Feature detection using electromagnetic tomography and neural networks

Electromagnetic tomography (EMT) or magnetic induction tomography is a recent imaging technique. Its purpose is to determine the distribution of electrical conductivity and magnetic permeability in a given region-of-interest (ROI). This distribution is computed from the measurements of the induced voltages at several coils placed around the ROI which are produced by the application of different excitation patterns. Usually images are obtained through inversion of the projection data. In this case, however, a different approach was used. The aim was to perform feature detection of anomalies of `a priori´ known objects. Neural network (NN) techniques were used to process the data. It is concluded that neural networks (NNs) have shown to have good discrimination properties over the features for which they had been trained. Even in a not very precise environment where the positioning of the objects was done by hand, the networks proved to be good feature extractors. The estimation approach has lead to simpler and faster network training. For a similar resolution, the network complexity is also much lower than for the classification approach. The connection of EMT and NNs seems very much promising when one aims not to obtain an image, but to get information on the variability of a particular feature of a given object. Thus, it may prove to be very effective for nondestructive testing of metallic objects. Although the present work has only addressed surface defects, the authors believe that similar results could be obtained in other situations if the scanner frequency was lowered