A statistics-based weight assignment in a Hopfield neural network for adaptive image restoration
Stuart William Perry, Ling Guan · 2002
This paper investigates the assignment of weights to a Hopfield-based neural network in adaptive image restoration. The network is given a range of possible weights which are functions of a constraint factor to suppress noise in the restored image. Two methods for choosing the optimal weights are investigated. The first method is the traditional gradient descent method based on choosing the constraint value which best minimizes the neural network energy function for each pixel during each iteration of the algorithm. It is shown that, contrary to our intuition, this method does not produce optimal results. We then propose a second method which is based on selecting each neurons constraint value by considering local image statistics before restoration is commenced. It is shown that this method compares favourably with other neural network image restoration techniques.