence instead of improve its quality. Myriad filter has a strong mathematical analysis and more powerful and efficient than median and mean filters. The paper proposes a recursive video enhancement using a robust multiframe SRR for applying on image sequences contaminated by any noise models at several noise powers. The proposed SRR framework is based on stochastic regularization with Myriad filter, which is used for removing outliers in the data and for measuring the difference between the projected estimating of the HR image and each LR image. For removing artifacts from the final answer and improving the rate of convergence, Tikhonov regularization is compulsively incorporated because of the SRR ill-pose condition. The performance of proposed method compared with classical SRR algorithms based on median and mean filter is demonstrated on a number of experiments under several noise models (such as Noiseless, AWGN, Poisson Noise, Salt&Pepper Noise and Speckle Noise) at different noise power. Both of the PSNR and virtual images are used to measure the quality of a reconstructed image sequence.