Time-frequency analysis as a tool for improving neural detectors for low probability of false alarm

This paper deals with the application of time-frequency analysis for transforming the received radar echoes in order to facilitate a neural network classification task. So as to compress the time-frequency representations maintaining most of the information, a feature extractor is designed. The proposed detector is compared with a single Multilayer Perceptron (MLP). The results show that time-frequency decompositions improve the performance of neural networks for slow fluctuating radar targets detection, specially for low values of Probability of False Alarm. The performance of the new detector is nearly independent on the Training-Signal-to-Noise-Ratio (TSNR) and the training initial conditions.