Fast robust image restoration using a new neural fusion method?

Abstract

This paper proposes a new neural fusion algorithm for fast robust image restoration without requiring the optimal regularization parameter. The new neural fusion algorithm is based on a new reduced dimension neural network (RDNN). The RDNN is guaranteed to obtain an optimal fusion weight. The proposed RDNN-based neural fusion algorithm uses only a very small solution space to compute the optimal fusion weight, unlike existing neural fusion algorithms with solution space dimension being grater than image size. Unlike current image restoration algorithms, the proposed neural fusion algorithm has a low-dimensional solution space Computed results show that the proposed new algorithm has a robust performance against non-Gaussian noise and can obtain a good image estimate at a fast speed by using the non-optimal regularization parameter.