Novel Variational Approach for Generalized Signal Dependent Noise Removal

Abstract

Image denoising is fundamental in image processing and computer vision field, where the assumption of a signal independent additive white Gaussian noise model dominates the most state-of-the-art denoising algorithms. In this paper, we propose a novel variational approach that addresses more challenging generalized signal-dependent noise model, which is more flexible and practical for real imaging scenarios. The generalized noise model is decomposed into two independent procedures by introducing an intermediate latent variable, with which the variational model is thus constructed. Moreover, the popular second-order total generalized variation regularizer is introduced into our variational model. To effectively solve the variational problem, we first simply the objective functional by substituting one of the corresponding Euler-Lagrange equations and thus eliminating the intermediate variable. A semi-explicit iterative scheme is present in the proximity operation so as to overcome the difficulty of the high nonlinearity involved in the data term when applying primal and dual hybrid gradient algorithm. Experimental results validate the effectiveness and efficiency of our method.