Shape classification of flaw indications in three-dimensional ultrasonic images

The rapid evolution of computing hardware technology now allows sophisticated software techniques to be employed which will aid the NDT data interpreter in the process of defect detection and classification. The paper describes an investigation into the area of three-dimensional ultrasonic image evaluation and, more specifically, the problem of characterising the shape of suspect flaw regions. A backpropagation neural network is used as the classifier for a series of four three-dimensional feature extraction methods which are individually assessed on two particular recognition problems. The optimum technique was determined for inclusion in an evaluation environment called the NDT Workbench, which has been designed for the processing of real data. Two acquired ultrasonic data sets are assessed using the best-performing classification method