Neural network architectures for real time detection and localization of multiple targets

Summary form only given. The authors present a number of neural network architectures for the detection and localization of multiple targets in a scene corrupted by background noise and clutter. This scene may be obtained as the output of a staring or a scanning sensor array. The problem is to detect the presence and fix the location of multiple targets in the scene. The architectures considered are being called fully connected, replicated subnets, difference processor subnets, and disjoint subnets. Each of the subnets may be considered a multilayer perceptron network using a backward error propagation training algorithm. One of the problems using neural networks is that the training time is a function of the size of the training set and the number of weights to be learned. In the fully connected network the whole scene is used as input to a single large multiple-layer neural network. The other architectures consist of an interconnection of a number of neural subnetworks of some type. These architectures differ in the way the subnetworks are interconnected and how the final decisions are made. One advantage of the proposed interconnected architectures is that the weights calculated for one subnet can be used for many other subnetworks. This feature will reduce the training time for the overall network.<>