Neuralizing target superresolution algorithms

Tatem et al. (2001) have designed a Hopfield network-based algorithm for superresolving discrete targets that are larger than the sample spacing of an image. The algorithm iteratively minimizes a criterion function that contains a sigmoidal activation term. We have altered their algorithm to bring it in line with Hopfield's original network by reducing the pseudotemperature of the sigmoid. We found that smaller values of the pseudotemperature lead to faster convergence to a solution and resulting solutions that are more accurate.