An Improved MSR-Based Data-Driven Detection Method Using Smoothing Pre-Processing

Abstract

The mean spectral radius (MSR) which indicates data correlations is used as a test statistic in data-driven detection approaches based on random matrix theory (RMT). To further improve the detection performance of MSR-based detectors, a smoothing pre-processing method is proposed in this letter. By performing a smoothing pre-processing step on the original sampled data, the divergence between the distributions of the MSR under different hypotheses will be increased, effectively improving the detection probabilities. The AR(1) model is used to illustrate the effect of the smoothing pre-processing coefficients on detection performance. The optimum smoothing coefficients under different AR(1) coefficients and the change of detection probability under certain choices of smoothing coefficients are analyzed. It is verified that the proposed smoothing pre-processing method can effectively improve detection performance by the simulation of low-frequency oscillation detection in colored noise under low signal-to-noise ratio and experiments on floating small target detection in sea clutter using IPIX datasets.