Independent component analysis and feature extraction techniques for NDT data

The challenge of obtaining relevant information about a specimen containing anomalies using external, noninvasive measurements is barely feasible without recent advances in signal processing. The usual two step nondestructive testing (NDT) problem (detection and reconstruction of the discontinuity) can indeed be interpreted as a pattern recognition task in which the feature extraction aspect is by far the most interesting from a scientific viewpoint. Various relevant feature extraction techniques are compared in this work, with the aim to find the most advantageous mapping that reduces the dimensionality of the input patterns while preserving the relevant information content about the discontinuity. The paper proposes a combination of various types of features to cope with different aspects of the problem.