It was shown in previous published work that the neural networks Hopfield model cam be an efficient tool in grey level image restoration by reganding the problem as a minimization of a two part cost measure, in which one component measures roughness, and the other measures distance from the original image. In this thesis, a multistage approach to the Hopfield network to restore bi-level images degraded by noise is considered where the problem of error minimization is addressed locally within each partition while the other partition is frozen. The natural choice for partitioning was into two stages, minimizing the odd data first followed by the even. A natural extension followed, which is splitting into four stages. Simulations were carried for different levels of noise, and different values of the regularization constant and the regularization matrices. The Multistage technique, in general, was proven successful in pushirig the error function to a deeper minima than the one reached by the classical single stage Hopfield model.

Date of this Version

September 1995