Robust Multidimensional Matched Subspace Classifiers Based on Weighted Least-Squares

We propose and design two classes of robust subspace classifiers for classification of multidimensional signals. Our classifiers are based on robust M-estimators and the least-median-of-squares principle, and we show that they may be unified as iterated reweighted oblique subspace classifiers. The performance of the proposed classifiers are demonstrated by two examples: noncoherent detection of space-time frequency-shift keying signals, and shape classification of partially occluded two-dimensional (2-D)_ objects. In both cases, the proposed robust subspace classifiers outperform the conventional subspace classifiers