Comparing Pretrained Image-Net CNN with a Siamese Architecture for Few-Shot Learning Applications in Radar Systems

Abstract

Over the years, the increase in electronic devices and innovation towards technological capabilities have resulted in an increase in traffic in the electromagnetic spectrum, thus making it harder for radar systems to distinguish multiple emitters with added interference. Traditional methods for classification, such as machine learning, prove to be a suitable solution for this problem, however these models require an enormous amount of data to train and evaluate. This experiment implements a Few-Shot learning framework and evaluates the performance of different Neural Network Architectures such as a standard Convolutional Neural Network, and a Siamese Network from a previous experiment. The experiment will utilize different kinds of hardware equipment. This includes the ZCU104 FPGA board, AD-FMCOMMS2-EBZ RF module, the Jetson TX2, and NVIDIA Titan RTX. The hardware equipment will be evaluated using performance metrics such as hardware acceleration, to find the best medium between computational power, acceleration speed, and evaluation accuracy.