Comparative study of feature mapping and selection for ATR: experiments on SAR data

Publisher Website
Google Scholar
Abstract
Most practical automatic target detection/recognition algorithms (ATR) need to incorporate feature mapping and extraction. The performance of the detection algorithms is strongly affected by the effectiveness of feature mapping. For this paper, the effectiveness is evaluated not only by the number of features needed to represent a signal for a given mean-square magnitude error of approximation, but also by the separability of those features from the clutter background. To investigate the effectiveness of a class of linear feature mapping and selection techniques, a comparative study is conducted in this paper using the data collected by MIT Lincoln Laboratory on the SRI Ultra Wideband (UWB) SAR sensor. This data collection was sponsored by ARPA. The feature mappings considered in this paper include conventional image transforms and modern multiresolution decompositions. The traditional image transforms, such as the discrete Fourier transform and the discrete cosine transform provide frequency domain information in the data, while the modern multiresolution decomposition, such as the discrete wavelet transform, are well-known for their capability to provide both spatial and frequency information. In the transform domain, the significant features, i.e., the features with the greatest energy, are extracted by a zonal-filtering with selected masks. The separability of the extracted target significant features from the natural clutter features are compared in terms of the estimated effective GSNR offered by the different feature mappings using the imaged vehicles and the clutter data obtained from the SRI radar. Finally the experimental receiver operating characteristic curves are also presented for the selected feature mappings.© (1994) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

This publication has 0 references indexed in Scilit: