D´etection automatique CFAR de cibles dans les SAR imageurs.

Abstract

In this PhD Thesis, we address the problem of automatic ship detection acquires from SAR (Synthetic Aperture Radar) images in complex marine environments. Assuming a non-Gaussian sea clutter with no prior knowledge about the presence or not of any clutter edge and/or interfering targets in the sliding reference window, we propose and analyze the detection performances of three CFAR (Constant False Alarm Rate) detectors in homogeneous and heterogeneous Log- normal or Weibull sea clutter. In doing this, we first analyze the QM-CFAR (Quantile Matching-Constant False Alarm Rate) detector in a Weibull background. This detector addresses the problem of fixed-point(s) censoring detector in multiple target situations. Specifically, assuming a non-stationary Weibull clutter with the presence or not of interfering targets, the QM and the MLE (Maximum Likelihood Estimator) are concomitantly used to allow the proposed detector to perform robustly in multiple target situations with a priori unknown Weibull parameters. MC (Monte-Carlo) simulations show that, compared to recent existing CFAR algorithms, the QM-CFAR detector provides robust and accurate estimates of the Weibull distribution parameters and achieves less degradation of the PD (Probability of Detection) in multiple target situations. Then, for the sake of reducing the effect of outliers on the SD (Standard Deviation) based detector, we suggest the use of the MAD (Median Absolute Deviation), as it is a robust and fast alternative to SD. The newly presented MAD-CFAR detector’s detection threshold can be computed straightforwardly; yielding a significant gain in the PD and processing time. Finally, we address the problem of lower and upper automatic censoring of unwanted samples from a rank ordered data of reference cells, i.e., bilateral or dual automatic censoring. To this end, we suggest the use of CFCR (Constant False Censoring Rate) and CFAR detection biparametric thresholds to censor lower and upper outliers. In doing this, we propose a novel estimator AML (Approximate Maximum Likelihood), which generates closed-form expressions of lognormal distribution parameters with no iterations needed. We showed that in a log-normal heterogeneous background, the AML-CFAR ship detector acquires a fair PFA (Probability of False Alarm) regulation, a high detection performance and a fair time cost with regard to the challenging state-of-the-art detectors.