Self partitioning backpropagation network for target recognition

Abstract

A method for qualifying the degree of noncooperation that exists among the target members of the training set is presented. Both the network architecture and the training algorithm are taken into consideration while computing non-cooperation measures. Based on these measures, the network automatically creates several topologically identical partitions. Each partition learns a subset of the targets. The partitioning takes place only when necessary and requires minima computation. Each partition is simple with only one hidden layer and one node in the output layer. A fusion network combines partial results to produce the final response. Simulation results indicate that the method is robust and capable of self organization to overcome the ill effects of non-cooperating targets in the training set. Thus the network complexity and training time are significantly reduced. The self partitioning neural network (SPNN) approach has been tested through extensive simulation using more than 15 sets of real ATR data provided by the US Army Missile Command. Each data set consisted of hundreds of images which were extracted by the target detection system from the sensor's field of view for further processing. The study has indicated that the SPNN approach has the potential for use in real-time target recognition applications.