APPLICATION OF DATA FUSION THEORY AND SUPPORT VECTOR MACHINE TO X-RAY CASTINGS INSPECTION
Résumé
In the field of fully automated X-ray castings inspection, an automatic defect detection and recognition software is necessary. The main issue of such a software is to detect all true defects, while avoiding any false alarm. The Intelligent System for Automated Radioscopy (ISAR) is a software developed at Fraunhofer EZRT, which has an excellent performance in terms of true defects detection. Our work is devoted to the improvement of the ISAR system in terms of reduction of the false alarms rate. The proposed method is based on Dempster-Shafer theory, also called Evidence theory, commonly used in the field of data fusion. A feature extraction step allows to get an array of parameters for each object (area, contrast, elongation, ...). The goal of the data fusion approach is the determination of the feature combination optimally classifying all potential defects. Before combining the features, it is necessary to translate them into a common value space, which is called a mass function in the Evidence theory. An automatic algorithm for attribution of masses for each feature value has been developed and implemented, based on the histogram of the feature and the usage of fuzzy sets. Very good results were obtained on the validation database. Several combinations of features are better than the actual ISAR system. Results have also been compared to Support Vector Machines, and performance is similar.