Recovery of Scattered Data Using Neural Network Approach
Mary Jaya V J Mary Jaya V J, TJPRC · International journal of computer science engineering and information technology research · 2017
The problem of detection and characterization of a dielectric cylinder inside a given investigation domain by means of the scattered field values can be represented in terms of the reconstruction of the dielectric characteristics of a generic body illuminated by an incident electromagnetic wave and by means of scattered field values available at some points in the space outside the investigation domain. A circularCylinder with unknown radius 'r', Dielectric permitivity 'r' and center position (x, y) which is illuminated by an incident TM wave produces a scattered electric field which is measured at some point on a circle centered at the axis origin.The unknown dielectric and geometric characteristics of the investigated object can be found by means of a neural network with a single hidden layer.The inputs to the network are the complex values of the scattered field, while the outputs are the unknown variables.The data inputs are trained using the two-layer perceptron training algorithm, the backpropogation algorithm.This algorithm is very slow, and so another version of BP algorithm, called Vogl's acceleration, which allows a fast and more precise convergence is proposed.But when dealing with larger dimension networks, the YPROP algorithm gives a further modification of the BP original procedure by allowing the acceleration and deceleration factors to change at each iteration.After the training step, the neural network is tested with many, differently taken, data.Finally, test results for networks devoted to the retrieval of t he dielectric characteristics, or the position, or both.In this work I'm trying to present the application of a neural network approach to the electromagnetic inverse scattering problem and also trying to show that a suitable network could be able to give a sufficiently accurate solution to the problem of retrieving the position, radius, and dielectric permitivity of an unknown cylinder illuminated by a TM wave, starting from the knowledge of the scattered field.The stability and accuracy of the results are to be studied, with particular stress on the possibility to build specialized networks devoted to the separate extraction of the dielectric and geometric characteristics of the scatterer.