Image / video denosing

Abstract

In this work the issue of removing Poisson noise from images and videos is addressed. The Poisson noise is dependent on signal intensity and requires more sophisticated algorithmsfor its removal. For high Signal to Noise Ratio many transformations exist that removepoisson noise by converting it into Gaussian after applying some transformation likeAnscombe or Fisz transform, however these transforms are not effective when photoncountare very small and yield poor results. So it is required to address Poisson noise directlywithout its conversion to Gaussian. In this research low photon count region is targetedwithout utilizing any transform and novel Poisson noise removal algorithm for color imagesand videos is proposedIn the first part Poisson denoising scheme in non local frame work based on theapproach of Gaussian mixture models is proposed for astronomical imaging. Proposedscheme employ the concept ofexponential principal component analysis and sparsity ofimage patches. It first converts the color space from RGB to YCbCr and applies K-means++clustering on luminance component only. The same cluster centers are used for chromaticcomponents to improve the computational efficiency. In videos information of both spatialand temporal correlations are taken in cubes of luminance component hence resulting in improved denoising. Simulation results on different images and videos verify the significanceof proposed scheme both in visual and quantitative manner.In the second part the above exponential representation framework is extended to dictionarylearning and sparse coding regime with special focus on the initialization strategy ofdictionary and incoherence between its atoms. The algorithm starts with the designing ofincoherent frames by efficient algorithm. These frames are then adapted to training datasetby exploiting their incoherent structure. Less used atoms are replaced by new atoms that fitthe training data better in an iterative manner. The resultant matrix (dictionary) is providedas input to Poisson dictionary learning algorithm instead of conventional random initial dictionary. Instead of k-means++, greedy algorithm is used for clustering that suits the nature of Poisson greedy sparse recovery algorithm. The proposed algorithm converge to similarresults as of conventional dictionary learning algorithm but in very few iteration and for samenumber of iterations it gives much better results.