Performance prediction of quantized SAR ATR algorithms

Abstract

Automatic target recognition (ATR) of synthetic aperture radar (SAR) target chips is a difficult problem, complicated by the number of nuisance parameters typically present in SAR imagery. This also complicates performance analysis because fully sampling the space of nuisance parameters in an evaluation dataset is intractable. ATR algorithms that first quantize pixel intensity values have been shown to be effective for SAR ATR due to hypothetically reducing the sensitivity to these nuisance parameters. Here we study the performance of two such algorithms, multinomial pattern matching and quantized grayscale matching, and compare them with the traditional mean squared error (MSE) template matching based classification approach. Our approach is to approximate the decision statistic of each algorithm as a Gaussian random variable (RV) parameterized by the noise power, or alternatively signal-to-noise ratio (SNR). This allows the analytic prediction of algorithm performance when the noise process of test images differs from that of the dataset used to train each algorithm, without having to rely on costly empirical simulation. We verify our results in simulations utilizing the AFRL Civilian Vehicle dataset.