The Role of Feature Enhanced Processing for Automatic Target Recognition using High Resolution Polarimetric SAR Data

The Role of Feature Enhanced Processing for Automatic Target Recognition using High Resolution Polarimetric SAR Data

van den Broek, A.C.
Steeghs, T.P.H.
Dekker, R.J.
TNO Defensie en Veiligheid

2005

We have studied the effect of feature enhanced processing on the discrimination of targets in highresolution polarimetric ISAR and SAR images. This is done by comparing feature-based classification results for original images and images which have been pre-processed to enhance target features. The data comprised four military targets: T72, ZSU23/4, T62, and BMP2. Images at a resolution of 10 cm have been extracted from the ISAR data for a complete aspect range of 360 degrees. The SAR images were taken from the MSTAR database with a resolution of 30 cm. These images have been processed in order to enhance the geometrical delineation of the targets or to enhance point scattering. We have composed feature vectors out of individual features, which were extracted from the original and the enhanced images. The feature vectors are divided into three categories: radiometric, geometric and polarimetric. A maximum likelihood classifier was used to obtain discrimination results. Knowledge about the aspect angle allows target discrimination per aspect angle interval, which will improve classification results. We have investigated the effect of feature enhanced processing on pose estimation. Pose estimation was obtained from the Radon transform of the original and enhanced imagery. We found that the features extracted from enhanced images give slightly better results compared to features extracted from the original images. For the high-resolution enhanced ISAR data reasonable discrimination (about 80%) was obtained compared to the enhanced MSTAR data (about 60%), when radiometric and geometric features are used. Using the polarimetric features the discrimination results could be improved to 85%. For aspect angle determination using the Radon transformation and for target discrimination per aspect angle interval no significant improvement was obtained using feature enhanced processing. The aspect angle could be determined up to an accuracy of 10-15 degrees, depending on the target. Discrimination per aspect interval using radiometric and geometric features shows results of 90% for the ISAR and 75% for the MSTAR images. Polarimetric information improves the results up to 97% for the ISAR images.

Physics

http://resolver.tudelft.nl/uuid:b38dedd3-6977-4ed3-816c-58fddba1c810

222476

MP-SET-80

NATO-RTO, Neuilly-sur-Seine

The RTO Sensors and Electronics Technology Panel (SET) Symposium held in the Banner Hall at the Defence Museum, Akershus Fortress in Oslo, Norway on 11-13 October 2004

paper 42

conference paper