Robust detectors for targets of unknown correlation in clutter

A multilayer perceptron (MLP) based detector is proposed for approximating the average likelihood ratio (ALR) detector in composite hypothesis-testing problems. The case of detecting colored gaussian targets with Gaussian autocorrelation function (ACF) and unknown one-lag correlation coefficient (rho_s), in colored Gaussian interference with Gaussian ACF is considered. The robustness of the likelihood ratio (LR) detector for known rho_s is studied, and and the ALR detector for rho_s varying uniformly in [0, 1] is formulated. Due to the complexity of the involved integral, neural network (NN) based solutions are proposed. NNs not only are capable of approximating the ALR detector, but the implemented approximation is expected to have lower computational cost that other numerical approximations, a very important characteristic in real-time applications. MLPs of different sizes have been trained using a quasi-Newton algorithm to minimize the cross-entropy error. Results prove that MLPs with one hidden layer with 17 neurons can implement very robust detectors.