Comparison of Descriptors for SAR ATR

Recent studies have reported very high automatic target recognition (ATR) rates of synthetic aperture radar (SAR) images based on the Moving and Stationary Target Acquisition and Recognition (MSTAR) database. However, one of the limitations of the MSTAR database is known correlation between clutter contribution in the training and testing sets. In this paper, we propose an ATR method that consists of a machine learning segmentation based on an evolutive Gaussian Mixture Model (GMM), followed by a feature based classification. The use of segmentation results in a reduction of clutter correlation and computational time. A comparison of performance between different descriptors is presented with special emphasis on binary features, such as Binary Robust Invariant Scalable Keypoints (BRISK). Results show the proposed method achieves a 93.41% probability of correct classification when 3 targets in the Standard Operating Condition (SOC) ATR dataset are tested. Results also show this approach is less sensitive to clutter than methods employing gradient features, such as Scale Invariant Feature Transform (SIFT) or Speeded Up Robust Features (SURF), and template methods.