Design of regularization filters with linear neural networks (image restoration)

Abstract

The authors propose a linear neural network (LNN) that is suitable for the implementation of least squares and regularized inversion problems. They apply this network to the design of regularized filters, which are commonly used in image restoration problems. The constrained least squares (CLS) filter and the robust CLS regularized filter are considered. The CLS regularized filter is implemented using the proposed LNN, whereas the robust CLS regularized filter is implemented using a nonlinear modification called quasi-LNN. Several examples of actual image restoration applications are presented, which are based on the simulation of the proposed filters. SPICE simulation results of an actual circuit are also presented. >