Inline inspection of pipelines by means of intelligent pigs usually results in large amounts of data that are analyzed offline by human experts. In order to increase the reliability of the data analysis process as well as to speed up analysis times methods of artificial intelligence such as neural networks have been used in the past with more or less success. The basic requirement for any technique to be used in practice is that no relevant features should be overlooked while keeping the false call rate as low as possible. For the task of automated analysis of in-line inspection data obtained from ultrasonic metal loss inspections, we have developed a two-stage approach. In a first step (called boxing), any defect candidates exceeding the specified size limits are recognized and described by a surrounding box. In the second step, all boxes from step 1 are analyzed yielding basically a relevant/non relevant decision. Each feature considered to be relevant is then classified according to a given set of feature classes. In order to efficiently perform step 2, we have adapted the SVM (support vector machines) algorithm which offers some important advantages compared to, for example, neural networks. We describe the approach applied, and examples as obtained from in-line inspection data are presented.