DCT-Based Deep Learning of Polarimetric Scattering From a Dielectric Cylinder

Electromagnetic wave interaction with cylindrical structures plays an important role in numerous research fields. For the efficient analysis of electromagnetic scattering by cylindrical components, we previously proposed a deep learning scheme which verified that with increased abstraction capability by deep neural network (DNN), it was able to handle the highly oscillatory scattering patterns to an acceptable degree. In this letter, we extend by focusing on more efficient data structural exploration by eliminating the redundancy inherent in the bistatic scattering pattern. Two schemes based on discrete cosine transform (DCT) compression are identified with increasing complexity from 2-D to 3-D DCT, realizing compression ratios of around 26 and 58, respectively, and correspondingly training is about 11 and 14 times more efficient than our previous work on equivalent graphics processing unit (GPU) requirements. Both compression-based schemes have shown good learning capabilities in terms of accuracy and generalization potential, as demonstrated by their predictions of the complicated polarimetric bistatic scattering patterns within and outside the entire dataset and by the extremely high
$R^{2}$
values and low root mean square errors (RMSEs) in the scatter plots. The physical considerations of reciprocity are achieved automatically. It is expected that the proposed compression-based deep learning schemes be useful in the design of future physically based inversion algorithms.