Polarization Effect on Feature Extraction and Object Classification by Deep Learning

This paper reports the polarization impacts on feature extraction and object classification from ultra-wideband ground penetrating radar (GPR) by deep learning. Three different models are trained by the same convolutional neural network (CNN) to compare the effects of the polarization on the feature extraction and classification accuracy. It is shown that the quad-pole images have some inclusive features compared to single-pole ones. Moreover, the classification accuracy improves by 13.3% due to the inclusive feature of the quad-pole images.