One of the more commonly used methods for weld defect recognition is the extraction of form and position features of the weld bead. Kato [5], in his work, extracted 10 features to develop an automatic weld bead radiograph inspection system. Also, Aoki [4] proposed to work with 10 weld defect features. The use of features for defect recognition requires considerably less information than when working with image pixels as classifier data input, which require very
large input space consequently increasing the complexity of the calculations. In this work, four features were used to build a set of nonlinear pattern classifier data inputs. These features are described in brief below while a detailed description of them may be found in earlier work [12].