Detection and classification of weld discontinuities in radiographic images (Part I: Supervised learning)

Abstract

Radiographic testing of weld joints is of great importance for verifying and maintaining weld quality. This work presents a new technique for the development Of an automatic or semiautomatic system for radiographic weld analysis. This technique uses gray level profiles transversal to weld beads in radiographic patterns. These profiles were processed to aid in the setup of nonlinear pattern classifiers developed by neural networks with algorithms by backpropagation of error. The classification accuracy was estimated via the average correctness of 10 randomly chosen test sets. The results presented a general accuracy of classification correctness of around 95% for the class patterns in the profiles that were used.