Improving Measurement Geometry for Accurate Classification of Scattered Field data

Abstract

We study the problem of object recognition/ classification using microwave scattering by examining the non-redundant information content in the scattered field data. A broad synthetic forward data is generated using the finite-difference frequency-domain (FDFD) method. A simple correlation coefficient based (CCB) measure is introduced to examine the diversity in the measured data from objects of different shapes and dielectric permittivity. This scattered data is further fed directly to a support vector machine (SVM) for binary classification. The classification results are seen to be consistent with what is expected from the CCB analysis. It is observed that the near-field scattering information is much richer and more suitable for accurate machine-based classification as compared to the far-field information in the presence of noise. The proposed scheme is promising in providing valuable insights on improving the measurement set up so that the sensed data becomes amenable for machine learning.