Determination of characteristic defect features for computer-aided image evaluation of welded joint radiographs

Automatic inspection of weld radiographs require a reliable detection and evaluation of the existing defects. Pretentious image processing is necessary to identify the right defect type after the defect detection is already finished. For that, features must be extracted from the image plane to describe the special characteristics of the different defect types. Which kind of features are most important for the recogniction of the defect types will be discussed in this papaer. All features, used in this investigation, are based upon geometric shape and gray level distribution of the defect representation in the radiographic image. The features are tested as single features and feature combinations on more than 500 samples of real weld defects. In the one-dimensional case, all features are evaluated by the Fisher discriminant and the Bhattacharyya coefficient. For evaluation in the multi-dimensional feature space the principal component transformation is additionally used. The tests consider different defect types and different configurations of defect classes. The results are presented in a final statistical interpretation. (orig.)