Detection and classification of weld continuities in radiographic images: Part II unsupervised learning.

Abstract

The importance of industrial radiography as a nondestructive testing method is unquestionable and has lasted more than half a century. This paper presents innovative non-supervised pattern recognition application methodologies, evaluating the formation of cross-sectional profile patterns of grays corresponding to typical welding discontinuities extracted from radiographic images. The techniques involve the development of networks of the modified adaptive resonance theory (ART) type as well as the phenomenologic study of the patterns of each class of discontinuity. The results obtained are pioneering in this kind of research and are quite promising, mainly in connection with the image processing techniques that aim to extract data from the radiographic weld bead and in the detection and classification of discontinuities by analyzing the cross-sectional profile of the gray level. This is the second of three articles describing the work done on using these profiles as inputs for the classifiers