Using a genetic algorithm to find an optimized pulse coupled neural network solution

Pulse Coupled Neural Networks (PCNNs) have been shown to be of value in image processing applications, especially at identifying features of small spatial extent at low signal to noise ratio. In our use of the PCNN, every pixel in a scene feeds a neuron in a fully connected lateral neural network. Nearest neighbor neurons contribute to the output of any given neuron using weights that link the neuron and its neighborhood in both a linear and a non-linear fashion. The network is pulsed, and the output of the network at each pulse is a binary mask of neurons that are active. Pulsing drives the network to evaluate its state. The multi-dimensionality and the non-linear nature of the network make selecting weights using trial and error a non-trivial problem. It is important that the desired features of the input are identified on a predictable pulse, a problem that has yet to be sufficiently addressed by proponents of the PCNN. Our method to overcome these problems is to use a Genetic Algorithm to select the set of PCNN coefficients which will identify the pixels of interest on a predetermined pulse. This method enables PCNNs to be trained, which is a novel capability and renders the method of use for applications.