Target recognition: Conventional and neural network approaches

Summary form only given. The application of both conventional and neural network methods to actual sensor data for target recognition is addressed. The problem of interest is the detection and classification of objects in multisensor images. The types of images used are forward looking infrared (FLIR), absolute and relative range, and doppler. First, conventional approaches to target recognition used at the Air Force Institute of Technology (AFIT) are reviewed. Next, the application of neural networks to target recognition at AFIT is examined. Two classification tests are reported. In the first, the task is to label an object either target or nontarget. The neural network approach achieved 89% accuracy on the training data and was 75% correct on the test data. In the second test, the network learns to classify objects as either tank, truck, or armoured personnel carrier. An accuracy of 98% on the training data was recorded and 85% on the test set. Improvements on the standard backpropagation training rule are also reported. A method which varies the step size in gradient descent is examined. A second-order method is also examined and compared to the standard backpropagation algorithm.<>