LPI Radar Signal Enhancement Based on Generative Adversarial Networks under Small Samples

With the widespread deployment of low probability of intercept (LPI) radar systems, signal processing of LPI waveforms is becoming a key technology in the modern electronics field. In this paper, a signal enhancement framework aimed at denoising and restoring noisy time-frequency images (TFIs) of LPI radar signals is proposed. The method applies generative adversarial networks (GANs) to this field and conducts training in the case of small samples. A reasonable loss function is designed to optimize the model of signal enhancement at the same time. Furthermore, we utilize several classifiers to prove the validity of the model. Simulation results on eight kinds of typical radar signals demonstrate that the noisy TFIs can be well recovered. And the subsequent classification accuracy is greatly improved by using plain convolutional neural network (CNN), residual network (Resnet), visual geometry group (VGG) network, or any other method.