Neural Network LFM Pulse Compression

Abstract

Matched filtering plays an important role in radar systems as the established pulse compression technique. This article puts forwards an alternative machine learning based technique for the matched filtering process assuming the incoming signal is oversampled. The aim is to replace the convolutional operation with a small fully connected feedforwarding neural network and attain an additional increase in the range resolution. The paper demonstrates how such a neural network design can be constructed and a practical training approach is presented. The results are compared against traditional matched filtering and target detection methods showing a clear advantage of trained neural networks for the pulse compression procedure and as a mean to construct inventive mismatched filters.