Automatic classification of weld defects in radiographic images

Abstract

There are two problems that affect the accuracy of defect classification for automated radiographic NDT. One is the poor generalisation of the classification method led by a small training sample or an improper classifier, and the other is the poor separability of the feature group. To solve the former, we propose a method based on the direct multiclass support vector machine (DMSVM) to classify the defect, which has good generalisation under the circumstances of a small training set. To tackle the latter, we suggest four new features (three of them are based on the defect region) to characterise the defect, which greatly improve the separability of the feature group. Three classifiers (one-versus-rest SVM, one-versus-one SVM and MLP neuron network) and a group of feathers are used to compare with the classifier and the feature group we proposed. The bootstrap estimate is used to estimate their performances. The experimental results demonstrate that the bootstrap accuracy estimate of DMSVM is 94.25%, which is higher than that achieved by the three compared classifiers. Moreover, the separability of the suggested feature group is equivalent to that of the counterpart but with a two-thirds size, and the computation time is cut by 22.17%.