Dielectric cylinders discrimination in a combined analysis of amplitude and phase information of synthesized time-domain response using three-component stacked neural network

The three-component stacked neural network containing two autoencoders and thesoftmax unit has shown its applicability in the analysis of amplitude and phase 2D imagesof dielectric cylinders obtained using synthesized time-domain response. Consideration ofamplitude information in the presence of high noise levels requires increasing the size ofnetwork layers, while phase information is unstable even to small deviations in the diame-ter of the classified cylinders. The current paper proposes a combined approach that in-volves the simultaneous analysis of amplitude and phase images of each cylinder, followedby determination of the general probability of its belonging to a particular class. It isdemonstrated that network training based on this approach allows one to take into consid-eration both a high noise level and comparable deviations in the diameter of the cylindersin a short training time and without the need to increase the number of neurons in the hid-den layers.