Options
2004
Conference Paper
Titel
ATR of battlefield targets by SAR - classification results using the public MSTAR dataset compared with a dataset by QinetiQ, UK
Abstract
The development of ATR algorithms and the comparison of different classification schemes is one of the main goals of the SET-053 group. The group mainly focuses on SAR images of stationary ground targets, in which the targets are detected. These single image chips form a databank for ATR evaluation and identification to which the classification schemes can be applied. Because of the inhomogeneous measured and modelled datasets of the different nations we start our evaluation with the public MSTAR dataset, which is used since many years for ATR evaluation and identification. In most of the publications dealing with the MSTAR dataset [1,2,3] classification rates between 97% and 100% could be reached due to the good quality of the chip images (good adjustment, centered, good signal/noise ratio, nearly exact scaling).But these results should not be overestimated because the image quality can decrease having real applications with targets in battlefield situations. We investigate the performance of simple classification approaches when the quality of the MSTAR dataset was degraded by adding noise, decentering the targets and introducing errors in the crossrange scaling. In addition we used a dataset from real field measurements which was made available to the SET- 053 group by QinetiQ,UK. As anticipated, the classification rates dropped considerable in all mentioned cases. Consequently changes in the feature extraction schemes were investigated which were able to improve the classification rates again. Additionally we analyze the influence of clutter and target shadow on the classification rate. In both datasets the classification rate decreases when we separate the target from clutter and shadow. This is a hint, that a strict separation and segmentation of target and clutter is necessary to classify the real target. Therefore the targets should be measured independently and, if possible, at different locations, so that the clutter doesn't correlate between the test and training data. The target shadow can be used for additional information dependent on the depression angle. By comparing different classifiers (Nearest neighbour, different types of SVMs, HNetEL) we can conclude that the main work is not choosing and applying the classifier, but concentrate more on the data collection, preprocessing and feature extraction process. Therefore in this paper the results of the different investigations concerning the preprocessing of the datasets will be presented. Main topics are the target centering, segmentation, clustering and the influence of the image resolution on the classification rate.