Interest of Data Fusion for Improvement of Classification in X-ray Inspection Ahmad OSMAN * , Ulf HASSLER * , Valerie KAFTANDJIAN ** * Fraunhofer Development Center X-ray Technology EZRT, Dr.-Mack-Str. 81, 90762 Fuerth, Germany ** National Institute of Applied Sciences INSA- Lyon, Bat. St. Exupery, 20 Avenue Albert EINSTEIN, 69621 Villeurbanne, France Abstract. Dempster-Shafer (DS) evidence theory is developed as an attempt to overcome the limitation of conventional probability theory by handling uncertain, imprecise and incomplete information. In this paper we present a classification system based on this fusion theory. The performance of this system is evaluated on 2D and 3D X-ray data. Obtained results are very promising and encourage us to use this system in other applications, namely for ultrasound data classification. 1. Introduction X-ray inspection is a traditional non destructive testing method used to thoroughly test industrial parts, such as aluminium castings in the automotive sector. Safety specifications and quality control task are the main focus of the inspection process. Digital image processing, computational intelligence and hardware progress allowed automating this task. While the detection of true defects is the objective, one main difficulty in X-ray inspection is the detection of false alarms (or false defects), especially if very small and low contrasted defects have to be detected. Therefore, reducing the rejection ratio of good parts without risking missing true defects is a serious challenge. The automatic detection and recognition of defects requires computerized image processing, image analysis, and decision process. The image processing step is critical to detect potential defects. During the image analysis step, features are extracted to be further used in classification between true defects TD and false defects FD. Our intervention is in the classification step, where we developed a specific approach based on data fusion theory to combine different sources of information together in order to improve the true classification rates of true defects and false alarms. To make fusion between different sources possible, a transition from the source space into a common space called “mass values” space takes place. We present a completely automatic mass value attribution procedure with no need for expert supervision. Performances of the classifier are quantified in terms of correct classification rates respectively for true defects and false alarms, and also using Receiver Operating Characteristics curves (ROC). The developed classification system, called Data Fusion Classifier (DFC), was used to classify defects from segmented 2D radiographic images and 3D-CT volumes where it gave in both cases highly precise and reliable decisions. International Symposium on Digital Industrial Radiology and Computed Tomography - We.4.1 Licence: http://creativecommons.org/licenses/by-nd/3.0 1