Neural Network Based False Alarm Controllable Detector for Weak Targets via A Differentiable NP Loss

Weak target detection has been a vital challenge due to the low signal-to-noise ratio. Traditional detectors based on statistical characteristics may suffer from the difficulty of precise model estimation. On the other hand, existing neural-network-based detectors are prone to failing to control a constant probability of false alarm (PFA) effectively. In this paper, a PFA-controllable detector via a convolution-enhanced neural network (NN) is designed to detect weak targets. Firstly, features of weak targets are extracted through the fractional Fourier transform to accumulate energy of weak targets. An enhanced convolutional NN (CNN) is then employed to further adaptively extract and learn the characteristics of the weak targets from these feature images. To mitigate the stringent requirement for training samples in the current PFA-controllable methods, we introduce a differentiable Neyman-Pearson (NP) loss function. The proposed CNN-based detector can use the loss function to achieve end-to-end training under a controllable PFA without extra manual intervention. Numerical examples verify the superiority of the designed NP loss function in terms of PFA controllability. Compared with existing feature-based detectors and the CFAR detector, the proposed detector gains a higher probability of detection with both simulation and measured datasets.