Automatic Target Recognition in SAR Imagery Using a MLP Neural Network
Abstract
In this report, a Multi Layer Perceptron (MLP) Neural Network is used for recognizing military ground vehicles imaged by Synthetic Aperture Radar (SAR). In particular, the classifier is applied to SAR images taken from the MSTAR (Moving and Stationary Target Acquisition and Recognition) data set, which has been made available to the public. Signatures are extracted from the imagery using a Fourier Transform method and features are selected to feed the neural network. A 4-layer (including input and output layers) Neural Network with 38 input nodes, 13 first hidden nodes, 11-second hidden nodes and 3 output nodes, is implemented for this task. Standard delta rule back-propagation algorithm has been used to train the neural network. The MLP neural network is evaluated according to the MSTAR standard evaluation criteria. Training of 3 vehicle classes occurs using a set of SAR images at a 17-degree depression angle with 0-360 degree azimuthal angles, while the testing set contains images at a 15-degree depression angle with 0-360 degree azimuthal angles. The testing set contains both target vehicles that belong to the 3 trained classes and confuser vehicles that do not. Results of MLP neural network evaluation are shown using Receiver Operating Characteristic (ROC) curves and Confusion Matrices.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 2002
- Accession Number
- ADA417194
Entities
People
- Nicholas M. Sandirasegaram
Organizations
- Defence Research and Development Canada