A Neyman-Pearson Criterion-Based Neural Network Detector for Maritime Radar

Abstract

A convolutional neural network (CNN) detector with fixed probability of false alarm (PFA) for application to non-coherent wide area surveillance (WAS) maritime radars is proposed. This detector is trained using a novel cost function-based on Neyman-Pearson (NP) criterion. The use of machine learning allows the detector to learn a complex non-linear model of sea clutter and obviates the need for specifying complex, likely intractable, target plus clutter statistical models. The NP-CNN is shown to perform better than a simple cell-averaging constant false alarm rate (CA-CFAR) statistical detector and a CNN trained using the cross-entropy cost function.